opentui(v6): defer per-block copy button (carved to its own PR)

The clickable per-block ⧉ copy chip under each message block is split out of
the engine PR (#42922) into its own issue + PR, to keep the engine PR focused.

Removes:
- logic/blockCopy.ts + its unit test (copyBlock + injectable writer)
- the CopyChip component and its two render sites in view/messageLine.tsx
  (flat message text + text parts); the wrapper boxes unwrap back to bare
  <text>/<Markdown>
- the `chips` height accounting in logic/window.ts (estimateMessageHeight +
  partLines added a phantom +1 line per block) and the arg passed from
  view/transcript.tsx; updated window.test.ts / displayModes.test.tsx /
  transcriptWindow.test.tsx expectations accordingly

Unaffected (intentionally kept — core, parity-critical): mouse-selection copy
/ Ctrl+C / copy-on-select (OSC52) in boundary/renderer.ts, and the /copy [n]
command in logic/copy.ts.

Verified: npm run check green (type-check + lint + 813 tests), acceptance greps
clean (no CopyChip/copyBlock/blockCopy left; selection-copy + /copy intact),
and a live tmux smoke confirms no ⧉ copy renders under messages.

Dockerfile

@@ -1,12 +1,14 @@
 FROM ghcr.io/astral-sh/uv:0.11.6-python3.13-trixie@sha256:b3c543b6c4f23a5f2df22866bd7857e5d304b67a564f4feab6ac22044dde719b AS uv_source
-# Node 22 LTS source stage. Debian trixie's bundled nodejs is pinned to 20.x
-# which reached EOL in April 2026 — we copy node + npm + corepack from the
-# upstream node:22 image instead so we can stay on a supported LTS without
-# waiting for Debian 14 (forky, ~mid-2027).  Bookworm-based slim image used
-# so the produced binary links against glibc 2.36, which runs cleanly on
-# our Debian 13 (trixie, glibc 2.41) runtime.  Bumping to a new Node major
-# is a one-line ARG change; see #4977.
-FROM node:22-bookworm-slim@sha256:7af03b14a13c8cdd38e45058fd957bf00a72bbe17feac43b1c15a689c029c732 AS node_source
+# Node 26 source stage. Debian trixie's bundled nodejs is pinned to 20.x
+# (EOL April 2026), so we copy node + npm + corepack from the upstream node:26
+# image instead.  Node 26 (Current; LTS promotion ~Oct 2026) is REQUIRED by the
+# native OpenTUI TUI engine, which loads its renderer via the experimental
+# `node:ffi` API that only exists on Node 26.3+ (the Ink engine + web build run
+# on it too).  Bookworm-based slim image used so the produced binary links
+# against glibc 2.36, which runs cleanly on our Debian 13 (trixie, glibc 2.41)
+# runtime.  The pinned tag ships v26.3.0.  Bumping Node is a one-line change here.
+# NOTE: verify the full image build + Ink/web/Playwright on Node 26 in CI.
+FROM node:26-bookworm-slim@sha256:79723b41edbedf595f62e943a9f8b0ba9af5b1e61045c5f8f59c2c02c1212a16 AS node_source
 FROM debian:13.4
 
 # Disable Python stdout buffering to ensure logs are printed immediately
@@ -90,7 +92,7 @@ RUN useradd -u 10000 -m -d /opt/data hermes
 
 COPY --chmod=0755 --from=uv_source /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
 
-# Node 22 LTS: copy the node binary plus the bundled npm + corepack JS
+# Node 26: copy the node binary plus the bundled npm + corepack JS
 # installs from the upstream image.  npm and npx are recreated as symlinks
 # because they're symlinks in the source image (and need to live on PATH).
 # See node_source stage at the top of the file for the version-bump
@@ -119,7 +121,7 @@ COPY ui-tui/packages/hermes-ink/ ui-tui/packages/hermes-ink/
 
 # `npm_config_install_links=false` forces npm to install `file:` deps as
 # symlinks instead of copies.  This is the default since npm 10+, which is
-# what the image ships now (via the node:22 source stage).  We set it
+# what the image ships now (via the node:26 source stage).  We set it
 # explicitly anyway as defense-in-depth: the previous Debian-bundled npm
 # 9.x defaulted to install-as-copy, which produced a hidden
 # node_modules/.package-lock.json that permanently disagreed with the root
@@ -181,8 +183,16 @@ RUN uv sync --frozen --no-install-project --extra all --extra messaging --extra
 # invalidate the (relatively slow) web + ui-tui build layer.
 COPY web/ web/
 COPY ui-tui/ ui-tui/
+COPY ui-opentui/ ui-opentui/
+# ui-opentui is the opt-in native OpenTUI engine (HERMES_TUI_ENGINE=opentui;
+# default stays Ink). .dockerignore strips its node_modules/dist, so install +
+# esbuild-build it here -> dist/main.js, then prune devDeps (esbuild/babel/
+# vitest); the runtime only needs the prod deps (the external @opentui/core +
+# its native blob -- the bundle inlines solid/effect). Build needs Node 26.3
+# (node:ffi floor), which this image ships.
 RUN cd web && npm run build && \
-    cd ../ui-tui && npm run build
+    cd ../ui-tui && npm run build && \
+    cd ../ui-opentui && npm install --no-audit --no-fund && npm run build && npm prune --omit=dev
 
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.

README.md

@@ -107,6 +107,8 @@ You can still bring your own keys per-tool whenever you want — the gateway is
 
 Hermes has two entry points: start the terminal UI with `hermes`, or run the gateway and talk to it from Telegram, Discord, Slack, WhatsApp, Signal, or Email. Once you're in a conversation, many slash commands are shared across both interfaces.
 
+> **TUI engine:** On supported hosts (Linux/macOS with Node 26.3+), the terminal UI defaults to the native **OpenTUI** engine, which the installer provisions for you. The legacy **Ink** engine remains the fallback — it's used automatically on Windows, Termux, or when the native engine can't run, and you can select it explicitly with `HERMES_TUI_ENGINE=ink hermes`. Ink is not going away; it's the kept fallback.
+
 | Action                         | CLI                                           | Messaging platforms                                                              |
 | ------------------------------ | --------------------------------------------- | -------------------------------------------------------------------------------- |
 | Start chatting                 | `hermes`                                      | Run `hermes gateway setup` + `hermes gateway start`, then send the bot a message |

agent/auxiliary_client.py

@@ -3079,23 +3079,20 @@ def _try_configured_fallback_chain(
         if not fb_provider or fb_provider.lower() == skip:
             continue
         fb_model = str(entry.get("model", "")).strip() or None
-        fb_base_url = str(entry.get("base_url", "")).strip() or None
-        fb_api_key = str(entry.get("api_key", "")).strip() or None
 
         label = f"fallback_chain[{i}]({fb_provider})"
 
         try:
-            fb_client = _resolve_single_provider(
-                fb_provider, fb_model, fb_base_url, fb_api_key)
+            fb_client, resolved_model = _resolve_fallback_entry(entry)
         except Exception:
-            fb_client = None
+            fb_client, resolved_model = None, None
 
         if fb_client is not None:
             logger.info(
                 "Auxiliary %s: %s on %s — configured fallback to %s (%s)",
-                task, reason, failed_provider, label, fb_model or "default",
+                task, reason, failed_provider, label, resolved_model or fb_model or "default",
             )
-            return fb_client, fb_model, label
+            return fb_client, resolved_model or fb_model, label
         tried.append(label)
 
     if tried:
@@ -3106,6 +3103,103 @@ def _try_configured_fallback_chain(
     return None, None, ""
 
 
+def _fallback_entry_api_key(entry: Dict[str, Any]) -> Optional[str]:
+    """Resolve inline or env-backed API key from a fallback-chain entry."""
+    explicit = str(entry.get("api_key") or "").strip()
+    if explicit:
+        return explicit
+    key_env = str(entry.get("key_env") or entry.get("api_key_env") or "").strip()
+    if key_env:
+        return os.getenv(key_env, "").strip() or None
+    return None
+
+
+def _resolve_fallback_entry(entry: Dict[str, Any]) -> Tuple[Optional[Any], Optional[str]]:
+    """Resolve one fallback entry through the central provider router."""
+    provider = str(entry.get("provider") or "").strip()
+    model = str(entry.get("model") or "").strip() or None
+    if not provider or not model:
+        return None, None
+    base_url = str(entry.get("base_url") or "").strip() or None
+    api_key = _fallback_entry_api_key(entry)
+    api_mode = str(entry.get("api_mode") or entry.get("transport") or "").strip() or None
+    return resolve_provider_client(
+        provider,
+        model=model,
+        explicit_base_url=base_url,
+        explicit_api_key=api_key,
+        api_mode=api_mode,
+    )
+
+
+def _try_main_fallback_chain(
+    task: Optional[str],
+    failed_provider: str = "",
+    reason: str = "error",
+) -> Tuple[Optional[Any], Optional[str], str]:
+    """Try the top-level main-agent fallback chain for an auxiliary call.
+
+    ``provider: auto`` auxiliary tasks should respect the user's declared
+    main fallback policy before dropping into Hermes' built-in discovery
+    chain. The top-level chain is read through ``get_fallback_chain`` so
+    both modern ``fallback_providers`` and legacy ``fallback_model`` entries
+    participate in the same order as the main agent.
+    """
+    try:
+        from hermes_cli.config import load_config
+        from hermes_cli.fallback_config import get_fallback_chain
+
+        chain = get_fallback_chain(load_config())
+    except Exception as exc:
+        logger.debug("Auxiliary %s: could not load main fallback chain: %s", task or "call", exc)
+        return None, None, ""
+
+    if not chain:
+        return None, None, ""
+
+    failed_norm = (failed_provider or "").strip().lower()
+    main_norm = (_read_main_provider() or "").strip().lower()
+    skip = {p for p in (failed_norm, main_norm, "auto") if p}
+    tried: List[str] = []
+
+    for i, entry in enumerate(chain):
+        if not isinstance(entry, dict):
+            continue
+        fb_provider = str(entry.get("provider") or "").strip()
+        fb_model = str(entry.get("model") or "").strip()
+        if not fb_provider or not fb_model:
+            continue
+        fb_norm = fb_provider.lower()
+        label = f"fallback_providers[{i}]({fb_provider})"
+        if fb_norm in skip:
+            tried.append(f"{label} (skipped)")
+            continue
+        if _is_provider_unhealthy(fb_norm):
+            _log_skip_unhealthy(fb_norm, task)
+            tried.append(f"{label} (unhealthy)")
+            continue
+        try:
+            fb_client, resolved_model = _resolve_fallback_entry(entry)
+        except Exception as exc:
+            logger.debug("Auxiliary %s: main fallback %s failed to resolve: %s", task or "call", label, exc)
+            fb_client, resolved_model = None, None
+        if fb_client is not None:
+            logger.info(
+                "Auxiliary %s: %s on %s — main fallback chain to %s (%s)",
+                task or "call", reason, failed_provider or "auto", label,
+                resolved_model or fb_model,
+            )
+            return fb_client, resolved_model or fb_model, fb_provider
+        tried.append(label)
+
+    if tried:
+        logger.debug(
+            "Auxiliary %s: main fallback chain exhausted (tried: %s)",
+            task or "call", ", ".join(tried),
+        )
+    return None, None, ""
+
+
 def _resolve_single_provider(
     provider: str,
     model: Optional[str] = None,
@@ -3116,16 +3210,19 @@ def _resolve_single_provider(
 
     Uses the existing provider resolution infrastructure where possible.
     """
-    # Reuse resolve_provider_client which handles provider→client mapping
+    # Reuse resolve_provider_client which handles provider→client mapping.
     client, resolved_model = resolve_provider_client(
         provider=provider,
         model=model,
-        base_url=base_url,
-        api_key=api_key,
+        explicit_base_url=base_url,
+        explicit_api_key=api_key,
     )
     return client
 
-def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Optional[OpenAI], Optional[str]]:
+def _resolve_auto(
+    main_runtime: Optional[Dict[str, Any]] = None,
+    task: Optional[str] = None,
+) -> Tuple[Optional[OpenAI], Optional[str]]:
     """Full auto-detection chain.
 
     Priority:
@@ -3223,7 +3320,22 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
                             main_provider, resolved or main_model)
                 return client, resolved or main_model
 
-    # ── Step 2: aggregator / fallback chain ──────────────────────────────
+    # ── Step 2: user-configured fallback policy ─────────────────────────
+    # In auto mode, respect the task-specific fallback chain first, then the
+    # main agent's top-level fallback_providers/fallback_model chain. The
+    # hardcoded provider discovery chain below is only the convenience default
+    # for users who have not declared a fallback policy.
+    if task:
+        fb_client, fb_model, _fb_label = _try_configured_fallback_chain(
+            task, main_provider or "auto", reason="main provider unavailable")
+        if fb_client is not None:
+            return fb_client, fb_model
+    fb_client, fb_model, _fb_label = _try_main_fallback_chain(
+        task, main_provider or "auto", reason="main provider unavailable")
+    if fb_client is not None:
+        return fb_client, fb_model
+
+    # ── Step 3: aggregator / fallback chain ──────────────────────────────
     tried = []
     for label, try_fn in _get_provider_chain():
         if _is_provider_unhealthy(label):
@@ -3344,6 +3456,7 @@ def resolve_provider_client(
     api_mode: str = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> Tuple[Optional[Any], Optional[str]]:
     """Central router: given a provider name and optional model, return a
     configured client with the correct auth, base URL, and API format.
@@ -3464,7 +3577,7 @@ def resolve_provider_client(
 
     # ── Auto: try all providers in priority order ────────────────────
     if provider == "auto":
-        client, resolved = _resolve_auto(main_runtime=main_runtime)
+        client, resolved = _resolve_auto(main_runtime=main_runtime, task=task)
         if client is None:
             return None, None
         # When auto-detection lands on a non-OpenRouter provider (e.g. a
@@ -4357,11 +4470,16 @@ def _client_cache_key(
     api_mode: Optional[str] = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> tuple:
     runtime = _normalize_main_runtime(main_runtime)
     runtime_key = tuple(runtime.get(field, "") for field in _MAIN_RUNTIME_FIELDS) if provider == "auto" else ()
+    # `auto` can now resolve through task-specific or main fallback policy,
+    # so the task participates in the cache key. Non-auto providers keep the
+    # old cache shape because the explicit provider/model tuple is sufficient.
+    task_key = (task or "") if provider == "auto" else ""
     pool_hint = _pool_cache_hint(provider, main_runtime=main_runtime)
-    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision, pool_hint)
+    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision, task_key, pool_hint)
 
 
 def _store_cached_client(cache_key: tuple, client: Any, default_model: Optional[str], *, bound_loop: Any = None) -> None:
@@ -4554,6 +4672,7 @@ def _get_cached_client(
     api_mode: str = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> Tuple[Optional[Any], Optional[str]]:
     """Get or create a cached client for the given provider.
 
@@ -4591,6 +4710,7 @@ def _get_cached_client(
         api_mode=api_mode,
         main_runtime=main_runtime,
         is_vision=is_vision,
+        task=task,
     )
     with _client_cache_lock:
         if cache_key in _client_cache:
@@ -4635,6 +4755,7 @@ def _get_cached_client(
         api_mode=api_mode,
         main_runtime=runtime,
         is_vision=is_vision,
+        task=task,
     )
     if client is not None:
         # For async clients, remember which loop they were created on so we
@@ -5140,7 +5261,7 @@ def call_llm(
             if not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
-                client, final_model = _get_cached_client("auto", main_runtime=main_runtime)
+                client, final_model = _get_cached_client("auto", main_runtime=main_runtime, task=task)
         if client is None:
             raise RuntimeError(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
@@ -5466,14 +5587,19 @@ def call_llm(
 
             # Fallback order (#26882, #26803):
             #   1. User-configured fallback_chain (per-task) if set
-            #   2. Main agent model (last-resort safety net)
-            # For auto users (no explicit aux provider), use the full
-            # auto-detection chain instead — its Step 1 IS the main agent
-            # model, so users on `auto` already get main-model fallback.
+            #   2. For auto: top-level main fallback_providers/fallback_model
+            #   3. For auto: built-in auxiliary discovery chain
+            #   4. For explicit aux providers: main agent model safety net
             fb_client, fb_model, fb_label = (None, None, "")
             if is_auto:
-                fb_client, fb_model, fb_label = _try_payment_fallback(
-                    resolved_provider, task, reason=reason)
+                fb_client, fb_model, fb_label = _try_configured_fallback_chain(
+                    task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_main_fallback_chain(
+                        task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_payment_fallback(
+                        resolved_provider, task, reason=reason)
             else:
                 fb_client, fb_model, fb_label = _try_configured_fallback_chain(
                     task, resolved_provider or "auto", reason=reason)
@@ -5636,7 +5762,7 @@ async def async_call_llm(
             if not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
-                client, final_model = _get_cached_client("auto", async_mode=True)
+                client, final_model = _get_cached_client("auto", async_mode=True, main_runtime=main_runtime, task=task)
         if client is None:
             raise RuntimeError(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
@@ -5904,13 +6030,19 @@ async def async_call_llm(
 
             # Fallback order (#26882, #26803):
             #   1. User-configured fallback_chain (per-task) if set
-            #   2. Main agent model (last-resort safety net)
-            # Auto users get the full auto-detection chain instead — its
-            # Step 1 IS the main agent model.
+            #   2. For auto: top-level main fallback_providers/fallback_model
+            #   3. For auto: built-in auxiliary discovery chain
+            #   4. For explicit aux providers: main agent model safety net
             fb_client, fb_model, fb_label = (None, None, "")
             if is_auto:
-                fb_client, fb_model, fb_label = _try_payment_fallback(
-                    resolved_provider, task, reason=reason)
+                fb_client, fb_model, fb_label = _try_configured_fallback_chain(
+                    task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_main_fallback_chain(
+                        task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_payment_fallback(
+                        resolved_provider, task, reason=reason)
             else:
                 fb_client, fb_model, fb_label = _try_configured_fallback_chain(
                     task, resolved_provider or "auto", reason=reason)

apps/desktop/src/app/session/hooks/use-prompt-actions.ts

@@ -58,6 +58,7 @@ import { clearSessionTodos } from '@/store/todos'
 
 import type {
   ClientSessionState,
+  BrowserManageResponse,
   FileAttachResponse,
   HandoffFailResponse,
   HandoffRequestResponse,
@@ -1141,6 +1142,81 @@ export function usePromptActions({
           } catch (err) {
             renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
           }
+        },
+        // /browser connect|disconnect|status manages the live CDP connection on
+        // the gateway host, mirroring the TUI's browser.manage RPC. It mutates
+        // BROWSER_CDP_URL (and may launch Chrome) in the gateway process — only
+        // meaningful when that process runs on this machine, so it's gated to
+        // local connections. A remote gateway would act on the wrong host.
+        browser: async ctx => {
+          const resolved = await withSlashOutput(ctx)
+
+          if (!resolved) {
+            return
+          }
+
+          const { render: renderSlashOutput, sessionId } = resolved
+
+          if ($connection.get()?.mode === 'remote') {
+            renderSlashOutput(
+              '/browser manages a Chromium-family browser on the gateway host — only available when connected to a local gateway.'
+            )
+
+            return
+          }
+
+          const [rawAction = 'status', ...rest] = ctx.arg.trim().split(/\s+/).filter(Boolean)
+          const cmdAction = rawAction.toLowerCase()
+
+          if (!['connect', 'disconnect', 'status'].includes(cmdAction)) {
+            renderSlashOutput(
+              'usage: /browser [connect|disconnect|status] [url] · persistent: set browser.cdp_url in config.yaml'
+            )
+
+            return
+          }
+
+          const url = cmdAction === 'connect' ? rest.join(' ').trim() || 'http://127.0.0.1:9222' : undefined
+
+          if (url) {
+            renderSlashOutput(`checking Chromium-family browser remote debugging at ${url}...`)
+          }
+
+          try {
+            const result = await requestGateway<BrowserManageResponse>('browser.manage', {
+              action: cmdAction,
+              session_id: sessionId,
+              ...(url && { url })
+            })
+
+            // Without a streamed session subscription, the gateway bundles its
+            // progress lines into `messages` — flush them inline.
+            result?.messages?.forEach(message => renderSlashOutput(message))
+
+            if (cmdAction === 'status') {
+              renderSlashOutput(
+                result?.connected
+                  ? `browser connected: ${result.url || '(url unavailable)'}`
+                  : 'browser not connected (try /browser connect <url> or set browser.cdp_url in config.yaml)'
+              )
+
+              return
+            }
+
+            if (cmdAction === 'disconnect') {
+              renderSlashOutput('browser disconnected')
+
+              return
+            }
+
+            if (result?.connected) {
+              renderSlashOutput('Browser connected to live Chromium-family browser via CDP')
+              renderSlashOutput(`Endpoint: ${result.url || '(url unavailable)'}`)
+              renderSlashOutput('next browser tool call will use this CDP endpoint')
+            }
+          } catch (err) {
+            renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
+          }
         }
       }
 

apps/desktop/src/app/types.ts

@@ -46,6 +46,12 @@ export interface SlashExecResponse {
   warning?: string
 }
 
+export interface BrowserManageResponse {
+  connected?: boolean
+  url?: string
+  messages?: string[]
+}
+
 export interface SessionSteerResponse {
   // 'queued' == accepted into the live turn's steer slot (injected at the next
   // tool-result boundary); 'rejected' == no live tool window, caller queues.

apps/desktop/src/lib/desktop-slash-commands.test.ts

@@ -52,6 +52,17 @@ describe('desktop slash command curation', () => {
     expect(desktopSlashUnavailableMessage('/personality')).toBeNull()
   })
 
+  it('treats /browser as an executable action command (local-gateway connect)', () => {
+    // /browser used to be terminal-only; it now resolves to a desktop action
+    // handler that routes browser.manage RPC when the gateway is local.
+    expect(isDesktopSlashCommand('/browser')).toBe(true)
+    expect(isDesktopSlashSuggestion('/browser')).toBe(true)
+    expect(desktopSlashUnavailableMessage('/browser')).toBeNull()
+    expect(resolveDesktopCommand('/browser')?.surface).toEqual({ kind: 'action', action: 'browser' })
+    // Bare /browser expands to its sub-action options in the popover.
+    expect(resolveDesktopCommand('/browser')?.args).toBe(true)
+  })
+
   it('allows aliases to execute without cluttering the popover', () => {
     expect(isDesktopSlashSuggestion('/reset')).toBe(false)
     expect(isDesktopSlashCommand('/reset')).toBe(true)

apps/desktop/src/lib/desktop-slash-commands.ts

@@ -30,6 +30,7 @@ export interface DesktopThemeCommandOption {
  */
 export type DesktopActionId =
   | 'branch'
+  | 'browser'
   | 'handoff'
   | 'help'
   | 'new'
@@ -103,6 +104,12 @@ const DESKTOP_COMMAND_SPECS: readonly DesktopCommandSpec[] = [
   { name: '/skin', description: 'Switch desktop theme or cycle to the next one', surface: action('skin'), args: true },
   { name: '/title', description: 'Rename the current session', surface: action('title') },
   { name: '/help', description: 'Show desktop slash commands', aliases: ['/commands'], surface: action('help') },
+  {
+    name: '/browser',
+    description: 'Manage browser CDP connection [connect|disconnect|status] (local gateway only)',
+    surface: action('browser'),
+    args: true
+  },
 
   // Overlay pickers
   { name: '/model', description: 'Switch the model for this session', surface: picker('model'), hidden: true },
@@ -142,7 +149,7 @@ const DESKTOP_COMMAND_SPECS: readonly DesktopCommandSpec[] = [
 // per reason beats 40 identical object literals.
 const NO_DESKTOP_SURFACE: Record<DesktopUnavailableReason, readonly string[]> = {
   terminal: [
-    '/browser', '/busy', '/clear', '/compact', '/config', '/copy', '/cron', '/details',
+    '/busy', '/clear', '/compact', '/config', '/copy', '/cron', '/details',
     '/exit', '/footer', '/gateway', '/gquota', '/history', '/image', '/indicator', '/logs',
     '/mouse', '/paste', '/platforms', '/plugins', '/quit', '/redraw', '/reload', '/restart',
     '/sb', '/set-home', '/sethome', '/snap', '/snapshot', '/statusbar', '/toolsets', '/update', '/verbose'

docs/ink-env-flags.md

@@ -0,0 +1,68 @@
+# Ink TUI — diagnostic environment flags
+
+Non-secret behavioral knobs for the Ink engine (`ui-tui/`). These are
+**environment overrides**, not `.env` secrets — set them in your shell for a
+session, or `export` them in your shell rc to make them sticky. They mirror the
+OpenTUI engine's flags (`docs/opentui-env-flags.md`) so a single switch covers
+both engines.
+
+| Flag | Default | What it does |
+|---|---|---|
+| `HERMES_TUI_DIAGNOSTICS` | off | Master diagnostics switch. Turning it on enables the developer/profiling surface across the TUI — including the memory self-sampler below. One `export HERMES_TUI_DIAGNOSTICS=1` in your shell rc covers **every** session you start, on **either** engine. |
+| `HERMES_TUI_MEMLOG` | = `HERMES_TUI_DIAGNOSTICS` | In-process 1Hz memory self-sampling (`ui-tui/src/lib/memlog.ts`) → `~/.hermes/logs/memwatch/<boot>-<pid>.jsonl`. Defaults to the master switch; set `=1` / `=0` to force it on/off independently. |
+
+## What the memory trace captures
+
+Each Ink session, when sampling is enabled, appends one JSON line per second to
+its own file under `~/.hermes/logs/memwatch/`, keyed by boot time + pid:
+
+```json
+{"t":1781514892,"rss_kb":92148,"heap_used_kb":7234,"external_kb":2378}
+```
+
+- `t` — unix seconds.
+- `rss_kb` — resident set size (the number that matters for the native-RSS-gap
+  story: rss climbing while heap stays flat is the #15141-class signal).
+- `heap_used_kb` — V8 heap in use.
+- `external_kb` — off-heap (buffers, native allocations).
+
+**Ink emits no `mounted` / `peak_mounted` field.** Those are OpenTUI's
+windowing dev counters; Ink has no windowing, so it logs the rss/heap/external
+core only. `memwatch-report.mjs` treats `mounted` as optional, so Ink lines
+aggregate cleanly alongside OpenTUI's.
+
+## Why this exists — cross-engine memory comparison
+
+The filename scheme, directory, and line schema are **byte-compatible with
+OpenTUI's collector** (`ui-opentui/src/boundary/memlog.ts`). Both engines write
+to the same `~/.hermes/logs/memwatch/` directory, so one aggregator reads both:
+
+```sh
+# enable on either/both engines (master switch covers both)
+export HERMES_TUI_DIAGNOSTICS=1
+HERMES_TUI_ENGINE=ink     hermes --tui   # Ink session → its own .jsonl
+HERMES_TUI_ENGINE=opentui hermes --tui   # OpenTUI session → its own .jsonl
+
+# fleet table across BOTH engines' sessions:
+cd ~/github/tui-bench && node memwatch-report.mjs
+```
+
+This is what makes a true side-by-side **real-world** memory arc possible —
+cold floor → load → plateau/leak — instead of comparing OpenTUI dogfood traces
+against an Ink harness with no equivalent data.
+
+## Cost & safety
+
+- ~50 bytes/s when on; one `process.memoryUsage()` + one short append per
+  second. The interval is **unref'd** — it never keeps the process alive.
+- 14-day retention: older traces are pruned (best-effort) at start.
+- **Every failure path disables the logger silently.** Diagnostics must never
+  break the TUI — this is the one place the "errors propagate" rule is
+  intentionally inverted, matching the OpenTUI collector.
+- Off by default: regular users write nothing.
+
+## Getting a meaningful trace
+
+A short scroll-through won't show growth. For a comparison against OpenTUI's
+4–5h sessions, drive a tool-heavy 2–3h Ink session as the floor (see
+`docs/plans/opentui-ink-asymmetry-note.md` for why the harness ≠ dogfood data).

docs/opentui-dev-handoff.md

@@ -0,0 +1,120 @@
+# Handoff — OpenTUI memory + UX, continuing on the canonical branch
+
+**You are continuing the Hermes OpenTUI engine work.** This is the base operating manual; the
+user (glitch) appends specific tasks on top. Read it, then read the repo docs it points to. It
+assumes NO prior transcript/memory.
+
+## Where things are
+
+- **Canonical branch: `feat/opentui-native-engine`** (the draft PR to main, #42922).
+  `feat/opentui-memory-window` is a synonym at the *same tip* — they were consolidated. Treat
+  native-engine as canonical; if you work from memory-window, periodically
+  `git push origin HEAD:feat/opentui-native-engine` to keep them in sync, or just use native-engine.
+- The native engine source is **`ui-opentui/`**; the legacy Ink engine is `ui-tui/` (shipping
+  default, untouched by this campaign). The Python gateway is `tui_gateway/`, launcher
+  `hermes_cli/main.py`.
+- **The worktree is often the user's LIVE global `hermes`** (`~/.local/bin/hermes` symlinks into a
+  worktree's `.venv`). Consequences: (1) NEVER leave the worktree in a half-merged/conflicted state
+  — a new `hermes` session would fail to build; (2) after you land source changes, rebuild
+  `dist/main.js` so the next session picks them up; (3) `hermes-stable` is the flip-back to the
+  stock `~/.hermes/hermes-agent` install if you need to bypass the worktree.
+- Backups of pre-merge branch states exist as `backup/*` refs (recoverable via `git reset`).
+
+## Runtime, build, gate (Node 26 — NOT Bun; the port is done)
+
+```sh
+export PATH="$HOME/.local/share/fnm/node-versions/v26.3.0/installation/bin:$PATH"
+cd ui-opentui && node scripts/build.mjs            # → dist/main.js (esbuild + Solid/JSX)
+HERMES_TUI_MOUSE=1 node --experimental-ffi --no-warnings dist/main.js   # launch; quit = double Ctrl+C
+cd ui-opentui && npm run check                      # THE GATE: prettier+eslint(typed)+vitest (~700). Judge by `echo $?`, never a piped tail.
+```
+
+Never run bun here. Never run `hermes update` in the worktree (it flips the branch — recovery is
+painful). Never broad-pkill tui_gateway (other live sessions). Host RAM ~15GB, often <5GB free —
+run benches SEQUENTIALLY (the harness already wraps SUTs in `systemd-run … MemoryMax=2G`).
+
+## The docs that are the source of truth (read, and KEEP UPDATED as you change things)
+
+- `docs/opentui-memory-story.md` — ELI5 of the whole memory architecture (primitives + every decision).
+- `docs/plans/opentui-transcript-windowing.md` — windowing design (S1 spacers, S2 append-time), the
+  `correctionIsLegal` zero-jank law, pre-registered gates, SHIPPED status + S3 backlog.
+- `docs/opentui-env-flags.md` — the consolidated env-flag ledger (master switch / user / dev / plumbing).
+- `docs/opentui-upstream-alignment.md` — forkless invariant, `boundary/` shim ledger, the per-release
+  OpenTUI upgrade playbook (native-yoga is coming upstream — re-tune windowing margins when it lands).
+- the bench suite (cells, harness, live-attach, memwatch) now lives in its own
+  repo: **tui-bench** (`github.com/NousResearch/tui-bench`); see its `README.md`.
+- `ui-opentui/README.md` — Node 26 onboarding (fnm setup that doesn't disturb other projects).
+- `docs/plans/ink-memory-adversarial-review.md` — Ink's memory weaknesses (F1–F10, the turnabout).
+- `docs/plans/gateway-death-forensics.md`, `docs/plans/workorder-2026-06-11-results.md`,
+  `docs/plans/rebase-from-main-spec.md` — forensics, the merge-bar verdict, the rebase plan.
+
+## Workflow (this is how the last 60+ commits were produced with ~zero rework)
+
+1. **Subagent-driven** (skill: `subagent-driven-development`): one implementer per task with a TIGHT
+   file fence ("you own exactly these files; `git diff --cached --stat` before commit, abort on
+   out-of-fence"), a mandatory `opentui` skill read FIRST for any renderable work, and a gate judged
+   by exit code. Verify the self-report YOURSELF (re-run the gate, read the riskiest hunks, check the
+   commit file-list) — a subagent "✅ done" is a claim, not a fact.
+2. **Adversarial review** after a task: a fresh read-only reviewer (Explore-type) with NAMED attack
+   surfaces. Then ADJUDICATE in code — reviewers over-flag; ~half of "blockers" don't survive a read.
+3. **Parallel implementers are safe ONLY with disjoint file fences.** Read-only recon agents
+   parallelize freely.
+4. **Live smoke catches what headless can't** — tmux + the `tmux-pane-screenshot` skill for real
+   colored frames. The demo: `node scripts/build.mjs scripts/demo.tsx .demo` then
+   `DEMO_TOTAL=2000 … node --experimental-ffi --no-warnings .demo/demo.js`.
+5. Commit format `opentui(v6): …`, **NO attribution lines**. The user's standing instruction is
+   "commit + push as you land things" — honor it; otherwise don't push without asking. Edit large
+   load-bearing files (the Python launcher, `store.ts`) DIRECTLY, never via subagent.
+
+## Dogfooding (the user works on this FROM the hermes TUI)
+
+`export HERMES_TUI_DIAGNOSTICS=1` in the shell rc turns on, for every session: the `/mem` +
+`/heapdump` slash commands, window-stats, and **fleet memory self-logging** to
+`~/.hermes/logs/memwatch/<boot>-<pid>.jsonl`. Aggregate all sessions with
+`node memwatch-report.mjs` from the **tui-bench** repo
+(`github.com/NousResearch/tui-bench`) (per-session baseline/peak/slope + SLOPE/PEAK/MOUNTED anomaly
+flags). Chase a flagged session with tui-bench's `live-attach.sh <pid> --heap`. The discipline: live
+anomaly → encode as a bench cell → fix → validate against live sessions again.
+
+## Current state (2026-06) + the ranked backlog
+
+Windowing SHIPPED: 2k-msg peak ~300MB (was 686; Ink 234), scroll p99 6ms, cap restored 1000→3000,
+determinism digest unchanged, peak mounted ~31 rows. Live sessions peak <200MB. The transcript is no
+longer the biggest lever — the ~160MB floor is ≈104MB Node+OpenTUI runtime + **≈55MB tool/skill
+catalogs hydrated at boot**. Ranked next levers:
+
+1. **W3 — 1GB V8 heap default** (small, ~free): set the unconstrained default in
+   `_resolve_tui_heap_mb`; both engines are Node now so both inherit it. Ink half = separate gated
+   commit (shipping engine). Measured −90MB at bench scale.
+2. **cg_peak harness fix** (small): the cgroup `memory.peak` field is polluted (shared across runs) —
+   reset/scope it before quoting tui-bench's `report.html` again. Trust `vmhwm_kb` + `samples[].rss_kb`.
+3. **New bench cells** (before W1, as its baselines): `resume-1900` (real p99 shape: time-to-first-
+   paint + post-hydration RSS) and `10MB-tool-output` (the F1 byte-unbounded class). Run BOTH engines.
+4. **Catalog lazy-load** (new, promoted by live data): don't hydrate 1,185 tools at boot — fetch on
+   picker-open. Attacks the ≈55MB floor; pays on EVERY session (median is 20 msgs). Likely cheaper
+   than W1.
+5. **W1 thin renderer** (structural, biggest): bodies live in the gateway (SQLite); TUI keeps ~300B
+   stubs + fetches bodies for the window only. Design the gateway windowed-read RPC FIRST. WATCH: `/copy`
+   and the ⧉ block-copy read store parts — they need a fetch-on-demand fallback or W1 ships a copy regression.
+6. **Standing**: when native-yoga OpenTUI ships, run the upgrade playbook (re-bench, re-tune margins,
+   audit the shim ledger). Three questions to relay to the OpenTUI maintainer are in the alignment doc.
+
+## What NOT to do
+- Don't copy opencode's 100-msg store cap (user's p90 session is 182 msgs — it would truncate normal use).
+- Don't reintroduce estimate-correction scroll jank (the user explicitly vetoed it; `correctionIsLegal` forbids it).
+- Don't cite the obsolete "~210MB bun renderer / +120MB" memory figures — pre-port, pre-windowing, wrong.
+- Don't push/PR without the standing OK; don't commit `.plans/` scratch unless asked.
+
+## Suggested skills
+(All available from the Hermes TUI agent too — this is the dogfooding surface. Curated to the load-bearing set, not the full ~40-skill catalog.)
+- `opentui-tui-engineering` — the workflow/architecture/pitfalls layer for `ui-opentui/` (just updated).
+- `hermes-tui-architecture` — the Hermes-specific TUI facts (launch pipeline, both engines; just updated).
+- `opentui` — the offline renderable-API doc set; mandatory `skill_view` before any view/renderable code.
+- `subagent-driven-development` — the process spine for parallel/heavy work.
+- `tmux-pane-screenshot` — real colored PNG of a tmux pane for visual verification (ported
+  into hermes skills 2026-06-13). Use: `bash ~/.hermes/skills/software-development/
+  tmux-pane-screenshot/scripts/tshot.sh <session:win.pane> out.png 2`, then Read the PNG.
+  `freeze` (~/go/bin) + the resvg rasterizer are shared/system-wide — works as-is.
+- `effect-ts` — for the Effect-at-boundary entry/lifecycle code.
+- `superpowers:brainstorming` — before committing to a memory-architecture design (e.g. W1's store split).
+- `systematic-debugging` — if a gate fails; root-cause before patching.

docs/opentui-env-flags.md

@@ -0,0 +1,81 @@
+# OpenTUI env flags — the consolidated ledger
+
+Every environment variable the OpenTUI TUI reads (grep-verified 2026-06-12),
+classified by who should ever touch it. The design rule shipped with this doc:
+**regular users see zero diagnostic surface by default; one master switch
+(`HERMES_TUI_DIAGNOSTICS=1`) turns all of it on when needed.**
+
+## 1. The master switch
+
+| var | default | effect |
+|---|---|---|
+| `HERMES_TUI_DIAGNOSTICS` | **off** | Enables the diagnostic slash commands (`/mem`, `/heapdump`). While off they're hidden from `/help` (client-side filter) and invoking them prints the enable hint rather than executing. They never appear in slash *completion* in either state — completion is gateway-driven and these are client-only commands the gateway doesn't know (an adversarial review confirmed there's no bypass path; if a SERVER command named `mem`/`heapdump` is ever added it must be gated gateway-side too — the client gate would shadow but not hide it). Also flips the *default* of `HERMES_TUI_WINDOW_STATS` to on. Not a secret — support flows are "relaunch with `HERMES_TUI_DIAGNOSTICS=1`". |
+
+## 2. User-facing configuration (fine to document publicly)
+
+| var | default | effect |
+|---|---|---|
+| `HERMES_TUI_ENGINE` | auto (`opentui` if Node≥26.3 + built, else `ink`) | Engine pick; also `display.tui_engine` in config.yaml. |
+| `HERMES_TUI_MOUSE` / `HERMES_TUI_MOUSE_TRACKING` / `HERMES_TUI_DISABLE_MOUSE` | on | Mouse support (wheel scroll, selection, click-to-expand). **Defers to Ink's env surface (`logic/env.ts` `resolveMouseEnabled`):** precedence is `HERMES_TUI_MOUSE_TRACKING` (toggle, force knob) > `HERMES_TUI_DISABLE_MOUSE=1` (legacy kill switch) > `HERMES_TUI_MOUSE` (OpenTUI-native alias, kept — also what the launcher sets) > default on. OpenTUI's renderer mouse is a single boolean, so Ink's granular off\|wheel\|buttons\|all collapses to on/off (the granular mode lives in `display.mouse_tracking` config). |
+| `HERMES_TUI_SCROLL_SPEED` (alias `CLAUDE_CODE_SCROLL_SPEED`) | native | Wheel-scroll speed multiplier (Ink parity). UNSET → OpenTUI's native scroll acceleration (untouched). A positive value (clamped to (0,20]) installs a constant-multiplier `ScrollAcceleration` on the transcript scrollbox (`view/transcript.tsx`). |
+| `HERMES_TUI_NO_CONFIRM` | off | Skip the destructive-action confirm step (`/clear`, `/new`) and run immediately (Ink parity, `NO_CONFIRM_DESTRUCTIVE`). Wired at the `confirm` seam (`entry/main.tsx`). |
+| `HERMES_TUI_MAX_MESSAGES` | ceiling | Scrollback rows kept in the TUI. Can LOWER the ceiling, never raise: 3000 with windowing, 1000 with windowing off (handle-table safety). |
+| `HERMES_TUI_TOOL_OUTPUT_LINES` | unlimited | Cap expanded tool-output lines (set a number to restore a cap). |
+| `HERMES_TUI_TOOL_OUTPUTS` | **on** | Keep rich tool-call OUTPUTS (full result body + raw result/args dicts). `=off` drops both the RENDER and the STORE of those bodies (Ink parity: only a one-line context preview + name/duration/error/diff survive) — the memory lever for the OpenTUI-vs-Ink retention asymmetry, and what the bench launches OpenTUI with for the fair engine-overhead comparison (W3). Diffs (file-edit) are KEPT either way. |
+| `HERMES_TUI_HEAP_MB` | cgroup-aware (default 8192) | V8 `--max-old-space-size` (MB) for BOTH engines. Highest precedence (then `display.tui_heap_mb` config, then the cgroup-75% fallback). Set it LOW for a low-mem session (still cgroup-clamped on top so it never exceeds the container); raise it to lift the ceiling. The low-mem opt-in signal that also arms `HERMES_TUI_PROACTIVE_GC` (W1). |
+| `HERMES_TUI_PROACTIVE_GC` | = low-`HERMES_TUI_HEAP_MB` (≤4096) | Idle-gated `global.gc()` for the low-mem path. Defaults ON only when a low heap cap is set (so the knobs compose); `=on`/`=off` forces it. Needs `--expose-gc` (the OpenTUI argv now carries it). Never runs mid-stream; tightens cadence above 400MB RSS but stays idle-gated. OpenTUI-only — Ink never GCs proactively (W2). |
+| `HERMES_TUI_COMPOSER_ROWS` | default rows | Composer height. |
+
+## 3. Escape hatches & tuning (dev-facing, individually settable)
+
+| var | default | effect |
+|---|---|---|
+| `HERMES_TUI_WINDOWING` | **on** | `0` = bit-exact pre-windowing renderer (every row mounts; cap clamps back to 1000). The A/B + regression escape hatch. |
+| `HERMES_TUI_WINDOW_IDLE_MS` | ~1000 | Idle-measure pulse cadence (the spacer-exactness march). Test knob. |
+| `HERMES_TUI_WINDOW_STATS` | = `HERMES_TUI_DIAGNOSTICS` | Exposes live/peak mounted-row counters (`globalThis.__hermesTuiWindowStats`) for tui-bench's live-attach reads. |
+| `HERMES_TUI_MEMLOG` | = `HERMES_TUI_DIAGNOSTICS` | In-process 1Hz memory self-sampling (`boundary/memlog.ts`) → `~/.hermes/logs/memwatch/<boot>-<pid>.jsonl` (rss/heap/external + mounted rows; 14-day retention). Fleet view: `node memwatch-report.mjs` from the tui-bench repo (`github.com/NousResearch/tui-bench`). The "monitor all my sessions" answer: one `export HERMES_TUI_DIAGNOSTICS=1` in your shell rc covers every session. |
+| `HERMES_TUI_LOG_LEVEL` / `HERMES_TUI_LOG_FILE` | engine defaults | Logging verbosity/destination (`/logs` reads the ring buffer regardless). Deliberately independent of the master switch — support often wants logs without the full diag surface. |
+| `HERMES_HEAPDUMP_ON_START` | off | Write one V8 heap snapshot at boot (Ink parity). A deliberate baseline-capture escape hatch that BYPASSES the diagnostics master switch; lands at `$HERMES_HOME/logs/opentui-heap-<ts>.heapsnapshot` and echoes the path as a system line (`entry/main.tsx`). |
+| `HERMES_TUI_NOTIFY` | on | Desktop-notification kill switch (`=0`/`false`/`off` silences the "waiting on you" pings). The ping itself goes through the renderer's native `triggerNotification` (protocol detection + tmux/Zellij wrapping); the window title is not gated by this. |
+
+## 4. Internal plumbing (set by the launcher/tui-bench/tests — humans never set these)
+
+| var | set by | effect |
+|---|---|---|
+| `HERMES_PYTHON`, `HERMES_PYTHON_SRC_ROOT`, `HERMES_CWD` | launcher / bench | Which gateway python + repo root + cwd the TUI spawns against (the bench's fake-gateway seam). |
+| `HERMES_TUI_ACTIVE_SESSION_FILE` | launcher/bench | Session handoff file. |
+| `HERMES_TUI_RESUME`, `HERMES_TUI_QUERY`, `HERMES_TUI_PROMPT`, `HERMES_TUI_IMAGE`, `HERMES_TUI_FAKE` | launcher/tests | Resume-at-boot; seeded prompt (`--tui "prompt"`: launcher sets `HERMES_TUI_QUERY`, the engine reads QUERY > the `HERMES_TUI_PROMPT` alias > a bare argv tail — `logic/env.ts` `startupPrompt`); seeded image PATH (`--image`: `HERMES_TUI_IMAGE`, `image.attach`ed before the prompt — `startupImage`, attach in `postSessionSetup`); fake-mode. |
+| `HERMES_AUTO_HEAPDUMP*` (`_COOLDOWN_MS`/`_MAX_BYTES`), `HERMES_HEAPDUMP_DIR`, `HERMES_HEAPDUMP_MAX_BYTES` | — | **NOT read by the OpenTUI engine (deliberate).** The engine ports Ink's #34095 silent-death early-WARNING (a transcript system line, `boundary/memoryMonitor.ts`) but NOT the auto heap-SNAPSHOT capture — the always-on memlog NDJSON trace is the diagnosis path, and its rss-vs-heap divergence is the better diagnostic for the native-RSS leak class (#15141) a V8 snapshot captures poorly. So the #41948 disk-fill safety set (gate/cooldown/byte-cap/dir) has no consumer here. `HERMES_HEAPDUMP_ON_START` (manual one-shot, §3) is the only heapdump knob the engine honors. |
+| `HERMES_TUI_RPC_TIMEOUT_MS`, `HERMES_TUI_STARTUP_TIMEOUT_MS` | tests/CI | Protocol timeouts. |
+| (`ui-tui` only) `HERMES_TUI_MEMSAMPLE_FD/MS` | bench | Ink fd-3 node sampler. |
+
+## 5. Ink flags NOT ported — handled natively or out of scope
+
+These exist on the legacy Ink TUI (`ui-tui/`) and are deliberately **not** read
+by the OpenTUI engine. Documented so a missing flag reads as a decision, not a gap.
+
+| Ink flag | why not ported |
+|---|---|
+| `HERMES_TUI_TRUECOLOR` | OpenTUI core does COLORTERM/truecolor detection natively — the Ink force-truecolor hack is a fork workaround we shed. |
+| `HERMES_TUI_FORCE_OSC52` | OpenTUI core owns OSC52 clipboard as a primitive; no fallback hint needed. |
+| `HERMES_TUI_INLINE` / `HERMES_TUI_TERMUX_MODE` / `HERMES_TUI_TERMUX_FAST_ECHO` | Termux/primary-buffer accommodations. OpenTUI's native FFI floor (Node ≥26.3 + `--experimental-ffi`) is absent on Termux, so those sessions stay on **Ink** — these are correctly N/A for the OpenTUI engine. |
+| `HERMES_TUI_FPS` | Ink FPS overlay; the OpenTUI equivalent is the diag/window-stats surface (`HERMES_TUI_WINDOW_STATS`). Not parity-critical. |
+| `HERMES_DEV_CREDITS` / `HERMES_DEV_PERF*` | Dev-only throwaway scaffolding (live-spend readout, perf logging) — not user parity. |
+| `HERMES_BIN` / `HERMES_TUI_GATEWAY_URL` / `HERMES_TUI_SIDECAR_URL` | External-CLI / remote-gateway-URL overrides. OpenTUI spawns its gateway via the Effect boundary (`liveGateway.ts`) and does not shell out to `hermes` or take an external gateway URL. |
+| `HERMES_VOICE` | Voice mode is tracked on the OpenTUI parity backlog separately, not here. |
+
+## How the pieces compose (the support script)
+
+- Regular user, normal day: zero flags, zero diagnostic commands visible.
+- "My TUI feels heavy" support flow: `HERMES_TUI_DIAGNOSTICS=1 hermes` → `/mem`
+  for the live numbers, `/heapdump` for a snapshot to attach, window stats
+  exposed for tui-bench's `live-attach.sh <pid>` to read.
+- Developer profiling: same master switch + the individual knobs
+  (`HERMES_TUI_WINDOWING=0` A/B, `WINDOW_IDLE_MS` tuning) as needed.
+- Anything in section 4 appearing in a user-facing doc is a bug.
+
+Gating implementation: `logic/env.ts` (`diagnosticsEnabled()`),
+`logic/slash.ts` (`DIAGNOSTIC_COMMANDS` — dispatch hint, help + completion
+filtering), `view/transcript.tsx` (stats default). Tests:
+`slash.test.ts` (gating both states), `utilityCommands.test.ts` (commands
+themselves, gate enabled suite-wide).

docs/opentui-memory-story.md

@@ -0,0 +1,207 @@
+# How the OpenTUI transcript got from 686MB to ~300MB — the full story
+
+*For: glitch. Branch: `feat/opentui-memory-window`. Everything here is measured,
+not vibes; every number has a result JSON in the **tui-bench** repo's `results/` (`github.com/NousResearch/tui-bench`).*
+
+---
+
+## 1. The cast of characters (the primitives, bottom-up)
+
+To understand where the memory went, you need to know who's holding it. Six
+layers, from the screen up:
+
+**The terminal grid.** Your terminal is a spreadsheet of character cells.
+Nobody pays per-message here — tmux holds ~5MB flat no matter how long the
+session is (we measured). The terminal is never the problem.
+
+**The OpenTUI native renderer (Zig).** A compiled library that owns the
+"frame buffer" — the grid of cells about to be painted. Every piece of text the
+TUI shows lives in a native **TextBuffer** (the characters + their colors),
+viewed through a **TextBufferView**, styled by a **SyntaxStyle**. Each of those
+is a **native handle** — a ticket into one global table that has only **65,535
+slots, total, ever** (16-bit indices — like a coat check with 65k hooks).
+Destroying a renderable returns its tickets, so the constraint is not "how much
+have you ever created" but **"how much is alive right now."**
+
+**Renderables.** OpenTUI's UI objects — `<text>`, `<box>`, `<markdown>`,
+`<code>`, `<scrollbox>`. One transcript row (a message with its tool calls,
+markdown, code blocks, copy chips) is a *tree* of these: **~16 text renderables
+≈ 47 native handles ≈ ~250–340KB of RSS, per row.** This is the number that
+drives everything. 1,400 mounted rows × 47 handles = table full = the crash we
+root-caused last week.
+
+**Yoga (the layout engine, WASM).** Every renderable also has a Yoga node —
+Yoga is the flexbox calculator that decides where boxes go. OpenTUI ships it
+compiled to **WebAssembly**, and WASM has a brutal property: its memory can
+**grow but never shrink** back to the OS. So the peak number of
+*simultaneously-mounted* renderables sets a high-water mark you pay **forever**,
+even after everything is destroyed. (Fun fact from this week's forensics: we
+spent two days believing Ink had this disease. It doesn't — our Ink fork swapped
+Yoga-WASM for a plain TypeScript port at fork creation. **We** are the ones
+running layout in WASM. The accusation was true; we just had the defendant
+wrong.)
+
+**Solid (the view framework).** Renders each store message into a row via
+`<For>`. The property we exploit: Solid mounts/unmounts *surgically* — remove a
+row from what the component returns and Solid destroys exactly that row's
+renderables (returning its handles and freeing its Yoga nodes), touching
+nothing else. No virtual-DOM diffing, no collateral re-renders.
+
+**V8 (the JavaScript engine) + the store.** The store keeps every message as JS
+strings/objects. V8's garbage collector is *lazy by design*: with the default
+8GB ceiling we launch with, it sees no reason to clean up aggressively, so RSS
+includes a lot of "collectible but not yet collected" garbage. Cheap to fix,
+worth real MB (measured below).
+
+**The scrollbox.** One detail that fooled everyone at some point:
+`viewportCulling` (on by default) skips *drawing* offscreen rows — but they stay
+fully **mounted**: handles held, Yoga nodes alive, memory paid. Culling saves
+paint time, not memory. That misunderstanding is half the reason the "rolling
+store cap" was expected to be enough, and wasn't.
+
+## 2. Why it was 686MB
+
+Simple arithmetic. The old TUI mounted **every message in the store** as a full
+renderable tree. 2,000 messages × ~16 renderables × (handles + Yoga nodes +
+text buffers + V8 objects) ≈ 670–690MB, growing ~300MB per 1,000 messages. And
+at ~1,400 rows the handle table filled: first a hard crash (exit 7), then —
+after our containment fix — survival with **unstyled text** past that point,
+plus a cap clamped from 3,000 rows down to 1,000 as the price of not crashing.
+
+Ink, meanwhile, sat at ~234MB at the same workload, because Ink only ever
+mounts the rows near your viewport (~84–400 live nodes). Its memory is the
+*data* plus some caches — not the *view*.
+
+## 3. The decisions, in order
+
+### Decision 1: virtualize the view, don't starve the store
+
+Two ways to cut view memory: keep fewer messages (opencode's answer — they keep
+100 and delete the rest from memory; transcript truth lives on their server), or
+keep all messages but only *materialize* the ones near the viewport. You vetoed
+the first (your p90 session is 182 messages — a 100-row store truncates normal
+sessions), so: **windowing**. Notably the OpenTUI devs confirmed this week that
+framework-level virtualization is the intended path — the engine doesn't ship
+it out of the box, and opencode never built it. We did.
+
+### Decision 2: exact heights, recorded at unmount — never estimates in your face
+
+This is the load-bearing idea, and it's where we beat Ink at its own game.
+
+The hard problem of any virtualized list: an unmounted row still needs to
+occupy its correct *height*, or the scrollbar lies and content jumps. Ink
+solves it by **guessing** heights and correcting after measurement — those
+corrections are precisely the 83–101ms scroll stutters you hate. You explicitly
+vetoed "estimate-correction jank" as a model.
+
+Our advantage: OpenTUI lays out with real, queryable heights. So when a row
+scrolls out of the window, we record its **exact laid-out height** (an
+`onSizeChange` hook fires inside layout, pre-paint) and replace the row with an
+empty `<box height={exactly-that}/>` — a **spacer**: one Yoga node, zero text
+buffers, zero native handles. Think of a bookshelf where books you're not
+reading are swapped for cardboard sleeves cut to *exactly* the book's
+thickness: the shelf never shifts, and you can't tell from across the room.
+
+The window is your viewport ± one viewport of margin (plus hysteresis so it
+doesn't thrash at the edges). Scroll near a spacer and the real row remounts —
+at the recorded height, so nothing moves.
+
+And one **law**, written into the code as `correctionIsLegal`: a spacer's
+height may only ever be corrected where you *cannot see it* — fully above the
+viewport (with the scroll position compensated in the same frame, so the world
+doesn't move) or fully below it. A correction that would shift visible content
+is forbidden, structurally. Jank isn't tuned down; it's outlawed.
+
+### Decision 3 (the S2 insight): adjudicate on *append*, not just on scroll
+
+S1 alone got 686 → 518MB. Why not more? Because of *when* windowing decided.
+S1 re-decided the window when you **scrolled**. But during a streaming burst —
+an agent turn dumping hundreds of rows — you don't scroll; rows arrive, each
+mounting fully, and only get demoted later. That transient pile-up is mostly
+invisible in steady-state numbers… except for Yoga-WASM, where **the transient
+peak is permanent** (memory never shrinks). The burst was quietly ratcheting
+the floor.
+
+S2 makes the window recompute on **transcript growth**: while you're pinned at
+the bottom, the window anchors to the content *bottom*, so a row that falls
+more than a margin behind the live edge becomes a spacer the moment it's
+measured — not whenever you next scroll. Measured result: across a 1,500-row
+burst, the peak number of simultaneously-mounted rows is **31**.
+
+Same trick for **resume**: opening a 2,000-message session used to mount all of
+it (transient peak again — paid forever). Now resume mounts only the bottom
+window; everything above starts as spacers using a line-count estimate, and an
+idle-time "measure march" quietly mounts ten rows at a time near the window
+edge, records their true heights, and swaps them back — all outside the
+viewport, all invisible by the law above.
+
+### Decision 4: rows that must never be windowed
+
+Windowing has to know what it's not allowed to touch:
+- **Streaming rows** — the native markdown renderer streams incrementally;
+  unmounting mid-stream would restart it visibly.
+- **The bottom 30 rows** — the region you actually live in.
+- **Rows under a mouse selection** — the review caught that a lingering
+  highlight originally froze windowing *forever* (memory regrowing silently).
+  Fixed: only an active drag pauses swaps, and selected rows get pinned, so
+  copy is byte-exact while everything else keeps windowing.
+
+### Decision 5: give back the scrollback (cap 1,000 → 3,000)
+
+The 1,000-row clamp existed only because mounted-rows == stored-rows and the
+handle table dies at ~1,400. With windowing, mounted ≈ 31 regardless of store
+size — so the cap went back to the originally-shipped 3,000. It's
+windowing-aware: the `HERMES_TUI_WINDOWING=0` escape hatch (which mounts
+everything again) keeps the safe 1,000.
+
+### Decision 6 (measured, not yet shipped as default): right-size the V8 heap
+
+Running the windowed TUI with a 512MB heap ceiling instead of 8GB forced V8 to
+actually collect: another −90MB with zero latency cost. That's queued as a
+launcher default change (~1GB), for both engines.
+
+## 4. The scoreboard
+
+At 2,000 messages (your real p99 session size — yes, we checked your DB:
+median session is 20 messages, p99 is 1,941):
+
+| | peak memory | scroll p99 (slowest 1-in-100) |
+|---|---|---|
+| OpenTUI before | 686MB | 16ms |
+| + S1 windowing | 518MB | 16ms |
+| + S2 append/resume windowing | **300–375MB** | **6ms** |
+| Ink (reference) | 229–246MB | ~100ms |
+
+At the **3,000-message stress** with the restored triple-size scrollback:
+**360MB, fully styled, scroll p99 8ms** — a workload that six days ago crashed
+the process, and three days ago survived only by dropping syntax colors.
+
+Scroll got *faster* because there are simply fewer live renderables to walk.
+The determinism gate stayed **byte-identical** — the windowed TUI's settled
+frame is provably the same pixels as before. And the live smoke (2,000-message
+session: full sweep to the top, resize storm, back to bottom) returned a frame
+pixel-identical to boot, with deep history fully syntax-highlighted — something
+the pre-windowing TUI literally could not do.
+
+## 5. What's honestly still open
+
+- The remaining ~60–120MB over Ink is mostly the **store's JS strings** and
+  process baseline — the view is no longer the problem. The structural fix is
+  the **thin renderer** (W1): bodies live in the Python gateway (which already
+  has them in SQLite); the TUI keeps ~300-byte stubs and fetches bodies only
+  for the window. That also fixes the class of problem neither engine handles
+  today: a single 10MB tool output.
+- Two accepted, documented limits: scrollbar-*jumping* deep into a freshly
+  resumed session can land on estimate-height rows that snap to true height as
+  they enter view (normal scrolling doesn't — the margin pre-measures; the idle
+  march erodes the exposure over time), and a tool you expanded, scrolled far
+  away from, then returned to will have re-collapsed (state is component-local;
+  hoisting it to the store is queued).
+- Everything is behind `HERMES_TUI_WINDOWING` (default on, `0` = bit-exact old
+  behavior) — a one-env escape hatch if anything feels off in real use.
+
+*Where to verify: the **tui-bench** repo's `results/` (`github.com/NousResearch/tui-bench`; every number above), the design+gates doc
+`docs/plans/opentui-transcript-windowing.md`, tests in
+`ui-opentui/src/test/window.test.ts` and `transcriptWindow.test.tsx` (the
+zero-jank invariants are literal assertions: identical scrollHeight windowed
+vs not, byte-stable frames across corrections).*

docs/opentui-native-engine.md

@@ -0,0 +1,432 @@
+# OpenTUI native engine — PR documentation
+
+**Branch:** `feat/opentui-native-engine` · **Base:** `origin/main` (merged in; HEAD is at `~main`)
+**New engine root:** `ui-opentui/` (Node 26 + `@opentui/core` 0.4.1 + `@opentui/solid`, Effect at the boundary)
+**Legacy engine root:** `ui-tui/` (React + the `@hermes/ink` fork at `ui-tui/packages/hermes-ink/`)
+
+> This is the canonical in-repo doc for the PR. The companion interactive HTML
+> write-up (`~/projects/opentui-perf-writeup/index.html`) is the case/benchmark
+> deep-dive; this doc is the reviewable text version + the four things review
+> actually needs: **(1) the LoC reduction math, (2) the measured perf deltas,
+> (3) the real UI divergence (with screenshots), (4) the non-core / kitchen-sink
+> change audit.**
+
+This PR adds a from-scratch native terminal UI built on OpenTUI, intended to
+replace the React/Ink TUI **and the Ink fork we maintain alone**. It currently
+ships as a parallel engine (Ink untouched, auto-fallback), selected by
+`HERMES_TUI_ENGINE` env > `display.tui_engine` config > auto (OpenTUI when the
+host is Node ≥ 26.3 with the built bundle, else Ink). **100% parity with the Ink
+TUI is the bar.**
+
+---
+
+## 1. Line-of-code reduction (the headline maintenance win)
+
+All counts are **git-tracked files only** (respects `.gitignore`; `dist/` and
+`node_modules/` are untracked and excluded). Measured live on this branch at
+`~HEAD`. "Code" = `.ts/.tsx/.js/.jsx` only; "total" includes config/json/md.
+
+### What gets *removed* when Ink is retired
+
+| Area | Files | Total lines | Code lines (ts/tsx/js) | Non-blank code |
+|---|---:|---:|---:|---:|
+| `ui-tui/src/` — Ink **consumer app** (our React/Ink view code) | 204 | 40,422 | 40,422 | 33,550 |
+| `ui-tui/packages/hermes-ink/` — **the fork** (`@hermes/ink`) | 148 | 28,167 | 28,113 | 23,718 |
+| **`ui-tui/` whole tree (tracked)** | **362** | **69,320** | **68,831** | **57,545** |
+
+The `ui-tui/` whole-tree number (69,320) also folds in a handful of build
+scripts, `.prettierrc`, `package.json`, etc. The two rows above it are the
+load-bearing split:
+
+- **The fork alone is 28,167 LOC across 148 files** — code we own and can never
+  sync from upstream. Upstream Ink v6.8.0 `src/` is ~7,259 LOC, so the fork's
+  renderer core is **~3.2× the size of stock Ink**. (Cross-checked against the
+  HTML write-up's `ink-fork-analysis.json`: 28,111 LOC / 148 files — the 56-line
+  delta is a single tracked JSON the file-level count includes.)
+- **The consumer app is another 40,422 LOC** — React components/hooks that only
+  exist to drive Ink.
+
+### What gets *added*
+
+| Area | Files | Total lines | Code lines | Non-blank code |
+|---|---:|---:|---:|---:|
+| `ui-opentui/src/` — new engine (app code **+ its own tests**) | 153 | 28,763 | 28,763 | 26,495 |
+|   ↳ non-test (app code only) | 97 | 16,628 | 16,628 | 15,450 |
+|   ↳ tests (`src/test/`) | 56 | 12,135 | 12,135 | 11,045 |
+| Tree-sitter grammars (`python`…`toml`) | 0 | 0 | 0 | 0 |
+| **`ui-opentui/` whole tree (tracked)** | **~170** | **~34,800** | **29,614** | **27,283** |
+
+> Tree-sitter grammars carry **zero repo lines**: the engine declares the 10
+> extra grammars as remote URLs (`src/boundary/parsers.manifest.json`) and
+> OpenTUI fetches+caches each `.wasm`/`.scm` on first use into
+> `~/.hermes/cache/opentui-parsers/` (à la opencode, which vendors none). An
+> earlier revision vendored them as 37,302 checked-in binary lines (10 `.wasm` +
+> 10 `.scm`); that's gone — code lines and total lines now move together.
+
+### The net reduction (code lines, the honest comparison)
+
+| Comparison | Removed (ts/tsx/js) | Added (ts/tsx/js) | Net change |
+|---|---:|---:|---:|
+| **Incl. fork** — retire all of `ui-tui/` vs add `ui-opentui/src` | −68,831 | +28,763 | **−40,068 LOC (−58%)** |
+| **Incl. fork, app-vs-app** (exclude both test suites) | −56,463¹ | +16,628 | **−39,835 LOC (−71%)** |
+| **Excl. fork** — only the Ink *consumer app* vs new engine | −40,422 | +28,763 | **−11,659 LOC (−29%)** |
+| **The fork in isolation** (the unsyncable liability we shed) | −28,113 | — | **−28,113 code lines deleted outright (28,167 incl. its 1 config file)** |
+
+¹ `ui-tui/src` non-test = 28,350 LOC + fork (≈ all 28,113 code lines are non-test;
+it carries only ~54 config lines) = 56,463. (`ui-tui/src` carries 80 test files /
+12,072 LOC; the new engine carries 56 test files / 12,135 LOC.)
+
+**Read it this way:**
+
+- **The cleanest single number: ~−40k code lines net** (retire all of `ui-tui/`,
+  add `ui-opentui/src`). That is a **~58% reduction in the TUI's
+  hand-maintained surface**, and it *includes* the new engine's full 56-file test
+  suite.
+- **The most important number is the fork: −28,167 LOC of unsyncable engine
+  code** disappears. That is the load-bearing maintenance win — it's not just
+  fewer lines, it's lines we are the *sole* maintainer of (own reconciler, ANSI
+  parser, scrollbox, selection/OSC52, hand-rolled memory eviction, Yoga binding).
+- **Even excluding the fork** — i.e. if you imagine upstream Ink were free — the
+  app rewrite is still a net reduction (−11,659 LOC) because the new engine
+  mounts OpenTUI built-ins instead of hand-building components.
+
+### Caveat on the comparison (keep it honest for review)
+
+- These are **whole-tree retirements vs a single source dir add.** If/when Ink is
+  deleted, the `ui-tui/` `package.json`, lockfile, and build scripts go too; the
+  table counts `ui-tui/src` + the fork as the apples-to-apples "hand-maintained
+  TS" figure.
+- **Tree-sitter grammars are NOT vendored.** The 10 extra grammars are declared
+  as remote URLs (`src/boundary/parsers.manifest.json`); OpenTUI fetches each
+  `.wasm`/`.scm` on first use of a language and caches it under
+  `~/.hermes/cache/opentui-parsers/` (profile-aware, set via
+  `HERMES_TUI_PARSER_CACHE` by the launcher). Registration does **zero** network;
+  the fetch is lazy and off the boot critical path, and an unreachable
+  GitHub/air-gapped env degrades that language to plain text — never a throw. This
+  replaces an earlier revision that vendored 37k binary lines, so the repo no
+  longer grows on disk for syntax highlighting. (Trade-off: first-use-per-language
+  needs network to `github.com`/`raw.githubusercontent.com`; pre-seed the cache in
+  a Docker build if you need offline highlighting.)
+- Python/backend LoC is **not** part of this reduction: `tui_gateway/` (~12k LOC)
+  is **shared by both engines** and stays. See §4.
+
+---
+
+## 2. Performance (CPU / latency / memory)
+
+Measured with the `tui-bench` harness driving **both engines on a real PTY
+120×40**, fake gateway feeding deterministic events, `/proc`-sampled identically,
+each SUT under `systemd-run --scope -p MemoryMax=2G -p MemorySwapMax=0`,
+sequential with a load-gate + 10s cooldown. Determinism gate **GREEN**, 71 result
+files, 0 cell errors, 3 reps/cell, `@opentui/core` 0.4.1 native-yoga
+(`libopentui.so`, no `yoga.wasm`). Every number traces to a `summary.<field>` in
+a result dir. Source: `~/projects/opentui-html/bench-numbers.json` (frozen
+2026-06-14, build under test `1ddf7a102` + WIP).
+
+### Scorecard
+
+| Dimension | Winner | Margin | Source cell |
+|---|---|---|---|
+| Streaming frame rate | **OpenTUI** | **~3×** (43 vs 14 fps) | `cpu800.frame_pacing` |
+| Streaming smoothness (interframe p95) | **OpenTUI** | **40ms vs ~220ms** (no ¼-second stalls) | `cpu800.frame_pacing` |
+| Scroll CPU | **OpenTUI** | **~2.7× cheaper** (134–155 vs 403–416 ticks) | `scroll3000.scroll.cpu_ticks` |
+| Cold-start floor | **OpenTUI** | ~97–103 vs ~107–109 MB | `startup.vmhwm_kb` |
+| Session-create latency | **OpenTUI** | ~151–177 vs ~204–229 ms | `startup.session_create_ms` |
+| First-byte paint | Ink | ~93 vs ~122 ms | `startup.first_byte_ms` |
+| Memory @ small/typical | Ink | OpenTUI +30–50 MB | `mem50/100/300.vmhwm` |
+| Memory @ heavy tool output | **OpenTUI** | **crossover** (258–265 vs 280–290 MB) | `results-fat-mem-*` |
+| Layout reflow latency | **Ink** | **~0ms vs ~13ms** (OpenTUI's one honest loss) | `resize3000.resize.reflow_ms` |
+
+### The honest reading
+
+- **OpenTUI wins everything you feel continuously** — frame rate (~3×), scroll
+  CPU (~2.7×), and smoothness (no 200ms hitches; p95 40ms vs ~220ms). This is the
+  lead. The single most user-perceptible difference is the stall-free stream.
+- **Memory: lead with smoothness, not raw RSS.** Ink is lighter at small/typical
+  sizes (OpenTUI carries a ~102 MB irreducible Node+V8+`libopentui.so` floor, so
+  it sits +30–50 MB above Ink there). But it **crosses over** under heavy tool
+  output (mem300: 258–265 MB OpenTUI vs 280–290 MB Ink) because windowing beats
+  Ink's mount-every-row. Real-world: 20 memwatch sessions show a flat ~108 MB
+  floor and ~0 MB/h on long sessions (one 15h session, 0 MB/h; one 4.4h session
+  plateaus flat at ~237 MB with mounted rows pinned at 33).
+- **The one outright loss is layout reflow** (~13ms p50 vs Ink's ~0ms; under a
+  resize storm OpenTUI degrades to ~14fps/~197ms vs Ink ~26fps/~100ms). Heavier
+  native renderables vs Ink's string nodes. This is a real, quantified
+  optimization target — **not** a regression vs current behavior, and **not** the
+  "halved 0.4.0→0.4.1" delta (we measured the absolute 12–15ms only; do not quote
+  "halved" from this run).
+- **The memory fix is engine-agnostic** — a rolling display cap
+  (`HERMES_TUI_MAX_MESSAGES=3000` default) that is display-only and never touches
+  the model's context. Uncapped is a stress config, not real usage (10k msgs
+  uncapped: 793 MB; capped sessions are flat MB/h).
+- **Gut-check vs upstream/opencode: no bugs.** Exactly one frame callback
+  (early-exits cheaply), zero `writeToScrollback` for the transcript (one sticky
+  `<scrollbox>` + reactive `<For>`), native `<markdown streaming>` byte-for-byte
+  parity with live opencode, no reactive-read-outside-tracking-scope (the #1 Solid
+  trap). Source: `docs/plans/opentui-gutcheck-verification.md`.
+
+Full methodology + every cell: see the HTML write-up's benchmark sections and
+`docs/plans/opentui-endgame-benchmark-report.md`.
+
+---
+
+## 3. UI parity — and where the two engines genuinely diverge visually
+
+100% *feature* parity is the bar (matrix in §6), but the two engines are **not**
+visually identical. The Ink TUI renders the transcript as a **box-drawing tree**;
+OpenTUI renders it **flat and marker-based**. This is a deliberate design
+divergence, captured in `ui-opentui/src/view/messageLine.tsx`:
+
+> *"the view is a dark room and gold is the single lamp — it sits on the NEWEST
+> answer's `⚕` and the user's `❯`, nowhere else (older assistant glyphs demote to
+> grey: they merely happened)."*
+
+Real screenshots (saved under `docs/research/opentui-screenshots/`), captured live
+on a real PTY 120×40 via the `tmux-pane-screenshot` workflow — **same session
+resumed in both engines** where possible.
+
+### Legacy Ink — `docs/research/opentui-screenshots/ink-transcript.png`
+
+![Ink transcript](research/opentui-screenshots/ink-transcript.png)
+
+- **Box-drawing tree layout.** Each turn is a nested structure: `└─ Response`,
+  `└─ ▾ Tool calls (1)`, `   └─ ● Terminal("…")` — explicit corner rails and
+  disclosure triangles.
+- **`┊` dotted quote-bar** prefixes assistant prose.
+- **Tool calls collapse by default** behind a `▾ Tool calls (N)` disclosure,
+  nested one rail deeper.
+- **Whole assistant message tinted gold/amber** (body text is colored, not just
+  the marker).
+- Right-edge scrollbar: thin `│` track + `┃`/orange thumb.
+- Status bar: `─ ready │ opus 4.8 fast high │ 0/1m │ [░░░░░░] 0% │ 25s │ voice off │ 1 session ─ ~`
+  — leading dash, pipe-delimited fields, trailing `~`.
+- **No top header bar.**
+
+### New OpenTUI — `docs/research/opentui-screenshots/opentui-transcript.png` (+ `opentui-toolcall.png`)
+
+![OpenTUI transcript](research/opentui-screenshots/opentui-transcript.png)
+
+![OpenTUI tool call](research/opentui-screenshots/opentui-toolcall.png)
+
+- **Flat, marker-based layout.** No tree rails. Assistant = `⚕` (caduceus, gold
+  only on the newest answer), user = `❯` (gold chevron + gold text). Older
+  assistant glyphs demote to grey.
+- **Neutral body text.** Gold is reserved for markers and inline-code accents;
+  prose is grey/white (the "single lamp" rule), so the screen reads calmer than
+  Ink's all-amber blocks.
+- **Tool calls render inline, expanded, on one header line:**
+  `⚕ ▶ delegate_task  Run the shell command `…`  (/agents to monitor)  · 41s  (11 lines)`
+  — marker, `▶` collapse triangle, bold tool name, grey arg preview, hint,
+  `· duration`, `(N lines)` — and the result flows flat directly below (no nesting
+  rail). Per-tool renderers exist (`view/tools/registry.tsx`) — bash/file+diff/
+  read/search/skill/clarify/todo each render differently, not a uniform dump.
+- **Per-block `⧉ copy` affordance** on a quiet footer line under every settled
+  assistant block and user prompt (click → copies that block's source).
+- **Top header bar:** `⚕ Hermes Agent · opentui · ready` + a gold horizontal rule
+  (Ink has none).
+- Status bar (real backend): `● claude-fable-5 │ [▒▒▒] 4% │ …/lively-thrush/hermes-agent (feat/opentui-native-engine)`
+  — green status dot, model, context/token bar, **right-pinned cwd + branch**.
+
+### Divergence summary table
+
+| Aspect | Ink (legacy) | OpenTUI (new) |
+|---|---|---|
+| Transcript structure | Box-drawing **tree** (`└─`, rails) | **Flat**, indented, marker-based |
+| Assistant marker | `└─ Response` rail + `┊` quote-bar | `⚕` caduceus glyph |
+| User marker | (rail) | `❯` gold chevron |
+| Assistant body color | Tinted gold/amber | Neutral grey/white (gold = accents only) |
+| Tool calls | Collapsed `▾ Tool calls (N)`, nested | Inline expanded header + flat result |
+| Per-tool rendering | Largely uniform | Dedicated renderers per tool |
+| Copy affordance | `/copy` command | `/copy` **+ per-block `⧉ copy`** |
+| Header bar | None | `⚕ Hermes Agent · opentui · ready` + rule |
+| Status bar | `─`/`│`-delimited, trailing `~` | dot + bars + right-pinned cwd/branch |
+
+**For review:** the divergence is intentional (a design pass, not an accident),
+but it means "drop-in replacement" is true at the *feature* level, not the
+*pixel* level. A user switching engines will immediately notice the flatter,
+calmer transcript. Worth calling out explicitly so the swap isn't sold as
+visually invisible.
+
+---
+
+## 4. Non-core / kitchen-sink change audit (what review should scrutinize)
+
+Full report: **`docs/research/opentui-noncore-change-audit.md`** (file-by-file,
+commit-by-commit, with `file:line` evidence). Summary below.
+
+This PR's net footprint vs `origin/main` (two-dot diff = exactly this PR's adds,
+no main work re-included):
+
+| Bucket | Files | Net diff |
+|---|---:|---:|
+| UI (`ui-opentui/`, the engine + tests) | 197 | +36,001 / −1 |
+| Docs | 8 | +1,164 / −0 |
+| **Other (the review-flag surface)** | **28** | **+3,218 / −204** |
+
+The 28 "other" files are the only place this PR touches shared Hermes core. They
+classify as:
+
+### ✅ CORE-OPENTUI-NECESSARY (the engine can't work without these; Ink path provably untouched)
+
+- **`hermes_cli/main.py`** (+382/−5) — dual-engine launcher (engine resolution,
+  Node 26 / fnm detection, `_make_opentui_argv`, heap override). Default falls
+  back to Ink unless the host is OpenTUI-ready (`main.py:1685`); OpenTUI is
+  dispatched *around* the Ink bootstrap, never through it (`main.py:1914-1922`).
+- **`scripts/install.sh`** (+78/−1) — `install_opentui` stage, **strictly
+  best-effort** (every failure returns 0; falls back to Ink; Windows/Termux
+  skipped). Ink install path unchanged.
+- **`Dockerfile`** (+21/−11) — Node 22→**26** bump (required by the `node:ffi`
+  renderer) + `ui-opentui` build step. Opt-in; Ink build line preserved. **Caveat:
+  the Node major bump affects the whole image (Ink + web + Playwright)** — the
+  diff self-flags "verify the full image build on Node 26 in CI."
+- **`hermes_cli/_parser.py`** (+16/−2) — bare `--resume` → OpenTUI session picker;
+  `--resume <id>` unchanged.
+- **`tui_gateway/server.py`** (+612/−40) — predominantly opt-in RPCs/fields the
+  new engine calls (`session.peek`, `session.list` filters, `startup.catalog`,
+  `diff_unified`, window-title, skin keys). Each is gated so **the Ink path is
+  byte-for-byte unchanged** (`server.py:3930`, `:4254`, `:10447`). *Note:* this
+  file also carries some of the cost-accounting code (below) — separable.
+
+> `tui_gateway/` (~12k LOC Python) is **shared by both engines** and is **not**
+> removed when Ink is retired. Only the `ui-tui/` frontend tree goes.
+
+### 🚩 FLAG FOR REVIEW — Category C, separable from an OpenTUI PR
+
+These do **not** need to ship with the engine and a reviewer should ask to split
+them out:
+
+1. **Provider-reported-cost accounting** (commits `85546bb9e` + `364b93a4b` +
+   `e01b04de4`) — a coherent feature spanning **11 files**: `agent/usage_pricing.py`,
+   `plugins/model-providers/openrouter/__init__.py`,
+   `agent/transports/chat_completions.py`, `agent/agent_init.py`, `run_agent.py`,
+   `agent/conversation_loop.py`, `agent/account_usage.py`, `hermes_state.py`,
+   `gateway/slash_commands.py`, the cost half of `cli.py`, and the
+   `_get_usage`/`_compact_usage_text` blocks of `tui_gateway/server.py` (+ 5 test
+   files). Strongest evidence: commit `85546bb9e` *"gateway: capture real
+   provider-reported cost (openrouter usage accounting)"* — a provider-accounting
+   rework, not a renderer.
+2. **`plugins/model-providers/openrouter/__init__.py`** — sends
+   `usage:{include:true}`, a provider request-shape change affecting *all*
+   interfaces, not just the TUI (`openrouter/__init__.py:85-90` cites the
+   OpenRouter usage-accounting docs).
+3. **Worktree lock / dirty-tree preservation** (commit `94765e48f`,
+   `cli.py` + `tests/cli/test_worktree.py`, ~145 lines) — git-worktree lifecycle
+   safety plumbing with **zero TUI references** (`cli.py:1391-1545`, `:1635-1713`).
+4. **`tools/clarify_tool.py`** (+16/−4) — docstring/schema-description-only fix
+   (commit `16e408f3f`); applies to every interface, trivially separable.
+
+### ✅ Conversation-loop / role-alternation / prompt-cache correctness verdict: **NO RISK**
+
+Verified: none of `run_agent.py`, `agent/conversation_loop.py`,
+`agent/agent_init.py`, `agent/transports/chat_completions.py` touch
+message-role alternation or the prompt-cache prefix. The
+`conversation_loop.py` added lines grep clean for
+`cache_control|alternation|prompt_cach|api_messages`; the cache/alternation
+machinery (`:57`, `:660-674`, `:759`) is untouched; the PR's insertion at
+`:1809-1879` is purely additive cost bookkeeping after `cost_result`. **Prompt
+caching and strict role alternation are preserved.**
+
+---
+
+## 5. What this does and does NOT fix
+
+**Fixes (structurally, by replacing the rendering substrate):** the renderer bug
+class — layout/scroll/input/copy/mouse/markdown/resize — plus the
+hand-maintained memory-eviction problem (windowing + Solid keyed `<For>`
+unmount→`destroy()`→`free()`), and several long-open feature requests (mouse,
+collapsible tool calls, session title/status bar, double-ESC, chronological
+thinking/tool ordering).
+
+**Does NOT fix:** the gateway is unchanged — the biggest single hotspot file in
+triage is `tui_gateway/server.py`, and whole bug clusters are gateway/Python-side
+(WS write-timeout/RPC pool, MCP-failure startup freezes, shell.exec denylist).
+The engine swap addresses rendering/input/scroll/memory; **gateway bugs ride
+along.** The Effect-boundary hardening does make those failures *visible* (typed
+events → system lines instead of a frozen spinner) and the TUI auto-heals
+(crash → backoff → respawn → resume, capped 3/60s).
+
+---
+
+## 6. Feature parity matrix (vs the Ink TUI)
+
+Verbatim, detailed, surface-by-surface with `file:line` evidence:
+**`docs/plans/opentui-ink-parity-matrix.md`** (interactive/filterable version in
+the HTML write-up). Headline state:
+
+| Surface | State |
+|---|---|
+| Transcript rendering (scrollbox, markdown, code, diffs, collapsible tools, reasoning, chronological order, windowing) | **full parity (9/9)** |
+| Blocking prompts (approval/clarify/sudo/secret/confirm) | **full parity (5/5)** |
+| Theming (skins, light/dark, ANSI-256 norm) | **full parity** |
+| Mouse / copy (tracking, selection, multi-click, OSC52, click-to-expand, wheel accel) | **full parity** |
+| Resilience (crash auto-heal + resume) | **parity++ (exponential backoff)** |
+| Composer / input | near parity — **missing: external editor (Ctrl+G → `$EDITOR`)**; ghost-text autosuggest partial |
+| Slash commands | core parity — **missing: `/setup`, `/redraw`, `/plugins`, `/voice`**; `/undo` prefill + `/image` partial |
+| Status bar / header chrome | almost all closed — **missing: MCP-servers panel, profile-in-prompt** |
+| Agent surfaces | most shipped — **missing: voice indicators, browser/CDP indicator** |
+| Utility commands | **missing: `/redraw`, `/setup`**; rest present |
+
+> The original PR-draft gap list was **substantially stale** — the WIP since
+> shipped context %/token bar, cost, compressions, duration, update banner, todos
+> panel, activity feed, notifications, background-task indicator, **and per-tool
+> renderers** (the "every tool renders the same" claim is false:
+> `view/tools/registry.tsx` has dedicated renderers).
+
+### Genuinely-remaining parity gaps
+
+- [ ] **External editor (Ctrl+G → `$EDITOR`)** — highest-impact missing composer affordance
+- [ ] MCP-servers detail panel; profile-in-prompt marker
+- [ ] Voice indicators (listening/transcribing/REC/STT) + `/voice`
+- [ ] Browser/CDP connection indicator + `/browser`
+- [ ] `/setup` wizard handoff, `/redraw`, `/plugins` hub
+- [ ] Draggable scrollbar; sticky-prompt line
+- [ ] `/undo` prefill into composer; model-picker persist-global toggle; skills-hub install/manage
+
+---
+
+## 7. Rollout, runtime & risks
+
+- **Runtime:** plain Node 26 (FFI floor 26.3+) — one runtime, no Bun. (Note: the
+  upstream OpenTUI docs say "requires Bun"; this engine deliberately runs on Node
+  26's experimental `node:ffi` instead — that's the load-bearing runtime decision.)
+- **Rollback:** Ink is untouched and remains the fallback; reverting is a launcher
+  decision, not a code revert.
+- **Default-engine selection:** auto-picks OpenTUI only when the host is genuinely
+  set up (Node ≥ 26.3 + built bundle), else Ink; explicit env/config bypasses the
+  probe.
+- **Known sharp edges:** `libopentui.so` native-lib distribution (P1 upstream:
+  copies can fill `/tmp`); the Dockerfile Node major bump needs full-image CI
+  verification; tree-sitter grammars are fetched from GitHub on first use and
+  cached in `~/.hermes/cache/opentui-parsers/` — air-gapped hosts get plain-text
+  highlighting until the cache is pre-seeded (the fetch never blocks boot and
+  never throws).
+
+## 8. Try it
+
+```bash
+hermes                              # auto-selects OpenTUI when the host supports it
+HERMES_TUI_ENGINE=opentui hermes    # force the native engine
+HERMES_TUI_ENGINE=ink hermes        # force the legacy Ink engine
+# preview standalone (no backend), Node 26:
+cd ui-opentui && npm install
+node scripts/build.mjs scripts/demo.tsx .demo
+DEMO_TOTAL=120 HERMES_TUI_MAX_MESSAGES=80 \
+  node --experimental-ffi --no-warnings .demo/demo.js   # inside a TTY
+```
+
+Requires Node 26.3+. On older Node / Windows / Termux it auto-falls-back to Ink.
+
+---
+
+## Appendix — source-of-truth files in this repo
+
+| Topic | File |
+|---|---|
+| Non-core change audit (full) | `docs/research/opentui-noncore-change-audit.md` |
+| Feature parity matrix (verbatim) | `docs/plans/opentui-ink-parity-matrix.md` |
+| Benchmark report | `docs/plans/opentui-endgame-benchmark-report.md` |
+| Gut-check verification | `docs/plans/opentui-gutcheck-verification.md` |
+| Ink↔OpenTUI capture asymmetry | `docs/plans/opentui-ink-asymmetry-note.md` |
+| UI screenshots | `docs/research/opentui-screenshots/{ink,opentui}-*.png` |
+| PR description (prose) | `docs/pr-description-main-doc.md` |
+| Interactive write-up | `~/projects/opentui-perf-writeup/index.html` (out-of-repo) |

docs/opentui-upstream-alignment.md

@@ -0,0 +1,73 @@
+# Upstream alignment — how we inherit OpenTUI's performance work for free
+
+Context (maintainer, 2026-06-11): opencode's 100-message cap was a November-era
+performance workaround, since obsoleted; the **next OpenTUI version ships
+native yoga** (≥2× layout performance, more improvements building on it);
+opencode does not use virtualization.
+
+## The invariant that makes alignment free
+
+**We are forkless and public-API-only.** The windowing layer (S1+S2) drives the
+STOCK `<scrollbox>` through documented surface only — `onSizeChange`,
+`setFrameCallback`, `scrollTop`/`viewport`/`scrollHeight`, Solid `<Show>`
+mount/unmount. Zero patches to `@opentui/core`. Every upstream release
+therefore drops in by bumping three pinned versions in `ui-opentui/package.json`
+(`@opentui/{core,keymap,solid}`, currently 0.4.0). Keep it that way: any new
+code that needs core behavior goes through a `boundary/` wrapper, never a
+patched dependency.
+
+## What native yoga changes for us (and what it doesn't)
+
+- **Kills the WASM ratchet** (grow-only linear memory → freeable native
+  allocations). This retro-justifies S2 less, but S2's append-time windowing
+  remains correct: transient mounted peaks still cost handles and RSS.
+- **Does NOT obsolete windowing.** The binding constraint is the 65,535-slot
+  native handle table: ~47 handles/row × 3,000 stored rows ≈ 141k handles —
+  over the table at ANY layout speed. Windowing is what makes the 3,000-row
+  scrollback possible; yoga's backend is irrelevant to that math.
+- **Makes windowing feel even better**: 2× layout = cheaper margin remounts =
+  smaller window margins viable and less exposure for the one accepted limit
+  (estimate-height snap under scrollbar jumps). After the bump, re-tune margin/
+  hysteresis against the scroll cell.
+
+## The shim ledger (delete-on-upstream-fix; all in `ui-opentui/src/boundary/`)
+
+| shim | what it papers over | delete when |
+|---|---|---|
+| `ffiSafe.ts` | u32 draw coords go negative under Node FFI (Bun silently wraps) — ERR_INVALID_ARG_VALUE loop | upstream clamps, or Node FFI path is officially supported |
+| `nativeHandles.ts` | SyntaxStyle exhaustion crashes mid-mount; degrade-to-unstyled | handle table widened (INDEX_BITS>16) or per-kind tables |
+| `renderer.ts` exit-signal guard | core 0.4.0 treats SIGPIPE (clipboard spawn) as an exit signal; its own uncaughtException handler allocates a handle and dies (exit-7 masking) | both fixed upstream |
+| `clipboard.ts` hardening | same SIGPIPE incident class | with the above |
+
+Each is (a) isolated, (b) inert if upstream fixes the behavior, (c) worth
+reporting upstream — four concrete, reproduced, root-caused issues. Filing them
+is the cheapest alignment lever we have: it converts our workarounds into
+upstream regression tests. (Needs glitch's go-ahead — public repo activity.)
+
+## The upgrade playbook (per upstream release)
+
+1. Branch `chore/opentui-X.Y.Z`, bump the three pins, `npm ci`.
+2. `npm run check` (648 tests; the windowing invariants — identical
+   scrollHeight ON/OFF, byte-stable frames across corrections — are literal
+   assertions and will catch behavioral drift).
+3. Bench acceptance, sequential: `--cell gate` (determinism digest; EXPECT a
+   new digest if upstream changed rendering — eyeball the frame, re-bless),
+   `--cell mem3000 --msgs 2000` + `--cell scroll --msgs 3000` vs current
+   numbers (300–375MB / p99 6–8ms), `--cell pipeline` (frame pacing ≥22fps).
+4. Shim audit: try each boundary shim OFF; delete the ones upstream fixed.
+5. Live tmux smoke (scroll sweep / resize / selection-copy), screenshots.
+6. Windowing re-tune if layout got faster: margins up or hysteresis down,
+   re-run scroll cell, keep p99 ≤ 17ms gate.
+
+The bench suite IS the upgrade contract — it's exactly the harness that lets
+us take every upstream improvement within a day of release, with proof.
+
+## Questions worth relaying to the maintainer
+
+1. Any plan to widen the 16-bit native handle table (or split per-kind)?
+   That's our hard ceiling, independent of yoga.
+2. Is the Node `--experimental-ffi` path on their support radar, or Bun-only?
+   (Native yoga adds new FFI surface; we run Node.)
+3. Would they take the windowing layer's core-agnostic pieces (exact-height
+   spacer pattern, correction-legality rule) as a documented recipe or
+   framework-level utility? We have it production-shaped with tests.

docs/plans/opentui-background-activity.md

@@ -0,0 +1,150 @@
+# OpenTUI — Background Activity: agents inspection, background panel, notifications + density
+
+**Status:** SPEC (brainstormed with glitch 2026-06-13) · target branch `feat/opentui-native-engine`
+**Hard constraint:** TUI-LAYER ONLY (`ui-opentui/`). **Zero changes to `tui_gateway/server.py` or
+`run_agent.py` core.** Build only on gateway events/RPCs that already exist. Everything below was
+feasibility-checked against the live gateway surface (see "Gateway surface" §).
+
+## Why
+
+Dogfeedback (screenshots `iznq/qxpe/rpiw/rplj`):
+1. **Agents dashboard is too crowded** (`rplj`) — master rows dump each subagent's full multi-line
+   prompt; the trace pane is squished. Inspection + transcript reading is "not great."
+2. **Background processes are basically invisible** (`qxpe`) — completions leak into the transcript
+   as plain lines that read like model output; no panel, no badge, notifications are non-existent.
+3. **Input zone is too crowded** (`rpiw`) — status bar + composer + agents tray + completion menu +
+   shell note stack under the transcript.
+
+## Design decisions (from the brainstorm)
+
+- **Two SEPARATE surfaces, ONE shared substrate.** Background *agents* (delegated subagents) and
+  background *work* (detached runs + OS processes) are visually/feature-wise distinct, but share the
+  underlying tracking + notification + badge plumbing.
+- **Notifications are multi-channel** on every relevant state change:
+  - **(C) inline card** in the transcript — a distinct, colored, collapsed *system card*, clearly
+    NOT model output (replaces today's plain-line leak).
+  - **(A) ambient badge** — a live count in chrome (status-bar `bg:`/the `⚡ N agents` tray) that
+    flashes on change; you pull-to-inspect. Stays visible while things run.
+  - **OSC desktop** — reuse the EXISTING `boundary/termChrome.ts` (`notify`, OSC 9/99/777, already
+    focus-gated so it only fires when the terminal is blurred).
+- **Agents surface = inspection only.** No foregrounding / "become the subagent" (that would change
+  core subagent UX — explicitly out of scope). Scannable list + a faithful render of the *already-
+  tracked* live activity (goal/model/reasoning/tool calls/progress/summary). No new fetch.
+- **Background surface = view + stop.** List runs + OS processes with status/uptime; cancel a run
+  (`session.interrupt`/`subagent.interrupt`); **stop-all** OS processes (`process.stop`). Per-process
+  kill and per-process logs are NOT exposed as RPCs → out of scope under the no-core rule (noted).
+- **Input density is in scope** (own phase).
+
+## Gateway surface we build on (verified — all already exist)
+
+| Need | Mechanism (existing) |
+|---|---|
+| Background-run lifecycle | `prompt.background` (start), `background.complete` (event) |
+| Notifications | `notification.show` / `notification.clear` events — payload `{text, level, kind, ttl_ms, key, id}` |
+| Subagent stream | `subagent.spawn_requested/start/thinking/tool/progress/complete` events (store already consumes) |
+| List OS processes | `agents.list` RPC → `{processes:[{session_id, command, status, uptime_seconds}]}` |
+| Stop OS processes | `process.stop` RPC → `kill_all()` (**all**, not per-process) |
+| Cancel a run / subagent | `session.interrupt`, `subagent.interrupt` |
+| List active sessions/runs | `session.active_list`, `session.status` |
+| Subagent trace (archived) | `spawn_tree.list/load` (already used by `/replay`) |
+| OSC desktop notify | `boundary/termChrome.ts` `notify(TermNotification)` |
+
+**Honest limits (no-core constraint):** OS processes get list + stop-all only — no per-process kill
+(`process_registry.kill_process` exists but isn't an RPC) and no per-process log tail
+(`read_log` isn't an RPC). If the no-core rule is ever relaxed, each is a ~5-line additive `@method`.
+
+## Architecture (Approach 1 — substrate-first)
+
+```
+gateway events ──► store: backgroundActivity slice ──► derived counts/state
+                          │                                  │
+                          ├─► notificationDispatcher ─────────┼─► (C) inline card  (transcript)
+                          │     (card + badge + OSC)          ├─► (A) ambient badge (statusBar/tray)
+                          │                                   └─► OSC via termChrome.notify
+                          ├─► Surface 1: AgentsDashboard (revamp) — list + rich activity pane
+                          └─► Surface 2: BackgroundPanel (new)    — runs + processes, stop
+```
+
+### Shared substrate (the "underneath" both surfaces use)
+
+- **`logic/backgroundActivity.ts`** (new) — pure model + reducers. Types:
+  - `BackgroundRun` (from `prompt.background`/`background.complete`/`session.active_list`):
+    `{ id, label, status: 'running'|'complete'|'failed'|'cancelled', startedAt, summary? }`
+  - `BackgroundProcess` (from `agents.list`): `{ sessionId, command, status, uptimeSeconds }`
+  - `Notification` (from `notification.show`): `{ id, key?, text, level, kind, ttlMs?, at }`
+  - Pure helpers: `applyNotification`, `clearNotification(key)`, counts (`runningCount`),
+    `mergeProcessList`, dedupe by `key`/`id`. Fully unit-testable (no renderer).
+- **`store.ts`** — a `backgroundActivity` slice + event handlers for `notification.show/clear`,
+  `background.complete`, and a polled `agents.list` snapshot (poll only while a panel/badge is live,
+  or piggyback existing cadence). Existing `subagent.*` handling is untouched.
+- **`logic/notificationDispatcher.ts`** (new, pure) — given a state-change, decide the channels:
+  returns `{ card?: SystemCard, badge: delta, osc?: TermNotification }`. The boundary calls
+  `termChrome.notify` for the OSC part; the store appends the card + bumps the badge.
+
+### Surface 1 — Agents inspection overlay (revamp `view/overlays/agentsDashboard.tsx`)
+
+- **Master list rows = ONE line each:** `<statusGlyph> <truncated goal (truncRight to width)> · <model>`.
+  No multi-line prompt dump. Selected row highlighted (existing `▸` + accent).
+- **Detail pane = faithful activity transcript** of the selected agent, styled like the main
+  transcript (not flat dumped lines): goal+model header, then the trace rendered by *type*
+  (reasoning / tool-call+result / progress / final summary), newest last, sticky-bottom, PgUp/PgDn.
+  - Requires giving `SubagentInfo.trace` light typing (`{ kind:'tool'|'reasoning'|'progress'|'summary', text }`)
+    instead of `string[]`, populated where `subagent.*` events are reduced. Internal data-shape
+    change only; no gateway change.
+- Keep Esc/q close, ↑↓ select. Reuse theme + `truncRight` from statusBar.
+
+### Surface 2 — Background panel (new `view/overlays/backgroundPanel.tsx`)
+
+- **Two sections:** *Runs* (background agent runs) and *Processes* (OS processes from `agents.list`).
+- Each row: status glyph + label/command (truncated) + uptime/elapsed + status.
+- **Actions:** `↑↓` select; on a *run* → `c` cancel (`session.interrupt`/`subagent.interrupt`);
+  global **stop-all processes** (`x` → `process.stop`, confirm). Esc/q close.
+- **Access:** new client slash `/bg` (alias `/background`, `/jobs`) in `logic/slash.ts` CLIENT set →
+  `store.openBackgroundPanel()`. Also reachable from the ambient badge.
+- Poll `agents.list` on open + on a light interval while open; stop polling on close.
+
+### Notifications (the (C)+(A)+OSC wiring)
+
+- **(C) inline card** — a new transcript element `view/notificationCard.tsx`: a bordered/colored,
+  `selectable:false` system card keyed by `notification.id`, level-tinted (`info/warn/error`),
+  collapsed to one line by default with the `kind` + `text`; clearable by `notification.clear` key.
+  Appended into the message stream as a distinct row type (NOT a plain `system` text line). Replaces
+  the current plain-line leak. (`/details` interplay: cards are chrome, always shown, never windowed.)
+- **(A) ambient badge** — `statusBar.tsx` `bg: N` segment (already reserved) bound to
+  `runningCount()`; the `agentsTray.tsx` count already exists — extend it to "agents + background."
+  Flash/recolor on a fresh notification (brief).
+- **OSC** — on `notification.show` with a terminal level (complete/failed), call
+  `termChrome.notify({title, body})` (already focus-gated). No new escape-sequence code.
+
+### Input-zone density pass (`view/composer.tsx` / `view/App.tsx`)
+
+- Audit what stacks under the transcript and collapse/gate: the `⚡ N agents` tray line folds into
+  the ambient badge (shrinks one line); ensure the shell-mode note, completion menu, and status bar
+  don't co-stack more than necessary. Concrete rules decided with a tmux density pass (ASCII-mocked,
+  approved) — kept minimal; no behavior change, just fewer competing chrome lines.
+
+## Phases (implementation order — each gated + tmux-smoked + committed)
+
+- **P1 — Notification substrate** (`backgroundActivity.ts` + `notificationDispatcher.ts` + store
+  slice + `notificationCard.tsx` + badge wiring + OSC call). Highest visible win; the shared core.
+- **P2 — Agents inspection revamp** (`agentsDashboard.tsx` + typed `trace`). De-crowds `rplj`.
+- **P3 — Background panel** (`backgroundPanel.tsx` + `/bg` + actions). New surface.
+- **P4 — Input density pass.** Folds the tray into the badge; trims co-stacked chrome.
+
+## Testing / gates (per phase)
+
+- **Pure logic** (`backgroundActivity`, `notificationDispatcher`, slash `/bg` routing,
+  trace-typing) → vitest unit tests, TDD where natural.
+- **Views** → headless frame tests (`renderProbe`) for the card, the de-crowded dashboard row
+  format, the background panel sections; + **live tmux smoke** (`tmux-pane-screenshot`) for each
+  surface using a seeded-store harness (the `uxSmoke` pattern: `store.apply`/`applyInfo`/
+  `commitSnapshot` + canned events).
+- **Gate** `cd ui-opentui && npm run check` green (judge by real exit, not a piped tail) after each
+  phase; rebuild `dist/main.js`; commit `opentui(v6): …` (no attribution) and push per standing instr.
+
+## Out of scope (explicit)
+
+- Foregrounding / "becoming" a subagent (B/C from the brainstorm) — would change core subagent UX.
+- Per-process kill + per-process log tail for OS processes — needs additive gateway RPCs (no-core veto).
+- "Collect result into transcript" for finished runs — deferred (Q6=B, view+stop only).
+- Any change to `tui_gateway/server.py` / `run_agent.py`.

docs/plans/opentui-composer-ux-9.md

@@ -0,0 +1,248 @@
+# Plan — OpenTUI composer/UX batch (10 features)
+
+> **STATUS: SHIPPED (2026-06-13).** All 10 features implemented, gate green
+> (ui-opentui 714 tests + 316 gateway + 25 cost tests), F5/F6 verified live via
+> tmux screenshot. Commits: `f4dacc68e` (F1/F2/F7/F8/F8b/F9/F10), `20d516ae9`
+> (F4/F5/F6), `9aa5e54be` (F3). Decisions taken: **D1 = cursor-aware onType**
+> (threaded `ta.cursorOffset`); **D2 = chrome cost is Nous-header-only via a new
+> `nous_header_cost_usd`, `/usage` page kept full via `real_session_cost_usd`**.
+> F10 (right-pinned cwd) was added mid-session by the user.
+
+**Branch:** `feat/opentui-native-engine` · **Engine:** `ui-opentui/` (Node 26)
+**Gate:** `cd ui-opentui && PATH="$HOME/.local/share/fnm/node-versions/v26.3.0/installation/bin:$PATH" npm run check` → exit 0.
+
+## TL;DR
+
+Nine UX fixes for the native composer + clarify prompt. **8 of 9 are front-end-only**
+in `ui-opentui/`; only F3 (cost) touches the Python gateway. Every backend the new
+behaviour needs (`shell.exec`, `complete.path` with `@file:`/`@folder:`/fuzzy) **already
+exists** — most of this is client wiring, not new RPC surface. No new core tools, no new
+`HERMES_*` env vars, no prompt-cache impact (composer/prompt are client-render only).
+
+| # | Symptom | Fix site | Backend |
+|---|---|---|---|
+| F1 | bare `/` opens the modal | `logic/slash.ts:115` `planCompletion` | none |
+| F2 | `/abs/path` text triggers slash | `logic/slash.ts:115` + `logic/skillMatch.ts` | none |
+| F3 | cost wrong / shows for non-Nous | `tui_gateway/server.py` + `agent/usage_pricing.py` | gateway |
+| F4 | can't paste until composer focused | `view/composer.tsx` onPaste/focus | none |
+| F5 | clarify ugly (no wrap, weak diff, "Other" is a row) | `view/prompts/clarifyPrompt.tsx` rewrite | none |
+| F6 | clarify arrows scroll the transcript | same rewrite (preventDefault) | none |
+| F7 | slash highlight/menu dies after line 1 | `logic/slash.ts:114` | none |
+| F8 | file mention dies after line 1 | `logic/slash.ts:114` | none |
+| F8b | `@` should be the ONLY file-mention trigger | `logic/slash.ts:93` `isPathLike` | none |
+| F9 | `!cmd` → run bash, show result | `entry/main.tsx` submit + new system render | uses existing `shell.exec` |
+
+---
+
+## F1 + F2 + F7 + F8 + F8b — the completion trigger (`logic/slash.ts`)
+
+All five live in one ~10-line function, `planCompletion` (slash.ts:113-121). Current:
+
+```ts
+export function planCompletion(text: string): CompletionPlan | null {
+  if (text.includes('\n')) return null                                   // ← F7/F8 die here
+  if (text.startsWith('/')) return { from: 0, method: 'complete.slash', params: { text } } // ← F1/F2
+  const word = /(\S+)$/.exec(text)?.[1]
+  if (word && isPathLike(word)) { ... complete.path ... }                // ← F8b: too many triggers
+  return null
+}
+```
+
+### F1/F2 — slash only for a real command token
+- A bare `/` (no char yet) must **not** query. Require `/` + at least one name char.
+- A `/abs/path` (slash followed by a path with more `/`) is **not** a command — it's
+  text. The slash menu should only fire when the FIRST token matches the command
+  grammar (`/[A-Za-z0-9][\w.-]*` — the `NAME_RE` already in skillMatch.ts:51, which
+  excludes `/`). `/usr/bin` fails NAME_RE → no slash menu.
+- Concretely: replace `text.startsWith('/')` with: the text starts with `/`, and the
+  first whitespace-delimited token after the `/` is non-empty AND matches `NAME_RE`
+  (i.e. `/m`, `/model foo` → yes; `/`, `/usr/bin`, `/./x` → no). Reuse `slashTokens`
+  /`NAME_RE` from skillMatch.ts so the trigger and the highlighter share one grammar.
+
+### F7/F8 — completion must survive newlines (shift+enter)
+- `if (text.includes('\n')) return null` is the bug. It was a blunt guard so a multi-line
+  paste wouldn't spam path-completion. The right rule operates on the **current line /
+  current token at the cursor**, not the whole buffer.
+- The composer passes the full `plainText` to `onType`. We don't currently pass the
+  cursor offset. **Decision D1 (below):** either (a) thread the cursor offset into
+  `onType` and complete the token under the cursor, or (b) cheap interim — slice to the
+  **last line** (`text.slice(text.lastIndexOf('\n')+1)`) and run the existing logic on
+  that. (a) is correct (mid-buffer edits), (b) is 1 line and covers the reported case
+  (typing at the end on line N). Recommend (a) for correctness; it also future-proofs
+  @-mention mid-line.
+- Slash *highlighting* (skillMatch.ts `slashTokens`) **already scans multi-line text
+  correctly** (it iterates the whole string, newline-aware via `nativeCharOffset`). So
+  F7's "highlighting stopped" is really the same `planCompletion` newline bail starving
+  the menu; the highlight token itself still styles. Verify in the live smoke.
+
+### F8b — `@` is the only mention trigger
+- `isPathLike` (slash.ts:93) currently returns true for `@`, `~`, `./`, `../`, `/`, or
+  any word containing `/`. The user wants **`@`-only** (drop `~`/`./`/bare paths as
+  mention triggers). Narrow it to `word.startsWith('@')`.
+- The gateway `complete.path` (server.py:8543) already special-cases `@` richly
+  (`@file:`, `@folder:`, `@diff`, `@staged`, `@url:`, `@git:`, fuzzy basename search).
+  Its `~`/`./` branches become dead trigger paths from this TUI — leave the gateway code
+  (Ink still uses the path forms; it's shared) but stop emitting those queries from
+  ui-opentui. **No gateway change.**
+- Net: typing `@` (even bare) opens the mention menu via the `@`-bare branch at
+  server.py:8555. Picking splices `@file:rel/path` etc. (existing accept path,
+  `completionFrom` honoured).
+
+**Tests:** extend `test/slash.test.ts` — `planCompletion('/')` → null; `planCompletion('/usr/bin')`
+→ null; `planCompletion('/model')` → complete.slash; multi-line `"a\n/mod"` → complete.slash
+on the trailing token; `"~/foo"` / `"./x"` → null (no longer path-like); `"@foo"` → complete.path.
+Keep them as behaviour assertions, not snapshots.
+
+---
+
+## F3 — cost: Nous-portal headers only (`tui_gateway` + `agent/usage_pricing.py`)
+
+**Current:** `_get_usage` (server.py:2157-2167) sets `cost_usd` from
+`real_session_cost_usd(agent)` (usage_pricing.py:887), which sums **two** provider-reported
+sources:
+1. `agent.session_actual_cost_usd` — OpenRouter `usage.cost` accumulator.
+2. `agent.get_credits_spent_micros()` — Nous `x-nous-credits-*` header delta.
+
+The TUI already **hides** the cost segment when `cost_usd` is absent (statusBar.tsx:241-243,
+`costText` returns '' when `costUsd === undefined`) — so this is purely "which sources count."
+
+**User's intent (F3):** cost should come **only from the Nous portal headers**; suppress it
+for every other route (cache-token pricing is unreliable across the model long tail).
+
+**Change:** make the OpenRouter accumulator source conditional on the route being Nous, OR
+drop source #1 entirely so only the header delta (source #2) feeds `cost_usd`. Source #2 is
+intrinsically Nous-only (the header only exists on Nous-portal responses), so dropping #1
+achieves "Nous-header-only" with one edit.
+
+> **DECISION D2 (needs glitch's confirm):** Drop OpenRouter's `session_actual_cost_usd`
+> source from `real_session_cost_usd`? Trade-off: OpenRouter's `usage.cost` is itself
+> *provider-reported* (the real charged number, not a Hermes estimate), so OR users lose an
+> accurate readout. But it removes the cache-token guesswork the user is worried about and
+> matches "only via the headers when using nous portal" literally.
+> **Recommended default (implementing unless told otherwise):** gate source #1 so it only
+> contributes when the active route is the Nous portal (base_url == nous inference api),
+> else it's dropped. This keeps the segment Nous-only AND avoids touching shared OR/CLI
+> behaviour for the `/usage` page. If even Nous-route OR-accumulator is unwanted, collapse
+> to header-only.
+
+**Scope guard:** `real_session_cost_usd` is also consumed by `/usage` page rendering
+(server.py:2237) and DB usage totals. Prefer a NEW, status-bar-specific helper
+(e.g. `nous_header_cost_usd(agent)`) wired only into `_get_usage`'s `cost_usd`, leaving the
+`/usage` accounting page untouched — so we don't regress the full cost report. Confirm with
+the gate + a gateway unit test (`tui_gateway` tests) that a non-Nous session yields no
+`cost_usd`.
+
+---
+
+## F4 — paste while composer unfocused (`view/composer.tsx`)
+
+**Current:** the global keyboard handler reclaims focus on a *printable keystroke*
+(`isPrintableKey`, composer.tsx:415-417). A **bracketed-paste event is not a keystroke** —
+it arrives at `onPaste` only if the textarea is focused, so an unfocused composer drops it;
+the user has to click/type first.
+
+**Fix:** the renderer delivers paste through the focused renderable. Two options:
+- (a) Keep focus on the composer more aggressively (opencode keeps the prompt focused via a
+  reactive effect). Risky — fights transcript scroll focus.
+- (b) **Recommended:** handle paste at the renderer/global level. Check whether OpenTUI
+  exposes a global paste hook (`renderer.on('paste')` or a keyboard event with
+  `key.name === 'paste'` / a paste event type). If a global paste signal exists, on paste:
+  `ta.focus()` then route the bytes into the existing `onPaste` logic (image / placeholder /
+  insert). **Must verify the API in the `opentui` skill before coding** (skill_view
+  references/docs). If only the focused-renderable paste exists, fall back to (a) scoped:
+  refocus the composer whenever no overlay/prompt is open and focus drifted (a
+  `createEffect` watching focus + `store.state.prompt`/overlay state).
+
+**Verify in live smoke** (tmux + tmux-pane-screenshot): scroll the transcript to drop focus,
+then paste — text must land without a prior click.
+
+---
+
+## F5 + F6 — clarify prompt rewrite (`view/prompts/clarifyPrompt.tsx`)
+
+Screenshot `/tmp/screenshots/SCR-20260613-iznq.png` confirms: long options run off the right
+edge (no wrap), options differ only by `▶`/`—` glyphs (no numbers, weak), and "✎ Other…" is
+a `<select>` row that *switches* to an input on Enter rather than being an inline input.
+
+**Current:** one native `<select>` over `[...choices, {Other}]` (clarifyPrompt.tsx:61-75).
+Native `<select>` doesn't wrap long rows and (F6) doesn't `preventDefault` arrows, so they
+leak to the transcript scrollbox.
+
+**Rewrite plan (verify renderable API in `opentui` skill first):**
+- Replace native `<select>` with a **custom keyboard-driven list** (a `For` over options +
+  a `selected` signal + `useKeyboard` with `key.preventDefault()` on up/down/enter — same
+  pattern the composer's `routeMenuKey` uses; F6 fixed by preventDefault so arrows never
+  reach the scrollbox).
+- **Wrapping (F5):** render each option as a `<text>` that wraps to the box width (no fixed
+  single-line). Indent continuation lines under the option label. Confirm `<text>` soft-wrap
+  behaviour in the opentui skill (it wraps by default within a flex box of bounded width).
+- **Differentiation (F5):** number every option `1.` `2.` … (digit hotkeys optional, nice-to-
+  have), and give the selected row the themed `selectionBg` + accent fg (the composer's
+  `completionCurrentBg` model), not just a glyph. Number + background + accent = three signals.
+- **Inline custom answer (F5):** render the `<input>` **inside the same screen, always
+  present** as the last "row" (an `Other:` labeled input), instead of an item that toggles.
+  Selecting/focusing it lets the user type; Enter in it submits the free text. Keep the
+  existing `clarify.respond {answer}` wiring. Arrow-down past the last choice lands on the
+  input; arrow-up from the input returns to the list (focus handoff like the composer↔tray).
+- Keep Esc/Ctrl+C → cancel (clarifyPrompt.tsx:31-33).
+
+**Reference:** opencode's selection/list components in `~/github/opencode/packages/tui` for
+the wrap + highlight + hotkey idiom; the composer dropdown (composer.tsx:441-458) for the
+in-repo highlight/selectable pattern.
+
+**Tests:** `test/render.test.tsx`-style headless frame — long option wraps (frame contains the
+tail of a long choice on a 2nd line), selected row shows numbered + highlighted, custom input
+present in the same frame, arrow keys don't change scrollTop (assert transcript scroll
+unchanged), Enter on a choice → onAnswer(choice), Enter in input → onAnswer(typed).
+
+---
+
+## F9 — `!cmd` runs bash (`entry/main.tsx` + a system render)
+
+**Backend exists:** `shell.exec` (server.py:10301) runs the command (30s timeout, dangerous/
+hardline-command guards, returns `{stdout, stderr, code}`).
+**Ink parity reference:** `ui-tui/src/app/useSubmission.ts:291` — `full.startsWith('!')` →
+`shellExec(full.slice(1).trim())` → appends a user line `!cmd` + a system line with output;
+the prompt glyph flips while the buffer starts with `!` (appLayout.tsx:178).
+
+**Plan (ui-opentui):**
+- In the entry `submit` (main.tsx:517-520), add a branch BEFORE the slash check:
+  `if (text.startsWith('!')) { runShell(text.slice(1).trim()); return }`.
+- `runShell(cmd)`: `store.pushUser('!' + cmd)` (echo the invocation in the transcript), then
+  `gateway.request('shell.exec', { command: cmd })`; on resolve, `store.pushSystem` the
+  combined `stdout`/`stderr` (or the error message / non-zero `code`); on reject,
+  pushSystem the error. Detached `runFork` like `submitPrompt`. No session turn, no model call.
+- Empty `!` (just the bang) → no-op (or a hint), matching Ink.
+- **Optional polish (parity, not required):** flip the composer prompt glyph (or tint) while
+  the buffer starts with `!`, like Ink's appLayout. Low-risk; do only if cheap.
+
+**Tests:** entry-level/logic test that a `!`-prefixed submit routes to `shell.exec` (not
+`prompt.submit`), and the system line renders stdout. Mirror the slashMenu.test harness
+(fake gateway capturing the method).
+
+---
+
+## Sequencing & fences (subagent-driven; disjoint files)
+
+Parallel-safe groups (disjoint file fences):
+1. **slash trigger** — `logic/slash.ts` (+ `logic/skillMatch.ts` reuse) + `test/slash.test.ts`. (F1/F2/F7/F8/F8b)
+2. **clarify** — `view/prompts/clarifyPrompt.tsx` + a clarify test. (F5/F6)
+3. **shell-exec** — `entry/main.tsx` (edit DIRECTLY — load-bearing) + system render + test. (F9)
+4. **paste focus** — `view/composer.tsx` (edit directly; verify opentui paste API first). (F4)
+5. **cost** — `tui_gateway/server.py` + `agent/usage_pricing.py` + gateway test. (F3) — Python, isolated.
+
+`entry/main.tsx` and `store.ts` are edited directly, never via subagent (handoff rule).
+Each renderable change: `skill_view(opentui, references/docs/...)` FIRST. Verify every
+subagent self-report (re-run `npm run check` exit code, read the diff).
+
+## Open decisions (need glitch)
+- **D1 (F7/F8):** thread cursor offset into `onType` (correct) vs. last-line slice (cheap)?
+  Recommend cursor offset.
+- **D2 (F3):** drop OpenRouter cost source entirely, or gate it to the Nous route? Recommend
+  Nous-route gate via a status-bar-only helper, leaving `/usage` accounting intact.
+
+## Invariants to preserve
+- Per-conversation prompt caching untouched (all client-render or post-hoc gateway usage).
+- No new `HERMES_*` env var (these are behaviour, not secrets).
+- Strict no change-detector tests — assert behaviour/invariants.
+- Don't regress the `/usage` accounting page when narrowing the chrome cost source.

docs/plans/opentui-usage-notice-chrome.md

@@ -0,0 +1,217 @@
+# OpenTUI — usage/credits notice in the composer chrome
+
+**Status:** spec (not started) · **Engine:** `ui-opentui/` · **Author:** glitch · 2026-06-14
+
+## Goal
+
+Render the gateway's **usage / credits notices** as a persistent, level-tinted
+**chrome banner pinned at the top of the input zone** (directly above the status
+bar), with the same lifecycle the Ink engine already has — sticky vs TTL,
+mid-turn hold + turn-end reveal, and "flash-and-yield" for the usage bands.
+
+Today the OpenTUI engine **receives** these notices but mis-renders them as
+scrolling inline transcript cards with no lifecycle. This spec fixes that without
+touching the gateway or the agent (the data already flows correctly).
+
+## What already exists (verified)
+
+### The wire (source of truth — do NOT change)
+The gateway emits one event for every notice, snake_case payload:
+
+```
+notification.show   payload { text, level, kind, ttl_ms, key, id }   # tui_gateway/server.py:2878
+notification.clear  payload { key }                                  # tui_gateway/server.py:2890
+```
+
+These come from `AgentNotice` (`agent/credits_tracker.py:177`). The credits
+policy (`evaluate_credits_notices`, `agent/credits_tracker.py:245`) emits exactly
+four notices — the full catalog this feature renders:
+
+| `key`                 | `text` (already glyphed by policy)              | `level`   | `kind`   | `ttl_ms` | lifecycle      |
+|-----------------------|-------------------------------------------------|-----------|----------|----------|----------------|
+| `credits.usage`       | `⚠/• Credits N% used · $X cap` (bands 50/75/90) | info/warn | `sticky` | —        | flash-and-yield |
+| `credits.grant_spent` | `• Grant spent · $X top-up left`                | info      | `sticky` | —        | flash-and-yield |
+| `credits.depleted`    | `✕ Credit access paused · run /usage for balance` | error   | `sticky` | —        | sticky          |
+| `credits.restored`    | `✓ Credit access restored`                      | success   | `ttl`    | `8000`   | TTL self-expire |
+
+**Load-bearing facts:**
+- `text` is **already glyphed** (⚠ • ✕ ✓) by the Python policy — the renderer
+  **must not** prepend another glyph. It only tints by `level`.
+- `level` includes **`success`** (green) — a level the current OpenTUI parser
+  silently drops to `info`.
+- `kind` is the **lifecycle marker** (`sticky` | `ttl`), NOT a display label.
+  `id` == `key` (stable per kind, not unique per emission).
+- Notices are **reconciled**: the policy emits `to_clear` (a `notification.clear`)
+  then `to_show`. A band change clears `credits.usage` then re-shows it.
+
+### The Ink reference behavior (what we're matching)
+`ui-tui/src/app/turnController.ts` + `appChrome.tsx`:
+- `showNotice` (`:181`): if **busy**, hold in `pendingNotice` (latest-wins);
+  if idle, apply now.
+- `applyNotice` (`:213`): set the visible notice; for `kind: 'ttl'` with
+  `ttl_ms > 0`, arm a self-expiry timer (clearing any prior timer first).
+- `clearNotice(key)` (`:198`): drop the visible **and** pending notice only when
+  the key matches (a stale clear must not wipe a newer notice).
+- `flushPendingNotice` (`:245`): at **turn end** (only the real end sites) apply
+  the held notice — its TTL clock starts here, when it first becomes visible.
+- **Flash-and-yield** (`startMessage`, `:917`): at **turn start**, if the visible
+  notice's key is `credits.usage` or `credits.grant_spent`, clear it — "show
+  once, then get out of the way." `credits.depleted` and others stay sticky. The
+  Python `active` latch keeps the key so it won't re-fire next turn.
+- Session reset clears all notice state so session A's notice can't bleed into B.
+- Color by level: `error→error`, `warn→warn`, `success→statusGood`,
+  `info→accent` (`noticeColor`, `appChrome.tsx:192`).
+
+### The OpenTUI side (what we change)
+- `notification.show` → `parseNotification` → `pushNotification` → **inline card**
+  in the transcript (`store.ts:832`, `notificationCard.tsx`). All kinds, no
+  lifecycle. The Option B process-completion card (`kind: 'process.complete'`)
+  and `background.complete` (`kind: 'background task complete'`) also use this
+  path — **they must keep working unchanged.**
+- `parseNotification` coerces `level` to `info|warn|error` only
+  (`backgroundActivity.ts:48`) — drops `success`.
+- Store carries `lastNotification` (OSC seam), `bgTasks`; **no** `notice` slot.
+- Theme has `accent`, `warn`, `error`, `ok`/`statusGood`, `muted`
+  (`logic/theme.ts`) — `success` maps to `statusGood`.
+- Input zone layout (`view/App.tsx:140-211`): a top-bordered column —
+  `<StatusBar>` → composer `<Switch>` → `<AgentsTray>`. The new banner mounts at
+  `App.tsx:144`, **directly above `<StatusBar>`** (the topmost line of the chrome).
+- Turn lifecycle hooks: `case 'message.start'` (`store.ts:779`, sets
+  `info.running = true`) and `case 'message.complete'` (`store.ts:811`, sets
+  `info.running = false`). `clearTranscript` (`store.ts:631`) is the reset site.
+- `Date.now()` is used freely in the store (`:877`) — `setTimeout` for TTL is fine.
+
+## The one design decision: routing
+
+`kind` is the discriminator. **`notification.show` with `kind === 'sticky'` or
+`kind === 'ttl'` → the new chrome-notice path; every other kind → the existing
+inline-card path, untouched.** This mirrors Ink's `Notice.kind: 'sticky' | 'ttl'`
+exactly, and the credits policy sets `kind` to one of those for all four notices,
+while the process/background cards use label-strings (`process.complete`,
+`background task complete`) that are neither — so they stay inline cards. No
+gateway change, no key-prefix sniffing.
+
+**Divergence from Ink (intentional):** Ink hides the notice while busy because the
+FaceTicker shares its one status slot. OpenTUI's busy face (`StatusLine`) lives in
+the transcript area, so the banner has a **dedicated row** and stays visible
+through a turn (a depletion warning shouldn't vanish mid-turn). We still **hold
+new notices** that arrive mid-turn (`pendingNotice`) and reveal them at turn end —
+matching Ink's "don't pop a fresh banner mid-stream" intent.
+
+## Implementation
+
+### Phase 1 — parser + type (`logic/backgroundActivity.ts`)
+1. Widen `ActivityNotification.level` to `'info' | 'warn' | 'error' | 'success'`.
+2. `coerceLevel`: also accept `'success'` (still fall back to `'info'`).
+3. Add `export function isChromeNotice(n: ActivityNotification): boolean` →
+   `n.kind === 'sticky' || n.kind === 'ttl'`.
+4. `parseNotification` already maps `ttl_ms → ttlMs` and preserves `key`/`id` —
+   no shape change beyond the widened level.
+
+**Tests** (`backgroundActivity.test.ts` or `notificationCard.test.tsx`):
+`success` survives parse; `kind: 'ttl'` + `ttl_ms` → `ttlMs`; `isChromeNotice`
+true for sticky/ttl, false for `process.complete`/`''`.
+
+### Phase 2 — store lifecycle (`logic/store.ts`)
+Add state + a private (non-reactive) timer handle in `createSessionStore`:
+- `notice: ActivityNotification | null` (visible chrome notice) — new state field,
+  init `null`.
+- `pendingNotice: ActivityNotification | null` — held mid-turn, init `null`.
+- `let noticeTimer: ReturnType<typeof setTimeout> | undefined` (closure var).
+
+Functions (port of `turnController`):
+- `showNotice(n)`: `state.info.running ? setState('pendingNotice', n) : applyNotice(n)`
+  (latest-wins — assigning replaces any prior pending).
+- `applyNotice(n)`: clear `noticeTimer`; `setState('notice', n)`; if
+  `n.kind === 'ttl' && n.ttlMs && n.ttlMs > 0`, arm `setTimeout(n.ttlMs)` that
+  clears `notice` only if `state.notice?.id === n.id` (defensive guard).
+- `clearNotice(key)`: if `state.pendingNotice?.key === key` → null it; if
+  `state.notice?.key === key` → clear timer + null `notice`.
+- `flushPendingNotice()`: if `state.pendingNotice` → `applyNotice` it, null pending.
+- `clearNoticeState()`: null `notice` + `pendingNotice`, clear timer.
+
+Wire into the event reducer:
+- `notification.show` (`store.ts:832`): route —
+  `const n = parseNotification(...); if (!n) break; if (isChromeNotice(n)) showNotice(n); else pushNotification(n)`.
+  (Still record `lastNotification` for the OSC seam in **both** paths — extract
+  the `setState('lastNotification', {...n})` so a chrome notice also pings a
+  blurred terminal, matching the inline-card behavior.)
+- `notification.clear` (`store.ts:837`): call **both** `clearNotificationCards(key)`
+  (cards) **and** `clearNotice(key)` (chrome) — a key only ever lives in one, so
+  calling both is safe and avoids guessing.
+- `message.start` (`store.ts:779`): flash-and-yield — if
+  `state.notice?.key === 'credits.usage' || === 'credits.grant_spent'` →
+  `clearNotice(state.notice.key)`. (Do this **before** flipping `running` true so
+  the read is clean.)
+- `message.complete` (`store.ts:811`): call `flushPendingNotice()` (after the
+  `running = false` set, so a held notice reveals on the now-idle bar).
+- `clearTranscript` (`store.ts:631`) and any session-switch reset:
+  `clearNoticeState()`.
+
+Export `notice` via the store's state and `showNotice`/`clearNotice` if a test or
+future slash command needs them.
+
+**Tests** (`statusNotice.test.ts`, new):
+- idle `showNotice` → `state.notice` set, no card pushed.
+- routing: `notification.show` `kind:'sticky'` → `notice` set, **no** transcript
+  card; `kind:'process.complete'` → card pushed, `notice` still null.
+- mid-turn hold: `message.start` → `showNotice` → `notice` stays null,
+  `pendingNotice` set → `message.complete` → `notice` revealed.
+- `clearNotice` by key drops visible + pending; non-matching key is a no-op.
+- TTL: `kind:'ttl', ttlMs:50` auto-clears (vitest fake timers).
+- flash-and-yield: visible `credits.usage` cleared on `message.start`;
+  `credits.depleted` persists across a start/complete cycle.
+- `clearTranscript` resets `notice` + `pendingNotice`.
+- `success` notice keeps its level.
+
+### Phase 3 — view (`view/noticeBanner.tsx` + `App.tsx`)
+New `NoticeBanner` (sibling style to `notificationCard.tsx`):
+- Props: `notice: ActivityNotification | null`, plus terminal width for truncation.
+- `<Show when={notice}>` — renders nothing when null.
+- One row, `flexShrink: 0`, `paddingLeft: 1`, `selectable={false}`.
+- Text rendered **verbatim** (glyph already present), tinted by level:
+  `error→error`, `warn→warn`, `success→statusGood`, `info→accent`.
+- Truncate to width with `truncRight` (`logic/truncate.ts`) so a long notice can
+  never push the composer or wrap.
+
+Mount in `App.tsx:144`, the first child of the top-bordered input zone, directly
+above `<StatusBar store={...} />`:
+```tsx
+<box border={['top']} ...>
+  <NoticeBanner notice={props.store.state.notice} />   {/* new */}
+  <StatusBar store={props.store} />
+  ...
+```
+
+**Tests** (`noticeBanner.test.tsx`, frame): renders the text without adding a
+glyph; warn→warn color, success→statusGood color; truncates at narrow width;
+renders an empty frame when `notice` is null.
+
+### Phase 4 — parity verification + docs
+- `npm run check` green (prettier + eslint + vitest).
+- Headless frame dump: a `credits.usage` warn banner above the status bar; a
+  `credits.depleted` error banner surviving a turn; a `credits.restored` success
+  banner that disappears after its TTL.
+- tmux smoke per `docs/opentui-dev-handoff.md` (inject the three notices via the
+  test harness / a scripted gateway event; screenshot the chrome).
+- Cross-check the four-notice catalog renders identically in tone to Ink's
+  `appChromeStatusRule` (color-by-level, no double glyph, truncation).
+
+## Non-goals
+- No gateway/agent changes — the wire and the policy are the source of truth.
+- No new notice kinds — render exactly the four the policy emits.
+- The inline-card path (process/background completions) is **unchanged**.
+- No status-bar segment changes — the banner is its own row above the bar.
+
+## Risk / footguns
+- **Schema decode-at-boundary**: `notification.show` payload is a loose Record
+  read by `parseNotification`, not strict-decoded — a wrong-typed field won't blank
+  the bar (unlike `applyInfo`). Keep the loose reads.
+- **createStore reference-aliasing**: store `notice` and `pendingNotice` distinct
+  objects; when applying pending, it's already its own object — don't alias it to
+  `lastNotification`. (See `[[solid-createstore-reference-aliasing]]`.)
+- **Timer leak**: `clearNoticeState` must clear `noticeTimer`; ensure session
+  reset and store dispose clear it so a TTL callback can't fire into a dead store.
+- **Routing regression**: assert in tests that `process.complete` /
+  `background task complete` still produce **cards**, not banners — the whole
+  feature hinges on the `kind` discriminator.

gateway/platforms/base.py

@@ -77,6 +77,13 @@ def _thread_metadata_for_source(source, reply_to_message_id: str | None = None)
     return metadata
 
 
+def _mark_notify_metadata(metadata: dict | None) -> dict:
+    """Clone metadata and mark a user-visible reply as notify-worthy."""
+    notify_metadata = dict(metadata) if metadata else {}
+    notify_metadata["notify"] = True
+    return notify_metadata
+
+
 def _reply_anchor_for_event(event) -> str | None:
     """Return reply_to id for platforms that need reply semantics.
 
@@ -3889,7 +3896,7 @@ class BasePlatformAdapter(ABC):
                     chat_id=event.source.chat_id,
                     content=_text,
                     reply_to=_reply_anchor_for_event(event),
-                    metadata=thread_meta,
+                    metadata=_mark_notify_metadata(thread_meta),
                 )
                 if _eph_ttl > 0 and _r.success and _r.message_id:
                     self._schedule_ephemeral_delete(
@@ -3995,7 +4002,7 @@ class BasePlatformAdapter(ABC):
                             chat_id=event.source.chat_id,
                             content=_text,
                             reply_to=_reply_anchor_for_event(event),
-                            metadata=_thread_meta,
+                            metadata=_mark_notify_metadata(_thread_meta),
                         )
                         if _eph_ttl > 0 and _r.success and _r.message_id:
                             self._schedule_ephemeral_delete(
@@ -4045,7 +4052,7 @@ class BasePlatformAdapter(ABC):
                                 chat_id=event.source.chat_id,
                                 content=_text,
                                 reply_to=_reply_anchor_for_event(event),
-                                metadata=_thread_meta,
+                                metadata=_mark_notify_metadata(_thread_meta),
                             )
                             if _eph_ttl > 0 and _r.success and _r.message_id:
                                 self._schedule_ephemeral_delete(
@@ -4268,6 +4275,12 @@ class BasePlatformAdapter(ABC):
                         )
                         text_content = _recovered
 
+                # Final user-visible content (text, TTS, media, files) gets
+                # the existing notify=True marker. Clone once so typing/status
+                # metadata stays unmarked and progress bubbles remain
+                # thread-strict.
+                _final_thread_metadata = _mark_notify_metadata(_thread_metadata)
+
                 # Auto-TTS: if voice message, generate audio FIRST (before sending text)
                 # Gated via ``_should_auto_tts_for_chat``: fires when the chat has
                 # an explicit ``/voice on|tts`` opt-in OR when ``voice.auto_tts`` is
@@ -4307,7 +4320,7 @@ class BasePlatformAdapter(ABC):
                             chat_id=event.source.chat_id,
                             audio_path=_tts_path,
                             caption=telegram_tts_caption,
-                            metadata=_thread_metadata,
+                            metadata=_final_thread_metadata,
                         )
                         _tts_caption_delivered = bool(
                             telegram_tts_caption and getattr(tts_result, "success", False)
@@ -4322,23 +4335,11 @@ class BasePlatformAdapter(ABC):
                 if text_content and not _tts_caption_delivered:
                     logger.info("[%s] Sending response (%d chars) to %s", self.name, len(text_content), event.source.chat_id)
                     _reply_anchor = _reply_anchor_for_event(event)
-                    # Mark final response messages for notification delivery.
-                    # Platform adapters that support per-message notification
-                    # control (e.g. Telegram's disable_notification) use this
-                    # flag to override silent-mode and ensure the final
-                    # response triggers a push notification.
-                    # Clone to avoid mutating the metadata shared with the
-                    # typing-indicator task (which must remain unmarked).
-                    if _thread_metadata is not None:
-                        _thread_metadata = dict(_thread_metadata)
-                        _thread_metadata["notify"] = True
-                    else:
-                        _thread_metadata = {"notify": True}
                     result = await self._send_with_retry(
                         chat_id=event.source.chat_id,
                         content=text_content,
                         reply_to=_reply_anchor,
-                        metadata=_thread_metadata,
+                        metadata=_final_thread_metadata,
                     )
                     _record_delivery(result)
 
@@ -4367,7 +4368,7 @@ class BasePlatformAdapter(ABC):
                         await self.send_multiple_images(
                             chat_id=event.source.chat_id,
                             images=images,
-                            metadata=_thread_metadata,
+                            metadata=_final_thread_metadata,
                             human_delay=human_delay,
                         )
                     except Exception as batch_err:
@@ -4409,7 +4410,7 @@ class BasePlatformAdapter(ABC):
                         await self.send_multiple_images(
                             chat_id=event.source.chat_id,
                             images=_batch,
-                            metadata=_thread_metadata,
+                            metadata=_final_thread_metadata,
                             human_delay=human_delay,
                         )
                     except Exception as batch_err:
@@ -4424,19 +4425,19 @@ class BasePlatformAdapter(ABC):
                             media_result = await self.send_voice(
                                 chat_id=event.source.chat_id,
                                 audio_path=media_path,
-                                metadata=_thread_metadata,
+                                metadata=_final_thread_metadata,
                             )
                         elif ext in _VIDEO_EXTS:
                             media_result = await self.send_video(
                                 chat_id=event.source.chat_id,
                                 video_path=media_path,
-                                metadata=_thread_metadata,
+                                metadata=_final_thread_metadata,
                             )
                         else:
                             media_result = await self.send_document(
                                 chat_id=event.source.chat_id,
                                 file_path=media_path,
-                                metadata=_thread_metadata,
+                                metadata=_final_thread_metadata,
                             )
 
                         if not media_result.success:
@@ -4454,13 +4455,13 @@ class BasePlatformAdapter(ABC):
                             await self.send_video(
                                 chat_id=event.source.chat_id,
                                 video_path=file_path,
-                                metadata=_thread_metadata,
+                                metadata=_final_thread_metadata,
                             )
                         else:
                             await self.send_document(
                                 chat_id=event.source.chat_id,
                                 file_path=file_path,
-                                metadata=_thread_metadata,
+                                metadata=_final_thread_metadata,
                             )
                     except Exception as file_err:
                         logger.error("[%s] Error sending local file %s: %s", self.name, file_path, file_err)

gateway/platforms/email.py

@@ -678,8 +678,13 @@ class EmailAdapter(BasePlatformAdapter):
         image_url: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        """Send an image URL as part of an email body."""
+        """Send an image URL as part of an email body.
+
+        ``metadata`` is accepted to honor the base-class contract; the
+        email body send doesn't use it.
+        """
         text = caption or ""
         text += f"\n\nImage: {image_url}"
         return await self.send(chat_id, text.strip(), reply_to)

gateway/platforms/whatsapp.py

@@ -846,13 +846,20 @@ class WhatsAppAdapter(WhatsAppBehaviorMixin, BasePlatformAdapter):
         image_url: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        """Download image URL to cache, send natively via bridge."""
+        """Download image URL to cache, send natively via bridge.
+
+        ``metadata`` is accepted to honor the base-class contract — the
+        batch sender ``send_multiple_images`` passes it through to every
+        send path. The bridge media call doesn't use it, matching the
+        sibling overrides (send_video / send_voice / send_document).
+        """
         try:
             local_path = await cache_image_from_url(image_url)
             return await self._send_media_to_bridge(chat_id, local_path, "image", caption)
         except Exception:
-            return await super().send_image(chat_id, image_url, caption, reply_to)
+            return await super().send_image(chat_id, image_url, caption, reply_to, metadata)
 
     async def send_image_file(
         self,
@@ -1136,6 +1143,15 @@ class WhatsAppAdapter(WhatsAppBehaviorMixin, BasePlatformAdapter):
             body = data.get("body", "")
             if data.get("isGroup"):
                 body = self._clean_bot_mention_text(body, data)
+
+            # If this is a reply, include the quoted message text so the agent
+            # knows exactly what the user is responding to (fixes "approve" context issue)
+            quoted_text = str(data.get("quotedText") or "").strip()
+            if quoted_text and data.get("hasQuotedMessage"):
+                # Truncate long quoted text to keep prompts reasonable
+                if len(quoted_text) > 300:
+                    quoted_text = quoted_text[:297] + "..."
+                body = f"[Replying to: \"{quoted_text}\"]\n{body}"
             MAX_TEXT_INJECT_BYTES = 100 * 1024
             if msg_type == MessageType.DOCUMENT and cached_urls:
                 for doc_path in cached_urls:

gateway/run.py

@@ -413,6 +413,57 @@ def _resolve_progress_thread_id(platform: Any, source_thread_id: Any, event_mess
     return None
 
 
+def _has_platform_display_override(user_config: dict, platform_key: str, setting: str) -> bool:
+    """Return True when display.platforms.<platform> explicitly sets setting."""
+    display = user_config.get("display") if isinstance(user_config, dict) else None
+    if not isinstance(display, dict):
+        return False
+    platforms = display.get("platforms")
+    if not isinstance(platforms, dict):
+        return False
+    platform_cfg = platforms.get(platform_key)
+    return isinstance(platform_cfg, dict) and setting in platform_cfg
+
+
+def _resolve_gateway_display_bool(
+    user_config: dict,
+    platform_key: str,
+    setting: str,
+    *,
+    default: bool = False,
+    platform: Any = None,
+    require_platform_override_for: set[Any] | None = None,
+) -> bool:
+    """Resolve a boolean display setting with optional platform-only opt-in.
+
+    Some display features expose assistant scratch text rather than deliberate
+    user-facing output.  For high-noise threaded chat surfaces such as
+    Mattermost, a global opt-in is too broad: they must be enabled with an
+    explicit display.platforms.<platform>.<setting> override.
+    """
+    current_platform = _gateway_platform_value(platform or platform_key)
+    platform_only = {
+        _gateway_platform_value(candidate)
+        for candidate in (require_platform_override_for or set())
+    }
+    if (
+        current_platform in platform_only
+        and not _has_platform_display_override(user_config, platform_key, setting)
+    ):
+        return False
+
+    from gateway.display_config import resolve_display_setting
+
+    value = resolve_display_setting(user_config, platform_key, setting, default)
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        return value.strip().lower() in {"true", "yes", "1", "on"}
+    if value is None:
+        return bool(default)
+    return bool(value)
+
+
 def _telegramize_command_mentions(text: str, platform: Any) -> str:
     """Rewrite slash-command mentions to Telegram-valid command names.
 
@@ -8989,17 +9040,24 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                     source, session_entry, reason="agent-result-compression",
                 )
 
-            # Prepend reasoning/thinking if display is enabled (per-platform)
+            # Prepend reasoning/thinking if display is enabled (per-platform).
+            # Mattermost requires explicit per-platform opt-in because this is
+            # scratch text, not ordinary final-answer content.
             try:
-                from gateway.display_config import resolve_display_setting as _rds
-                _show_reasoning_effective = _rds(
+                _show_reasoning_effective = _resolve_gateway_display_bool(
                     _load_gateway_config(),
                     _platform_config_key(source.platform),
                     "show_reasoning",
-                    getattr(self, "_show_reasoning", False),
+                    default=bool(getattr(self, "_show_reasoning", False)),
+                    platform=source.platform,
+                    require_platform_override_for={Platform.MATTERMOST},
                 )
             except Exception:
-                _show_reasoning_effective = getattr(self, "_show_reasoning", False)
+                _show_reasoning_effective = (
+                    False
+                    if source.platform == Platform.MATTERMOST
+                    else getattr(self, "_show_reasoning", False)
+                )
             if _show_reasoning_effective and response and not _intentional_silence:
                 last_reasoning = agent_result.get("last_reasoning")
                 if last_reasoning:
@@ -13635,18 +13693,32 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
         # in chat platforms while opting into concise mid-turn updates.
         interim_assistant_messages_enabled = (
             source.platform != Platform.WEBHOOK
-            and bool(
-                resolve_display_setting(
-                    user_config,
-                    platform_key,
-                    "interim_assistant_messages",
-                    True,
-                )
+            and _resolve_gateway_display_bool(
+                user_config,
+                platform_key,
+                "interim_assistant_messages",
+                default=True,
+                platform=source.platform,
+                require_platform_override_for={Platform.MATTERMOST},
             )
         )
-        
+        # thinking_progress is independent — if enabled, we need the progress
+        # queue even when tool_progress is off (thinking relay uses same infra).
+        # Mattermost requires a per-platform opt-in: global scratch-text display
+        # is too easy to leak into busy public threads.
+        _thinking_enabled = _resolve_gateway_display_bool(
+            user_config,
+            platform_key,
+            "thinking_progress",
+            default=False,
+            platform=source.platform,
+            require_platform_override_for={Platform.MATTERMOST},
+        )
+        needs_progress_queue = tool_progress_enabled or _thinking_enabled
+
+
         # Queue for progress messages (thread-safe)
-        progress_queue = queue.Queue() if tool_progress_enabled else None
+        progress_queue = queue.Queue() if needs_progress_queue else None
         last_tool = [None]  # Mutable container for tracking in closure
         last_progress_msg = [None]  # Track last message for dedup
         repeat_count = [0]  # How many times the same message repeated
@@ -13752,6 +13824,24 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                     logger.debug("tool-progress onboarding hint failed: %s", _hint_err)
                 return
 
+            # "_thinking" is assistant scratch text between tool calls.  It
+            # is never ordinary tool progress: only relay it when the platform
+            # explicitly opted into thinking_progress.  Handle both legacy
+            # callback shapes: ("_thinking", text) and
+            # ("reasoning.available", "_thinking", text, ...).
+            if event_type == "_thinking" or tool_name == "_thinking":
+                if not _thinking_enabled:
+                    return
+                thinking_text = preview if tool_name == "_thinking" else tool_name
+                msg = f"💬 {thinking_text}" if thinking_text else None
+                if msg:
+                    progress_queue.put(msg)
+                return
+
+            # If tool_progress is off, only _thinking passes through (above).
+            # Regular tool calls are suppressed.
+            if not tool_progress_enabled:
+                return
 
             # Only act on tool.started events (ignore tool.completed, reasoning.available, etc.)
             if event_type not in {"tool.started",}:
@@ -14783,6 +14873,10 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
 
             agent.clarify_callback = _clarify_callback_sync
 
+            # Show assistant thinking between tool calls — independent of
+            # tool_progress mode. Mattermost needs an explicit per-platform
+            # opt-in so global scratch-text display does not leak into threads.
+            agent.thinking_progress = _thinking_enabled
             # Store agent reference for interrupt support
             agent_holder[0] = agent
             # Capture the full tool definitions for transcript logging

gateway/stream_consumer.py

@@ -197,6 +197,30 @@ class GatewayStreamConsumer:
         # this response and route through edit-based for graceful degradation.
         self._draft_failures = 0
 
+    def _metadata_for_send(
+        self,
+        *,
+        final: bool = False,
+        expect_edits: bool = False,
+    ) -> dict | None:
+        """Return per-send metadata for stream-created messages.
+
+        Mattermost treats notify-worthy sends as user-visible final content
+        when deciding whether a broken thread root may fall back flat.  Preview
+        and progress sends keep their original metadata and remain thread-strict.
+
+        ``expect_edits`` preserves the upstream Telegram streaming contract:
+        preview messages that may be edited later must stay on the editable
+        legacy send path, while fresh/fallback final sends can still use richer
+        final-message delivery.
+        """
+        meta = dict(self.metadata) if self.metadata else {}
+        if expect_edits:
+            meta["expect_edits"] = True
+        if final:
+            meta["notify"] = True
+        return meta or None
+
     @property
     def already_sent(self) -> bool:
         """True if at least one message was sent or edited during the run."""
@@ -513,7 +537,11 @@ class GatewayStreamConsumer:
                         chunks_delivered = False
                         reply_to = self._message_id or self._initial_reply_to_id
                         for chunk in chunks:
-                            new_id = await self._send_new_chunk(chunk, reply_to)
+                            new_id = await self._send_new_chunk(
+                                chunk,
+                                reply_to,
+                                final=got_done,
+                            )
                             if new_id is not None and new_id != reply_to:
                                 chunks_delivered = True
                         self._accumulated = ""
@@ -749,7 +777,13 @@ class GatewayStreamConsumer:
         # Strip trailing whitespace/newlines but preserve leading content
         return cleaned.rstrip()
 
-    async def _send_new_chunk(self, text: str, reply_to_id: Optional[str]) -> Optional[str]:
+    async def _send_new_chunk(
+        self,
+        text: str,
+        reply_to_id: Optional[str],
+        *,
+        final: bool = False,
+    ) -> Optional[str]:
         """Send a new message chunk, optionally threaded to a previous message.
 
         Returns the message_id so callers can thread subsequent chunks.
@@ -758,15 +792,11 @@ class GatewayStreamConsumer:
         if not text.strip():
             return reply_to_id
         try:
-            meta = dict(self.metadata) if self.metadata else {}
-            # This chunk becomes the next edit target — adapters that support
-            # rich final sends (Telegram) must keep it on the editable path.
-            meta["expect_edits"] = True
             result = await self.adapter.send(
                 chat_id=self.chat_id,
                 content=text,
                 reply_to=reply_to_id,
-                metadata=meta,
+                metadata=self._metadata_for_send(final=final, expect_edits=True),
             )
             if result.success and result.message_id:
                 self._message_id = str(result.message_id)
@@ -885,7 +915,7 @@ class GatewayStreamConsumer:
                 result = await self.adapter.send(
                     chat_id=self.chat_id,
                     content=chunk,
-                    metadata=self.metadata,
+                    metadata=self._metadata_for_send(final=True),
                 )
                 if result.success:
                     break
@@ -1242,7 +1272,7 @@ class GatewayStreamConsumer:
             result = await self.adapter.send(
                 chat_id=self.chat_id,
                 content=text,
-                metadata=self.metadata,
+                metadata=self._metadata_for_send(final=True),
             )
         except Exception as e:
             logger.debug("Fresh-final send failed, falling back to edit: %s", e)
@@ -1532,7 +1562,10 @@ class GatewayStreamConsumer:
                     chat_id=self.chat_id,
                     content=text,
                     reply_to=self._initial_reply_to_id,
-                    metadata={**(self.metadata or {}), "expect_edits": True},
+                    metadata=self._metadata_for_send(
+                        final=finalize,
+                        expect_edits=True,
+                    ),
                 )
                 if result.success:
                     if result.message_id:

hermes_cli/_parser.py

@@ -145,8 +145,16 @@ def build_top_level_parser():
         "--resume",
         "-r",
         metavar="SESSION",
+        # nargs="?" + const=True: bare `--resume` parses to the sentinel True,
+        # which `hermes --tui` turns into the session picker
+        # (HERMES_TUI_RESUME=picker). `--resume <id|title>` is unchanged.
+        nargs="?",
+        const=True,
         default=None,
-        help="Resume a previous session by ID or title",
+        help=(
+            "Resume a previous session by ID or title. With --tui, bare "
+            "--resume (no argument) opens the session picker."
+        ),
     )
     parser.add_argument(
         "--continue",
@@ -301,8 +309,14 @@ def build_top_level_parser():
         "--resume",
         "-r",
         metavar="SESSION_ID",
+        # Same bare-flag picker sentinel as the top-level --resume.
+        nargs="?",
+        const=True,
         default=argparse.SUPPRESS,
-        help="Resume a previous session by ID (shown on exit)",
+        help=(
+            "Resume a previous session by ID (shown on exit). With --tui, "
+            "bare --resume opens the session picker."
+        ),
     )
     chat_parser.add_argument(
         "--continue",

hermes_cli/main.py

@@ -1640,8 +1640,286 @@ def _find_bundled_tui(hermes_cli_dir: Path | None = None) -> Path | None:
     return bundled if bundled.is_file() else None
 
 
+def _config_tui_engine_early() -> str | None:
+    """Read ``display.tui_engine`` from config via a minimal YAML read.
+
+    Returns the configured engine string, or ``None`` when unset/unreadable so the
+    caller can apply the availability-gated default. Mirrors
+    :func:`_config_default_interface_early`.
+    """
+    try:
+        home = os.environ.get("HERMES_HOME")
+        cfg_path = (
+            os.path.join(home, "config.yaml")
+            if home
+            else os.path.join(os.path.expanduser("~"), ".hermes", "config.yaml")
+        )
+        if os.path.exists(cfg_path):
+            import yaml as _yaml_eng
+
+            with open(cfg_path, encoding="utf-8") as _f:
+                raw = _yaml_eng.safe_load(_f) or {}
+            disp = raw.get("display", {})
+            if isinstance(disp, dict):
+                eng = disp.get("tui_engine")
+                if isinstance(eng, str) and eng.strip():
+                    return eng.strip().lower()
+    except Exception:
+        pass
+    return None
+
+
+def _resolve_tui_engine() -> str:
+    """Which TUI engine to launch: "ink" (default) or "opentui".
+
+    Precedence: ``HERMES_TUI_ENGINE`` env > ``display.tui_engine`` config >
+    (OpenTUI when this host can run it — Node >= 26.3 + the built package — else Ink).
+    The OpenTUI engine runs on Node 26.3+ via the experimental ``node:ffi`` renderer,
+    which is not validated on Windows or Termux — a request for "opentui" there falls
+    back to "ink" with a notice so a stale flag never strands the user on an engine
+    that can't start.
+    """
+    env = (os.environ.get("HERMES_TUI_ENGINE") or "").strip().lower()
+    # Explicit choice (env > config) wins; otherwise default to OpenTUI when this
+    # host is genuinely set up for it (Node >= 26.3 + the built bundle), else Ink.
+    engine = env or _config_tui_engine_early() or ("opentui" if _opentui_available() else "ink")
+    if engine != "opentui":
+        return "ink"
+
+    # opentui requested — gate on platform support.
+    unsupported = sys.platform.startswith("win") or _is_termux_startup_environment()
+    if unsupported:
+        if not os.environ.get("HERMES_QUIET"):
+            where = "Windows" if sys.platform.startswith("win") else "Termux"
+            print(
+                f"HERMES_TUI_ENGINE=opentui is not supported on {where} "
+                f"(needs Node 26.3+ with experimental FFI) — falling back to the Ink engine.",
+                file=sys.stderr,
+            )
+        return "ink"
+    return "opentui"
+
+
+NODE26_MIN_VERSION = (26, 3, 0)
+
+
+def _node_version_tuple(node_bin: str) -> tuple[int, int, int] | None:
+    """Return (major, minor, patch) for a node binary, or ``None`` if unreadable."""
+    try:
+        out = subprocess.run([node_bin, "--version"], capture_output=True, text=True, timeout=5)
+    except Exception:
+        return None
+    if out.returncode != 0:
+        return None
+    raw = (out.stdout or "").strip().lstrip("v").split("-", 1)[0]
+    parts = raw.split(".")
+    try:
+        return (int(parts[0]), int(parts[1]), int(parts[2]))
+    except (IndexError, ValueError):
+        return None
+
+
+def _fnm_node26_candidates() -> list[str]:
+    """Node binaries from fnm's installed versions, newest first.
+
+    fnm keeps each version at ``<FNM_DIR>/node-versions/v<X.Y.Z>/installation/
+    bin/node`` (default ``FNM_DIR``: ``$XDG_DATA_HOME/fnm`` or ``~/.local/share/
+    fnm``; macOS Homebrew also uses ``~/Library/Application Support/fnm``). When
+    the *active* node is older than 26.3 — e.g. the user's fnm default is on
+    v25 — the right 26.x is still installed and usable; surface it so OpenTUI
+    works without the user re-aliasing their global default. Version-sorted so
+    the newest qualifying node wins.
+    """
+    roots: list[Path] = []
+    fnm_dir = os.environ.get("FNM_DIR")
+    if fnm_dir:
+        roots.append(Path(fnm_dir))
+    xdg = os.environ.get("XDG_DATA_HOME")
+    if xdg:
+        roots.append(Path(xdg) / "fnm")
+    roots.append(Path.home() / ".local" / "share" / "fnm")
+    roots.append(Path.home() / "Library" / "Application Support" / "fnm")
+
+    seen: set[Path] = set()
+    found: list[tuple[tuple[int, int, int], str]] = []
+    for root in roots:
+        versions_dir = root / "node-versions"
+        if versions_dir in seen or not versions_dir.is_dir():
+            continue
+        seen.add(versions_dir)
+        try:
+            entries = list(versions_dir.iterdir())
+        except OSError:
+            continue
+        for entry in entries:
+            node_bin = entry / "installation" / "bin" / "node"
+            if not (node_bin.is_file() and os.access(node_bin, os.X_OK)):
+                continue
+            # Trust the directory name for sorting; the real probe happens in
+            # the caller (a renamed/symlinked dir still gets version-checked).
+            name = entry.name.lstrip("v").split("-", 1)[0]
+            parts = name.split(".")
+            try:
+                ver = (int(parts[0]), int(parts[1]), int(parts[2]))
+            except (IndexError, ValueError):
+                ver = (0, 0, 0)
+            found.append((ver, str(node_bin)))
+    found.sort(key=lambda pair: pair[0], reverse=True)
+    return [path for _, path in found]
+
+
+def _node26_bin_or_none() -> str | None:
+    """Resolve a Node >= 26.3.0 binary (no exit — a probe), or ``None``.
+
+    Order: ``HERMES_NODE`` override > ``node`` on PATH > newest fnm-installed
+    version. Each is gated on the real ``--version`` being >= 26.3.0. OpenTUI's
+    native renderer loads via the experimental ``node:ffi`` API that only exists
+    on Node 26.3+, so an older Node is treated as "not available" — but an
+    installed-yet-inactive 26.x (common when fnm's default is on an older line)
+    is discovered and used so the engine still launches.
+    """
+    candidates: list[str] = []
+    env_node = os.environ.get("HERMES_NODE")
+    if env_node and os.path.isfile(env_node) and os.access(env_node, os.X_OK):
+        candidates.append(env_node)
+    path = shutil.which("node")
+    if path:
+        candidates.append(path)
+    candidates.extend(_fnm_node26_candidates())
+    for cand in candidates:
+        ver = _node_version_tuple(cand)
+        if ver is not None and ver >= NODE26_MIN_VERSION:
+            return cand
+    return None
+
+
+def _node26_bin() -> str:
+    """Resolve Node >= 26.3.0 for the OpenTUI engine, or exit with a clear message.
+
+    Use :func:`_node26_bin_or_none` for a non-fatal availability probe.
+    """
+    node = _node26_bin_or_none()
+    if node is not None:
+        return node
+    print(
+        "Node.js >= 26.3.0 not found — the OpenTUI TUI engine needs it for the "
+        "experimental node:ffi renderer.\n"
+        "Install Node 26.3+ (e.g. via fnm/nvm) or set HERMES_NODE=/path/to/node, "
+        "or unset HERMES_TUI_ENGINE to use the default Ink engine.",
+        file=sys.stderr,
+    )
+    sys.exit(1)
+
+
+def _opentui_npm() -> str:
+    """Resolve npm (ships with Node) to build the OpenTUI bundle, or exit."""
+    npm = shutil.which("npm")
+    if npm:
+        return npm
+    print(
+        "npm not found — needed to build the OpenTUI engine bundle.\n"
+        "Install Node 26.3+ (it ships npm), or unset HERMES_TUI_ENGINE for Ink.",
+        file=sys.stderr,
+    )
+    sys.exit(1)
+
+
+def _opentui_available() -> bool:
+    """Whether the OpenTUI engine can actually launch on this host.
+
+    True only when the platform is supported (not Windows/Termux), a Node >= 26.3
+    binary resolves (the node:ffi floor), AND the v2 package is BUILT
+    (``dist/main.js``) with its ``node_modules`` installed. This gates the DEFAULT
+    engine: a host genuinely set up for OpenTUI defaults to it; everyone else stays
+    on Ink. An explicit ``HERMES_TUI_ENGINE`` env or ``display.tui_engine`` config
+    choice bypasses this probe (and triggers an on-demand build).
+    """
+    if sys.platform.startswith("win") or _is_termux_startup_environment():
+        return False
+    if _node26_bin_or_none() is None:
+        return False
+    pkg = PROJECT_ROOT / "ui-opentui"
+    built = pkg / "dist" / "main.js"
+    return built.is_file() and (pkg / "node_modules" / "@opentui").is_dir()
+
+
+def _make_opentui_argv(tui_dev: bool) -> tuple[list[str], Path]:
+    """Argv for the native OpenTUI engine under Node 26 (no Bun).
+
+    Builds the Solid + Effect-at-boundary engine (``ui-opentui``) with esbuild
+    (``npm run build`` → ``dist/main.js``) when the bundle is missing (or always, in
+    ``--dev``), then launches it on Node with the experimental FFI flag:
+
+        node --experimental-ffi --no-warnings dist/main.js
+
+    ``--no-warnings`` keeps the ExperimentalWarning off the TUI's stderr. Returns the
+    argv and the package cwd.
+
+    The spawned ``tui_gateway`` resolves its Python from ``HERMES_PYTHON_SRC_ROOT``
+    (the caller sets it to ``PROJECT_ROOT``); the built bundle's own fallback also
+    walks up to the checkout root, so the gateway resolves correctly either way.
+    """
+    app_dir = PROJECT_ROOT / "ui-opentui"
+    entry_src = app_dir / "src" / "entry" / "main.tsx"
+    if not entry_src.is_file():
+        print(
+            f"OpenTUI v2 engine entry not found at {entry_src}.\n"
+            f"Unset HERMES_TUI_ENGINE to use the default Ink engine.",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+
+    node = _node26_bin()
+
+    # The esbuild build needs the package's node_modules (esbuild + the @opentui
+    # packages + the native blob). Without them the build/launch dies cryptically.
+    if not (app_dir / "node_modules" / "@opentui").is_dir():
+        print(
+            f"OpenTUI engine dependencies are not installed in {app_dir}.\n"
+            f"Run:  (cd {app_dir} && npm install)\n"
+            f"Or unset HERMES_TUI_ENGINE to use the default Ink engine.",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+
+    built = app_dir / "dist" / "main.js"
+    if tui_dev or not built.is_file():
+        npm = _opentui_npm()
+        if not os.environ.get("HERMES_QUIET"):
+            print("Building the OpenTUI engine…", file=sys.stderr)
+        result = subprocess.run(
+            [npm, "run", "build"],
+            cwd=str(app_dir),
+            capture_output=True,
+            text=True,
+        )
+        if result.returncode != 0:
+            combined = f"{result.stdout or ''}{result.stderr or ''}".strip()
+            preview = "\n".join(combined.splitlines()[-30:])
+            print("OpenTUI engine build failed.", file=sys.stderr)
+            if preview:
+                print(preview, file=sys.stderr)
+            sys.exit(1)
+
+    # --expose-gc (parity with Ink, main.py ~1909): makes `global.gc()` a real
+    # callable so the OpenTUI engine's GC hooks (W2 proactive idle GC; /heapdump)
+    # work instead of being silent no-ops. MUST be an argv flag — Node rejects
+    # --expose-gc in NODE_OPTIONS (see the heap-cap injection below).
+    return [node, "--experimental-ffi", "--no-warnings", "--expose-gc", str(built)], app_dir
+
+
 def _make_tui_argv(tui_dir: Path, tui_dev: bool) -> tuple[list[str], Path]:
-    """TUI: --dev → tsx src; else node dist (HERMES_TUI_DIR prebuilt or esbuild)."""
+    """TUI: --dev → tsx src; else node dist (HERMES_TUI_DIR prebuilt or esbuild).
+
+    Dual-engine: when ``HERMES_TUI_ENGINE``/``display.tui_engine`` selects the
+    native OpenTUI engine, dispatch to ``_make_opentui_argv`` (Node 26 + its own
+    esbuild build) BEFORE the Ink Node bootstrap — the OpenTUI engine resolves its
+    own Node >= 26.3 and builds its own bundle, so it must not be routed through
+    ``_ensure_tui_node`` / the Ink prebuilt-dir logic.
+    """
+    if _resolve_tui_engine() == "opentui":
+        return _make_opentui_argv(tui_dev)
+
     _ensure_tui_node()
 
     def _node_bin(bin: str) -> str:
@@ -1877,6 +2155,57 @@ def _read_cgroup_memory_limit() -> Optional[int]:
     return None
 
 
+def _config_tui_heap_mb_early() -> int | None:
+    """Read ``display.tui_heap_mb`` from config via a minimal YAML read.
+
+    Returns the configured V8 heap cap in MB, or ``None`` when unset/unreadable.
+    Mirrors :func:`_config_tui_engine_early`. A non-secret behavioral setting, so
+    it lives in ``config.yaml`` (NOT a ``HERMES_*`` env / the NODE_OPTIONS bridge,
+    which is denylisted) — the ``HERMES_TUI_HEAP_MB`` env is only the per-launch
+    override on top of this.
+    """
+    try:
+        home = os.environ.get("HERMES_HOME")
+        cfg_path = (
+            os.path.join(home, "config.yaml")
+            if home
+            else os.path.join(os.path.expanduser("~"), ".hermes", "config.yaml")
+        )
+        if os.path.exists(cfg_path):
+            import yaml as _yaml_heap
+
+            with open(cfg_path, encoding="utf-8") as _f:
+                raw = _yaml_heap.safe_load(_f) or {}
+            disp = raw.get("display", {})
+            if isinstance(disp, dict):
+                val = disp.get("tui_heap_mb")
+                if isinstance(val, bool):  # guard: YAML true/false is an int subclass
+                    return None
+                if isinstance(val, int) and val > 0:
+                    return val
+                if isinstance(val, str) and val.strip().isdigit():
+                    n = int(val.strip())
+                    if n > 0:
+                        return n
+    except Exception:
+        pass
+    return None
+
+
+def _resolve_tui_heap_override() -> int | None:
+    """The user's explicit V8 heap cap (MB), or ``None`` for the default path.
+
+    Precedence: ``HERMES_TUI_HEAP_MB`` env > ``display.tui_heap_mb`` config
+    (matches the ``HERMES_TUI_ENGINE`` env-first pattern). Honored by BOTH engines
+    via the shared ``NODE_OPTIONS`` injection. A positive integer wins; anything
+    else (unset/garbage/non-positive) falls through to the cgroup-aware default.
+    """
+    env_val = os.environ.get("HERMES_TUI_HEAP_MB", "").strip()
+    if env_val.isdigit() and int(env_val) > 0:
+        return int(env_val)
+    return _config_tui_heap_mb_early()
+
+
 def _resolve_tui_heap_mb(default_mb: int = 8192) -> int:
     """Pick a V8 ``--max-old-space-size`` (MB) that fits the container.
 
@@ -1885,7 +2214,16 @@ def _resolve_tui_heap_mb(default_mb: int = 8192) -> int:
     cgroup limit so the heap + non-heap RSS stays under the cgroup ceiling,
     clamped to a sane floor (1536MB — below this V8 GC-thrashes and the TUI
     is barely usable).  Never exceeds ``default_mb``.
+
+    An explicit ``HERMES_TUI_HEAP_MB`` env / ``display.tui_heap_mb`` config
+    override REPLACES the 8192 default (D3): setting it low is the low-mem opt-in,
+    setting it high raises the ceiling. The cgroup-fit clamp still applies on top
+    so an override never exceeds what the container can hold — a low override is
+    honored as-is, a too-high one is still trimmed to ~75% of the cgroup limit.
     """
+    override = _resolve_tui_heap_override()
+    if override is not None:
+        default_mb = override
     limit = _read_cgroup_memory_limit()
     if not limit:
         return default_mb
@@ -1902,7 +2240,8 @@ def _resolve_tui_heap_mb(default_mb: int = 8192) -> int:
 
 
 def _launch_tui(
-    resume_session_id: Optional[str] = None,
+    # str session id, the bare-`--resume` picker sentinel True, or None.
+    resume_session_id: "Optional[str | bool]" = None,
     tui_dev: bool = False,
     model: Optional[str] = None,
     provider: Optional[str] = None,
@@ -1921,6 +2260,14 @@ def _launch_tui(
     """Replace current process with the TUI."""
     tui_dir = PROJECT_ROOT / "ui-tui"
 
+    # Bare `--resume` arrives as the argparse sentinel True: open the TUI
+    # resume picker instead of resuming a specific session id. Normalize it
+    # here so everything downstream (exit summary, env forwarding) keeps
+    # seeing either a real session id string or None.
+    resume_picker = resume_session_id is True
+    if resume_picker:
+        resume_session_id = None
+
     import tempfile
 
     env = os.environ.copy()
@@ -1934,11 +2281,31 @@ def _launch_tui(
     )
     os.close(active_session_fd)
     env["HERMES_TUI_ACTIVE_SESSION_FILE"] = active_session_file
+    # Tree-sitter grammar cache for the OpenTUI engine: grammars are fetched
+    # from GitHub on first use and cached here (profile-aware). Unset → OpenTUI
+    # falls back to its XDG default ($XDG_DATA_HOME/opentui). See
+    # ui-opentui/src/boundary/parsers.ts.
+    try:
+        from hermes_cli.config import get_hermes_home
+
+        env["HERMES_TUI_PARSER_CACHE"] = str(
+            get_hermes_home() / "cache" / "opentui-parsers"
+        )
+    except Exception:
+        logger.debug("Failed to resolve OpenTUI parser cache dir", exc_info=True)
     env["HERMES_PYTHON_SRC_ROOT"] = os.environ.get(
         "HERMES_PYTHON_SRC_ROOT", str(PROJECT_ROOT)
     )
     env.setdefault("HERMES_PYTHON", sys.executable)
     env.setdefault("HERMES_CWD", os.getcwd())
+    # The TUI subprocess is launched with cwd=<engine package dir> (so its
+    # build/resolution works), which means the gateway it spawns would otherwise
+    # auto-detect THAT dir as the workspace (chrome bar showed "ui-opentui" no
+    # matter where you ran hermes). TERMINAL_CWD is the gateway's canonical
+    # launch-dir channel (_completion_cwd) — set it to the real cwd here so the
+    # session, chrome bar, and terminal tool all anchor to where you actually
+    # are. Worktree mode overrides it to the worktree path below.
+    env.setdefault("TERMINAL_CWD", os.getcwd())
     env.setdefault("NODE_ENV", "development" if tui_dev else "production")
 
     wt_info = None
@@ -2015,6 +2382,11 @@ def _launch_tui(
     # --expose-gc is *not* added here: Node rejects it in NODE_OPTIONS
     # ("--expose-gc is not allowed in NODE_OPTIONS") and refuses to start.
     # It is passed as a direct argv flag in _make_tui_argv() instead.
+    #
+    # Both TUI engines run on Node/V8 now — Ink, and the native OpenTUI engine
+    # (Node 26 + node:ffi). So --max-old-space-size (a V8/Node flag) applies to
+    # both. (Pre-Node-26 the OpenTUI engine ran on Bun/JavaScriptCore, which has
+    # no such flag; that gate is gone now that the engine is Node.)
     _tokens = env.get("NODE_OPTIONS", "").split()
     if not any(t.startswith("--max-old-space-size=") for t in _tokens):
         _tokens.append(f"--max-old-space-size={_resolve_tui_heap_mb()}")
@@ -2027,7 +2399,11 @@ def _launch_tui(
     # resolved for this invocation; direct `node ui-tui/dist/entry.js` users can
     # still set HERMES_TUI_RESUME themselves.
     env.pop("HERMES_TUI_RESUME", None)
-    if resume_session_id:
+    if resume_picker:
+        # Bare --resume: tell the TUI to open the resume picker before any
+        # session.create (create is lazy, so nothing is wasted).
+        env["HERMES_TUI_RESUME"] = "picker"
+    elif resume_session_id:
         env["HERMES_TUI_RESUME"] = resume_session_id
 
     argv, cwd = _make_tui_argv(tui_dir, tui_dev)
@@ -2136,6 +2512,18 @@ def cmd_chat(args):
     """Run interactive chat CLI."""
     use_tui = _resolve_use_tui(args)
 
+    # Bare `--resume` (argparse sentinel True) opens the TUI resume picker —
+    # `_launch_tui` translates it to HERMES_TUI_RESUME=picker. The classic
+    # REPL has no picker overlay, so point at the equivalents instead of
+    # silently resuming something the user didn't choose.
+    if getattr(args, "resume", None) is True and not use_tui:
+        print("Bare --resume opens the session picker, which requires the TUI.")
+        print(
+            "Use 'hermes --tui --resume', 'hermes --resume <id|title>', "
+            "'hermes -c', or 'hermes sessions browse'."
+        )
+        sys.exit(2)
+
     # Resolve --continue into --resume with the latest session or by name
     continue_val = getattr(args, "continue_last", None)
     if continue_val and not getattr(args, "resume", None):
@@ -2161,9 +2549,10 @@ def cmd_chat(args):
                 print(f"No previous {kind} session found to continue.")
                 sys.exit(1)
 
-    # Resolve --resume by title if it's not a direct session ID
+    # Resolve --resume by title if it's not a direct session ID. The bare
+    # picker sentinel (True) is not a name — leave it for _launch_tui.
     resume_val = getattr(args, "resume", None)
-    if resume_val:
+    if resume_val and resume_val is not True:
         resolved = _resolve_session_by_name_or_id(resume_val)
         if resolved:
             args.resume = resolved

scripts/install.sh

@@ -268,7 +268,7 @@ emit_manifest() {
     if [ "$INCLUDE_DESKTOP" = true ]; then
         desktop_stage='{"name":"desktop","title":"Build desktop app","category":"runtime","needs_user_input":false},'
     fi
-    printf '%s' '{"protocol_version":1,"stages":[{"name":"prerequisites","title":"System prerequisites","category":"runtime","needs_user_input":false},{"name":"repository","title":"Download Hermes Agent","category":"runtime","needs_user_input":false},{"name":"venv","title":"Create Python virtual environment","category":"runtime","needs_user_input":false},{"name":"python-deps","title":"Install Python dependencies","category":"runtime","needs_user_input":false},{"name":"node-deps","title":"Install browser-tool dependencies","category":"runtime","needs_user_input":false},{"name":"path","title":"Install hermes command","category":"runtime","needs_user_input":false},{"name":"config","title":"Prepare config and skills","category":"configuration","needs_user_input":false},{"name":"setup","title":"Configure API keys and settings","category":"configuration","needs_user_input":true},{"name":"gateway","title":"Configure gateway service","category":"configuration","needs_user_input":true},'"$desktop_stage"'{"name":"complete","title":"Finish install","category":"runtime","needs_user_input":false}]}'
+    printf '%s' '{"protocol_version":1,"stages":[{"name":"prerequisites","title":"System prerequisites","category":"runtime","needs_user_input":false},{"name":"repository","title":"Download Hermes Agent","category":"runtime","needs_user_input":false},{"name":"venv","title":"Create Python virtual environment","category":"runtime","needs_user_input":false},{"name":"python-deps","title":"Install Python dependencies","category":"runtime","needs_user_input":false},{"name":"node-deps","title":"Install browser-tool dependencies","category":"runtime","needs_user_input":false},{"name":"opentui-engine","title":"Set up OpenTUI engine","category":"runtime","needs_user_input":false},{"name":"path","title":"Install hermes command","category":"runtime","needs_user_input":false},{"name":"config","title":"Prepare config and skills","category":"configuration","needs_user_input":false},{"name":"setup","title":"Configure API keys and settings","category":"configuration","needs_user_input":true},{"name":"gateway","title":"Configure gateway service","category":"configuration","needs_user_input":true},'"$desktop_stage"'{"name":"complete","title":"Finish install","category":"runtime","needs_user_input":false}]}'
     printf '\n'
 }
 
@@ -1980,6 +1980,76 @@ install_node_deps() {
     restore_dirty_lockfiles "$INSTALL_DIR"
 }
 
+# Provision the native OpenTUI engine on NODE 26.3+ (no Bun): `npm install` +
+# `npm run build` (esbuild → dist/main.js) in ui-opentui. The engine's
+# renderer loads via the experimental `node:ffi` API that only exists on Node
+# 26.3+. The launcher (hermes_cli/main.py:_opentui_available) only uses OpenTUI
+# when a Node >= 26.3 resolves AND the v2 package is built; otherwise it falls
+# back to the Ink engine. So this stage is STRICTLY best-effort: any failure
+# (unsupported platform, Node < 26.3, no network, install/build fails) logs a
+# warning and returns 0. A skipped OpenTUI setup just means the user gets Ink —
+# breaking the install would be far worse than skipping OpenTUI. Every sub-step
+# is guarded; this function never `exit`s and never returns non-zero.
+install_opentui() {
+    # node:ffi isn't validated on Windows/Termux — keep those hosts on Ink.
+    if [ "$OS" = "windows" ] || [ "$DISTRO" = "termux" ] || [ "$OS" = "android" ]; then
+        log_info "Skipping OpenTUI engine (unsupported platform) — using Ink."
+        return 0
+    fi
+
+    # Only meaningful if the v2 package is present in this checkout.
+    if [ ! -f "$INSTALL_DIR/ui-opentui/package.json" ]; then
+        log_info "Skipping OpenTUI engine (ui-opentui not present) — using Ink."
+        return 0
+    fi
+
+    log_info "Setting up OpenTUI engine (native TUI, Node 26.3+ / node:ffi)..."
+
+    # Resolve a Node >= 26.3.0 (the node:ffi floor): HERMES_NODE > node on PATH,
+    # version-checked. We do NOT install Node here — if one new enough isn't
+    # available the launcher cleanly falls back to Ink.
+    local node_bin=""
+    for cand in "${HERMES_NODE:-}" "$(command -v node 2>/dev/null || true)"; do
+        [ -n "$cand" ] && [ -x "$cand" ] || continue
+        if "$cand" -e 'const p=process.versions.node.split(".").map(Number); process.exit(p[0]>26||(p[0]===26&&p[1]>=3)?0:1)' 2>/dev/null; then
+            node_bin="$cand"
+            break
+        fi
+    done
+    if [ -z "$node_bin" ]; then
+        log_warn "OpenTUI engine setup skipped (needs Node >= 26.3.0; none found) — using the Ink engine. Install Node 26.3+ or set HERMES_NODE."
+        return 0
+    fi
+    log_success "Node found ($("$node_bin" --version 2>/dev/null || echo "unknown"))"
+
+    # npm ships with Node; the build (`node scripts/build.mjs`) runs fine on any
+    # recent Node — only the runtime needs 26.3, which the launcher re-checks.
+    local npm_bin
+    npm_bin="$(command -v npm 2>/dev/null || true)"
+    if [ -z "$npm_bin" ]; then
+        log_warn "OpenTUI engine setup skipped (npm not found) — using the Ink engine."
+        return 0
+    fi
+
+    cd "$INSTALL_DIR/ui-opentui" || { log_warn "OpenTUI engine setup skipped (cd failed) — using Ink."; return 0; }
+
+    # Pull deps (fetches the per-arch @opentui/core-<arch> native lib) then build
+    # the Node bundle (dist/main.js). Both idempotent.
+    log_info "Installing OpenTUI dependencies (npm install)..."
+    if ! "$npm_bin" install --no-audit --no-fund >/dev/null 2>&1; then
+        log_warn "OpenTUI engine setup skipped (npm install failed) — the Ink engine will be used."
+        return 0
+    fi
+    log_info "Building OpenTUI engine (npm run build)..."
+    if ! "$npm_bin" run build >/dev/null 2>&1; then
+        log_warn "OpenTUI engine setup skipped (build failed) — the Ink engine will be used."
+        return 0
+    fi
+
+    log_success "OpenTUI engine ready (opt-in: HERMES_TUI_ENGINE=opentui; default is Ink)."
+    return 0
+}
+
 run_setup_wizard() {
     if [ "$RUN_SETUP" = false ]; then
         log_info "Skipping setup wizard (--skip-setup)"
@@ -2636,6 +2706,12 @@ run_stage_body() {
             check_node
             install_node_deps
             ;;
+        opentui-engine)
+            detect_os
+            resolve_install_layout
+            require_install_dir
+            install_opentui
+            ;;
         path)
             detect_os
             resolve_install_layout
@@ -2743,6 +2819,7 @@ main() {
     setup_venv
     install_deps
     install_node_deps
+    install_opentui
     setup_path
     copy_config_templates
     run_setup_wizard

scripts/release.py

@@ -90,6 +90,7 @@ AUTHOR_MAP = {
     "290859878+synapsesx@users.noreply.github.com": "synapsesx",
     "157689911+itsflownium@users.noreply.github.com": "itsflownium",
     "dirtyren@users.noreply.github.com": "dirtyren",
+    "evansrory@gmail.com": "zimigit2020",
     "237263164+ft-ioxcs@users.noreply.github.com": "ft-ioxcs",
     "tharushkadinujaya05@gmail.com": "0xneobyte",
     "138671361+Veritas-7@users.noreply.github.com": "Veritas-7",

tests/agent/test_auxiliary_client.py

@@ -1653,6 +1653,37 @@ class TestAuxiliaryFallbackLayering:
         exc.status_code = 402
         return exc
 
+    def test_auto_provider_uses_task_then_main_chain_before_builtin_chain(self, monkeypatch):
+        """Auto aux call failures try per-task then top-level fallback before built-ins."""
+        primary_client = MagicMock()
+        primary_client.chat.completions.create.side_effect = self._make_payment_err()
+
+        main_chain_client = MagicMock()
+        main_chain_client.chat.completions.create.return_value = MagicMock(choices=[
+            MagicMock(message=MagicMock(content="from main fallback chain"))
+        ])
+
+        with patch("agent.auxiliary_client._get_cached_client",
+                   return_value=(primary_client, "qwen/qwen3.5-122b-a10b")), \
+             patch("agent.auxiliary_client._resolve_task_provider_model",
+                   return_value=("auto", None, None, None, None)), \
+             patch("agent.auxiliary_client._try_configured_fallback_chain",
+                   return_value=(None, None, "")) as mock_task_chain, \
+             patch("agent.auxiliary_client._try_main_fallback_chain",
+                   return_value=(main_chain_client, "inclusionai/ring-2.6-1t:free", "openrouter")) as mock_main_chain, \
+             patch("agent.auxiliary_client._try_payment_fallback") as mock_builtin_chain:
+            result = call_llm(
+                task="title_generation",
+                messages=[{"role": "user", "content": "hello"}],
+            )
+
+        assert main_chain_client.chat.completions.create.called
+        mock_task_chain.assert_called_once_with(
+            "title_generation", "auto", reason="payment error")
+        mock_main_chain.assert_called_once_with(
+            "title_generation", "auto", reason="payment error")
+        mock_builtin_chain.assert_not_called()
+
     def test_explicit_provider_uses_configured_chain_first(self, monkeypatch, caplog):
         """When a user has fallback_chain configured, it's tried BEFORE the main agent model."""
         monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")

tests/agent/test_auxiliary_main_first.py

@@ -118,6 +118,64 @@ class TestResolveAutoMainFirst:
         assert client is chain_client
         assert model == "google/gemini-3-flash-preview"
 
+    def test_main_unavailable_uses_task_fallback_chain_before_builtin_chain(self):
+        """Auto aux resolution honors auxiliary.<task>.fallback_chain before built-ins."""
+        task_client = MagicMock()
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="nvidia",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="qwen/qwen3.5-122b-a10b",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client",
+            return_value=(None, None),  # main provider has no client
+        ), patch(
+            "agent.auxiliary_client._try_configured_fallback_chain",
+            return_value=(task_client, "task-free-model", "fallback_chain[0](openrouter)"),
+        ) as mock_task_chain, patch(
+            "agent.auxiliary_client._try_main_fallback_chain",
+        ) as mock_main_chain, patch(
+            "agent.auxiliary_client._try_openrouter",
+        ) as mock_openrouter:
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto(task="title_generation")
+
+        assert client is task_client
+        assert model == "task-free-model"
+        mock_task_chain.assert_called_once_with(
+            "title_generation", "nvidia", reason="main provider unavailable")
+        mock_main_chain.assert_not_called()
+        mock_openrouter.assert_not_called()
+
+    def test_main_unavailable_uses_main_fallback_chain_before_builtin_chain(self):
+        """Auto aux resolution honors top-level fallback_providers before built-ins."""
+        main_fallback_client = MagicMock()
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="nvidia",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="qwen/qwen3.5-122b-a10b",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client",
+            return_value=(None, None),  # main provider has no client
+        ), patch(
+            "agent.auxiliary_client._try_configured_fallback_chain",
+            return_value=(None, None, ""),
+        ), patch(
+            "agent.auxiliary_client._try_main_fallback_chain",
+            return_value=(main_fallback_client, "inclusionai/ring-2.6-1t:free", "openrouter"),
+        ) as mock_main_chain, patch(
+            "agent.auxiliary_client._try_openrouter",
+        ) as mock_openrouter:
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto(task="title_generation")
+
+        assert client is main_fallback_client
+        assert model == "inclusionai/ring-2.6-1t:free"
+        mock_main_chain.assert_called_once_with(
+            "title_generation", "nvidia", reason="main provider unavailable")
+        mock_openrouter.assert_not_called()
+
     def test_no_main_config_uses_chain_directly(self):
         """No main provider configured → skip step 1, use chain (no regression)."""
         chain_client = MagicMock()

tests/gateway/test_mattermost.py

@@ -6,7 +6,10 @@ import pytest
 from unittest.mock import MagicMock, patch, AsyncMock
 
 from gateway.config import Platform, PlatformConfig
-from gateway.run import _resolve_progress_thread_id
+from gateway.run import (
+    _resolve_gateway_display_bool,
+    _resolve_progress_thread_id,
+)
 
 
 class TestMattermostProgressThreadRouting:
@@ -32,6 +35,97 @@ class TestMattermostProgressThreadRouting:
         ) is None
 
 
+class TestMattermostDisplayHygiene:
+    def test_mattermost_requires_platform_opt_in_for_interim_assistant_messages(self):
+        """Global interim commentary must not make Mattermost leak scratch notes."""
+        user_config = {"display": {"interim_assistant_messages": True}}
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "mattermost",
+            "interim_assistant_messages",
+            default=True,
+            platform=Platform.MATTERMOST,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is False
+
+    def test_mattermost_platform_opt_in_can_enable_interim_assistant_messages(self):
+        """Mattermost can still opt into commentary explicitly per platform."""
+        user_config = {
+            "display": {
+                "interim_assistant_messages": False,
+                "platforms": {
+                    "mattermost": {"interim_assistant_messages": True},
+                },
+            }
+        }
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "mattermost",
+            "interim_assistant_messages",
+            default=True,
+            platform=Platform.MATTERMOST,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is True
+
+    def test_mattermost_requires_platform_opt_in_for_thinking_progress(self):
+        """Global thinking_progress must not surface internal analysis in Mattermost."""
+        user_config = {"display": {"thinking_progress": True}}
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "mattermost",
+            "thinking_progress",
+            default=False,
+            platform=Platform.MATTERMOST,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is False
+
+    def test_mattermost_requires_platform_opt_in_for_show_reasoning(self):
+        """Global show_reasoning must not prepend scratch reasoning in Mattermost."""
+        user_config = {"display": {"show_reasoning": True}}
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "mattermost",
+            "show_reasoning",
+            default=False,
+            platform=Platform.MATTERMOST,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is False
+
+    def test_mattermost_platform_opt_in_can_enable_show_reasoning(self):
+        user_config = {
+            "display": {
+                "show_reasoning": False,
+                "platforms": {"mattermost": {"show_reasoning": True}},
+            }
+        }
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "mattermost",
+            "show_reasoning",
+            default=False,
+            platform=Platform.MATTERMOST,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is True
+
+    def test_global_thinking_progress_still_applies_to_other_platforms(self):
+        """The Mattermost guard must not silently neuter Telegram/other chats."""
+        user_config = {"display": {"thinking_progress": True}}
+
+        assert _resolve_gateway_display_bool(
+            user_config,
+            "telegram",
+            "thinking_progress",
+            default=False,
+            platform=Platform.TELEGRAM,
+            require_platform_override_for={Platform.MATTERMOST},
+        ) is True
+
+
 # ---------------------------------------------------------------------------
 # Platform & Config
 # ---------------------------------------------------------------------------
@@ -347,6 +441,24 @@ class TestMattermostSend:
         payload = self.adapter._api_post.call_args_list[0][0][1]
         assert payload["root_id"] == "root_post"
 
+    @pytest.mark.asyncio
+    async def test_progress_send_with_invalid_thread_root_never_falls_back_flat(self):
+        """Tool/status/progress bubbles must stay quiet when the thread is broken."""
+        self.adapter._reply_mode = "thread"
+        self.adapter._api_get = AsyncMock(return_value={"id": "bad_root", "root_id": ""})
+        self.adapter._api_post = AsyncMock(return_value={})
+
+        result = await self.adapter.send(
+            "channel_1",
+            "⚙️ terminal...",
+            metadata={"thread_id": "bad_root"},
+        )
+
+        assert result.success is False
+        assert self.adapter._api_post.call_count == 1
+        payload = self.adapter._api_post.call_args_list[0][0][1]
+        assert payload["root_id"] == "bad_root"
+
     @pytest.mark.asyncio
     async def test_send_api_failure(self):
         """When API returns error, send should return failure."""

tests/gateway/test_media_metadata_contract.py

@@ -0,0 +1,80 @@
+"""Contract: media-send overrides must accept the ``metadata`` kwarg.
+
+``BasePlatformAdapter.send_multiple_images`` passes ``metadata=metadata``
+to ``send_image`` / ``send_image_file`` / ``send_animation`` on every send.
+An override whose signature stops at ``reply_to`` raises ``TypeError:
+send_image() got an unexpected keyword argument 'metadata'`` at runtime —
+which is exactly how image delivery broke on WhatsApp and email.
+
+This mirrors ``test_discord_media_metadata.py`` but covers the two
+adapters that previously slipped, plus a best-effort sweep over every
+adapter that imports cleanly so the next slip is caught at test time.
+"""
+
+from __future__ import annotations
+
+import importlib
+import inspect
+
+import pytest
+
+
+def _accepts_metadata(method) -> bool:
+    params = inspect.signature(method).parameters
+    if "metadata" in params:
+        return True
+    # A ``**kwargs`` catch-all also absorbs metadata (the convention used by
+    # WhatsApp's send_video / send_voice / send_document overrides).
+    return any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+
+
+# (module, class) for the two adapters this fix targeted. These must import
+# in CI, so assert directly rather than skipping.
+@pytest.mark.parametrize(
+    "module_name, class_name",
+    [
+        ("gateway.platforms.whatsapp", "WhatsAppAdapter"),
+        ("gateway.platforms.email", "EmailAdapter"),
+    ],
+)
+def test_send_image_accepts_metadata(module_name, class_name):
+    cls = getattr(importlib.import_module(module_name), class_name)
+    assert _accepts_metadata(cls.send_image), (
+        f"{class_name}.send_image must accept 'metadata' (or **kwargs) — "
+        f"send_multiple_images passes it on every send"
+    )
+
+
+# Best-effort sweep across all shipped adapters. Modules whose optional
+# platform SDK isn't installed are skipped; an adapter that imports but
+# whose override drops metadata is a hard failure.
+_ALL_ADAPTERS = [
+    ("gateway.platforms.bluebubbles", "BlueBubblesAdapter"),
+    ("gateway.platforms.dingtalk", "DingTalkAdapter"),
+    ("gateway.platforms.discord", "DiscordAdapter"),
+    ("gateway.platforms.email", "EmailAdapter"),
+    ("gateway.platforms.feishu", "FeishuAdapter"),
+    ("gateway.platforms.matrix", "MatrixAdapter"),
+    ("gateway.platforms.mattermost", "MattermostAdapter"),
+    ("gateway.platforms.signal", "SignalAdapter"),
+    ("gateway.platforms.slack", "SlackAdapter"),
+    ("gateway.platforms.telegram", "TelegramAdapter"),
+    ("gateway.platforms.wecom", "WeComAdapter"),
+    ("gateway.platforms.weixin", "WeixinAdapter"),
+    ("gateway.platforms.whatsapp", "WhatsAppAdapter"),
+    ("gateway.platforms.yuanbao", "YuanbaoAdapter"),
+]
+
+
+@pytest.mark.parametrize("module_name, class_name", _ALL_ADAPTERS)
+def test_all_adapters_send_image_metadata_sweep(module_name, class_name):
+    try:
+        module = importlib.import_module(module_name)
+    except Exception as exc:  # optional platform dep not installed
+        pytest.skip(f"{module_name} not importable: {exc}")
+    cls = getattr(module, class_name, None)
+    if cls is None or "send_image" not in cls.__dict__:
+        pytest.skip(f"{class_name} has no send_image override")
+    assert _accepts_metadata(cls.send_image), (
+        f"{class_name}.send_image drops the 'metadata' kwarg"
+    )

tests/gateway/test_stream_consumer_thread_routing.py

@@ -106,6 +106,42 @@ class TestInitialReplyToId:
         assert call_kwargs["metadata"] == {**metadata, "expect_edits": True}
         assert metadata == {"thread_id": "omt_topic789"}
 
+    @pytest.mark.asyncio
+    async def test_final_first_send_marks_metadata_notify_true(self):
+        """Final streaming sends should use the existing notify=True marker."""
+        adapter = _make_adapter()
+        consumer = GatewayStreamConsumer(
+            adapter,
+            "chat_123",
+            metadata={"thread_id": "root_post_123"},
+            initial_reply_to_id="reply_post_456",
+        )
+
+        await consumer._send_or_edit("Final answer", finalize=True)
+
+        call_kwargs = adapter.send.call_args[1]
+        metadata = call_kwargs["metadata"]
+        assert metadata["thread_id"] == "root_post_123"
+        assert metadata["notify"] is True
+        assert "delivery_kind" not in metadata
+        assert "allow_flat_fallback" not in metadata
+
+    @pytest.mark.asyncio
+    async def test_nonfinal_first_send_does_not_mark_notify(self):
+        """Preview/interim streaming sends must not be notify-worthy."""
+        adapter = _make_adapter()
+        consumer = GatewayStreamConsumer(
+            adapter,
+            "chat_123",
+            metadata={"thread_id": "root_post_123"},
+            initial_reply_to_id="reply_post_456",
+        )
+
+        await consumer._send_or_edit("Preview", finalize=False)
+
+        metadata = adapter.send.call_args[1]["metadata"]
+        assert metadata == {"thread_id": "root_post_123", "expect_edits": True}
+
 
 class TestOverflowFirstMessage:
     """Verify thread routing is preserved when the first message overflows."""

tests/gateway/test_telegram_overflow_partial.py

@@ -134,7 +134,7 @@ async def test_stream_consumer_fallback_sends_tail_after_partial_overflow():
 
     adapter.send.assert_awaited_once()
     assert adapter.send.await_args.kwargs["content"] == "world"
-    assert adapter.send.await_args.kwargs["metadata"] == {"thread_id": "77"}
+    assert adapter.send.await_args.kwargs["metadata"] == {"thread_id": "77", "notify": True}
     adapter.delete_message.assert_not_awaited()
     assert consumer.final_response_sent is True
     assert consumer.final_content_delivered is True

tests/gateway/test_tts_media_routing.py

@@ -76,7 +76,7 @@ async def test_base_adapter_routes_telegram_flac_media_tag_to_document_sender(tm
     adapter.send_document.assert_awaited_once_with(
         chat_id="chat-1",
         file_path=str(media_file),
-        metadata=None,
+        metadata={"notify": True},
     )
     adapter.send_voice.assert_not_awaited()
 
@@ -95,7 +95,7 @@ async def test_base_adapter_routes_non_voice_telegram_ogg_media_tag_to_document_
     adapter.send_document.assert_awaited_once_with(
         chat_id="chat-1",
         file_path=str(media_file),
-        metadata=None,
+        metadata={"notify": True},
     )
     adapter.send_voice.assert_not_awaited()
 
@@ -116,7 +116,7 @@ async def test_base_adapter_routes_voice_tagged_telegram_ogg_media_tag_to_voice_
     adapter.send_voice.assert_awaited_once_with(
         chat_id="chat-1",
         audio_path=str(media_file),
-        metadata=None,
+        metadata={"notify": True},
     )
     adapter.send_document.assert_not_awaited()
 

tests/hermes_cli/test_opentui_node_resolution.py

@@ -0,0 +1,69 @@
+"""Node-26 resolution for the OpenTUI engine + the launch-cwd channel.
+
+Regression coverage for two ways the local TUI silently fell back to Ink /
+showed the wrong directory:
+
+1. fnm's active/default node was on an older line (v25) while a usable v26.3
+   sat installed-but-inactive — ``_node26_bin_or_none`` only checked
+   ``HERMES_NODE`` + ``which node`` and so reported "no node 26" → OpenTUI
+   unavailable → Ink fallback.
+2. ``TERMINAL_CWD`` (the gateway's launch-dir channel) was only exported in
+   worktree mode, so a normal launch let the gateway auto-detect the engine's
+   own package dir as the workspace.
+"""
+
+import os
+import stat
+
+import pytest
+
+import hermes_cli.main as main_mod
+
+
+def _fake_node(path, version: str) -> None:
+    """Write a stub `node` that prints `version` for `--version`."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(f'#!/bin/sh\necho "{version}"\n')
+    path.chmod(path.stat().st_mode | stat.S_IEXEC | stat.S_IXGRP | stat.S_IXOTH)
+
+
+class TestFnmNode26Discovery:
+    def test_discovers_inactive_v26_when_default_is_older(self, tmp_path, monkeypatch):
+        """A v26.3 installed under fnm is found even when PATH `node` is v25."""
+        fnm_dir = tmp_path / "fnm"
+        for ver in ("24.11.0", "25.9.0", "26.3.0"):
+            _fake_node(fnm_dir / "node-versions" / f"v{ver}" / "installation" / "bin" / "node", f"v{ver}")
+        monkeypatch.setenv("FNM_DIR", str(fnm_dir))
+        monkeypatch.delenv("HERMES_NODE", raising=False)
+        # PATH node is the too-old default (v25).
+        monkeypatch.setattr(main_mod.shutil, "which", lambda _b: str(
+            fnm_dir / "node-versions" / "v25.9.0" / "installation" / "bin" / "node"
+        ))
+
+        resolved = main_mod._node26_bin_or_none()
+        assert resolved is not None
+        assert "v26.3.0" in resolved  # newest qualifying, not the v25 default
+
+    def test_candidates_sorted_newest_first(self, tmp_path, monkeypatch):
+        fnm_dir = tmp_path / "fnm"
+        for ver in ("26.1.0", "26.4.0", "25.0.0"):
+            _fake_node(fnm_dir / "node-versions" / f"v{ver}" / "installation" / "bin" / "node", f"v{ver}")
+        monkeypatch.setenv("FNM_DIR", str(fnm_dir))
+        cands = main_mod._fnm_node26_candidates()
+        # Directory-name order: 26.4 before 26.1 before 25.0.
+        idx = [next(i for i, c in enumerate(cands) if f"v{v}" in c) for v in ("26.4.0", "26.1.0", "25.0.0")]
+        assert idx == sorted(idx)
+
+    def test_no_fnm_dir_is_empty_not_error(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("FNM_DIR", str(tmp_path / "does-not-exist"))
+        monkeypatch.setenv("XDG_DATA_HOME", str(tmp_path / "xdg-none"))
+        monkeypatch.setattr(main_mod.Path, "home", classmethod(lambda cls: tmp_path / "home-none"))
+        assert main_mod._fnm_node26_candidates() == []
+
+    def test_hermes_node_still_wins(self, tmp_path, monkeypatch):
+        """An explicit HERMES_NODE >= 26.3 takes precedence over fnm discovery."""
+        explicit = tmp_path / "explicit" / "node"
+        _fake_node(explicit, "v26.5.0")
+        monkeypatch.setenv("HERMES_NODE", str(explicit))
+        monkeypatch.setattr(main_mod.shutil, "which", lambda _b: None)
+        assert main_mod._node26_bin_or_none() == str(explicit)

tests/hermes_cli/test_tui_heap_sizing.py

@@ -99,6 +99,70 @@ class TestResolveTuiHeapMb:
         assert self._resolve(64 * GB) == 8192
 
 
+class TestHeapOverride:
+    """HERMES_TUI_HEAP_MB env / display.tui_heap_mb config override (W1/D3).
+
+    The override REPLACES the 8192 default; the cgroup-fit clamp still applies on
+    top so a too-high override can't exceed the container. Precedence: env > config.
+    """
+
+    def _resolve(self, limit_bytes, env=None, config_mb=None):
+        with mock.patch.object(m, "_read_cgroup_memory_limit", return_value=limit_bytes), \
+             mock.patch.object(m, "_config_tui_heap_mb_early", return_value=config_mb), \
+             mock.patch.dict(m.os.environ, env or {}, clear=False):
+            if env is None:
+                m.os.environ.pop("HERMES_TUI_HEAP_MB", None)
+            return m._resolve_tui_heap_mb()
+
+    def test_env_override_unconstrained(self):
+        # explicit low cap, no cgroup limit -> used as-is (the low-mem opt-in).
+        assert self._resolve(None, env={"HERMES_TUI_HEAP_MB": "256"}) == 256
+
+    def test_env_override_raises_ceiling(self):
+        # a higher-than-default cap is honored when unconstrained.
+        assert self._resolve(None, env={"HERMES_TUI_HEAP_MB": "16384"}) == 16384
+
+    def test_env_wins_over_config(self):
+        assert self._resolve(None, env={"HERMES_TUI_HEAP_MB": "512"}, config_mb=4096) == 512
+
+    def test_config_used_when_no_env(self):
+        assert self._resolve(None, config_mb=2048) == 2048
+
+    def test_override_still_cgroup_clamped(self):
+        # user asks for 16GB but the container is 4GB -> trimmed to 75% = 3072.
+        assert self._resolve(4 * GB, env={"HERMES_TUI_HEAP_MB": "16384"}) == 3072
+
+    def test_low_override_honored_under_big_container(self):
+        # a deliberately low cap is NOT raised by a roomy container.
+        assert self._resolve(16 * GB, env={"HERMES_TUI_HEAP_MB": "256"}) == 256
+
+    def test_garbage_env_falls_through_to_default(self):
+        assert self._resolve(None, env={"HERMES_TUI_HEAP_MB": "nope"}) == 8192
+
+    def test_nonpositive_env_falls_through(self):
+        assert self._resolve(None, env={"HERMES_TUI_HEAP_MB": "0"}) == 8192
+
+
+class TestExposeGcOnOpenTuiArgv:
+    """W1/D4: the OpenTUI engine argv must carry --expose-gc (parity with Ink) so
+    global.gc() is a real call, not a no-op."""
+
+    def test_opentui_argv_has_expose_gc(self, tmp_path):
+        app_dir = tmp_path / "ui-opentui"
+        (app_dir / "src" / "entry").mkdir(parents=True)
+        (app_dir / "src" / "entry" / "main.tsx").write_text("// entry")
+        (app_dir / "node_modules" / "@opentui").mkdir(parents=True)
+        (app_dir / "dist").mkdir()
+        (app_dir / "dist" / "main.js").write_text("// built")
+        with mock.patch.object(m, "PROJECT_ROOT", tmp_path), \
+             mock.patch.object(m, "_node26_bin", return_value="/usr/bin/node"):
+            argv, cwd = m._make_opentui_argv(tui_dev=False)
+        assert "--expose-gc" in argv
+        assert argv[0] == "/usr/bin/node"
+        assert argv[-1].endswith("dist/main.js")
+        assert cwd == app_dir
+
+
 class TestNodeOptionsTokenMerge:
     """The _launch_tui token-merge block must add the sized cap unless the user
     already supplied one, and must preserve unrelated NODE_OPTIONS flags."""

tests/hermes_cli/test_tui_npm_install.py

@@ -14,6 +14,20 @@ def main_mod():
     return m
 
 
+@pytest.fixture(autouse=True)
+def _pin_ink_engine(monkeypatch):
+    """These tests exercise the Ink/npm bootstrap inside ``_make_tui_argv``.
+
+    The dual-engine dispatch (``_resolve_tui_engine``) auto-selects the native
+    OpenTUI engine whenever ``ui-opentui/dist`` is built in the repo, which
+    would route ``_make_tui_argv`` away from the npm path under test. Pin the
+    engine to ink, mirroring test_tui_resume_flow.py.
+    """
+    import hermes_cli.main as m
+
+    monkeypatch.setattr(m, "_resolve_tui_engine", lambda: "ink")
+
+
 def _touch_ink(root: Path) -> None:
     ink = root / "node_modules" / "@hermes" / "ink" / "package.json"
     ink.parent.mkdir(parents=True, exist_ok=True)

tests/hermes_cli/test_tui_resume_flow.py

@@ -118,6 +118,82 @@ def test_cmd_chat_tui_resume_resolves_title_before_launch(monkeypatch, main_mod)
     assert captured["resume"] == "20260409_000000_aa11bb"
 
 
+def test_bare_resume_parses_to_picker_sentinel():
+    from hermes_cli._parser import build_top_level_parser
+
+    parser, _subparsers, _chat_parser = build_top_level_parser()
+
+    args = parser.parse_args(["--tui", "--resume"])
+    assert args.resume is True
+
+    args = parser.parse_args(["--resume", "abc123"])
+    assert args.resume == "abc123"
+
+    args = parser.parse_args(["chat", "--tui", "--resume"])
+    assert args.resume is True
+
+
+def test_cmd_chat_tui_bare_resume_skips_resolution_and_launches_picker(
+    monkeypatch, main_mod
+):
+    captured = {}
+
+    def fake_launch(resume_session_id=None, **kwargs):
+        captured["resume"] = resume_session_id
+        raise SystemExit(0)
+
+    def boom(_val):
+        raise AssertionError("bare --resume must not hit name/id resolution")
+
+    monkeypatch.setattr(main_mod, "_resolve_session_by_name_or_id", boom)
+    monkeypatch.setattr(main_mod, "_launch_tui", fake_launch)
+
+    with pytest.raises(SystemExit):
+        main_mod.cmd_chat(_args(resume=True))
+
+    assert captured["resume"] is True
+
+
+def test_cmd_chat_bare_resume_without_tui_exits_with_guidance(
+    monkeypatch, capsys, main_mod
+):
+    monkeypatch.setattr(main_mod, "_resolve_use_tui", lambda args: False)
+    monkeypatch.setattr(
+        main_mod,
+        "_launch_tui",
+        lambda *a, **kw: pytest.fail("must not launch the TUI"),
+    )
+
+    with pytest.raises(SystemExit) as exc:
+        main_mod.cmd_chat(_args(tui=False, resume=True))
+
+    assert exc.value.code == 2
+    out = capsys.readouterr().out
+    assert "requires the TUI" in out
+    assert "hermes --tui --resume" in out
+
+
+def test_launch_tui_sets_picker_env_for_bare_resume(monkeypatch, main_mod):
+    captured = {}
+
+    monkeypatch.setenv("HERMES_TUI_RESUME", "stale-missing-session")
+    monkeypatch.setattr(
+        main_mod,
+        "_make_tui_argv",
+        lambda tui_dir, tui_dev: (["node", "dist/entry.js"], Path(".")),
+    )
+    monkeypatch.setattr(
+        main_mod.subprocess,
+        "call",
+        lambda argv, cwd=None, env=None: captured.update({"env": env}) or 1,
+    )
+
+    with pytest.raises(SystemExit):
+        main_mod._launch_tui(resume_session_id=True)
+
+    assert captured["env"]["HERMES_TUI_RESUME"] == "picker"
+
+
 def test_cmd_chat_tui_passes_model_and_provider(monkeypatch, main_mod):
     captured = {}
 
@@ -1008,6 +1084,10 @@ def test_make_tui_argv_dev_prebuilds_hermes_ink(monkeypatch, main_mod, tmp_path)
     monkeypatch.setattr(main_mod, "_tui_need_npm_install", lambda _tui_dir: False)
     monkeypatch.delenv("HERMES_TUI_DIR", raising=False)
     monkeypatch.setattr(main_mod.shutil, "which", lambda bin_name: f"/usr/bin/{bin_name}")
+    # _make_tui_argv now dispatches on the TUI engine first; resolving "opentui"
+    # availability probes `node --version` (a subprocess.run this test would
+    # otherwise record). Pin the Ink engine — this test covers the Ink dev path.
+    monkeypatch.setattr(main_mod, "_resolve_tui_engine", lambda: "ink")
 
     calls = []
 

tests/run_agent/test_async_httpx_del_neuter.py

@@ -176,11 +176,16 @@ class TestClientCacheBoundedGrowth:
         """When the loop changes, the old entry should be replaced, not duplicated."""
         from agent.auxiliary_client import (
             _client_cache,
+            _client_cache_key,
             _client_cache_lock,
             _get_cached_client,
         )
 
-        key = ("test_replace", True, "", "", "", (), False, "")
+        key = _client_cache_key(
+            "test_replace",
+            async_mode=True,
+            task="",
+        )
 
         # Simulate a stale entry from a closed loop
         old_loop = asyncio.new_event_loop()

tests/test_tui_gateway_server.py

@@ -7,6 +7,7 @@ import time
 import types
 from datetime import datetime
 from pathlib import Path
+import pytest
 from unittest.mock import patch
 
 from hermes_constants import reset_hermes_home_override, set_hermes_home_override
@@ -380,6 +381,78 @@ def test_tui_verbose_tool_events_omit_details_when_redaction_fails(monkeypatch):
     assert "result_text" not in events[1][2]
 
 
+def test_tool_complete_emits_full_unified_diff(monkeypatch):
+    events: list[tuple[str, str, dict]] = []
+    monkeypatch.setattr(
+        server, "_emit", lambda event_type, sid, payload: events.append((event_type, sid, payload))
+    )
+    monkeypatch.setitem(
+        server._sessions,
+        "diff-test",
+        {"tool_progress_mode": "concise", "tool_started_at": {}, "edit_snapshots": {}},
+    )
+
+    diff = "--- a/x.py\n+++ b/x.py\n@@ -1 +1 @@\n-a = 1\n+a = 2\n"
+    result = json.dumps({"success": True, "diff": diff})
+    server._on_tool_complete("diff-test", "tool-1", "patch", {"mode": "replace", "path": "x.py"}, result)
+
+    assert events and events[0][0] == "tool.complete"
+    payload = events[0][2]
+    # the raw unified diff rides alongside the pretty/capped inline_diff
+    assert payload["diff_unified"] == diff
+    assert "inline_diff" in payload
+
+
+def test_verbose_result_text_drops_diff_echo_when_diff_unified_ships(monkeypatch):
+    # A tall edit's result JSON embeds the WHOLE diff; tail-capping that echo
+    # yields an unparseable JSON-looking fragment the TUI can't suppress
+    # reliably. When diff_unified ships, result_text must carry only the
+    # non-diff signal — small, parseable, never the diff echo.
+    events: list[tuple[str, str, dict]] = []
+    monkeypatch.setattr(
+        server, "_emit", lambda event_type, sid, payload: events.append((event_type, sid, payload))
+    )
+    monkeypatch.setitem(
+        server._sessions,
+        "diff-echo-test",
+        {"tool_progress_mode": "verbose", "tool_started_at": {}, "edit_snapshots": {}},
+    )
+
+    lines = "\n".join(f"+def fn_{i}() -> int: return {i}" for i in range(60))
+    diff = f"--- a/x.py\n+++ b/x.py\n@@ -1,0 +1,60 @@\n{lines}\n"
+    result = json.dumps(
+        {"success": True, "diff": diff, "files_modified": ["x.py"], "_warning": "stale read"}
+    )
+    server._on_tool_complete("diff-echo-test", "tool-1", "patch", {"mode": "patch"}, result)
+
+    payload = events[0][2]
+    assert payload["diff_unified"] == diff
+    text = payload["result_text"]
+    assert "[showing verbose tail" not in text  # small enough to dodge the cap
+    parsed = json.loads(text)  # parseable …
+    assert "diff" not in parsed  # … with the echo gone
+    assert parsed["_warning"] == "stale read"  # non-diff signal survives
+    # without diff_unified (non-edit tools) the result_text is untouched
+    assert server._result_sans_diff_echo("plain text result") == "plain text result"
+    assert server._result_sans_diff_echo('{"output": "x"}') == '{"output": "x"}'
+
+
+def test_cap_diff_unified_truncates_at_line_boundary():
+    line = "+" + "x" * 63  # 64 bytes per line incl. newline
+    diff = "\n".join([line] * 100)
+
+    capped = server._cap_diff_unified(diff, max_bytes=1000)
+
+    body, _, marker = capped.rpartition("\n")
+    assert marker.startswith("# … diff truncated (")
+    assert marker.endswith(" more bytes)")
+    assert len(body.encode("utf-8")) <= 1000
+    # cut on a line boundary: every surviving line is intact
+    assert all(l == line for l in body.split("\n"))
+    # under the cap → untouched
+    assert server._cap_diff_unified("small", max_bytes=1000) == "small"
+
+
 def test_dispatch_rejects_non_object_request():
     resp = server.dispatch([])
 
@@ -4162,6 +4235,51 @@ def test_session_info_includes_mcp_servers(monkeypatch):
     assert info["mcp_servers"] == fake_status
 
 
+def test_session_info_includes_session_title(monkeypatch):
+    """session.info carries the live session title (window-title chrome).
+
+    Resolution order mirrors _session_live_title: DB row wins, a queued
+    pending_title fills in before the row exists, "" until either lands.
+    """
+    agent = types.SimpleNamespace(tools=[], model="m", provider="p")
+
+    # No session at all -> "" (and never a crash).
+    assert server._session_info(agent, None)["title"] == ""
+
+    # pending_title before the DB row exists.
+    session = {"session_key": "k1", "pending_title": "rename the moon"}
+    monkeypatch.setattr(server, "_get_db", lambda: None)
+    assert server._session_info(agent, session)["title"] == "rename the moon"
+
+    # DB row wins over pending.
+    fake_db = types.SimpleNamespace(get_session_title=lambda key: "db title")
+    monkeypatch.setattr(server, "_get_db", lambda: fake_db)
+    assert server._session_info(agent, session)["title"] == "db title"
+
+
+def test_emit_title_refresh_pushes_session_info(monkeypatch):
+    """_emit_title_refresh emits a session.info for a live session and is a
+    silent no-op for unknown/agent-less sessions (it runs on the auto-title
+    worker thread -- it must never raise)."""
+    events = []
+    monkeypatch.setattr(
+        server, "_emit", lambda event_type, sid, payload: events.append((event_type, sid, payload))
+    )
+    monkeypatch.setattr(server, "_session_info", lambda agent, session: {"title": "t"})
+
+    agent = types.SimpleNamespace(tools=[], model="m", provider="p")
+    monkeypatch.setitem(server._sessions, "title-test", {"agent": agent, "session_key": "k"})
+
+    server._emit_title_refresh("title-test")
+    assert events == [("session.info", "title-test", {"title": "t"})]
+
+    # Unknown sid / no agent -> no emission, no exception.
+    server._emit_title_refresh("missing-sid")
+    monkeypatch.setitem(server._sessions, "agentless", {"session_key": "k2"})
+    server._emit_title_refresh("agentless")
+    assert len(events) == 1
+
+
 # ---------------------------------------------------------------------------
 # History-mutating commands must reject while session.running is True.
 # Without these guards, prompt.submit's post-run history write either
@@ -4562,6 +4680,90 @@ def test_respond_unpacks_sid_tuple_correctly():
         server._answers.pop("rid-x", None)
 
 
+# ---------------------------------------------------------------------------
+# Blocking prompts wait for the human (v6 north-star #5): _block with
+# timeout=None must never expire — interrupt/shutdown (_clear_pending)
+# are the only releases.
+# ---------------------------------------------------------------------------
+
+
+def _run_block_in_thread(monkeypatch, sid):
+    """Start _block(timeout=None) on a background thread; return
+    (thread, results, get_rid) where get_rid polls for the pending rid."""
+    monkeypatch.setattr(server, "_emit", lambda *a, **kw: None)
+    results: list[str] = []
+
+    def runner():
+        results.append(server._block("clarify.request", sid, {"question": "q"}))
+
+    t = threading.Thread(target=runner, daemon=True)
+    t.start()
+
+    def get_rid():
+        deadline = time.time() + 5
+        while time.time() < deadline:
+            with server._prompt_lock:
+                for rid, (owner, _ev) in server._pending.items():
+                    if owner == sid:
+                        return rid
+            time.sleep(0.01)
+        raise AssertionError("pending rid never appeared for sid=%s" % sid)
+
+    return t, results, get_rid
+
+
+def test_block_no_timeout_waits_for_delayed_answer(monkeypatch):
+    """_block(timeout=None) must keep blocking until the answer arrives —
+    no premature empty return."""
+    t, results, get_rid = _run_block_in_thread(monkeypatch, "sid_block_wait")
+    rid = get_rid()
+
+    # Answer after a short delay; _block must still be waiting.
+    time.sleep(0.3)
+    assert t.is_alive(), "_block returned before any answer was provided"
+    with server._prompt_lock:
+        server._answers[rid] = "green"
+        server._pending[rid][1].set()
+
+    t.join(timeout=5)
+    assert not t.is_alive()
+    assert results == ["green"]
+
+
+def test_clear_pending_releases_no_timeout_block(monkeypatch):
+    """_clear_pending(sid) must release a timeout=None _block with ''."""
+    t, results, get_rid = _run_block_in_thread(monkeypatch, "sid_block_clear")
+    get_rid()
+
+    server._clear_pending("sid_block_clear")
+    t.join(timeout=5)
+    assert not t.is_alive()
+    assert results == [""]
+
+
+def test_clear_pending_other_sid_does_not_release_block(monkeypatch):
+    """_clear_pending on an unrelated session must NOT release a pending
+    timeout=None _block (session scoping)."""
+    t, results, get_rid = _run_block_in_thread(monkeypatch, "sid_block_scoped")
+    rid = get_rid()
+
+    server._clear_pending("sid_some_other_session")
+    time.sleep(0.2)
+    assert t.is_alive(), (
+        "_clear_pending on another sid released a prompt owned by "
+        "sid_block_scoped — session scoping is broken"
+    )
+    assert not results
+
+    # Clean up: release properly so the thread joins.
+    with server._prompt_lock:
+        server._answers[rid] = "done"
+        server._pending[rid][1].set()
+    t.join(timeout=5)
+    assert not t.is_alive()
+    assert results == ["done"]
+
+
 # ---------------------------------------------------------------------------
 # /model switch and other agent-mutating commands must reject while the
 # session is running.  agent.switch_model() mutates self.model, self.provider,
@@ -5776,6 +5978,351 @@ def test_session_most_recent_handles_db_unavailable(monkeypatch):
     assert resp["result"]["session_id"] is None
 
 
+# ── session.list (resume-picker filters + widened projection) ───────
+
+
+def _picker_row(sid, source, started, **extra):
+    row = {
+        "id": sid,
+        "source": source,
+        "title": f"title-{sid}",
+        "preview": f"preview-{sid}",
+        "started_at": started,
+        "last_active": started,
+        "message_count": 3,
+        "ended_at": None,
+        "cwd": None,
+        "model": None,
+    }
+    row.update(extra)
+    return row
+
+
+class _PickerDB:
+    """list_sessions_rich stand-in honouring the kwargs session.list maps."""
+
+    def __init__(self, rows):
+        self.rows = rows
+        self.calls = []
+
+    def list_sessions_rich(
+        self,
+        *,
+        source=None,
+        limit=20,
+        offset=0,
+        order_by_last_active=False,
+        id_query=None,
+    ):
+        self.calls.append(
+            {
+                "source": source,
+                "limit": limit,
+                "offset": offset,
+                "order_by_last_active": order_by_last_active,
+                "id_query": id_query,
+            }
+        )
+        rows = [dict(r) for r in self.rows]
+        if source:
+            rows = [r for r in rows if r.get("source") == source]
+        if id_query:
+            needle = id_query.strip().lower()
+            rows = [r for r in rows if needle in r["id"].lower()]
+        if order_by_last_active:
+            rows.sort(key=lambda r: r.get("last_active") or 0, reverse=True)
+        else:
+            rows.sort(key=lambda r: r.get("started_at") or 0, reverse=True)
+        return rows[offset : offset + limit]
+
+
+def _picker_db():
+    return _PickerDB(
+        [
+            _picker_row(
+                "tui-1",
+                "tui",
+                100,
+                cwd="/home/u/proj",
+                model="nous/hermes-4",
+                ended_at=150.0,
+            ),
+            _picker_row("cli-1", "cli", 90),
+            _picker_row("cron-1", "cron", 80),
+            _picker_row("cron-2", "cron", 70),
+            _picker_row("tg-1", "telegram", 60),
+            _picker_row("tool-1", "tool", 50),
+        ]
+    )
+
+
+def test_session_list_no_params_keeps_legacy_behavior_with_widened_rows(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request({"id": "1", "method": "session.list", "params": {}})
+
+    rows = resp["result"]["sessions"]
+    # Legacy semantics: started_at DESC, `tool` denied, one DB fetch.
+    assert [r["id"] for r in rows] == ["tui-1", "cli-1", "cron-1", "cron-2", "tg-1"]
+    assert db.calls == [
+        {
+            "source": None,
+            "limit": 400,
+            "offset": 0,
+            "order_by_last_active": False,
+            "id_query": None,
+        }
+    ]
+    # Widened projection, None-safe for rows missing the new columns.
+    first, second = rows[0], rows[1]
+    assert first["cwd"] == "/home/u/proj"
+    assert first["model"] == "nous/hermes-4"
+    assert first["ended_at"] == 150.0
+    assert first["last_active"] == 100
+    assert second["cwd"] is None
+    assert second["model"] is None
+    assert second["ended_at"] is None
+    assert second["last_active"] == 90
+    # Legacy keys are still present and unchanged.
+    assert second["title"] == "title-cli-1"
+    assert second["preview"] == "preview-cli-1"
+    assert second["message_count"] == 3
+    assert second["source"] == "cli"
+
+
+def test_session_list_single_source_passes_through_to_db(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.list", "params": {"sources": ["tui"]}}
+    )
+
+    assert [r["id"] for r in resp["result"]["sessions"]] == ["tui-1"]
+    assert db.calls[0]["source"] == "tui"  # pushed into SQL, not Python-filtered
+
+
+def test_session_list_multi_source_filters_gateway_side(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request(
+        {
+            "id": "1",
+            "method": "session.list",
+            "params": {"sources": ["cron", "telegram"]},
+        }
+    )
+
+    assert [r["id"] for r in resp["result"]["sessions"]] == [
+        "cron-1",
+        "cron-2",
+        "tg-1",
+    ]
+    assert db.calls[0]["source"] is None  # multi-source: DB scan + gateway filter
+
+
+def test_session_list_sources_tool_stays_denied(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.list", "params": {"sources": ["tool"]}}
+    )
+
+    assert resp["result"]["sessions"] == []
+
+
+def test_session_list_query_maps_to_id_query_on_last_active_path(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.list", "params": {"query": "cron"}}
+    )
+
+    assert [r["id"] for r in resp["result"]["sessions"]] == ["cron-1", "cron-2"]
+    assert db.calls[0]["id_query"] == "cron"
+    assert db.calls[0]["order_by_last_active"] is True
+
+
+def test_session_list_offset_limit_paginate(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    def page(offset, limit):
+        resp = server.handle_request(
+            {
+                "id": "1",
+                "method": "session.list",
+                "params": {
+                    "sources": ["tui", "cli", "cron", "telegram"],
+                    "offset": offset,
+                    "limit": limit,
+                },
+            }
+        )
+        return [r["id"] for r in resp["result"]["sessions"]]
+
+    assert page(0, 2) == ["tui-1", "cli-1"]
+    assert page(2, 2) == ["cron-1", "cron-2"]
+    assert page(4, 2) == ["tg-1"]
+    assert page(6, 2) == []
+
+
+# ── session.peek (resume-picker Space preview) ───────────────────────
+
+
+class _PeekDB:
+    def __init__(self, session, messages):
+        self.session = session
+        self.messages = messages
+
+    def get_session(self, session_id):
+        if self.session and session_id == self.session["id"]:
+            return dict(self.session)
+        return None
+
+    def get_messages(self, session_id):
+        return [dict(m) for m in self.messages]
+
+
+def _peek_db():
+    session = {
+        "id": "sess-1",
+        "title": "picker demo",
+        "source": "tui",
+        "model": "nous/hermes-4",
+        "cwd": "/home/u/proj",
+        "started_at": 100.0,
+        "ended_at": 400.0,
+        "end_reason": "tui_shutdown",
+        "message_count": 6,
+        "actual_cost_usd": None,
+        "estimated_cost_usd": 0.42,
+    }
+    messages = [
+        {"id": 1, "role": "system", "content": "sys prompt", "timestamp": 100.0},
+        {"id": 2, "role": "user", "content": "first prompt", "timestamp": 110.0},
+        {"id": 3, "role": "assistant", "content": "first answer", "timestamp": 120.0},
+        {"id": 4, "role": "tool", "content": "tool output", "timestamp": 130.0},
+        {"id": 5, "role": "user", "content": "second prompt", "timestamp": 140.0},
+        {"id": 6, "role": "assistant", "content": "final answer", "timestamp": 150.0},
+    ]
+    return _PeekDB(session, messages)
+
+
+def test_session_peek_returns_metadata_and_head_tail(monkeypatch):
+    monkeypatch.setattr(server, "_get_db", _peek_db)
+
+    resp = server.handle_request(
+        {
+            "id": "1",
+            "method": "session.peek",
+            "params": {"session_id": "sess-1", "head": 1, "tail": 2},
+        }
+    )
+
+    result = resp["result"]
+    meta = result["session"]
+    assert meta == {
+        "id": "sess-1",
+        "title": "picker demo",
+        "source": "tui",
+        "model": "nous/hermes-4",
+        "cwd": "/home/u/proj",
+        "started_at": 100.0,
+        "ended_at": 400.0,
+        "end_reason": "tui_shutdown",
+        "message_count": 6,
+        "last_active": 150.0,  # last message timestamp, not ended_at
+        "cost_usd": 0.42,  # estimated fallback when actual is None
+    }
+    # Only displayable (user/assistant) messages, system/tool rows skipped.
+    assert [(m["role"], m["content"]) for m in result["head"]] == [
+        ("user", "first prompt")
+    ]
+    assert [(m["role"], m["content"]) for m in result["tail"]] == [
+        ("user", "second prompt"),
+        ("assistant", "final answer"),
+    ]
+    assert result["total_messages"] == 4
+    assert all(m["truncated"] is False for m in result["head"] + result["tail"])
+
+
+def test_session_peek_head_tail_never_overlap(monkeypatch):
+    db = _peek_db()
+    db.messages = db.messages[:3]  # system + user + assistant
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+
+    resp = server.handle_request(
+        {
+            "id": "1",
+            "method": "session.peek",
+            "params": {"session_id": "sess-1", "head": 2, "tail": 2},
+        }
+    )
+
+    result = resp["result"]
+    head_ids = [m["id"] for m in result["head"]]
+    tail_ids = [m["id"] for m in result["tail"]]
+    assert head_ids == [2, 3]
+    assert tail_ids == []  # both displayable rows already consumed by head
+    assert result["total_messages"] == 2
+
+
+def test_session_peek_does_not_build_an_agent(monkeypatch):
+    spawned = []
+    monkeypatch.setattr(server, "_get_db", _peek_db)
+    monkeypatch.setattr(
+        server, "_make_agent", lambda *a, **kw: spawned.append("make") or None
+    )
+    monkeypatch.setattr(
+        server, "_start_agent_build", lambda *a, **kw: spawned.append("build")
+    )
+    before_sessions = dict(server._sessions)
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.peek", "params": {"session_id": "sess-1"}}
+    )
+
+    assert "result" in resp
+    assert spawned == []
+    assert server._sessions == before_sessions  # no live session registered
+
+
+def test_session_peek_unknown_id_is_clean_error(monkeypatch):
+    monkeypatch.setattr(server, "_get_db", _peek_db)
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.peek", "params": {"session_id": "nope"}}
+    )
+
+    assert resp["error"]["code"] == 4007
+    assert resp["error"]["message"] == "session not found"
+
+
+def test_session_peek_requires_session_id(monkeypatch):
+    monkeypatch.setattr(server, "_get_db", _peek_db)
+
+    resp = server.handle_request({"id": "1", "method": "session.peek", "params": {}})
+
+    assert resp["error"]["code"] == 4006
+
+
+def test_session_peek_db_unavailable(monkeypatch):
+    monkeypatch.setattr(server, "_get_db", lambda: None)
+    monkeypatch.setattr(server, "_db_error", "locked")
+
+    resp = server.handle_request(
+        {"id": "1", "method": "session.peek", "params": {"session_id": "sess-1"}}
+    )
+
+    assert resp["error"]["code"] == 5046
+    assert "state.db unavailable" in resp["error"]["message"]
+
+
 # ── browser.manage ───────────────────────────────────────────────────
 
 
@@ -7223,6 +7770,84 @@ def test_sniff_image_ext_magic_and_filename():
     assert server._sniff_image_ext(b"\x89PNG", "photo.jpeg") == ".jpeg"
 
 
+def test_tool_complete_derives_error_from_result_convention(monkeypatch):
+    # The repo convention (agent.display._result_succeeded): a JSON-object
+    # result with a truthy string "error" — or success:false — IS a failure.
+    # The gateway surfaces it as payload["error"] so clients (OpenTUI ✗ state,
+    # Ink trail ✗) don't sniff the convention themselves.
+    events: list[tuple[str, str, dict]] = []
+    monkeypatch.setattr(
+        server, "_emit", lambda event_type, sid, payload: events.append((event_type, sid, payload))
+    )
+    monkeypatch.setitem(
+        server._sessions,
+        "err-test",
+        {"tool_progress_mode": "verbose", "tool_started_at": {}, "edit_snapshots": {}},
+    )
+
+    # 1) error-string result → flattened, capped error on the payload
+    server._on_tool_complete(
+        "err-test", "t1", "read_file", {"path": "/nope"},
+        json.dumps({"error": "File not found:\n   /nope"}),
+    )
+    assert events[-1][2]["error"] == "File not found: /nope"
+
+    # 2) success:false without error string → generic failure marker
+    server._on_tool_complete(
+        "err-test", "t2", "patch", {"path": "x"}, json.dumps({"success": False})
+    )
+    assert events[-1][2]["error"] == "tool reported failure"
+
+    # 3) plain-text result → NEVER a failure
+    server._on_tool_complete("err-test", "t3", "terminal", {"command": "ls"}, "file-a\nfile-b")
+    assert "error" not in events[-1][2]
+
+    # 4) JSON result with error: null / no error key → success
+    server._on_tool_complete(
+        "err-test", "t4", "web_search", {"q": "x"},
+        json.dumps({"results": [1, 2], "error": None}),
+    )
+    assert "error" not in events[-1][2]
+
+
+def test_session_list_reports_scan_cap_truncation(monkeypatch):
+    # The bounded multi-source scan must say so when the 10k safety cap stops
+    # it with the requested window unfilled — an empty page past the cap is
+    # "truncated", not "no more sessions".
+    class _DenyAllDB:
+        def __init__(self):
+            self.offsets = []
+
+        def list_sessions_rich(self, source=None, limit=20, offset=0, **kw):
+            self.offsets.append(offset)
+            return [
+                {"id": f"s{offset}-{i}", "source": "tool", "title": "", "preview": "",
+                 "started_at": 1, "message_count": 1}
+                for i in range(limit)
+            ]
+
+    db = _DenyAllDB()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+    resp = server.handle_request(
+        {"id": "1", "method": "session.list", "params": {"sources": ["tui", "cli"], "limit": 5}}
+    )
+    result = resp["result"]
+    assert result["sessions"] == []
+    assert result["truncated"] is True
+    assert max(db.offsets) <= 10_000
+
+
+def test_session_list_truncated_false_on_normal_paths(monkeypatch):
+    db = _picker_db()
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+    legacy = server.handle_request({"id": "1", "method": "session.list", "params": {}})
+    assert legacy["result"]["truncated"] is False
+    filtered = server.handle_request(
+        {"id": "2", "method": "session.list", "params": {"sources": ["tui"], "limit": 5}}
+    )
+    assert filtered["result"]["truncated"] is False
+
+
 def test_slash_worker_close_reaps_zombie_and_closes_fds():
     """A hung worker is SIGKILLed, the zombie reaped, all pipes closed — once."""
     calls = {k: 0 for k in ("terminate", "kill", "wait", "stdin", "stdout", "stderr")}
@@ -7497,3 +8122,51 @@ def test_reap_idle_sessions_closes_only_evictable(monkeypatch):
         assert closed == [("stale", "idle_timeout")]
     finally:
         server._sessions.clear()
+
+
+class TestProcessCompletionCard:
+    """_emit_process_completion_card surfaces a background-process completion to
+    the TUI as a notification.show card (Option B, glitch 2026-06-14) — in
+    addition to the agent turn the completion triggers."""
+
+    @staticmethod
+    def _capture(monkeypatch):
+        emitted: list = []
+        monkeypatch.setattr(server, "_emit", lambda event, sid, payload=None: emitted.append((event, sid, payload)))
+        return emitted
+
+    def test_completion_exit_zero_is_an_info_card(self, monkeypatch):
+        emitted = self._capture(monkeypatch)
+        server._emit_process_completion_card(
+            "s1", {"type": "completion", "session_id": "proc_1", "command": "sleep 20 && echo hi", "exit_code": 0}
+        )
+        assert len(emitted) == 1
+        event, sid, payload = emitted[0]
+        assert event == "notification.show"
+        assert sid == "s1"
+        assert payload["text"] == "sleep 20 && echo hi exited 0"
+        assert payload["level"] == "info"
+        assert payload["kind"] == "process.complete"
+        assert payload["key"] == "proc:proc_1"
+
+    def test_nonzero_exit_is_a_warn_card(self, monkeypatch):
+        emitted = self._capture(monkeypatch)
+        server._emit_process_completion_card("s1", {"type": "completion", "command": "build", "exit_code": 1, "session_id": "p2"})
+        assert emitted[0][2]["level"] == "warn"
+        assert emitted[0][2]["text"] == "build exited 1"
+
+    def test_watch_match_is_not_carded(self, monkeypatch):
+        emitted = self._capture(monkeypatch)
+        server._emit_process_completion_card("s1", {"type": "watch_match", "command": "tail -f log"})
+        assert emitted == []
+
+    def test_long_command_is_truncated(self, monkeypatch):
+        emitted = self._capture(monkeypatch)
+        server._emit_process_completion_card("s1", {"type": "completion", "command": "x" * 100, "exit_code": 0, "session_id": "p3"})
+        assert "…" in emitted[0][2]["text"]
+        assert len(emitted[0][2]["text"]) < 80
+
+    def test_missing_exit_code_says_finished(self, monkeypatch):
+        emitted = self._capture(monkeypatch)
+        server._emit_process_completion_card("s1", {"type": "completion", "command": "daemon", "session_id": "p4"})
+        assert emitted[0][2]["text"] == "daemon finished"

tui_gateway/server.py

@@ -176,6 +176,12 @@ _LONG_HANDLERS = frozenset(
     {
         "browser.manage",
         "cli.exec",
+        # model.options is network-bound (pricing fetch + Nous tier check via
+        # build_models_payload, ~seconds). The native TUI prefetches it right
+        # after session.create; on the main thread that stalls every fast-path
+        # RPC — notably complete.slash, so the first `/` dropdown after launch
+        # took seconds to paint.
+        "model.options",
         "plugins.manage",
         "session.branch",
         "session.compress",
@@ -1351,7 +1357,11 @@ def _enable_gateway_prompts() -> None:
 # ── Blocking prompt factory ──────────────────────────────────────────
 
 
-def _block(event: str, sid: str, payload: dict, timeout: int = 300) -> str:
+def _block(event: str, sid: str, payload: dict, timeout: int | None = None) -> str:
+    # Blocking prompts wait for the human; interrupt/shutdown (via
+    # _clear_pending) are the only releases — v6 north-star #5.  A
+    # timeout would orphan the TUI prompt and silently feed the agent
+    # an empty answer, so callers default to waiting forever.
     rid = uuid.uuid4().hex[:8]
     ev = threading.Event()
     with _prompt_lock:
@@ -1402,6 +1412,11 @@ def resolve_skin() -> dict:
             "banner_hero": skin.banner_hero,
             "tool_prefix": skin.tool_prefix,
             "help_header": (skin.branding or {}).get("help_header", ""),
+            # Native engines (Ink + OpenTUI) can now consume these too: spinner
+            # animation data (faces/verbs/wings) and per-tool emoji overrides.
+            # Additive + optional — old engines ignore unknown keys.
+            "spinner": skin.spinner or {},
+            "tool_emojis": skin.tool_emojis or {},
         }
     except Exception:
         return {}
@@ -2410,6 +2425,17 @@ def _session_info(agent, session: dict | None = None) -> dict:
         yolo = bool(_YOLO_MODE_FROZEN) or session_yolo or _get_approval_mode() == "off"
     except Exception:
         yolo = False
+    # Session title (DB row, falling back to a not-yet-applied pending_title).
+    # Drives client window-title chrome (OSC 0/2 in the native TUI); "" until
+    # the first exchange titles the session.
+    title = ""
+    if session is not None:
+        try:
+            title = _session_live_title(
+                session, str(session.get("session_key") or "")
+            )
+        except Exception:
+            title = ""
     info: dict = {
         "model": getattr(agent, "model", ""),
         "provider": getattr(agent, "provider", ""),
@@ -2422,6 +2448,7 @@ def _session_info(agent, session: dict | None = None) -> dict:
         "cwd": cwd,
         "branch": _git_branch_for_cwd(cwd),
         "personality": str(personality or ""),
+        "title": title,
         "running": bool((session or {}).get("running")),
         "desktop_contract": DESKTOP_BACKEND_CONTRACT,
         "version": "",
@@ -2534,6 +2561,53 @@ def _cap_tui_verbose_text(text: str) -> str:
     return f"{label}{tail}"
 
 
+# The FULL raw unified diff shipped on file-edit tool.complete (`diff_unified`,
+# for clients with a native diff renderer — ui-opentui). Unlike the verbose
+# trail above this is rendered COLLAPSED by default and only on edit tools, so
+# it gets a far larger budget; 512KB is still a hard ceiling so a runaway
+# multi-megabyte edit can't flood the pipe.
+_DIFF_UNIFIED_MAX_BYTES = 512 * 1024
+
+
+def _cap_diff_unified(diff: str, max_bytes: int = _DIFF_UNIFIED_MAX_BYTES) -> str:
+    raw = diff.encode("utf-8")
+    if len(raw) <= max_bytes:
+        return diff
+    head = raw[:max_bytes].decode("utf-8", errors="ignore")
+    # Truncate at a line boundary so the surviving diff stays parseable, then
+    # append an honest marker line (never send more than the cap + marker).
+    cut = head.rfind("\n")
+    if cut > 0:
+        head = head[:cut]
+    omitted = len(raw) - len(head.encode("utf-8"))
+    return f"{head}\n# … diff truncated ({omitted} more bytes)"
+
+
+def _result_sans_diff_echo(result: str) -> str:
+    """The file-edit result JSON minus its `diff` echo.
+
+    Used for verbose `result_text` when the FULL diff already ships as
+    `diff_unified`: a multi-KB diff echo inside the result JSON gets
+    tail-capped by `_cap_tui_verbose_text` into an unparseable JSON-looking
+    fragment that a client can neither render nor reliably suppress. The
+    native renderer shows the real diff, so result_text should carry only the
+    non-diff signal (success/files_modified/warnings/lsp_diagnostics).
+    Returns `result` unchanged when it isn't a JSON object with a `diff` key.
+    """
+    try:
+        data = json.loads(result)
+    except Exception:
+        return result
+    if not isinstance(data, dict) or "diff" not in data:
+        return result
+    try:
+        return json.dumps(
+            {k: v for k, v in data.items() if k != "diff"}, ensure_ascii=False
+        )
+    except Exception:
+        return result
+
+
 def _redact_tui_verbose_text(text: str) -> str:
     try:
         from agent.redact import redact_sensitive_text
@@ -2637,8 +2711,37 @@ def _on_tool_start(sid: str, tool_call_id: str, name: str, args: dict):
         _emit("tool.start", sid, payload)
 
 
+def _tool_error_from_result(result: str) -> str | None:
+    """Derive a tool-failure message from the repo's result convention.
+
+    Canon (same as ``agent.display._result_succeeded``): a JSON-object result
+    with a truthy string ``error`` key — or an explicit ``success: false`` —
+    means the tool failed.  Returned flattened + capped for a one-line header.
+    Conservative on purpose: non-JSON / non-dict results are NEVER failures
+    (plain-text output is normal success output).
+    """
+    try:
+        data = json.loads(result)
+    except Exception:
+        return None
+    if not isinstance(data, dict):
+        return None
+    err = data.get("error")
+    if isinstance(err, str) and err.strip():
+        return " ".join(err.split())[:300]
+    if data.get("success") is False:
+        return "tool reported failure"
+    return None
+
+
 def _on_tool_complete(sid: str, tool_call_id: str, name: str, args: dict, result: str):
     payload = {"tool_id": tool_call_id, "name": name, "args": args}
+    # Failure surfaced explicitly so clients don't have to sniff the result
+    # convention themselves (the TUI's ✗ state and Ink's trail ✗ both key off
+    # payload["error"], which was previously never set on this path).
+    _tool_err = _tool_error_from_result(result)
+    if _tool_err:
+        payload["error"] = _tool_err
     session = _sessions.get(sid)
     snapshot = None
     started_at = None
@@ -2655,10 +2758,6 @@ def _on_tool_complete(sid: str, tool_call_id: str, name: str, args: dict, result
     summary = _tool_summary(name, result, duration_s)
     if summary:
         payload["summary"] = summary
-    if _session_verbose(sid):
-        result_text = _tool_result_text(result)
-        if result_text:
-            payload["result_text"] = result_text
     if name == "todo":
         try:
             data = json.loads(result)
@@ -2680,7 +2779,32 @@ def _on_tool_complete(sid: str, tool_call_id: str, name: str, args: dict, result
             payload["inline_diff"] = "\n".join(rendered)
     except Exception:
         pass
-    if _tool_progress_enabled(sid) or payload.get("inline_diff"):
+    # Alongside the pretty-rendered/capped `inline_diff` (Ink consumes that), ship
+    # the RAW unified diff for clients with a native diff renderer (ui-opentui's
+    # file-tool view). Capped at _DIFF_UNIFIED_MAX_BYTES with an honest marker.
+    try:
+        from agent.display import extract_edit_diff
+
+        diff_unified = extract_edit_diff(
+            name,
+            result,
+            function_args=args,
+            snapshot=snapshot,
+        )
+        if diff_unified:
+            payload["diff_unified"] = _cap_diff_unified(diff_unified)
+    except Exception:
+        pass
+    if _session_verbose(sid):
+        # Computed AFTER diff_unified: when the full diff ships natively, the
+        # result_text drops the in-JSON diff echo (it would tail-cap into
+        # unparseable JSON-looking noise under the client's rendered diff).
+        result_text = _tool_result_text(
+            _result_sans_diff_echo(result) if payload.get("diff_unified") else result
+        )
+        if result_text:
+            payload["result_text"] = result_text
+    if _tool_progress_enabled(sid) or payload.get("inline_diff") or payload.get("diff_unified"):
         _emit("tool.complete", sid, payload)
 
 
@@ -2910,7 +3034,7 @@ def _wire_callbacks(sid: str):
     from tools.terminal_tool import set_sudo_password_callback
     from tools.skills_tool import set_secret_capture_callback
 
-    set_sudo_password_callback(lambda: _block("sudo.request", sid, {}, timeout=120))
+    set_sudo_password_callback(lambda: _block("sudo.request", sid, {}))
 
     def secret_cb(env_var, prompt, metadata=None):
         pl = {"prompt": prompt, "env_var": env_var}
@@ -3696,7 +3820,13 @@ def _coerce_message_text(content: Any) -> str:
     return str(content)
 
 
-def _history_to_messages(history: list[dict]) -> list[dict]:
+def _history_to_messages(history: list[dict], include_tool_output: bool = False) -> list[dict]:
+    # ``include_tool_output`` (opt-in; only the native/opentui engine passes it via
+    # session.resume) folds each tool's redacted+capped result + args into its row so
+    # a resumed transcript renders collapsible tool blocks identical to a live turn.
+    # OFF by default so the Ink path is byte-for-byte unchanged (its render tree showed
+    # the verbose trail expanded and OOM'd on big output — #34095; the native engine
+    # renders tools collapsed, so shipping the same capped tail is safe there).
     messages = []
     tool_call_args = {}
 
@@ -3724,9 +3854,13 @@ def _history_to_messages(history: list[dict]) -> list[dict]:
             tc_info = tool_call_args.get(tc_id) if tc_id else None
             name = (tc_info[0] if tc_info else None) or m.get("tool_name") or "tool"
             args = (tc_info[1] if tc_info else None) or {}
-            messages.append(
-                {"role": "tool", "name": name, "context": _tool_ctx(name, args)}
-            )
+            tool_msg = {"role": "tool", "name": name, "context": _tool_ctx(name, args)}
+            if include_tool_output:
+                if args:
+                    tool_msg["args"] = args
+                if content_text.strip():
+                    tool_msg["result_text"] = _redact_tui_verbose_text(content_text)
+            messages.append(tool_msg)
             continue
         # An assistant turn may carry only reasoning/thinking content with no
         # visible text (extended-thinking turns, thinking-only recovery
@@ -3966,6 +4100,23 @@ def _(rid, params: dict) -> dict:
 
 @method("session.list")
 def _(rid, params: dict) -> dict:
+    """List stored sessions for the resume picker / sidebar.
+
+    Optional params — omitting all of them keeps the legacy behaviour
+    (most-recently-started first, ``tool`` rows denied, default limit 200):
+
+      - ``sources``: list of source tags to include (e.g. ``["cli", "tui"]``).
+        Powers the picker's tab strip. The ``tool`` deny-list still applies
+        on top. A single-element list is pushed into SQL
+        (``list_sessions_rich(source=...)``); multi-element lists are
+        filtered gateway-side over a bounded scan (the DB layer only takes a
+        single ``source`` string).
+      - ``query``: case-insensitive *session-id* substring filter, mapped to
+        ``list_sessions_rich(id_query=...)``. The DB applies ``id_query``
+        only on its order-by-last-active path, so query results are ordered
+        by most recent activity. Title/preview search stays client-side.
+      - ``offset`` / ``limit``: pagination over the filtered list.
+    """
     db = _get_db()
     if db is None:
         return _db_unavailable_error(rid, code=5006)
@@ -3980,16 +4131,76 @@ def _(rid, params: dict) -> dict:
         # platform is added or a user names their own source.
         deny = frozenset({"tool"})
 
-        limit = int(params.get("limit", 200) or 200)
-        # Over-fetch modestly so per-source filtering doesn't leave us
-        # short; the compression-tip projection in ``list_sessions_rich``
-        # can also merge rows.
-        fetch_limit = max(limit * 2, 200)
-        rows = [
-            s
-            for s in db.list_sessions_rich(source=None, limit=fetch_limit)
-            if (s.get("source") or "").strip().lower() not in deny
-        ][:limit]
+        try:
+            limit = max(1, int(params.get("limit", 200) or 200))
+        except (TypeError, ValueError):
+            limit = 200
+        try:
+            offset = max(0, int(params.get("offset", 0) or 0))
+        except (TypeError, ValueError):
+            offset = 0
+        query = str(params.get("query") or "").strip()
+        raw_sources = params.get("sources")
+        sources: list = []
+        if isinstance(raw_sources, (list, tuple)):
+            sources = [
+                str(s).strip().lower() for s in raw_sources if str(s).strip()
+            ]
+
+        if not sources and not query and offset == 0:
+            # Legacy path (no filter params) — byte-for-byte today's
+            # behaviour. Over-fetch modestly so per-source filtering doesn't
+            # leave us short; the compression-tip projection in
+            # ``list_sessions_rich`` can also merge rows.
+            fetch_limit = max(limit * 2, 200)
+            rows = [
+                s
+                for s in db.list_sessions_rich(source=None, limit=fetch_limit)
+                if (s.get("source") or "").strip().lower() not in deny
+            ][:limit]
+            list_truncated = False
+        else:
+            # Filtered/paginated path. Single source pushes into SQL; the
+            # deny-list and multi-source filter run gateway-side, so keep
+            # scanning DB pages until the requested window is full (bounded
+            # by a generous safety cap so a pathological DB can't pin us).
+            source_arg = sources[0] if len(sources) == 1 else None
+            allowed = frozenset(sources) if sources else None
+            wanted = offset + limit
+
+            def _eligible(row: dict) -> bool:
+                src = (row.get("source") or "").strip().lower()
+                if src in deny:
+                    return False
+                return allowed is None or src in allowed
+
+            collected: list = []
+            db_offset = 0
+            page = max(wanted * 2, 200)
+            scan_capped = False
+            while len(collected) < wanted:
+                if db_offset >= 10_000:
+                    # Safety cap hit with the window still unfilled — report it
+                    # honestly (``truncated``) so the client can say "results
+                    # truncated" instead of silently serving an empty page.
+                    scan_capped = True
+                    break
+                batch = db.list_sessions_rich(
+                    source=source_arg,
+                    limit=page,
+                    offset=db_offset,
+                    # ``id_query`` only applies on the order-by-last-active
+                    # path; keep legacy started_at ordering when unfiltered.
+                    order_by_last_active=bool(query),
+                    id_query=query or None,
+                )
+                collected.extend(r for r in batch if _eligible(r))
+                if len(batch) < page:
+                    break
+                db_offset += page
+            rows = collected[offset : offset + limit]
+            list_truncated = scan_capped and len(collected) < wanted
+
         return _ok(
             rid,
             {
@@ -4001,15 +4212,133 @@ def _(rid, params: dict) -> dict:
                         "started_at": s.get("started_at") or 0,
                         "message_count": s.get("message_count") or 0,
                         "source": s.get("source") or "",
+                        # Picker row metadata (None-safe; older rows may
+                        # predate these columns).
+                        "cwd": s.get("cwd") or None,
+                        "last_active": s.get("last_active")
+                        or s.get("started_at")
+                        or 0,
+                        "ended_at": s.get("ended_at"),
+                        "model": s.get("model") or None,
                     }
                     for s in rows
-                ]
+                ],
+                # True when the bounded multi-source scan hit its safety cap
+                # before filling the requested window — the client should show
+                # "results truncated" instead of treating the page as final.
+                "truncated": list_truncated,
             },
         )
     except Exception as e:
         return _err(rid, 5006, str(e))
 
 
+@method("session.peek")
+def _(rid, params: dict) -> dict:
+    """DB-only preview of a stored session for the resume picker.
+
+    ``{session_id, head?, tail?}`` → session metadata plus the first ``head``
+    and last ``tail`` displayable messages (``user``/``assistant`` rows with
+    non-empty text content; tool spam and empty tool-call carriers are
+    skipped). Purely a read: no agent is constructed, no live session state
+    is created or switched — this powers the picker's Space preview, which
+    must stay cheap while the user scrolls.
+
+    Response shape::
+
+        {
+          "session": {id, title, source, model, cwd, started_at, ended_at,
+                      end_reason, message_count, last_active, cost_usd},
+          "head": [{id, role, content, truncated, timestamp}, ...],
+          "tail": [{...}],            # never overlaps head
+          "total_messages": <int>     # displayable (user/assistant) count
+        }
+    """
+    target = str(params.get("session_id") or "").strip()
+    if not target:
+        return _err(rid, 4006, "session_id required")
+    try:
+        head = max(0, int(params.get("head", 2) or 0))
+    except (TypeError, ValueError):
+        head = 2
+    try:
+        tail = max(0, int(params.get("tail", 2) or 0))
+    except (TypeError, ValueError):
+        tail = 2
+    db = _get_db()
+    if db is None:
+        return _db_unavailable_error(rid, code=5046)
+    try:
+        row = db.get_session(target)
+        if not row:
+            return _err(rid, 4007, "session not found")
+
+        msgs = db.get_messages(target)
+        last_active = (
+            (msgs[-1].get("timestamp") if msgs else None)
+            or row.get("started_at")
+            or 0
+        )
+
+        def _displayable(m: dict) -> bool:
+            if (m.get("role") or "") not in ("user", "assistant"):
+                return False
+            content = m.get("content")
+            if isinstance(content, str):
+                return bool(content.strip())
+            return bool(content)  # multimodal parts list
+
+        def _peek_msg(m: dict) -> dict:
+            content = m.get("content")
+            if not isinstance(content, str):
+                try:
+                    content = json.dumps(content, ensure_ascii=False)
+                except (TypeError, ValueError):
+                    content = str(content)
+            content = content or ""
+            return {
+                "id": m.get("id"),
+                "role": m.get("role") or "",
+                "content": content[:2000],
+                "truncated": len(content) > 2000,
+                "timestamp": m.get("timestamp"),
+            }
+
+        display = [m for m in msgs if _displayable(m)]
+        head_msgs = display[:head] if head else []
+        # Slice the remainder so head and tail never overlap, even when
+        # head + tail >= len(display).
+        tail_msgs = display[head:][-tail:] if tail else []
+
+        cost = row.get("actual_cost_usd")
+        if cost is None:
+            cost = row.get("estimated_cost_usd")
+
+        return _ok(
+            rid,
+            {
+                "session": {
+                    "id": row["id"],
+                    "title": row.get("title") or "",
+                    "source": row.get("source") or "",
+                    "model": row.get("model") or None,
+                    "cwd": row.get("cwd") or None,
+                    "started_at": row.get("started_at") or 0,
+                    "ended_at": row.get("ended_at"),
+                    "end_reason": row.get("end_reason"),
+                    "message_count": row.get("message_count") or 0,
+                    "last_active": last_active,
+                    "cost_usd": cost,
+                },
+                "head": [_peek_msg(m) for m in head_msgs],
+                "tail": [_peek_msg(m) for m in tail_msgs],
+                "total_messages": len(display),
+            },
+        )
+    except Exception as e:
+        return _err(rid, 5046, str(e))
+
+
 @method("session.most_recent")
 def _(rid, params: dict) -> dict:
     """Return the most recent human-facing session id, or ``None``.
@@ -4230,7 +4559,9 @@ def _(rid, params: dict) -> dict:
         display_history_prefix = display_history[
             : max(0, len(display_history) - len(history))
         ]
-        messages = _history_to_messages(display_history)
+        messages = _history_to_messages(
+            display_history, include_tool_output=bool(params.get("with_tool_output"))
+        )
         tokens = _set_session_context(target)
         try:
             # Pass the profile's db so the agent persists turns to the right
@@ -4397,6 +4728,22 @@ def _session_live_title(session: dict, key: str) -> str:
     return title
 
 
+def _emit_title_refresh(sid: str) -> None:
+    """Push a session.info refresh after a title change (pending-title
+    application, auto-title landing, or a session.title rename) so clients'
+    window-title chrome updates immediately. Thread-safe (auto-title calls
+    this from its worker thread; _emit serializes on the stdout lock).
+    Never raises."""
+    try:
+        session = _sessions.get(sid)
+        agent = (session or {}).get("agent")
+        if session is None or agent is None:
+            return
+        _emit("session.info", sid, _session_info(agent, session))
+    except Exception:
+        pass
+
+
 def _session_live_item(sid: str, session: dict, current_sid: str = "") -> dict:
     key = str(session.get("session_key") or sid)
     agent = session.get("agent")
@@ -4622,15 +4969,18 @@ def _(rid, params: dict) -> dict:
     title = (params.get("title", "") or "").strip()
     if not title:
         return _err(rid, 4021, "title required")
+    sid = str(params.get("session_id") or "")
     try:
         if db.set_session_title(key, title):
             session["pending_title"] = None
+            _emit_title_refresh(sid)
             return _ok(rid, {"pending": False, "title": title})
         # rowcount == 0 can mean "same value" as well as "missing row".
         # Queue only when the session row truly does not exist yet.
         existing_row = db.get_session(key)
         if existing_row:
             session["pending_title"] = None
+            _emit_title_refresh(sid)
             return _ok(
                 rid,
                 {
@@ -4639,6 +4989,7 @@ def _(rid, params: dict) -> dict:
                 },
             )
         session["pending_title"] = title
+        _emit_title_refresh(sid)
         return _ok(rid, {"pending": True, "title": title})
     except ValueError as e:
         return _err(rid, 4022, str(e))
@@ -5603,6 +5954,39 @@ def _notification_event_dedup_key(evt: dict) -> tuple:
     return (evt_sid, evt_type)
 
 
+def _emit_process_completion_card(sid: str, evt: dict) -> None:
+    """Surface a background-process COMPLETION to the TUI as a notification card,
+    in ADDITION to the agent turn it triggers. A bare `notify_on_complete` exit
+    otherwise reaches the TUI only as the agent's narration (the completion is fed
+    to the model as a synthetic prompt, never sent to the UI). Emitting a
+    ``notification.show`` lets the OpenTUI engine render a distinct inline card so
+    the user actually sees the process finish; the Ink engine treats it as a
+    notice. Additive — no existing behaviour changes. Completion events only
+    (watch matches aren't terminal); the dedup at the call sites ensures one card
+    per completion. (glitch 2026-06-14)"""
+    if evt.get("type", "completion") != "completion":
+        return
+    cmd = str(evt.get("command") or "process").strip().replace("\n", " ")
+    if len(cmd) > 60:
+        cmd = cmd[:59] + "…"
+    code = evt.get("exit_code")
+    if code is None:
+        text, level = f"{cmd} finished", "info"
+    else:
+        text = f"{cmd} exited {code}"
+        level = "info" if code == 0 else "warn"
+    _emit(
+        "notification.show",
+        sid,
+        {
+            "text": text,
+            "kind": "process.complete",
+            "level": level,
+            "key": f"proc:{evt.get('session_id', '')}",
+        },
+    )
+
+
 def _notification_poller_loop(
     stop_event: threading.Event, sid: str, session: dict
 ) -> None:
@@ -5661,6 +6045,7 @@ def _notification_poller_loop(
 
         rid = f"__notif__{int(time.time() * 1000)}"
         try:
+            _emit_process_completion_card(sid, evt)
             _emit("message.start", sid)
             _run_prompt_submit(rid, sid, session, text)
         except Exception as exc:
@@ -5704,6 +6089,7 @@ def _notification_poller_loop(
 
         rid = f"__notif__{int(time.time() * 1000)}"
         try:
+            _emit_process_completion_card(sid, evt)
             _emit("message.start", sid)
             _run_prompt_submit(rid, sid, session, text)
         except Exception as exc:
@@ -6039,10 +6425,20 @@ def _run_prompt_submit(rid, sid: str, session: dict, text: Any) -> None:
                         text,
                         raw,
                         session.get("history", []),
+                        # Auto-title lands on a background thread — refresh
+                        # session.info when it does so clients' window-title
+                        # chrome (OSC 0/2) updates without waiting for the
+                        # next turn. _emit is stdout-lock-guarded (thread-safe).
+                        title_callback=lambda _title: _emit_title_refresh(sid),
                     )
                 except Exception:
                     pass
 
+            # The pending title (applied synchronously above) is visible NOW —
+            # refresh session.info so window-title chrome picks it up.
+            if status == "complete" and _pending:
+                _emit_title_refresh(sid)
+
             # CLI parity: when voice-mode TTS is on, speak the agent reply
             # (cli.py:_voice_speak_response).  Only the final text — tool
             # calls / reasoning already stream separately and would be
@@ -6135,6 +6531,7 @@ def _run_prompt_submit(rid, sid: str, session: dict, text: Any) -> None:
                         break
                     session["running"] = True
                 try:
+                    _emit_process_completion_card(sid, _evt)
                     _emit("message.start", sid)
                     _run_prompt_submit(rid, sid, session, synth)
                 except Exception as _n_exc:
@@ -9930,6 +10327,67 @@ def _(rid, params: dict) -> dict:
         return _err(rid, 5031, str(e))
 
 
+@method("startup.catalog")
+def _(rid, params: dict) -> dict:
+    # Aggregate tools / skills / MCP servers for the native engine's startup panel
+    # (item 9). Opt-in RPC — only the opentui home screen calls it, so the Ink path
+    # is untouched. Each section is best-effort: a failing source yields an empty
+    # section rather than erroring the whole call.
+    tools: dict = {"total": 0, "toolsets": []}
+    try:
+        from toolsets import get_all_toolsets, get_toolset_info
+
+        # enabled toolsets for THIS session (or the config default), mirroring tools.list
+        session = _sessions.get(params.get("session_id", ""))
+        enabled = (
+            set(getattr(session["agent"], "enabled_toolsets", []) or [])
+            if session
+            else set(_load_enabled_toolsets() or [])
+        )
+        for name in sorted(get_all_toolsets().keys()):
+            info = get_toolset_info(name)
+            if not info:
+                continue
+            is_on = name in enabled if enabled else True
+            # the startup panel lists ENABLED toolsets with their tools (Ink parity)
+            tool_names = [str(t) for t in (info.get("resolved_tools") or [])]
+            tools["toolsets"].append(
+                {"name": name, "count": int(info["tool_count"]), "enabled": is_on, "tools": tool_names}
+            )
+            if is_on:
+                tools["total"] += int(info["tool_count"])
+    except Exception:
+        pass
+
+    skills: dict = {"total": 0, "categories": []}
+    try:
+        from hermes_cli.banner import get_available_skills
+
+        by_cat = get_available_skills() or {}
+        for cat in sorted(by_cat.keys()):
+            names = by_cat[cat] or []
+            skills["categories"].append({"name": cat, "count": len(names)})
+            skills["total"] += len(names)
+    except Exception:
+        pass
+
+    mcp_servers: list = []
+    try:
+        from hermes_cli.config import read_raw_config
+        from hermes_cli.tools_config import _parse_enabled_flag
+
+        raw_cfg = read_raw_config() or {}
+        servers = raw_cfg.get("mcp_servers")
+        if isinstance(servers, dict):
+            for name, cfg in servers.items():
+                if isinstance(cfg, dict) and _parse_enabled_flag(cfg.get("enabled", True), default=True):
+                    mcp_servers.append(str(name))
+    except Exception:
+        pass
+
+    return _ok(rid, {"tools": tools, "skills": skills, "mcp": {"servers": sorted(mcp_servers)}})
+
+
 @method("tools.show")
 def _(rid, params: dict) -> dict:
     try:

ui-opentui/.gitignore

@@ -0,0 +1,10 @@
+node_modules/
+dist/
+.repos/
+*.frame.txt
+*.ansi
+bun.lockb
+
+# the global ~/.gitignore_global `lib/` rule swallows our test harness — re-include it
+!src/test/lib/
+.bench/

ui-opentui/.node-version

@@ -0,0 +1 @@
+26.3

ui-opentui/.prettierrc

@@ -0,0 +1,11 @@
+{
+  "arrowParens": "avoid",
+  "bracketSpacing": true,
+  "endOfLine": "auto",
+  "printWidth": 120,
+  "semi": false,
+  "singleQuote": true,
+  "tabWidth": 2,
+  "trailingComma": "none",
+  "useTabs": false
+}

ui-opentui/README.md

@@ -0,0 +1,53 @@
+# ui-opentui — native OpenTUI engine for Hermes
+
+Solid + `@opentui/core` over Node FFI. Ink (`ui-tui/`) is the shipping default;
+this is the experimental engine (draft PR #42922).
+
+## Node 26 setup (required; will not touch your other projects)
+
+This package needs **Node ≥ 26.3** (`--experimental-ffi` floor). Everything
+else on this machine/repo can keep whatever Node it already uses — pin 26 to
+this directory only:
+
+```sh
+# 1. install fnm (skip if you have it; nvm/mise work too — see below)
+curl -fsSL https://fnm.vercel.app/install | bash
+# add to ~/.zshrc (or bashrc): eval "$(fnm env --use-on-cd --shell zsh)"
+
+# 2. install Node 26 SIDE BY SIDE (does NOT change your default)
+fnm install 26
+
+# 3. done — this directory has a .node-version (26.3), so `cd ui-opentui`
+#    auto-switches to 26 and leaving switches back. Do NOT run `fnm default 26`.
+node -v   # v26.x here; your old version everywhere else
+```
+
+No shell integration wanted (CI, scripts, one-off): `fnm exec --using 26 -- node ...`
+or invoke the absolute binary (`~/.local/share/fnm/node-versions/v26.*/installation/bin/node`).
+mise users: `mise use node@26` in this directory. nvm users: `nvm install 26`,
+plus an `.nvmrc` shim (`echo 26 > .nvmrc`) if you rely on auto-switching.
+
+### Gotchas
+
+- **Native modules are ABI-locked.** A `node_modules` installed under Node
+  20/22 will not load under 26 (and vice versa) — run `npm ci` (or
+  `npm rebuild`) after switching versions. Same applies to the **tui-bench** repo's node-pty (`github.com/NousResearch/tui-bench`).
+- **Global npm packages don't follow** between versions (per-version prefix);
+  reinstall the few you need, or don't use globals.
+- **Editor terminals** (Zed/VS Code) need the `fnm env` line in your shell rc;
+  the `.node-version` auto-switch then covers any shell that cd's here.
+- **Never run this package with bun** — the FFI seam and the Solid/JSX build
+  are Node-path only here.
+- `package.json` declares `engines.node >= 26.3`, so a wrong-Node `npm ci`
+  warns immediately.
+
+## Build & run
+
+```sh
+node scripts/build.mjs
+HERMES_TUI_MOUSE=1 node --experimental-ffi --no-warnings dist/main.js
+```
+
+Gates: `npm run check` (typecheck + lint + tests). Memory/perf benchmarks live
+in the **tui-bench** repo (`github.com/NousResearch/tui-bench`; see its README). Transcript windowing (memory architecture) is
+documented in `../docs/plans/opentui-transcript-windowing.md`.

ui-opentui/eslint.config.mjs

@@ -0,0 +1,103 @@
+import js from "@eslint/js"
+import tseslint from "typescript-eslint"
+import unusedImports from "eslint-plugin-unused-imports"
+
+export default tseslint.config(
+  {
+    // .bench/ and .demo/ are build artifacts (bench `nodes` cell and the
+    // smoke demo: `node scripts/build.mjs scripts/demo.tsx .demo`) — never lint.
+    ignores: ["node_modules/**", "dist/**", ".bench/**", ".demo/**", ".repos/**", "*.frame.txt", "*.ansi"],
+  },
+  js.configs.recommended,
+  ...tseslint.configs.recommendedTypeChecked,
+  {
+    files: ["**/*.ts", "**/*.tsx"],
+    languageOptions: {
+      parserOptions: {
+        projectService: true,
+        tsconfigRootDir: import.meta.dirname,
+      },
+    },
+    plugins: {
+      "unused-imports": unusedImports,
+    },
+    rules: {
+      // Boundary code bans these; the Solid view follows TS-strict but is not Effect.
+      "@typescript-eslint/no-explicit-any": "error",
+      "@typescript-eslint/consistent-type-imports": ["error", { prefer: "type-imports" }],
+      "@typescript-eslint/no-unused-vars": "off",
+      "@typescript-eslint/no-non-null-assertion": "error",
+      "unused-imports/no-unused-imports": "error",
+      "unused-imports/no-unused-vars": [
+        "error",
+        { vars: "all", varsIgnorePattern: "^_", args: "after-used", argsIgnorePattern: "^_" },
+      ],
+
+      // --- Type-aware, high-value: ON as ERROR ---
+      "@typescript-eslint/no-floating-promises": "error",
+      "@typescript-eslint/no-misused-promises": "error",
+      "@typescript-eslint/await-thenable": "error",
+
+      // --- Type-safety: ENFORCED as errors in our boundary/logic .ts code ---
+      // Production .ts is clean of the no-unsafe-* family (the loose-typed gateway
+      // payloads are Schema-decoded). The only sources are (a) *.tsx — @opentui/solid's
+      // JSX namespace types every component `return (<…>)` as `error`/unknown, a
+      // framework limitation disabled for views below — and (b) the test harness
+      // (loose render/effect fixtures + async mocks), exempt below. So we enforce ERROR.
+      "@typescript-eslint/no-unsafe-assignment": "error",
+      "@typescript-eslint/no-unsafe-member-access": "error",
+      "@typescript-eslint/no-unsafe-argument": "error",
+      "@typescript-eslint/no-unsafe-return": "error",
+      "@typescript-eslint/no-unsafe-call": "error",
+      "@typescript-eslint/no-base-to-string": "error",
+      "@typescript-eslint/restrict-template-expressions": "error",
+      "@typescript-eslint/no-unnecessary-type-assertion": "error",
+      "@typescript-eslint/require-await": "error",
+      // Defensive guards on untrusted runtime/gateway data: TS's narrowing doesn't
+      // model the wire, so "condition is always truthy" here is intentional armor,
+      // not dead code. Kept as a hint (warn), not a gate failure.
+      "@typescript-eslint/no-unnecessary-condition": "warn",
+    },
+  },
+  {
+    // @opentui/solid's custom JSX namespace types component returns as `error`/
+    // unknown, so EVERY `return (<…>)` in a view trips the no-unsafe-* family.
+    // That's a framework typing limitation, not unsafe app code — off for views.
+    files: ["**/*.tsx"],
+    rules: {
+      "@typescript-eslint/no-unsafe-return": "off",
+      "@typescript-eslint/no-unsafe-assignment": "off",
+      "@typescript-eslint/no-unsafe-member-access": "off",
+      "@typescript-eslint/no-unsafe-argument": "off",
+      "@typescript-eslint/no-unsafe-call": "off",
+    },
+  },
+  {
+    // Test helpers/fixtures: keep `!` on known-present data, and allow the loose
+    // render/effect harness casts + async mock signatures (they satisfy real
+    // Promise-returning interfaces with no body to await).
+    files: ["**/*.test.ts", "**/*.test.tsx", "src/test/lib/**"],
+    rules: {
+      "@typescript-eslint/no-non-null-assertion": "off",
+      "@typescript-eslint/no-unsafe-assignment": "off",
+      "@typescript-eslint/no-unsafe-member-access": "off",
+      "@typescript-eslint/no-unsafe-argument": "off",
+      "@typescript-eslint/no-unsafe-return": "off",
+      "@typescript-eslint/no-unsafe-call": "off",
+      "@typescript-eslint/no-unnecessary-type-assertion": "off",
+      "@typescript-eslint/require-await": "off",
+    },
+  },
+  {
+    // Build/config scripts (the eslint flat config, the esbuild build.mjs, the
+    // vitest config) are not part of the typed TS program, so the project service
+    // can't type them — disable type-aware linting there to avoid parser errors,
+    // and declare the Node globals they use (process, console, URL).
+    files: ["**/*.mjs", "*.config.ts"],
+    ...tseslint.configs.disableTypeChecked,
+    languageOptions: {
+      ...tseslint.configs.disableTypeChecked.languageOptions,
+      globals: { process: "readonly", console: "readonly", URL: "readonly", URLSearchParams: "readonly" },
+    },
+  },
+)

ui-opentui/package.json

@@ -0,0 +1,45 @@
+{
+  "name": "@hermes/ui-opentui",
+  "version": "0.0.0",
+  "private": true,
+  "type": "module",
+  "engines": {
+    "node": ">=26.3"
+  },
+  "description": "Native OpenTUI engine for Hermes (Solid + Effect-at-boundary, from scratch). Ink (ui-tui/) stays the shipping default.",
+  "scripts": {
+    "type-check": "tsc --noEmit",
+    "lint": "eslint .",
+    "lint:fix": "eslint . --fix",
+    "fmt": "prettier --write src",
+    "fix": "prettier --write src && eslint . --fix",
+    "build": "node scripts/build.mjs",
+    "start": "node --experimental-ffi --no-warnings dist/main.js",
+    "test": "vitest run",
+    "check": "bash scripts/check.sh",
+    "dev": "node scripts/build.mjs && node --experimental-ffi --no-warnings dist/main.js"
+  },
+  "dependencies": {
+    "@opentui/core": "0.4.1",
+    "@opentui/keymap": "0.4.1",
+    "@opentui/solid": "0.4.1",
+    "effect": "4.0.0-beta.78",
+    "fuzzysort": "^3.1.0",
+    "solid-js": "1.9.12"
+  },
+  "devDependencies": {
+    "@babel/core": "^7.29.7",
+    "@babel/preset-typescript": "^7.29.7",
+    "@effect/vitest": "^4.0.0-beta.78",
+    "@eslint/js": "^9",
+    "@types/node": "^24",
+    "babel-preset-solid": "^1.9.12",
+    "esbuild": "^0.28.0",
+    "eslint": "^9",
+    "eslint-plugin-unused-imports": "^4",
+    "prettier": "^3",
+    "typescript": "^5",
+    "typescript-eslint": "^8",
+    "vitest": "^4.1.8"
+  }
+}

ui-opentui/scripts/acceptance.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+# Single acceptance command for the Bun→Node-26 switchover (see
+# docs/plans/opentui-node26-build-spec.md). Proves, on a Node 26.3 host, that the
+# OpenTUI v2 engine runs WITHOUT Bun and at parity:
+#
+#   1. Node >= 26.3 present (the node:ffi floor); reports whether bun is on PATH
+#      (the engine must NOT need it).
+#   2. `npm run check`  — prettier + tsc + eslint + vitest (151+), all on Node.
+#   3. live-gateway transport smoke — spawns the real Python tui_gateway via the
+#      node:child_process client, asserts gateway.ready + session.create.
+#      (Skipped if no Hermes venv resolves — CI parity.)
+#   4. selection/markdown smoke in a real tmux TTY — asserts the native <markdown>
+#      (Tree-sitter) PAINTS under node --experimental-ffi and that a selection
+#      copies the RAW markdown source. (Skipped if tmux is unavailable.)
+#
+# Run:  cd ui-opentui && HERMES_PYTHON_SRC_ROOT=<checkout-root> bash scripts/acceptance.sh
+set -uo pipefail
+cd "$(dirname "$0")/.."
+
+# Absolute node, so a fresh tmux pane (which won't inherit our PATH / fnm shim)
+# runs the SAME Node 26.3, not the shell's default.
+NODE_BIN="$(command -v node || echo node)"
+
+pass=0; fail=0; skip=0
+ok()   { echo "  ✅ $1"; pass=$((pass+1)); }
+bad()  { echo "  ❌ $1"; fail=$((fail+1)); }
+note() { echo "  ⏭  $1"; skip=$((skip+1)); }
+
+echo "== [1/4] runtime: Node >= 26.3, Bun-free =="
+NODE_V="$(node -p 'process.versions.node' 2>/dev/null || echo 0.0.0)"
+node -e 'const [a,b]=process.versions.node.split(".").map(Number); process.exit(a>26||(a===26&&b>=3)?0:1)' \
+  && ok "node $NODE_V (>= 26.3)" || bad "node $NODE_V is below the 26.3 node:ffi floor"
+if command -v bun >/dev/null 2>&1; then
+  note "bun is on PATH ($(command -v bun)) — fine; the engine does not use it (proven below)"
+else
+  ok "no bun on PATH — single-runtime host"
+fi
+
+echo "== [2/4] check: prettier + tsc + eslint + vitest =="
+if bash scripts/check.sh >/tmp/accept-check.log 2>&1; then ok "check green ($(grep -c 'passed' /tmp/accept-check.log >/dev/null 2>&1; grep -oE '[0-9]+ passed' /tmp/accept-check.log | tail -1))"
+else bad "check failed — see /tmp/accept-check.log"; tail -20 /tmp/accept-check.log; fi
+
+echo "== [3/4] live-gateway transport smoke (real Python gateway, no Bun) =="
+if [ -n "${HERMES_PYTHON_SRC_ROOT:-}" ] || [ -x "../.venv/bin/python" ]; then
+  rm -rf .accept && node scripts/build.mjs src/test/liveGateway.smoke.ts .accept >/dev/null 2>&1
+  OUT="$(node --experimental-ffi --no-warnings .accept/liveGateway.smoke.js 2>&1)"
+  echo "$OUT" | grep -q "^PASS" && ok "$(echo "$OUT" | grep '^PASS')" || { echo "$OUT" | grep -qE "TRANSPORT ERROR|SKIP" && note "gateway smoke skipped (no python/model)" || bad "gateway smoke: $(echo "$OUT" | head -1)"; }
+  rm -rf .accept
+else
+  note "no HERMES_PYTHON_SRC_ROOT / venv — gateway smoke skipped"
+fi
+
+echo "== [4/4] selection/markdown smoke in a real tmux TTY (tree-sitter under FFI) =="
+if command -v tmux >/dev/null 2>&1; then
+  rm -rf .accept && node scripts/build.mjs src/test/selectionCopy.smoke.tsx .accept >/dev/null 2>&1
+  rm -f /tmp/accept-sel.json
+  S="accept-$$"
+  tmux kill-session -t "$S" 2>/dev/null
+  tmux new-session -d -s "$S" -x 120 -y 40
+  tmux send-keys -t "$S" "SEL_SMOKE_OUT=/tmp/accept-sel.json $NODE_BIN --experimental-ffi --no-warnings $PWD/.accept/selectionCopy.smoke.js; tmux wait-for -S $S" Enter
+  tmux wait-for "$S" 2>/dev/null || sleep 6
+  tmux kill-session -t "$S" 2>/dev/null
+  if node -e 'process.exit(require("/tmp/accept-sel.json").pass===true?0:1)' 2>/dev/null; then
+    ok "markdown painted + selection copied source (tree-sitter under node FFI)"
+  else
+    bad "selection/markdown smoke failed — see /tmp/accept-sel.json"; cat /tmp/accept-sel.json 2>/dev/null
+  fi
+  rm -rf .accept
+else
+  note "tmux not available — markdown smoke skipped (run it on a TTY host)"
+fi
+
+echo
+echo "== acceptance: $pass passed, $fail failed, $skip skipped =="
+[ "$fail" -eq 0 ] && { echo "ACCEPTANCE: PASS"; exit 0; } || { echo "ACCEPTANCE: FAIL"; exit 1; }

ui-opentui/scripts/build.mjs

@@ -0,0 +1,75 @@
+/**
+ * Build the OpenTUI v2 Solid app for Node 26 (no Bun).
+ *
+ * Mirrors OpenTUI's own Node recipe (`~/github/opentui/.../run-node26.mjs` +
+ * `packages/solid/scripts/solid-transform.ts`): apply babel-preset-solid in
+ * `generate:"universal"` mode with `moduleName:"@opentui/solid"` to every app
+ * .tsx/.jsx, and force solid-js to its CLIENT/universal build (the package's
+ * `node` export condition points at the SSR `server.js`, which lacks the
+ * reactive primitives the universal renderer needs).
+ *
+ * `@opentui/core` stays EXTERNAL: it resolves its per-arch native `libopentui.so`
+ * (and the tree-sitter worker) from its own package dir via `import.meta.url`;
+ * bundling it would break those paths.
+ *
+ * Run with the Node that will launch the app:
+ *   node scripts/build.mjs                       # → dist/main.js (app entry)
+ *   node scripts/build.mjs <entry.tsx> <outdir>  # build an arbitrary entry (smokes/spikes)
+ * Launch:
+ *   node --experimental-ffi --no-warnings dist/main.js
+ */
+import { readFile } from 'node:fs/promises'
+import { createRequire } from 'node:module'
+import { dirname, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+import { transformAsync } from '@babel/core'
+import tsPreset from '@babel/preset-typescript'
+import solidPreset from 'babel-preset-solid'
+import * as esbuild from 'esbuild'
+
+const require = createRequire(import.meta.url)
+const root = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+
+/** esbuild plugin that reproduces @opentui/solid's transform + solid-js resolution. */
+const opentuiSolid = {
+  name: 'opentui-solid',
+  setup(build) {
+    // App JSX (.tsx/.jsx, never node_modules) → babel-preset-solid (universal).
+    build.onLoad({ filter: /\.[cm]?[jt]sx$/ }, async args => {
+      if (args.path.includes('/node_modules/')) return null
+      const code = await readFile(args.path, 'utf8')
+      const out = await transformAsync(code, {
+        filename: args.path,
+        configFile: false,
+        babelrc: false,
+        presets: [[solidPreset, { moduleName: '@opentui/solid', generate: 'universal' }], [tsPreset]]
+      })
+      return { contents: out?.code ?? '', loader: 'js' }
+    })
+
+    // Force the universal/client solid-js build (node condition → server.js otherwise).
+    build.onResolve({ filter: /^solid-js$/ }, () => ({ path: require.resolve('solid-js/dist/solid.js') }))
+    build.onResolve({ filter: /^solid-js\/store$/ }, () => ({ path: require.resolve('solid-js/store/dist/store.js') }))
+  }
+}
+
+const [, , entryArg, outdirArg] = process.argv
+const entry = entryArg ? resolve(process.cwd(), entryArg) : resolve(root, 'src/entry/main.tsx')
+const outdir = outdirArg ? resolve(process.cwd(), outdirArg) : resolve(root, 'dist')
+
+await esbuild.build({
+  entryPoints: [entry],
+  outdir,
+  bundle: true,
+  format: 'esm',
+  platform: 'node',
+  target: 'node26',
+  splitting: true,
+  sourcemap: true,
+  logLevel: 'info',
+  // Native blob + tree-sitter worker resolve from @opentui/core's own dir at runtime.
+  external: ['@opentui/core', '@opentui/core/*'],
+  plugins: [opentuiSolid],
+  define: { 'process.env.OPENTUI_BUN_ONLY_EXAMPLES': '"false"' }
+})

ui-opentui/scripts/check.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Phase gate for the native OpenTUI engine (spec v4 §5). Runs the full headless
+# suite: format + type-check + lint + vitest (which includes the headless frame
+# gate via captureCharFrame). The agentic smoke (docs/plans/opentui-smoke.md) is
+# the live complement — run BOTH every phase.
+#
+# Runs entirely on Node 26.3 (no Bun). The OpenTUI native core loads via node:ffi
+# under --experimental-ffi; vitest passes that flag to its test forks (see
+# vitest.config.ts). Requires `node -v` == v26.3.x on PATH.
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+echo "== [1/4] format (prettier --check) =="
+npx prettier --check src
+
+echo "== [2/4] type-check =="
+npm run --silent type-check
+
+echo "== [3/4] lint =="
+npm run --silent lint
+
+echo "== [4/4] vitest (incl. headless frame gate) =="
+npm test
+
+echo "== check OK =="

ui-opentui/scripts/demo.tsx

@@ -0,0 +1,57 @@
+/**
+ * DEV DEMO — NOT a test, NOT production. Renders the bench fixture (lorem-ipsum +
+ * fat tool-turns from ./fixture.ts) in a REAL CliRenderer so you can attach over
+ * tmux, scroll, and eyeball the transcript + the rolling-cap truncation notice.
+ * No gateway is spawned (purely the fixture seeded into the store via the resume
+ * path), so typing won't reach a backend — it's for viewing/scrolling.
+ *
+ * Run (Node 26 — needs the esbuild/Solid transform, then --experimental-ffi):
+ *   node scripts/build.mjs scripts/demo.tsx .demo
+ *   node --experimental-ffi --no-warnings .demo/demo.js      # inside tmux (needs a TTY)
+ *   DEMO_TOTAL=200               fixture messages to seed (default 200)
+ *   HERMES_TUI_MAX_MESSAGES=80   cap → the "⤒ N earlier messages" notice fires
+ * Quit: Ctrl+C.
+ */
+import { createCliRenderer } from '@opentui/core'
+import { render } from '@opentui/solid'
+
+import { installMultiClickSelection } from '../src/boundary/multiClickSelect.ts'
+import { registerRemoteParsers } from '../src/boundary/parsers.ts'
+import { createSessionStore } from '../src/logic/store.ts'
+import { App } from '../src/view/App.tsx'
+import { ThemeProvider } from '../src/view/theme.tsx'
+import { materialize } from './fixture.ts'
+
+// Same grammar registration as the live entry so fixture code blocks highlight
+// (fetched+cached on first use; no HERMES_TUI_PARSER_CACHE here → OpenTUI default).
+registerRemoteParsers()
+
+const TOTAL = Number.parseInt(process.env.DEMO_TOTAL ?? '', 10) || 200
+
+const store = createSessionStore()
+store.apply({ type: 'gateway.ready' })
+store.setSessionId('demo-fixture-20260609')
+// Seed via the resume path so the cap slices + the `dropped` counter is set
+// (drives the truncation notice) exactly as a real `session.resume` would.
+store.beginBuffer()
+store.commitSnapshot(materialize(TOTAL))
+
+const renderer = await createCliRenderer({
+  externalOutputMode: 'passthrough',
+  targetFps: 60,
+  exitOnCtrlC: true,
+  useKittyKeyboard: {},
+  useMouse: true
+})
+// Same seam the live entry installs (boundary/renderer.ts) so the demo smokes
+// double-click word / triple-click line / drag-extend too.
+installMultiClickSelection(renderer)
+
+void render(
+  () => (
+    <ThemeProvider theme={() => store.state.theme}>
+      <App store={store} />
+    </ThemeProvider>
+  ),
+  renderer
+)

ui-opentui/scripts/fixture.ts

@@ -0,0 +1,292 @@
+/**
+ * DEV BENCH FIXTURE — NOT a test, NOT production code. A deterministic generator
+ * for a REALISTIC heavy session, consumed by `scripts/mem-bench.tsx`. Excluded
+ * from the vitest run (not a *.test.ts) and lint-clean.
+ *
+ * The old synthetic bench pushed tiny 3-delta turns (~5.5 mounted nodes each) —
+ * an unrealistic per-message cost. Real transcripts are LUMPY: an assistant turn
+ * is ONE `message` but a fat node subtree (markdown blocks + a reasoning block +
+ * several tool headers, each a multi-line result). That makes message-count a
+ * LOOSE proxy for memory, which is exactly what we're trying to quantify before
+ * picking a `HERMES_TUI_MAX_MESSAGES` default.
+ *
+ * Design: a turn is modeled as a small typed `TurnAction` union (user / system /
+ * gateway-event). The driver maps user→`pushUser`, system→`pushSystem`, and every
+ * gateway event through the SAME `apply()` reducer real usage takes — so the
+ * mounted result is identical to a live session. The same action stream also
+ * materializes a settled `Message[]` (via `materialize`) for the resume-path check
+ * (`commitSnapshot`). Everything is seeded by index (no `Math.random` —
+ * unavailable here), so a given `total` reproduces byte-for-byte.
+ */
+import type { GatewayEvent } from '../src/boundary/schema/GatewayEvent.ts'
+import { createSessionStore, type Message } from '../src/logic/store.ts'
+
+/** One scripted action in a turn: a composer push or a decoded gateway event. */
+type TurnAction =
+  | { kind: 'user'; text: string }
+  | { kind: 'system'; text: string }
+  | { kind: 'event'; event: GatewayEvent }
+
+/** A pool of lorem-ipsum words — varied content is selected by index from here. */
+const WORDS = [
+  'lorem',
+  'ipsum',
+  'dolor',
+  'sit',
+  'amet',
+  'consectetur',
+  'adipiscing',
+  'elit',
+  'sed',
+  'eiusmod',
+  'tempor',
+  'incididunt',
+  'labore',
+  'magna',
+  'aliqua',
+  'enim',
+  'minim',
+  'veniam',
+  'quis',
+  'nostrud',
+  'exercitation',
+  'ullamco',
+  'laboris',
+  'aliquip',
+  'commodo',
+  'consequat',
+  'duis',
+  'aute',
+  'irure',
+  'reprehenderit',
+  'voluptate',
+  'velit',
+  'esse',
+  'cillum',
+  'fugiat',
+  'nulla',
+  'pariatur',
+  'excepteur',
+  'occaecat',
+  'cupidatat',
+  'proident',
+  'sunt',
+  'culpa',
+  'officia',
+  'deserunt',
+  'mollit',
+  'anim'
+] as const
+
+/** Deterministic pseudo-word stream: pick from WORDS by a seeded index. */
+function word(seed: number, k: number): string {
+  return WORDS[(seed * 31 + k * 7) % WORDS.length] ?? 'lorem'
+}
+
+/** A lorem sentence of `n` words, capitalized + terminated. */
+function sentence(seed: number, n: number): string {
+  const parts: string[] = []
+  for (let k = 0; k < n; k++) parts.push(word(seed + k, k))
+  const text = parts.join(' ')
+  return text.charAt(0).toUpperCase() + text.slice(1) + '.'
+}
+
+/** A paragraph of `s` sentences (varying length by index). */
+function paragraph(seed: number, s: number): string {
+  const out: string[] = []
+  for (let i = 0; i < s; i++) out.push(sentence(seed + i * 13, 6 + ((seed + i) % 9)))
+  return out.join(' ')
+}
+
+/** N lorem-ipsum lines (for tool result bodies), each varying in length. */
+function lines(seed: number, n: number): string {
+  const out: string[] = []
+  for (let i = 0; i < n; i++) out.push(sentence(seed + i * 5, 4 + ((seed + i) % 11)))
+  return out.join('\n')
+}
+
+/** A markdown assistant body: paragraphs + a list + a fenced code block. */
+function assistantMarkdown(seed: number): string {
+  const lead = paragraph(seed, 1 + (seed % 3))
+  const bullets = [`- ${sentence(seed + 1, 5)}`, `- ${sentence(seed + 2, 7)}`, `- ${sentence(seed + 3, 4)}`].join('\n')
+  const code = [
+    '```ts',
+    `const x${seed % 7} = ${seed % 100}`,
+    `function f${seed % 5}() {`,
+    '  return x',
+    '}',
+    '```'
+  ].join('\n')
+  const tail = paragraph(seed + 17, 1 + ((seed + 1) % 2))
+  return `${lead}\n\n${bullets}\n\n${code}\n\n${tail}`
+}
+
+/** Tool names cycled by index (mirrors a real tool mix). */
+const TOOL_NAMES = ['terminal', 'read_file', 'edit_file', 'grep', 'web_search', 'write_file'] as const
+
+/** A tool.start + tool.complete pair for tool `t` in turn `seed`. */
+function toolEvents(seed: number, t: number): GatewayEvent[] {
+  const id = `tool-${seed}-${t}`
+  const name = TOOL_NAMES[(seed + t) % TOOL_NAMES.length] ?? 'terminal'
+  const variant = (seed + t) % 3
+  // short / capped-16-line / medium result bodies, mixing the render-cost cases.
+  // BENCH KNOB: HERMES_BENCH_TOOL_BODY_LINES overrides the body size to N lines
+  // for ALL tools — the "fat tool output" fixture (e.g. a `find /` dump) used to
+  // make W3's retention win measurable. UNSET = the original tiny bodies (so the
+  // default fixture stays byte-identical; existing benches are unaffected).
+  const fatLines = Number.parseInt(process.env.HERMES_BENCH_TOOL_BODY_LINES ?? '', 10)
+  const bodyLines = Number.isFinite(fatLines) && fatLines > 0 ? fatLines : variant === 0 ? 2 : variant === 1 ? 18 : 7
+  const resultText = lines(seed + t * 3, bodyLines)
+  const context = sentence(seed + t, 4)
+  // ~half the tools carry a multi-line args block (the expanded-view cost).
+  const withArgs = (seed + t) % 2 === 0
+  const start: GatewayEvent = {
+    type: 'tool.start',
+    payload: withArgs ? { tool_id: id, name, context, args_text: lines(seed + t, 5) } : { tool_id: id, name, context }
+  }
+  const complete: GatewayEvent = {
+    type: 'tool.complete',
+    payload: {
+      tool_id: id,
+      name,
+      result_text: resultText,
+      duration_s: 0.1 + ((seed + t) % 40) / 10,
+      args: { command: context, index: seed + t }
+    }
+  }
+  return [start, complete]
+}
+
+/** One USER message (1–4 lorem paragraphs; some very short, some RFC-sized). */
+function userText(seed: number): string {
+  const shape = seed % 7
+  if (shape === 0) return 'yes do that'
+  if (shape === 1) return 'ok'
+  if (shape === 6) {
+    // an RFC-sized pasted block: many paragraphs.
+    const out: string[] = []
+    for (let p = 0; p < 8; p++) out.push(paragraph(seed + p * 23, 4 + (p % 3)))
+    return out.join('\n\n')
+  }
+  const n = 1 + (seed % 4)
+  const out: string[] = []
+  for (let p = 0; p < n; p++) out.push(paragraph(seed + p * 11, 1 + ((seed + p) % 3)))
+  return out.join('\n\n')
+}
+
+/**
+ * Build the scripted actions for ONE turn. Most turns are a plain user+assistant
+ * exchange; a deterministic subset are tool-heavy (1–15 tool calls) or a system
+ * slash-output line. Returns the actions for the whole turn in order.
+ */
+function turnActions(turn: number): TurnAction[] {
+  const actions: TurnAction[] = []
+  // Occasional system slash-output line (≈ every 9th turn) instead of a user line.
+  if (turn % 9 === 4) {
+    actions.push({ kind: 'system', text: sentence(turn, 8) })
+    return actions
+  }
+
+  actions.push({ kind: 'user', text: userText(turn) })
+  actions.push({ kind: 'event', event: { type: 'message.start' } })
+
+  // Reasoning on ≈ every 3rd assistant turn.
+  if (turn % 3 === 0) {
+    actions.push({
+      kind: 'event',
+      event: {
+        type: 'reasoning.delta',
+        payload: { text: `**${sentence(turn, 3).replace(/\.$/, '')}**\n\n${paragraph(turn + 5, 2)}` }
+      }
+    })
+  }
+
+  // Leading text part.
+  actions.push({ kind: 'event', event: { type: 'message.delta', payload: { text: assistantMarkdown(turn) } } })
+
+  // Tool-heavy turns: ≈ every 4th assistant turn carries several tool calls,
+  // interleaved with a follow-up text part (the fat-turn stress case).
+  if (turn % 4 === 0) {
+    const toolCount = 1 + (turn % 15) // 1..15 tools
+    for (let t = 0; t < toolCount; t++) {
+      for (const ev of toolEvents(turn, t)) actions.push({ kind: 'event', event: ev })
+    }
+    actions.push({ kind: 'event', event: { type: 'message.delta', payload: { text: paragraph(turn + 31, 2) } } })
+  }
+
+  actions.push({ kind: 'event', event: { type: 'message.complete' } })
+  return actions
+}
+
+/** How many transcript ROWS a turn produces (user/system + at most one assistant). */
+export function rowsPerTurn(turn: number): number {
+  return turn % 9 === 4 ? 1 : 2
+}
+
+/** Apply ONE turn's actions to a store via the same paths real usage takes. */
+export function applyTurn(store: ReturnType<typeof createSessionStore>, turn: number): void {
+  for (const action of turnActions(turn)) {
+    if (action.kind === 'user') store.pushUser(action.text)
+    else if (action.kind === 'system') store.pushSystem(action.text)
+    else store.apply(action.event)
+  }
+}
+
+/**
+ * Drive at least `total` MESSAGES into the live store, calling `onSample(pushes)`
+ * each time the cumulative produced-row count crosses a `sampleEvery` boundary.
+ * `pushes` counts MESSAGES (rows produced, pre-cap), so the matrix samples on a
+ * raw message cadence regardless of the rolling cap.
+ */
+export function drive(
+  store: ReturnType<typeof createSessionStore>,
+  total: number,
+  sampleEvery: number,
+  onSample: (pushes: number) => void
+): number {
+  let pushed = 0
+  let nextSample = sampleEvery
+  let turn = 0
+  while (pushed < total) {
+    applyTurn(store, turn)
+    pushed += rowsPerTurn(turn)
+    turn++
+    while (pushed >= nextSample && nextSample <= total) {
+      onSample(Math.min(pushed, total))
+      nextSample += sampleEvery
+    }
+  }
+  return turn
+}
+
+/**
+ * Materialize the FULL settled `Message[]` for the resume path: replay the same
+ * action stream into a FRESH, EFFECTIVELY-UNCAPPED store and snapshot its rows.
+ * This guarantees the resume fixture is byte-identical to what the live push
+ * path produces (minus the rolling cap), so `commitSnapshot` mounts the real shape.
+ */
+export function materialize(total: number): Message[] {
+  // `uncappedFixture` bypasses the store's handle-safe cap CLAMP (an env value
+  // can no longer raise the cap past logic/store.ts HANDLE_SAFE_MAX_ROWS — the
+  // old env=MAX_SAFE_INTEGER trick would now silently truncate to 1000 rows).
+  // This store is never mounted into a renderer, so no native handles are at stake.
+  const store = createSessionStore({ uncappedFixture: true })
+  store.apply({ type: 'gateway.ready' })
+  let pushed = 0
+  let turn = 0
+  while (pushed < total) {
+    applyTurn(store, turn)
+    pushed += rowsPerTurn(turn)
+    turn++
+  }
+  // Deep-copy out of the solid store proxy into plain objects (the resume path
+  // takes a plain Message[]).
+  return store.state.messages.slice(0, total).map(cloneMessage)
+}
+
+/** Plain deep copy of a store Message (drop the solid proxy + streaming flag). */
+function cloneMessage(m: Message): Message {
+  const copy: Message = { role: m.role, text: m.text }
+  if (m.parts) copy.parts = m.parts.map(p => ({ ...p }))
+  return copy
+}

ui-opentui/scripts/mem-bench.tsx

@@ -0,0 +1,177 @@
+/**
+ * DEV BENCH — NOT a test, NOT production code. Throwaway memory-measurement
+ * harness for tuning the rolling `HERMES_TUI_MAX_MESSAGES` cap. Mounts the
+ * production `<App store={createSessionStore()}>` under the `@opentui/solid` test
+ * renderer and samples `process.memoryUsage()` + the mounted-renderable count +
+ * `getAllocatorStats().activeAllocations`, forcing `global.gc()` before each
+ * sample. Excluded from the test run (not a *.test.ts) and lint-clean.
+ *
+ * It pushes a REALISTIC heavy-session fixture (scripts/fixture.ts) — varied user
+ * turns + fat multi-part assistant turns (markdown + reasoning + several tool
+ * headers) — because per-message size varies hugely, so message-count is only a
+ * LOOSE memory proxy and we're choosing a cap default.
+ *
+ *   node scripts/build.mjs scripts/mem-bench.tsx .bench   # build once (Solid+TS → JS)
+ *   Uncapped:  MEM_BENCH_TOTAL=8000 HERMES_TUI_MAX_MESSAGES=100000 \
+ *     node --experimental-ffi --expose-gc --no-warnings .bench/mem-bench.js
+ *   Capped:    MEM_BENCH_TOTAL=8000 HERMES_TUI_MAX_MESSAGES=1500 \
+ *     node --experimental-ffi --expose-gc --no-warnings .bench/mem-bench.js
+ *
+ * Run each cap as a SEPARATE node invocation so the WASM/native heap starts fresh.
+ * The matrix loop:
+ *   for cap in 400 1500 3000 6000 100000; do \
+ *     MEM_BENCH_TOTAL=8000 HERMES_TUI_MAX_MESSAGES=$cap \
+ *       node --experimental-ffi --expose-gc --no-warnings .bench/mem-bench.js; done
+ *
+ * Signal: native `getAllocatorStats().activeAllocations` (the Zig-side allocator
+ * count — every live renderable/Yoga subtree contributes) and the recursive
+ * renderable descendant count under `renderer.root`. RSS is reported too but is
+ * noisy and grow-only (WASM linear memory never returns to the OS), so the
+ * meaningful comparison is the STEADY-STATE plateau: capped should flatten after
+ * ~CAP messages; uncapped should keep climbing.
+ *
+ * GC: forces `global.gc()` (synchronous) before each sample to measure RETAINED
+ * memory, not garbage — run Node with `--expose-gc` or the GC call is a no-op.
+ *
+ * RESUME PATH: after the live push matrix, builds the full fixture as a settled
+ * Message[] and `commitSnapshot`s it (the resume path), reporting mounted nodes +
+ * RSS — verifying the slice-before-set fix bounds resume mounting to ≤ cap.
+ */
+import { resolveRenderLib } from '@opentui/core'
+import type { Renderable } from '@opentui/core'
+import { testRender } from '@opentui/solid'
+
+import { createSessionStore } from '../src/logic/store.ts'
+import { App } from '../src/view/App.tsx'
+import { ThemeProvider } from '../src/view/theme.tsx'
+import { applyTurn, materialize, rowsPerTurn } from './fixture.ts'
+
+const lib = resolveRenderLib()
+
+const TOTAL = Number.parseInt(process.env.MEM_BENCH_TOTAL ?? '8000', 10)
+const SAMPLE_EVERY = Number.parseInt(process.env.MEM_BENCH_SAMPLE ?? '500', 10)
+const cap = process.env.HERMES_TUI_MAX_MESSAGES ?? '(default 400)'
+
+const MB = (bytes: number) => (bytes / 1024 / 1024).toFixed(1)
+
+/** Force a synchronous full GC to measure RETAINED memory. No-op without `node --expose-gc`. */
+const forceGc = (): void => {
+  const gc = (globalThis as { gc?: () => void }).gc
+  if (gc) gc()
+}
+
+/** Recursively count every Renderable under root (a proxy for live Yoga nodes). */
+function descendantCount(node: Renderable): number {
+  let n = 0
+  for (const child of node.getChildren()) n += 1 + descendantCount(child)
+  return n
+}
+
+async function main(): Promise<void> {
+  const store = createSessionStore()
+  store.apply({ type: 'gateway.ready' })
+
+  const setup = await testRender(
+    () => (
+      <ThemeProvider theme={() => store.state.theme}>
+        <App store={store} />
+      </ThemeProvider>
+    ),
+    { width: 100, height: 40, exitOnCtrlC: false }
+  )
+  await setup.renderOnce()
+  await setup.flush()
+
+  process.stdout.write(
+    `\n=== mem-bench (REALISTIC fixture)  cap=${cap}  total=${TOTAL}  sampleEvery=${SAMPLE_EVERY} ===\n`
+  )
+  process.stdout.write(
+    'pushes | msgs | rss(MB) | heapUsed(MB) | external(MB) | arrayBuf(MB) | activeAllocs | renderables\n'
+  )
+  process.stdout.write(
+    '-------+------+---------+--------------+--------------+--------------+--------------+------------\n'
+  )
+
+  async function sample(pushes: number): Promise<void> {
+    await setup.renderOnce()
+    await setup.flush()
+    forceGc() // synchronous, full GC — measure retained, not garbage
+    const m = process.memoryUsage()
+    const alloc = lib.getAllocatorStats()
+    const renderables = descendantCount(setup.renderer.root)
+    const cols = [
+      String(pushes).padStart(6),
+      String(store.state.messages.length).padStart(4),
+      MB(m.rss).padStart(7),
+      MB(m.heapUsed).padStart(12),
+      MB(m.external).padStart(12),
+      MB(m.arrayBuffers).padStart(12),
+      String(alloc.activeAllocations).padStart(12),
+      String(renderables).padStart(11)
+    ]
+    process.stdout.write(cols.join(' | ') + '\n')
+  }
+
+  await sample(0)
+  // Pump turns inline, sampling each time the cumulative produced-row count crosses
+  // a SAMPLE_EVERY boundary. Sampling is async (renderOnce/flush/gc), so it lives
+  // in the loop rather than a sync callback. Mounting is synchronous in Solid, so a
+  // render pass at the boundary reflects the just-pushed turns.
+  let pushed = 0
+  let nextSample = SAMPLE_EVERY
+  let turn = 0
+  while (pushed < TOTAL) {
+    applyTurn(store, turn)
+    pushed += rowsPerTurn(turn)
+    turn++
+    if (pushed >= nextSample) {
+      await sample(Math.min(pushed, TOTAL))
+      while (nextSample <= pushed) nextSample += SAMPLE_EVERY
+    }
+  }
+
+  // Tear down the live push tree BEFORE the resume path so its mounted nodes don't
+  // pollute the process-wide RSS the resume sample reads. (The renderable COUNT is
+  // already isolated per-renderer-root, but RSS is process-global.)
+  store.clearTranscript()
+  setup.renderer.destroy()
+  forceGc()
+
+  // ── RESUME PATH: build the full settled fixture and commitSnapshot it (the
+  // resume hydrate path). Verifies the slice-before-set fix bounds resume mounting
+  // to ≤ cap — mounting 8000 settled msgs at cap=1500 should mount ~1500-worth of
+  // rows, NOT 8000-worth. Done on a FRESH store + renderer so the live-push history
+  // above doesn't skew the count.
+  const resumeStore = createSessionStore()
+  resumeStore.apply({ type: 'gateway.ready' })
+  const resumeSetup = await testRender(
+    () => (
+      <ThemeProvider theme={() => resumeStore.state.theme}>
+        <App store={resumeStore} />
+      </ThemeProvider>
+    ),
+    { width: 100, height: 40, exitOnCtrlC: false }
+  )
+  await resumeSetup.renderOnce()
+  await resumeSetup.flush()
+
+  const fullFixture = materialize(TOTAL)
+  resumeStore.beginBuffer()
+  resumeStore.commitSnapshot(fullFixture)
+  await resumeSetup.renderOnce()
+  await resumeSetup.flush()
+  forceGc()
+  const rm = process.memoryUsage()
+  const ralloc = lib.getAllocatorStats()
+  const rrenderables = descendantCount(resumeSetup.renderer.root)
+  process.stdout.write('\n--- resume path (commitSnapshot of the full fixture) ---\n')
+  process.stdout.write(`fixture msgs built : ${fullFixture.length}\n`)
+  process.stdout.write(`mounted msgs (cap) : ${resumeStore.state.messages.length}\n`)
+  process.stdout.write(`mounted renderables: ${rrenderables}\n`)
+  process.stdout.write(`activeAllocations  : ${ralloc.activeAllocations}\n`)
+  process.stdout.write(`rss(MB)            : ${MB(rm.rss)}\n`)
+
+  resumeSetup.renderer.destroy()
+}
+
+await main()

ui-opentui/src/boundary/clipboard.ts

@@ -0,0 +1,126 @@
+/**
+ * Clipboard (item 1) — copy via OSC 52 (works over SSH/tmux) + a native platform
+ * command, and read a clipboard IMAGE for paste-to-attach. Ported/trimmed from
+ * opencode `clipboard.ts`. A boundary concern (spawns processes / writes stdout);
+ * everything is best-effort and never throws into the view.
+ */
+import { spawn } from 'node:child_process'
+import { existsSync } from 'node:fs'
+import { platform } from 'node:os'
+import { join } from 'node:path'
+
+/** Whether `cmd` resolves on PATH (cached). We DON'T spawn missing tools: a failed
+ *  spawn + writing to its dead stdin pipe raises EPIPE/SIGPIPE, and OpenTUI used to
+ *  treat SIGPIPE as a shutdown signal — i.e. a clipboard miss would quit the TUI.
+ *  Skipped on Windows (the built-in `clip` is always present; PATHEXT complicates
+ *  a filename probe). */
+const commandCache = new Map<string, boolean>()
+function commandExists(cmd: string): boolean {
+  if (platform() === 'win32') return true
+  const cached = commandCache.get(cmd)
+  if (cached !== undefined) return cached
+  const dirs = (process.env.PATH ?? '').split(':').filter(Boolean)
+  const found = dirs.some(dir => existsSync(join(dir, cmd)))
+  commandCache.set(cmd, found)
+  return found
+}
+
+/** Run a command, optionally piping `input` to stdin; resolve its stdout bytes.
+ *  Best-effort and crash-proof: every stream error (incl. EPIPE → SIGPIPE on a
+ *  clipboard tool that exits early) is swallowed so a failed copy never throws out
+ *  of the boundary or signals the process. */
+function run(cmd: string, args: string[] = [], input?: string): Promise<Buffer> {
+  return new Promise((resolve, reject) => {
+    let child
+    try {
+      child = spawn(cmd, args, { stdio: [input === undefined ? 'ignore' : 'pipe', 'pipe', 'ignore'] })
+    } catch (cause) {
+      reject(cause instanceof Error ? cause : new Error(String(cause)))
+      return
+    }
+    const out: Buffer[] = []
+    child.on('error', reject)
+    child.stdout?.on('error', () => {}) // a closed stdout pipe must not throw
+    child.stdout?.on('data', (c: Buffer) => out.push(c))
+    child.on('close', code => (code === 0 ? resolve(Buffer.concat(out)) : reject(new Error(`${cmd} exit ${code}`))))
+    if (input !== undefined && child.stdin) {
+      // Writing to a tool that died/closed early raises EPIPE (→ SIGPIPE). Swallow it.
+      child.stdin.on('error', () => {})
+      try {
+        child.stdin.end(input)
+      } catch {
+        // pipe already gone — nothing to flush
+      }
+    }
+  })
+}
+
+/** OSC 52 copy — the terminal puts `text` on the system clipboard (SSH/tmux-safe). */
+function writeOsc52(text: string): void {
+  if (!process.stdout.isTTY) return
+  const seq = `\x1b]52;c;${Buffer.from(text).toString('base64')}\x07`
+  // tmux/screen need the sequence wrapped in their passthrough escape.
+  process.stdout.write(process.env.TMUX || process.env.STY ? `\x1bPtmux;\x1b${seq}\x1b\\` : seq)
+}
+
+/** Native copy commands to try, in order, for the current platform. */
+function copyCandidates(): Array<[string, string[]]> {
+  const os = platform()
+  if (os === 'darwin') return [['pbcopy', []]]
+  if (os === 'win32') return [['clip', []]]
+  // linux: prefer Wayland, then X11 tools
+  const list: Array<[string, string[]]> = []
+  if (process.env.WAYLAND_DISPLAY) list.push(['wl-copy', []])
+  list.push(['xclip', ['-selection', 'clipboard']], ['xsel', ['--clipboard', '--input']])
+  return list
+}
+
+/** Copy `text` to the clipboard: OSC 52 (always) + the first native command that works. */
+export async function writeClipboard(text: string): Promise<void> {
+  writeOsc52(text) // primary path — SSH/tmux-safe, no subprocess
+  for (const [cmd, args] of copyCandidates()) {
+    if (!commandExists(cmd)) continue // never spawn a missing tool (avoids EPIPE/SIGPIPE)
+    try {
+      await run(cmd, args, text)
+      return
+    } catch {
+      // try the next candidate
+    }
+  }
+}
+
+/** Read a clipboard IMAGE as base64 PNG (for paste-to-attach); undefined if none. */
+export async function readClipboardImage(): Promise<{ data: string; mime: string } | undefined> {
+  const os = platform()
+  const tries: Array<[string, string[]]> = []
+  if (os === 'linux') {
+    if (process.env.WAYLAND_DISPLAY) tries.push(['wl-paste', ['-t', 'image/png']])
+    tries.push(['xclip', ['-selection', 'clipboard', '-t', 'image/png', '-o']])
+  } else if (os === 'darwin') {
+    tries.push(['pngpaste', ['-']]) // brew install pngpaste
+  } else if (os === 'win32') {
+    tries.push([
+      'powershell.exe',
+      [
+        '-NonInteractive',
+        '-NoProfile',
+        '-Command',
+        'Add-Type -AssemblyName System.Windows.Forms; $img=[System.Windows.Forms.Clipboard]::GetImage(); if($img){$ms=New-Object System.IO.MemoryStream; $img.Save($ms,[System.Drawing.Imaging.ImageFormat]::Png); [Console]::Out.Write([System.Convert]::ToBase64String($ms.ToArray()))}'
+      ]
+    ])
+  }
+  for (const [cmd, args] of tries) {
+    if (!commandExists(cmd)) continue // skip missing tools (no pointless failing spawns)
+    try {
+      const buf = await run(cmd, args)
+      if (buf.length) {
+        // powershell already returns base64 text; the others return raw PNG bytes.
+        const data = os === 'win32' ? buf.toString('utf8').trim() : buf.toString('base64')
+        if (data) return { data, mime: 'image/png' }
+      }
+    } catch {
+      // try the next candidate
+    }
+  }
+  return undefined
+}

ui-opentui/src/boundary/errors.ts

@@ -0,0 +1,29 @@
+/**
+ * Typed errors at the gateway boundary.
+ *
+ * Per spec v4 §3.4: internal errors use `Data.TaggedError`; wire/serializable
+ * errors use Schema-based tagged errors (added in Phase 1 alongside the
+ * GatewayEvent schema). Phase 0 ships the internal set the renderer/transport
+ * boundary needs.
+ *
+ * Boundary code yields these directly (`return yield* new FooError(...)`) — no
+ * throw / try-catch / Promise.catch / orDie.
+ */
+import { Data } from 'effect'
+
+/** The renderer (createCliRenderer) failed to acquire. */
+export class RendererError extends Data.TaggedError('RendererError')<{
+  readonly cause: unknown
+}> {}
+
+/** Could not resolve a usable Python interpreter for the gateway. */
+export class PythonResolutionError extends Data.TaggedError('PythonResolutionError')<{
+  readonly tried: ReadonlyArray<string>
+}> {}
+
+/** A JSON-RPC request to the gateway failed (timeout, transport down, rpc error). */
+export class GatewayError extends Data.TaggedError('GatewayError')<{
+  readonly method: string
+  readonly reason: 'timeout' | 'transport-down' | 'rpc-error'
+  readonly message: string
+}> {}

ui-opentui/src/boundary/ffiSafe.ts

@@ -0,0 +1,100 @@
+/**
+ * Node-FFI coordinate safety shim for @opentui/core 0.4.0.
+ *
+ * Root cause (live crash, ERR_INVALID_ARG_VALUE looping every frame): several
+ * OptimizedBuffer methods marshal x/y/width/height as **u32** in the FFI table
+ * (zig.ts: `bufferFillRect: ["u32","u32","u32","u32","u32","ptr"]`, same for
+ * `bufferDrawText` / `bufferSetCell*` / `bufferDrawChar`), while renderables
+ * pass RAW SCREEN COORDINATES — which go NEGATIVE inside a <scrollbox> when an
+ * element is partially scrolled above the viewport. Concretely:
+ * `LineNumberRenderable.renderSelf` does `buffer.fillRect(this.x + gutterWidth,
+ * this.y + i, …)` for diff added/removed line backgrounds, so expanding a tall
+ * `<diff showLineNumbers>` pinned to the scrollbox bottom rendered with
+ * `this.y < 0` and threw out of `CliRenderer.loop` on EVERY frame (frozen UI,
+ * console error spam) until a resize forced a fresh layout.
+ *
+ * Upstream-on-Bun this never throws: Bun's FFI silently WRAPS negatives to
+ * huge u32s and the native side bounds-checks them into a no-op. Node's
+ * experimental FFI (node:ffi) instead REJECTS the argument. Other draw entry
+ * points (`bufferDrawBox`, `bufferDrawTextBufferView`) already use i32 — which
+ * is why ordinary text/boxes scroll fine and only the diff gutter path crashed.
+ *
+ * Fix at the seam we own: clamp/skip BEFORE the FFI call.
+ *  - fillRect: clip the rect to the non-negative quadrant (the native side
+ *    already clips right/bottom against the buffer + scissor) and skip empties.
+ *  - drawText/setCell/setCellWithAlphaBlending/drawChar: skip when the origin
+ *    is negative (Bun-parity: those cells/rows are off-screen anyway).
+ *
+ * TODO(upstream): file/track an OpenTUI issue to widen these FFI params to i32
+ * (or clamp in core) — then this shim can be deleted.
+ */
+import { OptimizedBuffer, TextBufferView } from '@opentui/core'
+
+let installed = false
+
+/** Patch OptimizedBuffer's u32-coordinate methods to tolerate negative coords. Idempotent. */
+export function installFfiCoordSafety(): void {
+  if (installed) return
+  installed = true
+
+  const proto = OptimizedBuffer.prototype
+
+  // Prototype monkey-patching: extracting the original methods unbound is the
+  // point — they're re-invoked with `.call(this, …)` on the correct instance.
+  /* eslint-disable @typescript-eslint/unbound-method */
+
+  const origFillRect = proto.fillRect
+  proto.fillRect = function (this: OptimizedBuffer, x, y, width, height, bg) {
+    let x2 = Math.trunc(x)
+    let y2 = Math.trunc(y)
+    let w = Math.trunc(width)
+    let h = Math.trunc(height)
+    if (x2 < 0) {
+      w += x2
+      x2 = 0
+    }
+    if (y2 < 0) {
+      h += y2
+      y2 = 0
+    }
+    if (w <= 0 || h <= 0) return
+    origFillRect.call(this, x2, y2, w, h, bg)
+  }
+
+  const origDrawText = proto.drawText
+  proto.drawText = function (this: OptimizedBuffer, text, x, y, ...rest) {
+    if (x < 0 || y < 0) return
+    origDrawText.call(this, text, x, y, ...rest)
+  }
+
+  const origSetCell = proto.setCell
+  proto.setCell = function (this: OptimizedBuffer, x, y, ...rest) {
+    if (x < 0 || y < 0) return
+    origSetCell.call(this, x, y, ...rest)
+  }
+
+  const origSetCellAlpha = proto.setCellWithAlphaBlending
+  proto.setCellWithAlphaBlending = function (this: OptimizedBuffer, x, y, ...rest) {
+    if (x < 0 || y < 0) return
+    origSetCellAlpha.call(this, x, y, ...rest)
+  }
+
+  const origDrawChar = proto.drawChar
+  proto.drawChar = function (this: OptimizedBuffer, char, x, y, ...rest) {
+    if (x < 0 || y < 0) return
+    origDrawChar.call(this, char, x, y, ...rest)
+  }
+
+  // Same u32 marshaling on a different entry point: `textBufferViewSetViewport`
+  // takes x/y/width/height as u32, but `TextRenderable.onResize` feeds it the
+  // RAW transient layout size — observed NON-u32 (negative/NaN) mid-relayout
+  // while a shrinking list (the fuzzy picker filtering rows away) reflows. Bun
+  // wraps/coerces, node:ffi throws (`Argument 3 must be a uint32`). Coerce into
+  // the valid quadrant; a zero-sized viewport is the native side's own no-op.
+  const u32 = (v: number) => (Number.isFinite(v) ? Math.max(0, Math.trunc(v)) : 0)
+  const viewProto = TextBufferView.prototype
+  const origSetViewport = viewProto.setViewport
+  viewProto.setViewport = function (this: TextBufferView, x, y, width, height) {
+    origSetViewport.call(this, u32(x), u32(y), u32(width), u32(height))
+  }
+}

ui-opentui/src/boundary/gateway/GatewayService.ts

@@ -0,0 +1,29 @@
+/**
+ * GatewayService — the Effect-side transport boundary.
+ *
+ * Phase 0: the SHAPE only. The live layer (spawning the Python `tui_gateway`,
+ * JSON-RPC framing, Schema-decoding the wire union) lands in Phase 1
+ * (`boundary/gateway/liveGateway.ts`). For now the only implementation is
+ * `FakeGateway.layer` (entry/fakeGateway.ts), which the render/test harness uses.
+ *
+ * This is one of exactly two Effect<->Solid contact points: the Solid store
+ * subscribes via `subscribe(handler)` and the boundary pushes DECODED events in.
+ * Per spec v4 §1, the store/reducer themselves are plain Solid, never Effect.
+ */
+import { Context, type Effect } from 'effect'
+
+import type { GatewayError } from '../errors.ts'
+import type { GatewayEvent } from '../schema/GatewayEvent.ts'
+
+export interface GatewayServiceShape {
+  /** Push decoded gateway events into the Solid store. Returns an unsubscribe fn. */
+  readonly subscribe: (handler: (event: GatewayEvent) => void) => Effect.Effect<() => void>
+  /** Typed JSON-RPC request to the Python gateway. Fails with a typed GatewayError, never throws. */
+  readonly request: <A>(method: string, params: unknown) => Effect.Effect<A, GatewayError>
+  /** The active session id (for `approval.respond {session_id}`); undefined before a session exists. */
+  readonly sessionId: () => string | undefined
+}
+
+export class GatewayService extends Context.Service<GatewayService, GatewayServiceShape>()(
+  '@hermes-tui/GatewayService'
+) {}

ui-opentui/src/boundary/gateway/client.ts

@@ -0,0 +1,255 @@
+/**
+ * Low-level JSON-RPC-over-stdio client for the Python `tui_gateway` (spec v4 §4).
+ * Re-authored minimal (NOT the Ink client's 740-LOC attach-mode/buffering) but
+ * the WIRE CONTRACT is identical (verified against ui-tui/src/gatewayClient.ts +
+ * tui_gateway/server.py + entry.py + transport.py):
+ *
+ *  - spawn: `python -m tui_gateway.entry`, cwd=srcRoot, env={...process.env,
+ *    PYTHONPATH=srcRoot:…, HERMES_PYTHON_SRC_ROOT=srcRoot}, stdio piped.
+ *  - framing: newline-delimited compact JSON, BOTH directions, on ONE stdout.
+ *  - request:  {id:"r<n>", jsonrpc:"2.0", method, params} + "\n".
+ *  - response: {jsonrpc, id, result} | {jsonrpc, id, error:{code,message}} — match by id.
+ *  - event:    {jsonrpc, method:"event", params:{type, session_id?, payload?}} (NO id).
+ *  - handshake: child emits {event, params:{type:"gateway.ready", payload:{skin}}}
+ *    UNSOLICITED first; no subscribe RPC. Then client drives session.create /
+ *    session.resume / prompt.submit / *.respond.
+ *  - GOTCHA: session.resume/prompt.submit/slash.exec are LONG handlers — their
+ *    {id,result} arrives async, interleaved with events. Keep the pending map
+ *    authoritative; never assume in-order response delivery.
+ *
+ * Raw events are surfaced as `unknown` (the params object). The liveGateway
+ * layer Schema-decodes them once at the boundary (spec v4 §3.3); this client
+ * stays decode-agnostic so the transport and the schema evolve independently.
+ */
+import { spawn, type ChildProcessWithoutNullStreams } from 'node:child_process'
+
+import type { Log } from '../log.ts'
+import { resolvePython, resolveSrcRoot } from './python.ts'
+
+interface Pending {
+  resolve: (result: unknown) => void
+  reject: (error: Error) => void
+  method: string
+}
+
+export interface RawClientOptions {
+  readonly log: Log
+  /** Called with each server-pushed event's `params` object (still unknown — decoded upstream). */
+  readonly onEvent: (params: unknown) => void
+  /** Called when the child exits / errors (so the layer can reject pending + reconnect). */
+  readonly onExit?: (reason: string) => void
+}
+
+const REQUEST_TIMEOUT_MS = (() => {
+  const raw = Number.parseInt(process.env.HERMES_TUI_RPC_TIMEOUT_MS ?? '', 10)
+  return Number.isFinite(raw) && raw > 0 ? Math.max(5000, raw) : 120_000
+})()
+
+const STARTUP_TIMEOUT_MS = (() => {
+  const raw = Number.parseInt(process.env.HERMES_TUI_STARTUP_TIMEOUT_MS ?? '', 10)
+  return Number.isFinite(raw) && raw > 0 ? Math.max(2000, raw) : 20_000
+})()
+
+export class RawGatewayClient {
+  private proc: ChildProcessWithoutNullStreams | null = null
+  private pending = new Map<string, Pending>()
+  private reqId = 0
+  private stdinBuffer = ''
+  private startupTimer: ReturnType<typeof setTimeout> | undefined
+  private readonly log: Log
+  private readonly onEvent: (params: unknown) => void
+  private readonly onExit?: (reason: string) => void
+
+  constructor(options: RawClientOptions) {
+    this.log = options.log
+    this.onEvent = options.onEvent
+    if (options.onExit) this.onExit = options.onExit
+  }
+
+  /** Spawn the gateway child and begin reading frames. Idempotent. */
+  start(): void {
+    if (this.proc) return
+    const srcRoot = resolveSrcRoot()
+    const python = resolvePython(srcRoot)
+    const cwd = process.env.HERMES_CWD?.trim() || srcRoot
+    const env: Record<string, string> = { ...(process.env as Record<string, string>) }
+    env.PYTHONPATH = env.PYTHONPATH ? `${srcRoot}:${env.PYTHONPATH}` : srcRoot
+    env.HERMES_PYTHON_SRC_ROOT = srcRoot
+
+    this.log.info('gateway', 'spawning tui_gateway', { python, cwd, srcRoot })
+
+    const proc = spawn(python, ['-m', 'tui_gateway.entry'], {
+      cwd,
+      env,
+      stdio: ['pipe', 'pipe', 'pipe']
+    })
+    // Identity guard: a stale child's late exit/error must not act after a restart
+    // has already installed a new `this.proc` (else it'd null the live child).
+    // Nulling `this.proc` here makes a subsequent finish() a no-op (idempotent),
+    // covering the ENOENT case where 'error' fires and 'exit' does not.
+    const finish = (reason: string) => {
+      if (this.proc !== proc) return
+      this.log.warn('gateway', reason)
+      this.rejectAll(reason)
+      this.proc = null
+      this.onExit?.(reason)
+    }
+    proc.on('exit', (code, signal) => finish(`gateway exited (code=${code ?? 'null'} signal=${signal ?? 'null'})`))
+    proc.on('error', err => finish(`gateway spawn error: ${err instanceof Error ? err.message : String(err)}`))
+    this.proc = proc
+    this.readStdout(proc)
+    this.readStderr(proc)
+
+    // Startup-readiness watchdog: a child that hangs on import (wrong python /
+    // missing dep) never emits the unsolicited `gateway.ready` handshake, leaving
+    // a silent blank UI. Emit `gateway.start_timeout` so the store can surface a
+    // failure line + the captured stderr tail. Cleared on ready (dispatch) / stop.
+    // A recovery-respawn re-enters start(), so this re-arms per respawn — desired.
+    this.startupTimer = setTimeout(() => {
+      this.startupTimer = undefined
+      this.onEvent({
+        type: 'gateway.start_timeout',
+        payload: { message: `no gateway.ready within ${STARTUP_TIMEOUT_MS}ms` }
+      })
+    }, STARTUP_TIMEOUT_MS)
+  }
+
+  private readStdout(proc: ChildProcessWithoutNullStreams): void {
+    proc.stdout.setEncoding('utf8')
+    proc.stdout.on('data', (chunk: string) => {
+      this.stdinBuffer += chunk
+      let nl: number
+      while ((nl = this.stdinBuffer.indexOf('\n')) >= 0) {
+        const line = this.stdinBuffer.slice(0, nl)
+        this.stdinBuffer = this.stdinBuffer.slice(nl + 1)
+        if (line.trim()) this.dispatch(line)
+      }
+    })
+    proc.stdout.on('error', cause => this.log.error('gateway', 'stdout read loop failed', { cause: String(cause) }))
+  }
+
+  private readStderr(proc: ChildProcessWithoutNullStreams): void {
+    let buf = ''
+    proc.stderr.setEncoding('utf8')
+    proc.stderr.on('data', (chunk: string) => {
+      buf += chunk
+      let nl: number
+      while ((nl = buf.indexOf('\n')) >= 0) {
+        const line = buf.slice(0, nl)
+        buf = buf.slice(nl + 1)
+        if (line.trim()) {
+          this.log.debug('gateway.stderr', line)
+          // Surface as a synthetic gateway.stderr event (matches Ink).
+          this.onEvent({ type: 'gateway.stderr', payload: { line } })
+        }
+      }
+    })
+    // stderr pipe closing on exit is expected; ignore errors.
+    proc.stderr.on('error', () => {})
+  }
+
+  private dispatch(line: string): void {
+    let msg: unknown
+    try {
+      msg = JSON.parse(line)
+    } catch {
+      this.log.warn('gateway', 'unparseable frame', { preview: line.slice(0, 120) })
+      this.onEvent({ type: 'gateway.protocol_error', payload: { preview: line.slice(0, 120) } })
+      return
+    }
+    if (!msg || typeof msg !== 'object') return
+    const frame = msg as { id?: unknown; method?: unknown; params?: unknown; result?: unknown; error?: unknown }
+
+    // Response: has an id matching a pending request.
+    const pending = typeof frame.id === 'string' ? this.pending.get(frame.id) : undefined
+    if (typeof frame.id === 'string' && pending) {
+      const p = pending
+      this.pending.delete(frame.id)
+      if (frame.error) {
+        const err = frame.error as { code?: number; message?: string }
+        p.reject(new Error(err.message ?? `rpc error (${err.code ?? '?'})`))
+      } else {
+        p.resolve(frame.result)
+      }
+      return
+    }
+
+    // Event push: method === "event", no id. Surface params (decoded upstream).
+    if (frame.method === 'event' && frame.params && typeof frame.params === 'object') {
+      // Handshake arrived: cancel the startup-readiness watchdog. Narrow without
+      // `as` via `'type' in obj` + property access (the params record is loose).
+      if ('type' in frame.params && frame.params.type === 'gateway.ready') {
+        if (this.startupTimer) clearTimeout(this.startupTimer)
+        this.startupTimer = undefined
+      }
+      this.onEvent(frame.params)
+      return
+    }
+
+    this.log.warn('gateway', 'unroutable frame', { preview: line.slice(0, 120) })
+  }
+
+  /** Send a JSON-RPC request; resolves with `result` (long handlers reply async). */
+  request<A = unknown>(method: string, params: unknown): Promise<A> {
+    // Do NOT auto-start here: during the recovery backoff window `this.proc` is
+    // null, and a respawn here would BYPASS the backoff (the first spawn always
+    // comes from subscribe() → client.start()). A null proc rejects below.
+    const proc = this.proc
+    const stdin = proc?.stdin
+    if (!stdin) return Promise.reject(new Error('gateway not running'))
+
+    const id = `r${++this.reqId}`
+    const frame = JSON.stringify({ id, jsonrpc: '2.0', method, params: params ?? {} }) + '\n'
+
+    return new Promise<A>((resolve, reject) => {
+      const timer = setTimeout(() => {
+        if (this.pending.delete(id)) reject(new Error(`timeout: ${method}`))
+      }, REQUEST_TIMEOUT_MS)
+
+      this.pending.set(id, {
+        method,
+        resolve: result => {
+          clearTimeout(timer)
+          resolve(result as A)
+        },
+        reject: error => {
+          clearTimeout(timer)
+          reject(error)
+        }
+      })
+
+      try {
+        // Newline-delimited JSON to the child's stdin. Fire-and-forget: the write
+        // returns a backpressure boolean we intentionally ignore (frames are tiny
+        // and ordered; Node flushes the pipe itself).
+        stdin.write(frame)
+      } catch (cause) {
+        this.pending.delete(id)
+        clearTimeout(timer)
+        reject(cause instanceof Error ? cause : new Error(String(cause)))
+      }
+    })
+  }
+
+  private rejectAll(reason: string): void {
+    for (const p of this.pending.values()) p.reject(new Error(reason))
+    this.pending.clear()
+  }
+
+  /** Close stdin (EOF → child exits) and stop. */
+  stop(): void {
+    if (this.startupTimer) clearTimeout(this.startupTimer)
+    this.startupTimer = undefined
+    this.rejectAll('gateway stopping')
+    const stdin = this.proc?.stdin
+    if (stdin) {
+      try {
+        // Close stdin → child sees EOF and exits.
+        stdin.end()
+      } catch {
+        // already gone
+      }
+    }
+    this.proc = null
+  }
+}

ui-opentui/src/boundary/gateway/liveGateway.ts

@@ -0,0 +1,175 @@
+/**
+ * liveGateway — the GatewayService layer backed by the real Python `tui_gateway`
+ * (spec v4 §2/§3.2). Adapts RawGatewayClient to GatewayServiceShape:
+ *   - decodes each raw event ONCE with the GatewayEvent Schema
+ *     (decodeUnknownOption → unrecognized/malformed events skipped, never crash),
+ *   - coalesces decoded events on a 16ms debounce flushed inside Solid `batch()`
+ *     so a burst of deltas is ONE repaint (opencode sdk.tsx:54-80),
+ *   - tracks the session id (set from session.create/resume result) for
+ *     approval.respond {session_id},
+ *   - maps request failures to a typed GatewayError (never throws).
+ *
+ * The 16ms batch + `batch()` call is the boundary handing decoded events to
+ * Solid — one of the two approved Effect<->Solid contact points (spec v4 §1).
+ */
+import { Effect, Layer, Option, Schema } from 'effect'
+import { batch } from 'solid-js'
+
+import { backoffMs, planGatewayRecovery } from '../../logic/gatewayRecovery.ts'
+import { GatewayError } from '../errors.ts'
+import { getLog } from '../log.ts'
+import { GatewayEventSchema, type GatewayEvent } from '../schema/GatewayEvent.ts'
+import { GatewayService, type GatewayServiceShape } from './GatewayService.ts'
+import { RawGatewayClient } from './client.ts'
+
+const COALESCE_MS = 16
+
+const decodeEvent = Schema.decodeUnknownOption(GatewayEventSchema)
+
+function makeLiveGateway(): { service: GatewayServiceShape; stop: () => void } {
+  const log = getLog()
+  const handlers = new Set<(event: GatewayEvent) => void>()
+  let sessionId: string | undefined
+
+  // Auto-heal recovery state (driver below). `recoverSid` is the resume target
+  // carried across a respawn that died before gateway.ready; `recoveryAttempts`
+  // is the sliding crash-loop budget window; `restartTimer` is the pending
+  // backoff respawn (cleared on teardown so it can't fire post-stop).
+  let recoverSid: string | undefined
+  let recoveryAttempts: number[] = []
+  let restartTimer: ReturnType<typeof setTimeout> | undefined
+
+  // 16ms event coalescing → one batched repaint (opencode sdk.tsx model).
+  let queue: GatewayEvent[] = []
+  let timer: ReturnType<typeof setTimeout> | undefined
+  let last = 0
+
+  const flush = () => {
+    timer = undefined
+    if (queue.length === 0) return
+    const events = queue
+    queue = []
+    last = Date.now()
+    batch(() => {
+      for (const event of events) {
+        for (const handler of handlers) handler(event)
+      }
+    })
+  }
+
+  const enqueue = (event: GatewayEvent) => {
+    queue.push(event)
+    if (timer) return
+    // If we flushed recently (<16ms ago) batch with near-future events; else flush now.
+    if (Date.now() - last < COALESCE_MS) {
+      timer = setTimeout(flush, COALESCE_MS)
+    } else {
+      flush()
+    }
+  }
+
+  const onRawEvent = (params: unknown) => {
+    const decoded = decodeEvent(params)
+    if (Option.isNone(decoded)) {
+      const t = (params as { type?: unknown } | null)?.type
+      log.debug('gateway', 'skipped undecodable event', { type: typeof t === 'string' ? t : '(none)' })
+      return
+    }
+    enqueue(decoded.value)
+  }
+
+  // Recovery driver: on a child exit, clear the frozen spinner (via the store's
+  // gateway.exited case), then — under the crash-loop budget — respawn the child
+  // on exponential backoff. The post-respawn gateway.ready triggers the re-resume
+  // (driven from entry's subscribe callback). Hoisted so it can be passed to
+  // `new RawGatewayClient` below while itself referencing the `client` const —
+  // `client` is assigned by the time onExit ever fires at runtime.
+  function onExit(reason: string): void {
+    log.warn('gateway', 'transport exited', { reason })
+    // Clears the frozen spinner + shows status (store handles gateway.exited).
+    enqueue({ type: 'gateway.exited', payload: { reason } })
+    const plan = planGatewayRecovery(sessionId ?? null, recoverSid ?? null, recoveryAttempts, Date.now())
+    recoveryAttempts = plan.attempts
+    if (!plan.recover || plan.sid === null) {
+      enqueue({ type: 'error', payload: { message: 'gateway exited repeatedly — type /resume to retry' } })
+      return
+    }
+    recoverSid = plan.sid
+    const attempt = recoveryAttempts.length
+    const delay = backoffMs(attempt)
+    enqueue({ type: 'gateway.recovering', payload: { attempt, delay_ms: delay } })
+    if (restartTimer) clearTimeout(restartTimer)
+    restartTimer = setTimeout(() => {
+      restartTimer = undefined
+      client.start()
+    }, delay)
+  }
+
+  const client = new RawGatewayClient({
+    log,
+    onEvent: onRawEvent,
+    onExit
+  })
+
+  const service: GatewayServiceShape = {
+    subscribe: handler =>
+      Effect.sync(() => {
+        handlers.add(handler)
+        // Lazily spawn on first subscription so the child + its gateway.ready land.
+        client.start()
+        return () => {
+          handlers.delete(handler)
+        }
+      }),
+
+    request: <A>(method: string, params: unknown) =>
+      Effect.tryPromise({
+        try: () => client.request<A>(method, params),
+        catch: cause => {
+          const message = cause instanceof Error ? cause.message : String(cause)
+          const reason = message.startsWith('timeout:')
+            ? ('timeout' as const)
+            : message.includes('not running') || message.includes('stopping')
+              ? ('transport-down' as const)
+              : ('rpc-error' as const)
+          return new GatewayError({ method, reason, message })
+        }
+      }).pipe(
+        // Capture session id from create/resume results so approval.respond works.
+        Effect.tap(result =>
+          Effect.sync(() => {
+            if ((method === 'session.create' || method === 'session.resume') && result && typeof result === 'object') {
+              const sid = (result as { session_id?: unknown }).session_id
+              if (typeof sid === 'string') sessionId = sid
+            }
+          })
+        )
+      ),
+
+    sessionId: () => sessionId
+  }
+
+  // Clear a pending coalesce timer on teardown so a queued flush() can't fire
+  // batch()/handlers into a torn-down store after the layer scope releases.
+  const stop = () => {
+    if (timer) clearTimeout(timer)
+    timer = undefined
+    // Also kill any pending backoff respawn so it can't fire after teardown.
+    if (restartTimer) clearTimeout(restartTimer)
+    restartTimer = undefined
+    client.stop()
+  }
+  return { service, stop }
+}
+
+/**
+ * The live GatewayService layer (spawns + talks to the real Python tui_gateway).
+ * Scoped so the child process is stopped (stdin EOF → exit) on scope teardown —
+ * no orphaned gateway children when the renderer is destroyed.
+ */
+export const liveGatewayLayer: Layer.Layer<GatewayService> = Layer.effect(
+  GatewayService,
+  Effect.acquireRelease(Effect.sync(makeLiveGateway), ({ stop }) => Effect.sync(stop)).pipe(
+    Effect.map(({ service }) => service)
+  )
+)

ui-opentui/src/boundary/gateway/python.ts

@@ -0,0 +1,49 @@
+/**
+ * Python resolution for spawning the `tui_gateway` — mirrors Ink's
+ * `resolvePython` (ui-tui/src/gatewayClient.ts:45-64) EXACTLY so behavior is
+ * identical across engines (spec v4 §4). NEVER "probe any python".
+ *
+ * Order: HERMES_PYTHON / PYTHON env → $VIRTUAL_ENV (bin/python or
+ * Scripts/python.exe) → <root>/.venv → <root>/venv → bare `python3` (`python`
+ * on win32) on PATH. The source root is HERMES_PYTHON_SRC_ROOT (the launcher
+ * sets it) so the child resolves modules against the right checkout.
+ */
+import { existsSync } from 'node:fs'
+import { dirname, resolve } from 'node:path'
+
+export function resolvePython(root: string): string {
+  const configured = process.env.HERMES_PYTHON?.trim() || process.env.PYTHON?.trim()
+  if (configured) return configured
+
+  const venv = process.env.VIRTUAL_ENV?.trim()
+
+  const hit = [
+    venv && resolve(venv, 'bin/python'),
+    venv && resolve(venv, 'Scripts/python.exe'),
+    resolve(root, '.venv/bin/python'),
+    resolve(root, '.venv/bin/python3'),
+    resolve(root, 'venv/bin/python'),
+    resolve(root, 'venv/bin/python3')
+  ].find(p => p && existsSync(p))
+
+  return hit || (process.platform === 'win32' ? 'python' : 'python3')
+}
+
+/** The Hermes checkout root used as PYTHONPATH / HERMES_PYTHON_SRC_ROOT for the child. */
+export function resolveSrcRoot(): string {
+  const configured = process.env.HERMES_PYTHON_SRC_ROOT?.trim()
+  if (configured) return configured
+  // Fallback (no launcher env): walk up from this module to the Hermes checkout
+  // root — the dir holding the `hermes_cli` package / `pyproject.toml`. Bundle-
+  // agnostic, so it works whether running the source tree (.../src/boundary/gateway)
+  // or the built `dist/main.js`. (Under the real launcher this never runs — the
+  // launcher always sets HERMES_PYTHON_SRC_ROOT.)
+  let dir = import.meta.dirname
+  for (let i = 0; i < 8; i++) {
+    if (existsSync(resolve(dir, 'hermes_cli')) || existsSync(resolve(dir, 'pyproject.toml'))) return dir
+    const parent = dirname(dir)
+    if (parent === dir) break
+    dir = parent
+  }
+  return resolve(import.meta.dirname, '../../../../')
+}

ui-opentui/src/boundary/log.ts

@@ -0,0 +1,248 @@
+/**
+ * Log — TUI diagnostics sink (glitch: "v. important … hook into logs to figure
+ * out TUI state"). Design mirrors opencode's `util/log.ts` (levels + priority
+ * filter, scoped/child loggers, a `.time()` span helper) but adds a dual sink:
+ *
+ *   1. an in-memory RING BUFFER (queryable at runtime — a `/logs` overlay or a
+ *      test asserting TUI state transitions can read it live), AND
+ *   2. an append-only NDJSON FILE (default `~/.hermes/logs/opentui-v2.log`,
+ *      override via HERMES_TUI_LOG_FILE) so a live session is `tail -f`-able.
+ *
+ * The ring buffer is the key advantage over opencode's file-only logger: it lets
+ * us inspect engine state from inside the running TUI without leaving it.
+ *
+ * CRITICAL: OpenTUI HIJACKS `console.*` and stdout (opentui skill / gotcha) —
+ * logging to the terminal corrupts the rendered frame. So this NEVER touches
+ * console/stdout/stderr; file + ring only. It's the single approved logging path
+ * for the whole engine. Level filter via HERMES_TUI_LOG_LEVEL (default INFO).
+ */
+import { appendFileSync, mkdirSync, renameSync, statSync, unlinkSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { dirname, join } from 'node:path'
+
+import { Schema } from 'effect'
+
+// LogLevel is modeled schema-first (the schema-inferred-types idiom, mirroring
+// `boundary/schema/GatewayEvent.ts`): declare the literal union once and INFER
+// the TS type from it, so the two can never drift.
+export const LogLevelSchema = Schema.Literals(['debug', 'info', 'warn', 'error'])
+export type LogLevel = typeof LogLevelSchema.Type
+
+const PRIORITY: Record<LogLevel, number> = { debug: 0, info: 1, warn: 2, error: 3 }
+
+/**
+ * Serialize a value to JSON that NEVER throws. A caller-supplied `data` can hold
+ * a circular reference or a BigInt — plain `JSON.stringify` throws on both, which
+ * (in the file-write `catch` below) would flip `fileBroken` and kill ALL file
+ * logging for the session. Instead we degrade a bad payload to a placeholder:
+ *   - circular refs (tracked via a per-call `WeakSet` of seen objects) → '[Circular]'
+ *   - BigInt → `\`${n}n\`` (JSON has no bigint; keep it readable + reversible-ish)
+ * and wrap the whole thing so any other throw (e.g. a hostile `toJSON`) falls back
+ * to `String(value)`, then to '[unserializable]' if even that throws.
+ */
+export function safeStringify(value: unknown): string {
+  try {
+    const seen = new WeakSet<object>()
+    return JSON.stringify(value, (_key, val: unknown) => {
+      if (typeof val === 'bigint') return `${val}n`
+      if (typeof val === 'object' && val !== null) {
+        if (seen.has(val)) return '[Circular]'
+        seen.add(val)
+      }
+      return val
+    })
+  } catch {
+    try {
+      return String(value)
+    } catch {
+      return '[unserializable]'
+    }
+  }
+}
+
+export interface LogEntry {
+  readonly t: number // epoch ms
+  readonly level: LogLevel
+  readonly scope: string
+  readonly msg: string
+  readonly data?: unknown
+}
+
+const RING_LIMIT = 2000
+
+// Size-based rotation for the append-only NDJSON file (mirrors opencode's
+// keep-N model, but size- rather than time-keyed since we write one growing
+// file). When the live file crosses LOG_MAX_BYTES we shift
+// `.log` → `.log.1` → … → `.log.${LOG_KEEP}` (dropping the oldest) and resume on
+// a fresh empty `.log`. Rotation is best-effort: any failure leaves us writing
+// to the existing file (logging must never crash the engine).
+const LOG_MAX_BYTES = 5 * 1024 * 1024
+const LOG_KEEP = 5
+
+function defaultLogFile(): string {
+  const explicit = process.env.HERMES_TUI_LOG_FILE?.trim()
+  if (explicit) return explicit
+  return join(homedir(), '.hermes', 'logs', 'opentui-v2.log')
+}
+
+function defaultLevel(): LogLevel {
+  const raw = process.env.HERMES_TUI_LOG_LEVEL?.trim().toLowerCase()
+  return raw === 'debug' || raw === 'info' || raw === 'warn' || raw === 'error' ? raw : 'info'
+}
+
+/** A timing span — call `.stop()` (or `using` it) to log completion + duration. */
+export interface TimeSpan {
+  stop: () => void
+  [Symbol.dispose]: () => void
+}
+
+export class Log {
+  private ring: LogEntry[] = []
+  private file: string | null
+  private fileBroken = false
+  private minPriority: number
+  // Bytes in the live log file. Seeded from statSync on open (counter approach —
+  // we avoid a statSync on EVERY write); incremented by each line's byte length
+  // and reset to 0 after a rotation. Rotation triggers when this would cross
+  // LOG_MAX_BYTES, so the live file stays bounded without per-write fs stats.
+  private fileBytes = 0
+
+  constructor(file: string | null = defaultLogFile(), level: LogLevel = defaultLevel()) {
+    this.file = file
+    this.minPriority = PRIORITY[level]
+    if (this.file) {
+      try {
+        mkdirSync(dirname(this.file), { recursive: true })
+      } catch {
+        this.fileBroken = true
+      }
+      try {
+        this.fileBytes = statSync(this.file).size
+      } catch {
+        this.fileBytes = 0 // no existing file (or unreadable) → start the counter at 0
+      }
+    }
+  }
+
+  setLevel(level: LogLevel): void {
+    this.minPriority = PRIORITY[level]
+  }
+
+  /**
+   * Best-effort size-based rotation: `.log.${LOG_KEEP}` is dropped, every other
+   * `.log.N` shifts up, the live `.log` becomes `.log.1`, and the counter resets
+   * so writing continues on a fresh file. Any fs failure is swallowed and we keep
+   * writing to the existing file — rotation must never crash logging.
+   */
+  private rotate(file: string): void {
+    try {
+      try {
+        unlinkSync(`${file}.${LOG_KEEP}`)
+      } catch {
+        // oldest slot may not exist yet — fine
+      }
+      for (let i = LOG_KEEP - 1; i >= 1; i--) {
+        try {
+          renameSync(`${file}.${i}`, `${file}.${i + 1}`)
+        } catch {
+          // that slot may not exist yet — fine
+        }
+      }
+      renameSync(file, `${file}.1`)
+      this.fileBytes = 0
+    } catch {
+      // rotation failed (e.g. live file vanished) — leave the counter alone and
+      // keep appending to the existing path; better an oversized log than none.
+    }
+  }
+
+  private write(level: LogLevel, scope: string, msg: string, data?: unknown): void {
+    if (PRIORITY[level] < this.minPriority) return
+    const entry: LogEntry =
+      data === undefined ? { t: Date.now(), level, scope, msg } : { t: Date.now(), level, scope, msg, data }
+    this.ring.push(entry)
+    if (this.ring.length > RING_LIMIT) this.ring.shift()
+
+    if (this.file && !this.fileBroken) {
+      try {
+        const line = safeStringify(entry) + '\n'
+        if (this.fileBytes > 0 && this.fileBytes + Buffer.byteLength(line) > LOG_MAX_BYTES) this.rotate(this.file)
+        appendFileSync(this.file, line)
+        this.fileBytes += Buffer.byteLength(line)
+      } catch {
+        this.fileBroken = true // stop hammering a broken path; the ring keeps working
+      }
+    }
+  }
+
+  debug(scope: string, msg: string, data?: unknown): void {
+    this.write('debug', scope, msg, data)
+  }
+  info(scope: string, msg: string, data?: unknown): void {
+    this.write('info', scope, msg, data)
+  }
+  warn(scope: string, msg: string, data?: unknown): void {
+    this.write('warn', scope, msg, data)
+  }
+  error(scope: string, msg: string, data?: unknown): void {
+    this.write('error', scope, msg, data)
+  }
+
+  /** A logger bound to a fixed scope (opencode's tagged-logger ergonomics). */
+  child(scope: string): ScopedLog {
+    return new ScopedLog(this, scope)
+  }
+
+  /** Time an operation: logs `<msg> started` now and `<msg> completed` + duration on stop. */
+  time(scope: string, msg: string, data?: Record<string, unknown>): TimeSpan {
+    const started = Date.now()
+    this.info(scope, `${msg} started`, data)
+    const stop = () => this.info(scope, `${msg} completed`, { ...data, duration_ms: Date.now() - started })
+    return { stop, [Symbol.dispose]: stop }
+  }
+
+  /** Snapshot of the in-memory ring (newest last). For a `/logs` overlay or tests. */
+  tail(n = RING_LIMIT): LogEntry[] {
+    return n >= this.ring.length ? [...this.ring] : this.ring.slice(this.ring.length - n)
+  }
+
+  /** Where the file log is written (for surfacing in the UI / `/logs`). */
+  get filePath(): string | null {
+    return this.fileBroken ? null : this.file
+  }
+
+  clear(): void {
+    this.ring = []
+  }
+}
+
+/** A logger with a fixed scope — forwards to the parent Log. */
+export class ScopedLog {
+  constructor(
+    private readonly parent: Log,
+    private readonly scope: string
+  ) {}
+  debug(msg: string, data?: unknown): void {
+    this.parent.debug(this.scope, msg, data)
+  }
+  info(msg: string, data?: unknown): void {
+    this.parent.info(this.scope, msg, data)
+  }
+  warn(msg: string, data?: unknown): void {
+    this.parent.warn(this.scope, msg, data)
+  }
+  error(msg: string, data?: unknown): void {
+    this.parent.error(this.scope, msg, data)
+  }
+  time(msg: string, data?: Record<string, unknown>): TimeSpan {
+    return this.parent.time(this.scope, msg, data)
+  }
+}
+
+let _singleton: Log | null = null
+
+/** Module-singleton logger for the live engine. Tests construct their own `new Log(null)`. */
+export function getLog(): Log {
+  _singleton ??= new Log()
+  return _singleton
+}

ui-opentui/src/boundary/memlog.ts

@@ -0,0 +1,91 @@
+/**
+ * memlog — in-process 1Hz memory self-sampling to NDJSON.
+ *
+ * The fleet-monitoring answer to "attach live-attach.sh to all 5–10 of my
+ * sessions": instead of an external watcher chasing pids, every TUI session
+ * logs its OWN samples when enabled, keyed by pid + boot time, into
+ * `~/.hermes/logs/memwatch/`. Aggregate across sessions with
+ * the tui-bench repo's `memwatch-report.mjs` (github.com/NousResearch/tui-bench).
+ *
+ * Gating (docs/opentui-env-flags.md): `HERMES_TUI_MEMLOG` — defaults to the
+ * `HERMES_TUI_DIAGNOSTICS` master switch, individually overridable either way.
+ * One `export HERMES_TUI_DIAGNOSTICS=1` in a dev's shell rc therefore covers
+ * every session they ever start; regular users write nothing.
+ *
+ * Cost when on: one `process.memoryUsage()` + one short append per second
+ * (~60 bytes/s, ~5MB/day across ten busy sessions). The interval is unref'd —
+ * it never keeps the process alive. Every failure path disables the logger
+ * silently (diagnostics must never break the TUI). Retention: files older
+ * than 14 days are pruned at start, best-effort.
+ *
+ * Sample shape (one JSON object per line):
+ *   { t, rss_kb, heap_used_kb, external_kb, mounted, peak_mounted }
+ * `mounted`/`peak_mounted` come from the windowing DEV counters
+ * (logic/window.ts) — they update whenever windowing is active, independent
+ * of the WINDOW_STATS exposure flag.
+ */
+import { appendFileSync, mkdirSync, readdirSync, statSync, unlinkSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join } from 'node:path'
+
+import { diagnosticsEnabled, envFlag } from '../logic/env.ts'
+import { windowRowStats } from '../logic/window.ts'
+
+const RETENTION_DAYS = 14
+const SAMPLE_MS = 1000
+
+function memwatchDir(): string {
+  const home = process.env.HERMES_HOME?.trim()
+  const base = home && home.length > 0 ? home : join(homedir(), '.hermes')
+  return join(base, 'logs', 'memwatch')
+}
+
+function pruneOld(dir: string): void {
+  const cutoff = Date.now() - RETENTION_DAYS * 24 * 3600 * 1000
+  try {
+    for (const name of readdirSync(dir)) {
+      if (!name.endsWith('.jsonl')) continue
+      const p = join(dir, name)
+      try {
+        if (statSync(p).mtimeMs < cutoff) unlinkSync(p)
+      } catch {
+        /* best-effort */
+      }
+    }
+  } catch {
+    /* best-effort */
+  }
+}
+
+/** Start the self-sampler (no-op unless enabled). Returns a stop function. */
+export function startMemlog(): () => void {
+  if (!envFlag(process.env.HERMES_TUI_MEMLOG, diagnosticsEnabled())) return () => {}
+  try {
+    const dir = memwatchDir()
+    mkdirSync(dir, { recursive: true })
+    pruneOld(dir)
+    const boot = new Date().toISOString().replace(/[:.]/g, '').slice(0, 15)
+    const file = join(dir, `${boot}-${process.pid}.jsonl`)
+    const timer = setInterval(() => {
+      try {
+        const m = process.memoryUsage()
+        const w = windowRowStats()
+        const line = JSON.stringify({
+          t: Math.floor(Date.now() / 1000),
+          rss_kb: Math.floor(m.rss / 1024),
+          heap_used_kb: Math.floor(m.heapUsed / 1024),
+          external_kb: Math.floor(m.external / 1024),
+          mounted: w.mounted,
+          peak_mounted: w.peakMounted
+        })
+        appendFileSync(file, line + '\n')
+      } catch {
+        clearInterval(timer) // a failing diagnostic must not retry forever
+      }
+    }, SAMPLE_MS)
+    timer.unref?.()
+    return () => clearInterval(timer)
+  } catch {
+    return () => {}
+  }
+}

ui-opentui/src/boundary/memoryMonitor.ts

@@ -0,0 +1,46 @@
+/**
+ * memoryMonitor — the early-warning BOUNDARY (touches node:process; the pure
+ * threshold/growth logic lives in logic/memoryMonitor.ts).
+ *
+ * Ports the high-value #34095 silent-death early-warning from Ink
+ * (`ui-tui/src/lib/memoryMonitor.ts`) to the OpenTUI engine, and ONLY that:
+ *   - NO auto heap-snapshot capture (the #41948 disk-fill bug class is not
+ *     re-imported — the always-on memlog NDJSON trace is the diagnosis path,
+ *     and its rss-vs-heap divergence is the better diagnostic for the native
+ *     RSS-leak class a V8 snapshot captures poorly).
+ *   - NO Ink cache eviction (Solid disposes out-of-window rows; windowing +
+ *     proactiveGc already cover memory pressure).
+ *
+ * It polls `process.memoryUsage()` on a 10s unref'd interval and, when the
+ * pure detector fires, surfaces a single transcript system line so the user
+ * SEES "memory climbing fast" before Node OOMs under the exit threshold. This
+ * is ON by default (unlike memlog/heapdump): it's a user-facing safety
+ * heads-up, not a diagnostic dump, and costs one memoryUsage() read per 10s
+ * with zero disk. Every failure path disables silently (a diagnostic must
+ * never break the TUI — the one place the "errors propagate" rule is
+ * intentionally inverted, matching memlog/proactiveGc).
+ */
+import { createWarnState, evaluateWarn, warnLine } from '../logic/memoryMonitor.ts'
+
+/** Sample cadence — matches Ink's monitor (10s, unref'd). */
+const SAMPLE_MS = 10_000
+
+/**
+ * Start the early-warning watcher. `emitWarn` receives the ready-to-show system
+ * line on the (one-shot) tick growth looks abnormal. Returns a stop function.
+ * The interval is unref'd so it never keeps the process alive.
+ */
+export function startMemoryMonitor(emitWarn: (line: string) => void): () => void {
+  const state = createWarnState()
+  const timer = setInterval(() => {
+    try {
+      const { heapUsed, rss } = process.memoryUsage()
+      const { fire, growthBytes } = evaluateWarn(state, heapUsed)
+      if (fire) emitWarn(warnLine(heapUsed, rss, growthBytes))
+    } catch {
+      clearInterval(timer) // a failing diagnostic must not retry forever
+    }
+  }, SAMPLE_MS)
+  timer.unref?.()
+  return () => clearInterval(timer)
+}

ui-opentui/src/boundary/multiClickSelect.ts

@@ -0,0 +1,130 @@
+/**
+ * Multi-click selection — double-click selects the word, triple-click the
+ * line, drag after either extends by word/line with the clicked span held
+ * (boundary shim in the ffiSafe.ts / nativeHandles.ts mold).
+ *
+ * Why a shim: @opentui/core's renderer knows only press-drag character
+ * selection — `processSingleMouseEvent` calls `startSelection(renderable,x,y)`
+ * on a fresh left press and `updateSelection(renderable,x,y)` per drag step,
+ * with no click-count concept. Wrapping those two INSTANCE methods is the
+ * narrowest seam that adds multi-click without forking core: the press wrapper
+ * counts clicks (Ink's 500ms / 1-cell chain) and, on a multi-click, seeds the
+ * selection with the word/line span instead of a point; the drag wrapper snaps
+ * the focus to word/line bounds and flips the selection anchor to whichever
+ * end of the held span faces away from the pointer.
+ *
+ * Word/line bounds come from the presented frame (`currentRenderBuffer`'s
+ * char grid — the same buffer `captureCharFrame` reads in tests), so what
+ * highlights is exactly the run of characters the user sees. All wrapped paths
+ * degrade to core's plain character selection when anything is off (no
+ * buffer, destroyed renderer, out-of-bounds click) — selection must never
+ * throw out of the mouse pipeline.
+ */
+import type { CliRenderer } from '@opentui/core'
+
+import type { AnchorSpan, Point, ScreenText } from '../logic/multiClick.ts'
+import { comparePoints, createClickCounter, extendedSelection, lineSpanAt, wordSpanAt } from '../logic/multiClick.ts'
+
+/** The renderable surface the shim needs (anchor tracking reads live x/y). */
+interface AnchorRenderable {
+  readonly x: number
+  readonly y: number
+}
+
+/** The private renderer surface the shim wraps (runtime-verified shapes). */
+interface RendererSeam {
+  startSelection(renderable: AnchorRenderable, x: number, y: number): void
+  updateSelection(
+    renderable: AnchorRenderable | undefined,
+    x: number,
+    y: number,
+    options?: { finishDragging?: boolean }
+  ): void
+  currentRenderBuffer: {
+    width: number
+    height: number
+    buffers: { char: Uint32Array }
+  }
+}
+
+/** Adapt the presented frame to the pure logic's ScreenText; null when the
+ *  buffer is unreadable (mid-teardown/resize) → degrade to char selection. */
+function presentedFrame(seam: RendererSeam): ScreenText | null {
+  try {
+    const buffer = seam.currentRenderBuffer
+    const chars = buffer.buffers.char
+    const width = buffer.width
+    if (width <= 0 || buffer.height <= 0) return null
+    return {
+      width,
+      height: buffer.height,
+      codepointAt: (x, y) => chars[y * width + x] ?? 0
+    }
+  } catch {
+    return null
+  }
+}
+
+/**
+ * Native selection semantics (probed empirically, scratch test 2026-06-11):
+ * per-renderable native selection keeps the anchor from the initial
+ * `setLocalSelection` — the anchor args of later `updateLocalSelection` calls
+ * are IGNORED, so moving the anchor requires restarting the selection. And the
+ * selection is caret-style at the focus end: a forward selection covers cells
+ * `[anchor, focus)` (focus cell excluded) while a backward one covers
+ * `[focus, anchor]` (both included). Inclusive cell spans therefore translate
+ * to: forward focus = `hi + 1`, backward focus = `lo` exactly.
+ */
+function forwardFocusX(anchor: Point, focus: Point): number {
+  return comparePoints(focus, anchor) >= 0 ? focus.x + 1 : focus.x
+}
+
+/** Install the multi-click wrappers on a live renderer instance. */
+export function installMultiClickSelection(renderer: CliRenderer): void {
+  const seam = renderer as unknown as RendererSeam
+  const nextClickCount = createClickCounter()
+
+  // The held span while a multi-click selection is live: cleared by the next
+  // single click (which starts a plain char selection). `anchor` mirrors the
+  // selection's current anchor end so drag steps only rebind it on a flip.
+  let held: { span: AnchorSpan; renderable: AnchorRenderable; anchor: Point } | null = null
+
+  const coreStart = seam.startSelection.bind(renderer)
+  const coreUpdate = seam.updateSelection.bind(renderer)
+
+  seam.startSelection = (renderable, x, y) => {
+    held = null
+    const clicks = nextClickCount(x, y, Date.now())
+    const screen = clicks >= 2 ? presentedFrame(seam) : null
+    const span = screen ? (clicks === 2 ? wordSpanAt(screen, x, y) : lineSpanAt(screen, y)) : null
+    if (!span) {
+      coreStart(renderable, x, y)
+      return
+    }
+    // Seed anchor at the span start, focus past its end (forward caret) — one
+    // start+update pair, exactly the calls a real press-then-drag would make.
+    coreStart(renderable, span.lo.x, span.lo.y)
+    coreUpdate(renderable, span.hi.x + 1, span.hi.y)
+    held = {
+      span: { ...span, kind: clicks === 2 ? 'word' : 'line' },
+      renderable,
+      anchor: span.lo
+    }
+  }
+
+  seam.updateSelection = (renderable, x, y, options) => {
+    const screen = held ? presentedFrame(seam) : null
+    if (!held || !screen) {
+      coreUpdate(renderable, x, y, options)
+      return
+    }
+    const { anchor, focus } = extendedSelection(held.span, screen, x, y)
+    if (anchor.x !== held.anchor.x || anchor.y !== held.anchor.y) {
+      // The anchor end flipped across the held span — native selection anchors
+      // are fixed at set time (see forwardFocusX note), so restart it there.
+      coreStart(held.renderable, anchor.x, anchor.y)
+      held = { ...held, anchor }
+    }
+    coreUpdate(renderable, forwardFocusX(anchor, focus), focus.y, options)
+  }
+}

ui-opentui/src/boundary/nativeHandles.ts

@@ -0,0 +1,110 @@
+/**
+ * Native handle-table exhaustion safety for @opentui/core 0.4.0 — sibling of
+ * the ffiSafe.ts coordinate shim (same class of fix: harden OUR side of the
+ * Node-FFI seam, TODO(upstream) to delete).
+ *
+ * Root cause (bench crash: every otui mem3000 cell died at ≈3000 lumpy fixture
+ * messages, exit 7, ~880MB RSS — far below the 2GB cgroup cap): the native
+ * core indexes EVERY object — TextBuffer, TextBufferView, SyntaxStyle,
+ * OptimizedBuffer, … — through ONE global handle registry with 16-bit slot
+ * indices (core `src/zig/handles.zig`: `INDEX_BITS = 16` → `MAX_SLOTS = 65535`,
+ * slot 0 reserved). Measured on this install: exactly 65,534 live handles, the
+ * 65,535th `createSyntaxStyle()` fails; `destroy()` does recycle slots, so
+ * exhaustion means LIVE objects.
+ *
+ * Every `TextBufferRenderable` burns THREE slots at construction
+ * (`TextBufferRenderable.ts:77-80`: `TextBuffer.create()` +
+ * `TextBufferView.create()` + `SyntaxStyle.create()`). The mount-everything
+ * transcript hits the wall at ≈1,400 store rows (≈21.8k text renderables ×3 ≈
+ * 65.5k handles): the next mount throws `Failed to create SyntaxStyle`
+ * (zig.ts:4554) out of a Solid mount effect → uncaught → the renderer's OWN
+ * `uncaughtException` handler (renderer.ts `handleError`) calls
+ * `console.show()`, which allocates the console-overlay `OptimizedBuffer` —
+ * needing ANOTHER slot — so the handler itself throws `Failed to create
+ * optimized buffer: WxH` and Node dies with exit 7 (fatal error in the
+ * uncaughtException handler), MASKING the real error. (The exception-handler
+ * guard lives in renderer.ts `guardRendererErrorHandlers`.)
+ *
+ * Why we can't just SHARE one SyntaxStyle across renderables (the obvious
+ * 3→2 fix): the per-buffer style is load-bearing. The native styled-text path
+ * (text-buffer.zig `setStyledText`) registers each chunk's color by NAME —
+ * "chunk0", "chunk1", … — into the buffer's OWN syntax style, and
+ * registration is name-keyed-overwrite (syntax-style.zig `putStyle`: existing
+ * name → overwrite that id's definition). A shared style would have every
+ * styled `<text>` overwrite every other one's chunk colors (live highlights
+ * reference style IDS, re-resolved at render). So pooling is unsound at our
+ * layer; the table pressure itself is bounded by the store row cap
+ * (logic/store.ts, clamped to a handle-safe ceiling) until #27 lands
+ * renderable-weight-aware capping/virtualization.
+ *
+ * What THIS shim does: makes style allocation failure DEGRADE instead of
+ * throwing out of mount/render. `SyntaxStyle.create()` on a full table
+ * returns a DETACHED style (handle 0 = the native INVALID_HANDLE):
+ *  - JS-side styling still works — markdown/code chunk colors come from
+ *    `getStyle`/`mergeStyles`, which read the instance's JS `styleDefs` map
+ *    (see core lib/tree-sitter-styled-text.ts), never the native handle;
+ *  - every native call on handle 0 is already a safe no-op in zig (acquire
+ *    fails → early return), and `textBuffer.setSyntaxStyle(detached)` passes
+ *    ptr 0 which the native side treats as "no style" — buffer-level styled
+ *    -text highlights are skipped, i.e. that text renders unstyled;
+ *  - `destroy()` on a detached style is a native no-op (beginDestroy(0)).
+ *
+ * TODO(upstream): file an OpenTUI issue — (a) a global 64k handle table with a
+ * 3-slot cost per text renderable is too small for transcript-style TUIs;
+ * (b) allocation failure throws out of the render loop with no degrade path;
+ * (c) `handleError` allocates (console overlay) and so crashes on the very
+ * condition it is reporting, masking the root cause with exit 7.
+ */
+import { SyntaxStyle, resolveRenderLib, type SyntaxStyleHandle } from '@opentui/core'
+
+import { getLog } from './log.ts'
+
+/** The native side's INVALID_HANDLE — every FFI entry point no-ops on it. */
+const DETACHED: SyntaxStyleHandle = 0 as never
+
+let installed = false
+let warnedExhausted = false
+
+/** Build a SyntaxStyle backed by NO native handle: JS-side styleDefs/merge
+ *  caches fully functional, all native calls safe no-ops (handle 0). */
+function detachedSyntaxStyle(): SyntaxStyle {
+  return new SyntaxStyle(resolveRenderLib(), DETACHED)
+}
+
+/**
+ * Patch `SyntaxStyle.create` (the static the core's own TextBufferRenderable
+ * constructor calls — @opentui/core is external, one shared class object) so
+ * native handle-table exhaustion degrades to a detached, unstyled-but-inert
+ * style instead of throwing out of a Solid mount effect. Idempotent.
+ *
+ * @param factory test seam — inject a failing allocator to exercise the
+ *                degrade path (defaults to the real `SyntaxStyle.create`).
+ */
+export function installSyntaxStyleDegrade(factory?: () => SyntaxStyle): void {
+  if (installed) return
+  installed = true
+
+  const origCreate = factory ?? SyntaxStyle.create.bind(SyntaxStyle)
+
+  SyntaxStyle.create = function create(): SyntaxStyle {
+    try {
+      return origCreate()
+    } catch (cause) {
+      if (!warnedExhausted) {
+        warnedExhausted = true
+        try {
+          getLog().error(
+            'native',
+            'SyntaxStyle allocation failed — native handle table exhausted; degrading to unstyled',
+            {
+              cause: String(cause)
+            }
+          )
+        } catch {
+          // logging is best-effort inside a degrade path
+        }
+      }
+      return detachedSyntaxStyle()
+    }
+  }
+}

ui-opentui/src/boundary/parsers.manifest.json

@@ -0,0 +1,65 @@
+{
+  "comment": "Tree-sitter grammars beyond @opentui/core's bundled 5 (ts/js/markdown/markdown_inline/zig). NOT vendored — declared as remote URLs and fetched+cached at runtime by OpenTUI's TreeSitterClient into HERMES_TUI_PARSER_CACHE (see src/boundary/parsers.ts). This file is the URL source of truth. Grammar versions are pinned via the release tag in each wasm URL; .scm highlight queries follow opencode's per-language source choices (parser-repo queries for python/html where nvim-treesitter's are parser-incompatible, nvim-treesitter master otherwise). cpp deliberately dropped (3.28 MB, ~half the old vendored bundle). Aliases are belt-and-braces: core's extToFiletype/infoStringToFiletype already normalize py->python etc., but a literal alias filetype reaching the client still resolves. To add/refresh a grammar, edit the wasm tag + the highlights URL here — no binaries land in the repo.",
+  "parsers": [
+    {
+      "filetype": "python",
+      "aliases": ["py"],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-python/releases/download/v0.23.6/tree-sitter-python.wasm",
+      "highlights": "https://github.com/tree-sitter/tree-sitter-python/raw/refs/heads/master/queries/highlights.scm"
+    },
+    {
+      "filetype": "rust",
+      "aliases": ["rs"],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-rust/releases/download/v0.23.2/tree-sitter-rust.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/rust/highlights.scm"
+    },
+    {
+      "filetype": "go",
+      "aliases": [],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-go/releases/download/v0.23.4/tree-sitter-go.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/go/highlights.scm"
+    },
+    {
+      "filetype": "bash",
+      "aliases": ["sh", "shell", "zsh"],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-bash/releases/download/v0.23.3/tree-sitter-bash.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/bash/highlights.scm"
+    },
+    {
+      "filetype": "json",
+      "aliases": [],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-json/releases/download/v0.24.8/tree-sitter-json.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/json/highlights.scm"
+    },
+    {
+      "filetype": "c",
+      "aliases": ["h"],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-c/releases/download/v0.23.5/tree-sitter-c.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/c/highlights.scm"
+    },
+    {
+      "filetype": "html",
+      "aliases": [],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-html/releases/download/v0.23.2/tree-sitter-html.wasm",
+      "highlights": "https://github.com/tree-sitter/tree-sitter-html/raw/refs/heads/master/queries/highlights.scm"
+    },
+    {
+      "filetype": "css",
+      "aliases": [],
+      "wasm": "https://github.com/tree-sitter/tree-sitter-css/releases/download/v0.23.2/tree-sitter-css.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/css/highlights.scm"
+    },
+    {
+      "filetype": "yaml",
+      "aliases": ["yml"],
+      "wasm": "https://github.com/tree-sitter-grammars/tree-sitter-yaml/releases/download/v0.7.1/tree-sitter-yaml.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/yaml/highlights.scm"
+    },
+    {
+      "filetype": "toml",
+      "aliases": [],
+      "wasm": "https://github.com/tree-sitter-grammars/tree-sitter-toml/releases/download/v0.7.0/tree-sitter-toml.wasm",
+      "highlights": "https://raw.githubusercontent.com/nvim-treesitter/nvim-treesitter/refs/heads/master/queries/toml/highlights.scm"
+    }
+  ]
+}

ui-opentui/src/boundary/parsers.ts

@@ -0,0 +1,100 @@
+/**
+ * Extra Tree-sitter grammar registration — the syntax-highlighting language
+ * expansion (docs/plans/opentui-syntax-highlighting-languages.md).
+ *
+ * @opentui/core@0.4.x bundles only a handful of grammars (ts/js/markdown/
+ * markdown_inline/zig); everything else renders plain text. The cure is the
+ * public `addDefaultParsers()` API fed with REMOTE grammar URLs — OpenTUI's
+ * TreeSitterClient fetches each `.wasm`/`.scm` lazily on first use of a
+ * filetype and caches it under the client's `dataPath`. We do NOT vendor any
+ * binaries (cf. opencode, which checks in zero `.wasm`/`.scm` and lets OpenTUI
+ * fetch+cache). The grammar set + its URLs live in `parsers.manifest.json`.
+ *
+ * Cache location: `HERMES_TUI_PARSER_CACHE` (set by the Python launcher to
+ * `~/.hermes/cache/opentui-parsers/`, profile-aware via get_hermes_home). When
+ * unset (dev/demo/CI), we leave OpenTUI's default data path
+ * (`$XDG_DATA_HOME/opentui` → `~/.local/share/opentui`) untouched.
+ *
+ * `setDataPath()` on the GLOBAL client must run BEFORE the client initializes
+ * (it only mutates `options.dataPath` until init, then the worker boots with
+ * it). `addDefaultParsers()` must run BEFORE the first `<code>`/`<markdown>`
+ * mount (they grab the global client lazily and trigger init). The entry
+ * imports + calls `registerRemoteParsers()` at module load, ahead of renderer
+ * acquisition, so both orderings hold.
+ *
+ * Offline behavior: registration itself does NO network (it only declares the
+ * URL configs). The fetch happens on first highlight of a given language; if it
+ * fails (air-gapped / GitHub unreachable), OpenTUI degrades that filetype to
+ * plain text — never a throw. A registration error likewise degrades the whole
+ * extra set to plain text.
+ */
+import { getTreeSitterClient } from '@opentui/core'
+import { addDefaultParsers } from '@opentui/core'
+
+import manifest from './parsers.manifest.json'
+import { getLog } from './log.ts'
+
+interface ManifestParser {
+  readonly filetype: string
+  readonly aliases: readonly string[]
+  readonly wasm: string
+  readonly highlights: string
+}
+
+/** The registered parser configs (exported shape for tests/diagnostics). */
+export interface RegisteredParser {
+  filetype: string
+  aliases?: string[]
+  wasm: string
+  queries: { highlights: string[] }
+}
+
+/** The cache dir for fetched grammar assets, or undefined to use OpenTUI's
+ *  default ($XDG_DATA_HOME/opentui). The launcher sets this per-profile. */
+export function parserCacheDir(): string | undefined {
+  const dir = (process.env.HERMES_TUI_PARSER_CACHE ?? '').trim()
+  return dir.length ? dir : undefined
+}
+
+/** Build the remote parser configs from the manifest. Pure — no network, no
+ *  filesystem; just declares the URL configs OpenTUI fetches lazily. */
+export function remoteParsers(): RegisteredParser[] {
+  const configs: RegisteredParser[] = []
+  for (const parser of (manifest as { parsers: ManifestParser[] }).parsers) {
+    if (!parser.wasm || !parser.highlights) continue
+    configs.push({
+      filetype: parser.filetype,
+      ...(parser.aliases.length ? { aliases: [...parser.aliases] } : {}),
+      wasm: parser.wasm,
+      queries: { highlights: [parser.highlights] }
+    })
+  }
+  return configs
+}
+
+/** Point the global tree-sitter client's cache at our profile dir, then
+ *  register the remote grammars with core's global default-parser list.
+ *  Returns what was registered (empty on any failure — plain-text fallback). */
+export function registerRemoteParsers(): RegisteredParser[] {
+  try {
+    const cache = parserCacheDir()
+    if (cache) {
+      // Must precede the client's lazy initialize() (first <code>/<markdown>
+      // mount). Pre-init this only mutates options.dataPath; the returned
+      // promise resolves immediately (no worker yet) so we don't await it.
+      void getTreeSitterClient().setDataPath(cache)
+    }
+    const parsers = remoteParsers()
+    if (!parsers.length) {
+      getLog().warn('parsers', 'no remote tree-sitter grammars declared — extras render plain', {})
+      return []
+    }
+    addDefaultParsers(parsers)
+    return parsers
+  } catch (cause) {
+    getLog().warn('parsers', 'tree-sitter registration failed — extras render plain', {
+      cause: String(cause)
+    })
+    return []
+  }
+}

ui-opentui/src/boundary/proactiveGc.ts

@@ -0,0 +1,95 @@
+/**
+ * proactiveGc — opt-in, idle-gated `global.gc()` for the low-mem path (W2).
+ *
+ * GATED on the low-mem opt-in: only runs when the user set a LOW
+ * `HERMES_TUI_HEAP_MB` (the same knob W1 threads into `--max-old-space-size`).
+ * Default / unconstrained sessions do NOTHING — no proactive GC, no divergence
+ * from Ink on the default path (Ink never calls gc proactively; it only exposes
+ * it for heapdumps, so W2 is OpenTUI-only by design — spec D5).
+ *
+ * TRIGGER MODEL — idle-gated, never mid-stream:
+ *   - A low-frequency timer ticks every IDLE_MS. On each tick it calls
+ *     `global.gc()` ONLY when (a) a turn is NOT streaming (`isStreaming()` false
+ *     — so we never pause mid-render/mid-reply) and (b) at least one full idle
+ *     window has passed since the last activity (stream end / explicit touch).
+ *   - If RSS crosses RSS_EAGER_KB (>400MB) the cadence tightens (EAGER_MS) but
+ *     it STILL waits for idle — eagerness shortens the window, never bypasses it.
+ *   - `--expose-gc` (W1) makes `global.gc` real; without it this is a silent
+ *     no-op (we detect and disable). The timer is unref'd — it never keeps the
+ *     process alive — and every failure path disables silently (a GC helper must
+ *     never break the TUI).
+ *
+ * Reuses `process.memoryUsage().rss` (same read as memlog) for the >400MB check.
+ */
+import { envFlag } from '../logic/env.ts'
+
+/** Below this heap cap (MB) we treat the session as low-mem opt-in. 8192 is the
+ *  default; anyone who set a cap materially under it wants tight memory. */
+const LOW_MEM_HEAP_MB = 4096
+/** Idle window (ms): time since last activity before a GC is allowed. */
+const IDLE_MS = 8000
+/** Tightened idle window once RSS is high. */
+const EAGER_MS = 3000
+/** RSS (KB) above which GC becomes eager (still idle-gated). 400MB. */
+const RSS_EAGER_KB = 400 * 1024
+
+/** The configured heap cap in MB from the W1 knob, or null when unset/garbage.
+ *  The Python launcher reads the same env; the child inherits it, so the Node
+ *  side can read it directly to know whether low-mem mode is active. */
+function configuredHeapMb(): number | null {
+  const v = (process.env.HERMES_TUI_HEAP_MB ?? '').trim()
+  if (!/^\d+$/.test(v)) return null
+  const n = Number.parseInt(v, 10)
+  return n > 0 ? n : null
+}
+
+/** Whether proactive GC should run: a low heap cap is set AND gc is exposed.
+ *  `HERMES_TUI_PROACTIVE_GC` can force it on/off, but defaults to the low-mem
+ *  signal so the knob composes (spec D9: independent knobs). */
+export function proactiveGcEnabled(heapMb: number | null = configuredHeapMb()): boolean {
+  const lowMem = heapMb !== null && heapMb <= LOW_MEM_HEAP_MB
+  return envFlag(process.env.HERMES_TUI_PROACTIVE_GC, lowMem) && typeof global.gc === 'function'
+}
+
+/**
+ * Start the idle-gated proactive GC watcher. `isStreaming` reports whether a
+ * turn is mid-flight (read from the store's `info.running`). Returns a stop
+ * function and a `touch()` to mark fresh activity (e.g. on keypress / stream
+ * start) so the idle clock resets. No-op (returns inert handles) when disabled.
+ */
+export function startProactiveGc(isStreaming: () => boolean): { stop: () => void; touch: () => void } {
+  if (!proactiveGcEnabled()) return { stop: () => {}, touch: () => {} }
+  const gc = global.gc
+  if (typeof gc !== 'function') return { stop: () => {}, touch: () => {} }
+
+  let lastActivity = Date.now()
+  const touch = () => {
+    lastActivity = Date.now()
+  }
+
+  // Tick at the eager cadence; the idle-window check (below) does the real
+  // gating, so a high-RSS session reacts within EAGER_MS while a calm one still
+  // waits the full IDLE_MS. One cheap rss read + compare per tick.
+  const timer = setInterval(() => {
+    try {
+      if (isStreaming()) {
+        // mid-stream: defer entirely and keep the clock fresh so a GC can't fire
+        // the instant the stream ends — it waits a full idle window after.
+        lastActivity = Date.now()
+        return
+      }
+      const rssKb = Math.floor(process.memoryUsage().rss / 1024)
+      const window = rssKb > RSS_EAGER_KB ? EAGER_MS : IDLE_MS
+      if (Date.now() - lastActivity < window) return
+      gc()
+      // After a collection, reset the clock so we don't GC every tick — the next
+      // one waits another full idle window.
+      lastActivity = Date.now()
+    } catch {
+      clearInterval(timer) // a failing GC helper must not retry forever
+    }
+  }, EAGER_MS)
+  timer.unref?.()
+
+  return { stop: () => clearInterval(timer), touch }
+}

ui-opentui/src/boundary/renderer.ts

@@ -0,0 +1,234 @@
+/**
+ * Renderer lifecycle — the Effect-side resource boundary (spec v4 §3.1).
+ *
+ * `acquireRelease(createCliRenderer)` so the renderer is always destroyed on
+ * scope exit; a `Deferred` resolved on the renderer's "destroy" event lets the
+ * entry block until the user quits. Mirrors opencode `app.tsx:177` /
+ * `:185-225`.
+ *
+ * No throw / try-catch here: acquisition failure surfaces as a typed
+ * `RendererError` via `Effect.tryPromise`'s `catch`.
+ */
+import { createCliRenderer, type CliRenderer, type KeyEvent, type Selection } from '@opentui/core'
+import { Deferred, Effect } from 'effect'
+
+import { RendererError } from './errors.ts'
+import { installFfiCoordSafety } from './ffiSafe.ts'
+import { getLog } from './log.ts'
+import { installMultiClickSelection } from './multiClickSelect.ts'
+import { installSyntaxStyleDegrade } from './nativeHandles.ts'
+
+// Node-FFI seam: clamp negative draw coordinates BEFORE the u32 FFI marshaling
+// (see ffiSafe.ts — scrolled-out <diff> line backgrounds crashed the render loop).
+installFfiCoordSafety()
+// Native handle-table seam: SyntaxStyle allocation failure (global 65,534-slot
+// registry exhausted) degrades to an unstyled detached style instead of throwing
+// out of a Solid mount effect (see nativeHandles.ts for the full root cause).
+installSyntaxStyleDegrade()
+
+/**
+ * The text a finished selection copies: the RENDERED text the user highlighted,
+ * verbatim (`getSelectedText()` does correct same-line merging). Markdown markers
+ * are concealed in the pretty render, so a partial selection cannot recover source —
+ * this copies exactly what was highlighted (the `/copy` command gives full source).
+ * Total by construction — a copy must NEVER throw out of an input/event handler
+ * (that would tear down the render loop).
+ */
+function selectionCopyText(selection: Selection): string {
+  try {
+    return selection.getSelectedText()
+  } catch (cause) {
+    getLog().warn('copy', 'getSelectedText failed', { cause: String(cause) })
+    return ''
+  }
+}
+
+export interface RendererOptions {
+  /** Mouse tracking on/off (from decoded display config). */
+  readonly mouse: boolean
+  /** When true, a blocking prompt owns Ctrl+C (cancel) — the global quit is suppressed (gotcha §8 #6). */
+  readonly isBlocked?: () => boolean
+  /**
+   * Ctrl+C handler (item 11). When set, it OWNS Ctrl+C while not blocked — the
+   * entry's state machine decides interrupt-the-turn vs quit. When omitted, the
+   * default is an immediate `renderer.destroy()` (quit).
+   */
+  readonly onCtrlC?: () => void
+  /**
+   * Copy a mouse selection (item 1). When there's a live selection, Ctrl+C copies
+   * it (this callback) instead of interrupting/quitting — opencode's selection
+   * key precedence (`app.tsx:388`). Receives the rendered text the user highlighted.
+   */
+  readonly onCopySelection?: (text: string) => void
+}
+
+/**
+ * Acquire a CliRenderer inside the current scope and register its release.
+ * Returns the renderer plus a Deferred that resolves when the renderer is
+ * destroyed (user quit) — `await` it to keep the entry alive.
+ */
+export const acquireRenderer = Effect.fn('Renderer.acquire')(function* (options: RendererOptions) {
+  const renderer = yield* Effect.acquireRelease(
+    Effect.tryPromise({
+      try: async () => {
+        // Snapshot process error listeners so we can guard exactly the ones the
+        // renderer installs (its `handleError` — see guardRendererErrorHandlers).
+        const preexisting = snapshotErrorListeners()
+        const created = await createCliRenderer({
+          // Root canvas: TRANSPARENT by default — the terminal's own background
+          // shows through (do not paint a "default dark" canvas; glitch hated
+          // it). A skin's explicit ui_bg lands reactively via the header's
+          // theme effect (view/header.tsx setBackgroundColor).
+          // scrollbox clips growing output → no terminal-scrollback corruption (gotcha §8 #2).
+          externalOutputMode: 'passthrough',
+          targetFps: 60,
+          // Don't let core's uncaught-error handler call the ALLOCATING
+          // `console.show()` (0.4.1 public option; defaults true-in-dev). That
+          // call needs a native handle, so under handle-table exhaustion — the
+          // very condition being reported — it throws and exit-7-masks the
+          // original error (the bench mem3000 postmortem). Disabling it removes
+          // that failure mode at the source; `guardRendererErrorHandlers` below
+          // stays as belt-and-suspenders (honest logging if any core handler
+          // still throws), but is no longer load-bearing for the exit-7 mask.
+          openConsoleOnError: false,
+          // prompts own Ctrl+C → deny/cancel (gotcha §8 #6); the global quit is gated on !blocked.
+          exitOnCtrlC: false,
+          // OpenTUI's default exitSignals include SIGPIPE + SIGBUS, and its handler
+          // calls renderer.destroy() — so a broken clipboard pipe (writeClipboard
+          // spawning xclip/wl-copy that dies) raises SIGPIPE and QUITS THE TUI on
+          // copy. SIGPIPE/SIGBUS are not shutdown intents; restrict to the genuine
+          // termination signals so a stray pipe error can never tear down the UI.
+          exitSignals: ['SIGINT', 'SIGTERM', 'SIGQUIT', 'SIGHUP'],
+          useKittyKeyboard: {},
+          useMouse: options.mouse
+        })
+        guardRendererErrorHandlers(created, preexisting)
+        // Editor-grade mouse selection: double-click word, triple-click line,
+        // drag extends with the clicked span held (see multiClickSelect.ts).
+        installMultiClickSelection(created)
+        return created
+      },
+      catch: cause => new RendererError({ cause })
+    }),
+    renderer => Effect.sync(() => destroyRenderer(renderer))
+  )
+
+  const shutdown = yield* Deferred.make<void>()
+  renderer.once('destroy', () => {
+    Deferred.doneUnsafe(shutdown, Effect.void)
+  })
+
+  // Global quit on Ctrl+C. `exitOnCtrlC:false` hands Ctrl+C to us as a key event
+  // (not SIGINT), so destroying here fires 'destroy' → resolves `shutdown` → the
+  // entry scope closes → finalizers run: renderer teardown + the gateway layer's
+  // `client.stop()` EOFs the Python child's stdin so it exits (no orphan). When a
+  // blocking prompt is up, it owns Ctrl+C (→ deny/cancel) so we suppress the quit
+  // (gotcha §8 #6) — the prompt's own handler sends the cancel reply.
+  const isBlocked = options.isBlocked ?? (() => false)
+  renderer.keyInput.on('keypress', (key: KeyEvent) => {
+    if (!(key.ctrl && key.name === 'c') || renderer.isDestroyed) return
+    // Copy a live mouse selection first (item 1) — takes precedence over the
+    // interrupt/quit machine and over a blocking prompt's cancel.
+    if (options.onCopySelection) {
+      const selection = renderer.getSelection()
+      const text = selection ? selectionCopyText(selection) : ''
+      if (text) {
+        options.onCopySelection(text)
+        renderer.clearSelection()
+        return
+      }
+    }
+    if (isBlocked()) return // a blocking prompt owns Ctrl+C (→ deny/cancel)
+    if (options.onCtrlC) options.onCtrlC()
+    else renderer.destroy()
+  })
+
+  // Copy-on-select (item 1 parity with free-code/Ink): the renderer's "selection"
+  // event fires ONCE when a free-form mouse selection COMPLETES (drag finish);
+  // auto-copy the spanned selectable text. Unlike the Ctrl+C path above we do NOT
+  // clearSelection() — the highlight persists so the user sees what was copied and
+  // Ctrl+C still works on it. `writeClipboard` is idempotent, so both paths writing
+  // the same text is harmless (no double-write bug). `CliRenderer extends
+  // EventEmitter`, so `on('selection', …)` is untyped → annotate `selection`.
+  const onCopy = options.onCopySelection
+  if (onCopy) {
+    renderer.on('selection', (selection: Selection) => {
+      const text = selectionCopyText(selection)
+      if (text) onCopy(text)
+    })
+  }
+
+  return { renderer, shutdown } as const
+})
+
+/** Best-effort renderer teardown; never throws out of the finalizer. */
+function destroyRenderer(renderer: CliRenderer): void {
+  try {
+    if (!renderer.isDestroyed) renderer.destroy()
+  } catch {
+    // teardown is best-effort; a failed destroy must not mask the real exit cause.
+  }
+}
+
+// ── honest-crash guard for the renderer's process error handlers ─────────────
+//
+// CliRenderer installs its own `uncaughtException`/`unhandledRejection` handler
+// (`handleError`: console.error + console.show()). `console.show()` ALLOCATES —
+// the console-overlay OptimizedBuffer needs a native handle — so under native
+// handle-table exhaustion (the very condition being reported, see
+// nativeHandles.ts) the handler itself throws `Failed to create optimized
+// buffer: WxH`, and Node kills the process with exit 7, MASKING the original
+// error (this is exactly the bench mem3000 postmortem). Wrap the listeners the
+// renderer added so a handler failure is logged honestly and the original
+// error stays the story; while the renderer is alive the process keeps running
+// (core's own contract: handled uncaught exceptions don't exit).
+
+type ProcessErrorEvent = 'uncaughtException' | 'unhandledRejection'
+const PROCESS_ERROR_EVENTS: readonly ProcessErrorEvent[] = ['uncaughtException', 'unhandledRejection']
+type ErrorListener = (...args: unknown[]) => void
+
+// Node's typings don't accept the union event name in listeners/on/removeListener
+// overloads — view the process emitter through a minimal untyped seam.
+const proc = process as unknown as {
+  listeners(event: ProcessErrorEvent): ErrorListener[]
+  on(event: ProcessErrorEvent, listener: ErrorListener): void
+  removeListener(event: ProcessErrorEvent, listener: ErrorListener): void
+}
+
+function snapshotErrorListeners(): ReadonlyMap<ProcessErrorEvent, ReadonlySet<unknown>> {
+  return new Map(PROCESS_ERROR_EVENTS.map(event => [event, new Set(proc.listeners(event))]))
+}
+
+/** Re-wrap the error listeners `createCliRenderer` added (delta vs the snapshot)
+ *  so an exception INSIDE them can never exit-7-mask the original error. */
+function guardRendererErrorHandlers(
+  renderer: CliRenderer,
+  preexisting: ReadonlyMap<ProcessErrorEvent, ReadonlySet<unknown>>
+): void {
+  for (const event of PROCESS_ERROR_EVENTS) {
+    const before = preexisting.get(event)
+    for (const listener of proc.listeners(event)) {
+      if (before?.has(listener)) continue
+      proc.removeListener(event, listener)
+      const guarded: ErrorListener = (...args) => {
+        // After teardown the renderer can no longer report anything — rethrow so
+        // the ORIGINAL error reaches Node's default fatal path unmasked.
+        if (renderer.isDestroyed) throw args[0]
+        try {
+          listener(...args)
+        } catch (handlerFailure) {
+          try {
+            const original = args[0]
+            getLog().error('renderer', 'core error handler crashed while reporting an uncaught error', {
+              original: original instanceof Error ? (original.stack ?? original.message) : String(original),
+              handlerFailure: String(handlerFailure)
+            })
+          } catch {
+            // logging is best-effort — never throw out of an exception handler.
+          }
+        }
+      }
+      proc.on(event, guarded)
+    }
+  }
+}

ui-opentui/src/boundary/runtime.ts

@@ -0,0 +1,17 @@
+/**
+ * Runtime composition — the single edge where layers are provided and the
+ * program is run (spec v4 §3.1). Layers are provided HERE by the caller
+ * (the launcher entry), never inside components. Mirrors opencode
+ * `cli/tui/layer.ts:6` + `cli/cmd/tui.ts` runMain.
+ */
+import { Layer } from 'effect'
+
+import type { GatewayService } from './gateway/GatewayService.ts'
+
+/**
+ * The application layer. Phase 0 takes the GatewayService layer as a parameter
+ * so the entry can choose Fake (dev/test) or — from Phase 1 — the live
+ * `tui_gateway`-spawning layer. Compose additional boundary services
+ * (Config, Theme-with-IO) here as they land.
+ */
+export const makeAppLayer = (gateway: Layer.Layer<GatewayService>) => Layer.mergeAll(gateway)

ui-opentui/src/boundary/schema/GatewayEvent.ts

@@ -0,0 +1,259 @@
+/**
+ * GatewayEvent — the wire event union, modeled as an Effect Schema and decoded
+ * ONCE at the transport boundary (spec v4 §3.3). Mirrors Ink's
+ * `ui-tui/src/gatewayTypes.ts:509-587` (discriminant = `type`).
+ *
+ * beta.78 API (verified vs .d.ts): variants are `Schema.Struct` with a
+ * `Schema.Literal` `type`, combined with `Schema.Union([...]).pipe(
+ * Schema.toTaggedUnion("type"))`. Optional fields use `Schema.optionalKey`
+ * (exact-optional under exactOptionalPropertyTypes). Decode unknown wire JSON
+ * with `Schema.decodeUnknownOption` so an UNRECOGNIZED `type` yields `Option.none`
+ * and is skipped — a stray event never tears down the stream.
+ *
+ * Types are INFERRED from the schema (`typeof X["Type"]`), never hand-declared.
+ */
+import { Schema } from 'effect'
+
+const Str = Schema.String
+const opt = Schema.optionalKey
+
+// ── Skin (mirror GatewaySkin in ui-tui/src/gatewayTypes.ts) ───────────
+export const GatewaySkinSchema = Schema.Struct({
+  banner_hero: opt(Str),
+  banner_logo: opt(Str),
+  branding: opt(Schema.Record(Str, Str)),
+  colors: opt(Schema.Record(Str, Str)),
+  help_header: opt(Str),
+  tool_prefix: opt(Str),
+  // Spinner animation data (faces/verbs/wings) — mixed array/tuple shapes, kept
+  // loose at the boundary; the spinner component narrows what it reads. tool_emojis
+  // is a per-tool glyph override map. Both additive + optional (back-compat).
+  spinner: opt(Schema.Record(Str, Schema.Unknown)),
+  tool_emojis: opt(Schema.Record(Str, Str))
+})
+export type GatewaySkinDecoded = typeof GatewaySkinSchema.Type
+
+// ── Variant schemas (one per wire `type`) ─────────────────────────────
+// lifecycle
+const GatewayReady = Schema.Struct({
+  type: Schema.Literal('gateway.ready'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ skin: opt(GatewaySkinSchema) }))
+})
+const SkinChanged = Schema.Struct({
+  type: Schema.Literal('skin.changed'),
+  session_id: opt(Str),
+  payload: opt(GatewaySkinSchema)
+})
+const SessionInfoEvent = Schema.Struct({
+  type: Schema.Literal('session.info'),
+  session_id: opt(Str),
+  // SessionInfo is large + evolving; keep it loose at the boundary (Record),
+  // the chrome phase narrows the fields it actually reads.
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+
+// streaming text
+const MessageStart = Schema.Struct({ type: Schema.Literal('message.start'), session_id: opt(Str) })
+const MessageDelta = Schema.Struct({
+  type: Schema.Literal('message.delta'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ text: opt(Str), rendered: opt(Str) }))
+})
+const MessageComplete = Schema.Struct({
+  type: Schema.Literal('message.complete'),
+  session_id: opt(Str),
+  // `usage` carries the post-turn token/context totals → refreshes the status bar
+  // (item 14). Kept loose (Record) — the chrome reader narrows what it needs.
+  payload: opt(Schema.Struct({ text: opt(Str), rendered: opt(Str), usage: opt(Schema.Record(Str, Schema.Unknown)) }))
+})
+
+// reasoning / thinking — toTaggedUnion needs ONE literal per member, so the
+// reasoning.delta/reasoning.available pair is two structs sharing a shape.
+const ReasoningShape = {
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ text: opt(Str), verbose: opt(Schema.Boolean) }))
+}
+const ReasoningDelta = Schema.Struct({ type: Schema.Literal('reasoning.delta'), ...ReasoningShape })
+const ReasoningAvailable = Schema.Struct({ type: Schema.Literal('reasoning.available'), ...ReasoningShape })
+const ThinkingDelta = Schema.Struct({
+  type: Schema.Literal('thinking.delta'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ text: opt(Str) }))
+})
+
+// tools
+const ToolStart = Schema.Struct({
+  type: Schema.Literal('tool.start'),
+  session_id: opt(Str),
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+const ToolComplete = Schema.Struct({
+  type: Schema.Literal('tool.complete'),
+  session_id: opt(Str),
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+const ToolProgress = Schema.Struct({
+  type: Schema.Literal('tool.progress'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ name: opt(Str), preview: opt(Str) })
+})
+const ToolGenerating = Schema.Struct({
+  type: Schema.Literal('tool.generating'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ name: opt(Str) })
+})
+
+// blocking prompts (deadlock-critical — Phase 3 renders these)
+const ClarifyRequest = Schema.Struct({
+  type: Schema.Literal('clarify.request'),
+  session_id: opt(Str),
+  payload: Schema.Struct({
+    choices: opt(Schema.NullOr(Schema.Array(Str))),
+    question: opt(Str),
+    request_id: Str
+  })
+})
+const ApprovalRequest = Schema.Struct({
+  type: Schema.Literal('approval.request'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ command: Str, description: Str })
+})
+const SudoRequest = Schema.Struct({
+  type: Schema.Literal('sudo.request'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ request_id: Str })
+})
+const SecretRequest = Schema.Struct({
+  type: Schema.Literal('secret.request'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ env_var: Str, prompt: Str, request_id: Str })
+})
+
+// chrome / agent
+const StatusUpdate = Schema.Struct({
+  type: Schema.Literal('status.update'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ kind: opt(Str), text: opt(Str) }))
+})
+const NotificationShow = Schema.Struct({
+  type: Schema.Literal('notification.show'),
+  session_id: opt(Str),
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+const NotificationClear = Schema.Struct({
+  type: Schema.Literal('notification.clear'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ key: opt(Str) }))
+})
+const VoiceStatus = Schema.Struct({
+  type: Schema.Literal('voice.status'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ state: opt(Schema.Literals(['idle', 'listening', 'transcribing'])) }))
+})
+const VoiceTranscript = Schema.Struct({
+  type: Schema.Literal('voice.transcript'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ no_speech_limit: opt(Schema.Boolean), text: opt(Str) }))
+})
+const BrowserProgress = Schema.Struct({
+  type: Schema.Literal('browser.progress'),
+  session_id: opt(Str),
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+const BackgroundComplete = Schema.Struct({
+  type: Schema.Literal('background.complete'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ task_id: Str, text: Str })
+})
+const ReviewSummary = Schema.Struct({
+  type: Schema.Literal('review.summary'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ text: opt(Str) }))
+})
+const SubagentShape = { session_id: opt(Str), payload: Schema.Record(Str, Schema.Unknown) }
+const SubagentSpawnRequested = Schema.Struct({ type: Schema.Literal('subagent.spawn_requested'), ...SubagentShape })
+const SubagentStart = Schema.Struct({ type: Schema.Literal('subagent.start'), ...SubagentShape })
+const SubagentThinking = Schema.Struct({ type: Schema.Literal('subagent.thinking'), ...SubagentShape })
+const SubagentTool = Schema.Struct({ type: Schema.Literal('subagent.tool'), ...SubagentShape })
+const SubagentProgress = Schema.Struct({ type: Schema.Literal('subagent.progress'), ...SubagentShape })
+const SubagentComplete = Schema.Struct({ type: Schema.Literal('subagent.complete'), ...SubagentShape })
+
+// transport errors
+const ErrorEvent = Schema.Struct({
+  type: Schema.Literal('error'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ message: opt(Str) }))
+})
+const GatewayStderr = Schema.Struct({
+  type: Schema.Literal('gateway.stderr'),
+  session_id: opt(Str),
+  payload: Schema.Struct({ line: Str })
+})
+const GatewayStartTimeout = Schema.Struct({
+  type: Schema.Literal('gateway.start_timeout'),
+  session_id: opt(Str),
+  payload: Schema.Record(Str, Schema.Unknown)
+})
+const GatewayProtocolError = Schema.Struct({
+  type: Schema.Literal('gateway.protocol_error'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ preview: opt(Str) }))
+})
+// gateway lifecycle recovery (auto-heal): the child exited (crash/kill) and the
+// transport is respawning+resuming the session. Surfaced so the frozen spinner
+// clears and the user sees the in-flight reply was lost (see store cases).
+const GatewayExited = Schema.Struct({
+  type: Schema.Literal('gateway.exited'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ reason: opt(Str), code: opt(Schema.Number), signal: opt(Str) }))
+})
+const GatewayRecovering = Schema.Struct({
+  type: Schema.Literal('gateway.recovering'),
+  session_id: opt(Str),
+  payload: opt(Schema.Struct({ attempt: opt(Schema.Number), delay_ms: opt(Schema.Number) }))
+})
+
+// ── The union ─────────────────────────────────────────────────────────
+export const GatewayEventSchema = Schema.Union([
+  GatewayReady,
+  SkinChanged,
+  SessionInfoEvent,
+  MessageStart,
+  MessageDelta,
+  MessageComplete,
+  ReasoningDelta,
+  ReasoningAvailable,
+  ThinkingDelta,
+  ToolStart,
+  ToolComplete,
+  ToolProgress,
+  ToolGenerating,
+  ClarifyRequest,
+  ApprovalRequest,
+  SudoRequest,
+  SecretRequest,
+  StatusUpdate,
+  NotificationShow,
+  NotificationClear,
+  VoiceStatus,
+  VoiceTranscript,
+  BrowserProgress,
+  BackgroundComplete,
+  ReviewSummary,
+  SubagentSpawnRequested,
+  SubagentStart,
+  SubagentThinking,
+  SubagentTool,
+  SubagentProgress,
+  SubagentComplete,
+  ErrorEvent,
+  GatewayStderr,
+  GatewayStartTimeout,
+  GatewayProtocolError,
+  GatewayExited,
+  GatewayRecovering
+]).pipe(Schema.toTaggedUnion('type'))
+
+/** The decoded, typed event. Inferred from the schema — never hand-declared. */
+export type GatewayEvent = typeof GatewayEventSchema.Type

ui-opentui/src/boundary/schema/SessionInfo.ts

@@ -0,0 +1,113 @@
+/**
+ * SessionInfo + Catalog decoders — the decode-at-boundary idiom (spec v4 §3.3),
+ * mirroring GatewayEvent.ts. These two payloads are UNTRUSTED loose JSON from the
+ * Python `tui_gateway` (`session.info` event / `session.create`/`resume` result
+ * `info`, and the `startup.catalog` RPC result), so they are decoded ONCE with an
+ * Effect Schema instead of hand-rolled `as`-cast readers.
+ *
+ * Decode with `Schema.decodeUnknownOption`: a malformed/partial payload yields
+ * `Option.none` and the caller falls back to an empty patch / leaves the catalog
+ * unset — a stray shape never crashes the reducer.
+ *
+ * Wire field names are verified against `tui_gateway/server.py`:
+ *   - session.info  → `_session_info()` (server.py:~1798): top-level `model`,
+ *     `reasoning_effort`, `fast`, `cwd`, `branch`, `running`, `profile_name`,
+ *     `update_behind` (Optional[int] — null until the prefetched check lands),
+ *     `update_command`, `mcp_servers` (list of {name,transport,connected,tools}
+ *     dicts from `get_mcp_status()`), plus a nested `usage` (`_get_usage()`,
+ *     server.py:~1683) carrying `context_used`, `context_max`,
+ *     `context_percent`, `compressions` (context_* only present when the
+ *     compressor knows a context length) and `cost_usd` (only when the pricing
+ *     estimate succeeds).
+ *   - startup.catalog → `@method("startup.catalog")` (server.py:~8521):
+ *     `{ tools:{total, toolsets:[{name,count,enabled,tools}]},
+ *        skills:{total, categories:[{name,count}]}, mcp:{servers:[]} }`.
+ *
+ * These schemas are used PURELY as decoders; they do NOT Effect-ify the store's
+ * reactivity or control flow (Solid stays the runtime — spec v4 §1).
+ */
+import { Schema } from 'effect'
+
+const Str = Schema.String
+const Num = Schema.Number
+const Bool = Schema.Boolean
+const opt = Schema.optionalKey
+
+// ── session.info / session.create.info ────────────────────────────────
+// Context/usage numbers arrive nested under `usage`; the same names may also
+// appear at the top level depending on the RPC vs event path (the reader prefers
+// `usage.context_*`, then the top-level fallback). All keys are optional — a
+// `session.info` patch only carries the fields that actually changed.
+const UsageSchema = Schema.Struct({
+  context_used: opt(Num),
+  context_max: opt(Num),
+  context_percent: opt(Num),
+  compressions: opt(Num),
+  cost_usd: opt(Num)
+})
+
+export const SessionInfoPatchSchema = Schema.Struct({
+  model: opt(Str),
+  reasoning_effort: opt(Str),
+  fast: opt(Bool),
+  cwd: opt(Str),
+  branch: opt(Str),
+  // session title ("" until the first exchange titles it) — drives the
+  // terminal window-title chrome (OSC 0/2 via renderer.setTerminalTitle).
+  title: opt(Str),
+  running: opt(Bool),
+  // status-bar chrome extras (Epic 1.3): update banner, profile badge, MCP count.
+  // `update_behind` is null on the wire until the async update check resolves.
+  update_behind: opt(Schema.NullOr(Num)),
+  update_command: opt(Str),
+  profile_name: opt(Str),
+  mcp_servers: opt(Schema.Array(Schema.Unknown)),
+  // top-level context fallback (used when there's no nested `usage`)
+  context_used: opt(Num),
+  context_max: opt(Num),
+  context_percent: opt(Num),
+  compressions: opt(Num),
+  usage: opt(UsageSchema)
+})
+export type SessionInfoPatchDecoded = typeof SessionInfoPatchSchema.Type
+
+/** Decode a loose session.info payload → `Option<SessionInfoPatchDecoded>`. */
+export const decodeSessionInfoPatch = Schema.decodeUnknownOption(SessionInfoPatchSchema)
+
+// ── startup.catalog ───────────────────────────────────────────────────
+// Mirrors the `Catalog` interface in store.ts. `enabled` defaults to true at the
+// reader (an absent flag means on), so it stays optional here.
+const ToolsetSchema = Schema.Struct({
+  name: opt(Str),
+  count: opt(Num),
+  enabled: opt(Bool),
+  tools: opt(Schema.Array(Schema.Unknown))
+})
+const CategorySchema = Schema.Struct({
+  name: opt(Str),
+  count: opt(Num)
+})
+
+export const CatalogSchema = Schema.Struct({
+  tools: opt(
+    Schema.Struct({
+      total: opt(Num),
+      toolsets: opt(Schema.Array(ToolsetSchema))
+    })
+  ),
+  skills: opt(
+    Schema.Struct({
+      total: opt(Num),
+      categories: opt(Schema.Array(CategorySchema))
+    })
+  ),
+  mcp: opt(
+    Schema.Struct({
+      servers: opt(Schema.Array(Schema.Unknown))
+    })
+  )
+})
+export type CatalogDecoded = typeof CatalogSchema.Type
+
+/** Decode a loose startup.catalog result → `Option<CatalogDecoded>`. */
+export const decodeCatalog = Schema.decodeUnknownOption(CatalogSchema)

ui-opentui/src/boundary/schema/SessionPeek.ts

@@ -0,0 +1,55 @@
+/**
+ * SessionPeek decoder — decode-at-boundary (house rule) for the `session.peek`
+ * RPC result (tui_gateway/server.py `@method("session.peek")`, shipped with the
+ * resume-picker gateway half, commit 529d8084b). The response powers the
+ * picker's Space preview:
+ *
+ *   { session: {id, title, source, model, cwd, started_at, ended_at,
+ *               end_reason, message_count, last_active, cost_usd},
+ *     head: [{id, role, content(≤2000), truncated, timestamp}, …],
+ *     tail: [same — never overlaps head],
+ *     total_messages: int }
+ *
+ * Wire nullability per the server: `model`/`cwd`/`ended_at`/`end_reason`/
+ * `cost_usd` are `None` when unknown; message `id`/`timestamp` come straight
+ * off DB rows (left loose). Decoded with `Schema.decodeUnknownOption` — a
+ * malformed payload yields `Option.none` and the preview pane shows its
+ * honest "preview unavailable" line instead of crashing the overlay.
+ */
+import { Schema } from 'effect'
+
+const Str = Schema.String
+const Num = Schema.Number
+const opt = Schema.optionalKey
+
+const PeekMessageSchema = Schema.Struct({
+  role: opt(Str),
+  content: opt(Str),
+  truncated: opt(Schema.Boolean),
+  timestamp: opt(Schema.NullOr(Schema.Unknown))
+})
+
+export const SessionPeekSchema = Schema.Struct({
+  session: opt(
+    Schema.Struct({
+      id: opt(Str),
+      title: opt(Schema.NullOr(Str)),
+      source: opt(Schema.NullOr(Str)),
+      model: opt(Schema.NullOr(Str)),
+      cwd: opt(Schema.NullOr(Str)),
+      started_at: opt(Schema.NullOr(Num)),
+      ended_at: opt(Schema.NullOr(Num)),
+      end_reason: opt(Schema.NullOr(Str)),
+      message_count: opt(Schema.NullOr(Num)),
+      last_active: opt(Schema.NullOr(Num)),
+      cost_usd: opt(Schema.NullOr(Num))
+    })
+  ),
+  head: opt(Schema.Array(PeekMessageSchema)),
+  tail: opt(Schema.Array(PeekMessageSchema)),
+  total_messages: opt(Num)
+})
+export type SessionPeekDecoded = typeof SessionPeekSchema.Type
+
+/** Decode a loose session.peek result → `Option<SessionPeekDecoded>`. */
+export const decodeSessionPeek = Schema.decodeUnknownOption(SessionPeekSchema)

ui-opentui/src/boundary/termChrome.ts

@@ -0,0 +1,127 @@
+/**
+ * Terminal chrome seam — window title (OSC 0/2) + desktop notifications
+ * through the renderer's native primitives.
+ *
+ * Why the renderer and not process.stdout: the zig side owns the terminal —
+ * `setTerminalTitle` and `triggerNotification` are native FFI calls and
+ * `writeOut` serializes raw control bytes with frame presentation, so chrome
+ * writes can never tear a frame.
+ *
+ * Notifications go through the native `renderer.triggerNotification(message,
+ * title)` (zig `lib.triggerNotification`), NOT a hand-rolled OSC 9/99/777 spray.
+ * The zig side does what raw OSC can't: authoritative protocol detection
+ * (query > heuristic) so it picks the ONE protocol the terminal speaks, **tmux
+ * DCS passthrough wrapping** (raw OSC is silently eaten by tmux), and Zellij
+ * OSC-99 enforcement. It returns `false` when no protocol was detected.
+ *
+ * Focus suppression: core parses mode-1004 focus reports (`ESC[I`/`ESC[O`)
+ * and re-emits them as renderer `focus`/`blur` events — notifications are
+ * skipped while the terminal reports focused (you're already looking at it).
+ * Native `triggerNotification` does NOT do focus suppression, so it stays our
+ * policy here. Terminals that never report focus leave the state at the
+ * assumed-focused initial value… which would swallow every notification, so
+ * the FIRST blur is what arms suppression: until a blur arrives we treat focus
+ * as unknown and notify unconditionally (worst case: a redundant ping while
+ * focused).
+ *
+ * Everything here is total — chrome must never throw into the render loop
+ * or a teardown path.
+ */
+import type { CliRenderer } from '@opentui/core'
+
+import type { TermNotification } from '../logic/termChrome.ts'
+import {
+  notifyEnabled,
+  sanitizeOscText,
+  TITLE_STACK_RESTORE,
+  TITLE_STACK_SAVE,
+  windowTitleFor
+} from '../logic/termChrome.ts'
+import { getLog } from './log.ts'
+
+/** What the view layer needs from the chrome seam (DI-friendly for tests). */
+export interface TerminalChromeSeam {
+  /** Set the window title from the session title (undefined → generic). */
+  readonly setTitle: (sessionTitle: string | undefined) => void
+  /** Announce "waiting on you" to the hosting terminal (no-op while focused). */
+  readonly notify: (notification: TermNotification) => void
+}
+
+/** The renderer surface the seam writes through (runtime-verified shapes). */
+interface RendererSeam {
+  setTerminalTitle(title: string): void
+  /** Native desktop notification (protocol detection + tmux/Zellij wrapping). */
+  triggerNotification(message: string, title?: string): boolean
+  writeOut(chunk: string): void
+  on(event: 'focus' | 'blur', listener: () => void): unknown
+  once(event: 'destroy', listener: () => void): unknown
+  readonly isDestroyed: boolean
+}
+
+/** Install the chrome seam on a live renderer. Idempotent per renderer use —
+ *  the entry calls it once, right next to the render bridge. */
+export function installTerminalChrome(renderer: CliRenderer): TerminalChromeSeam {
+  const seam = renderer as unknown as RendererSeam
+  const notificationsOn = notifyEnabled()
+
+  // unknown (null) until the terminal proves it reports focus; then boolean.
+  let focused: boolean | null = null
+  try {
+    seam.on('focus', () => {
+      focused = true
+    })
+    seam.on('blur', () => {
+      focused = false
+    })
+  } catch (cause) {
+    getLog().warn('chrome', 'focus tracking unavailable', { cause: String(cause) })
+  }
+
+  // Bracket our title ownership: save the user's title now, restore on quit.
+  // Best-effort — terminals without the XTWINOPS title stack ignore both.
+  writeRaw(seam, TITLE_STACK_SAVE)
+  seam.once('destroy', () => writeRaw(seam, TITLE_STACK_RESTORE, { evenIfDestroyed: true }))
+
+  let lastTitle = ''
+  return {
+    setTitle: sessionTitle => {
+      const title = windowTitleFor(sessionTitle)
+      if (title === lastTitle) return
+      lastTitle = title
+      try {
+        if (!seam.isDestroyed) seam.setTerminalTitle(title)
+      } catch (cause) {
+        getLog().warn('chrome', 'setTerminalTitle failed', { cause: String(cause) })
+      }
+    },
+    notify: notification => {
+      if (!notificationsOn || focused === true) return
+      // Map our {title:'Hermes', body:'finished — …'} → native (message, title):
+      // native API takes the BODY as the message and the heading as the title.
+      const title = sanitizeOscText(notification.title)
+      const body = sanitizeOscText(notification.body ?? '')
+      if (!title) return
+      const message = body || title
+      try {
+        if (!seam.isDestroyed) seam.triggerNotification(message, title)
+      } catch (cause) {
+        getLog().warn('chrome', 'triggerNotification failed', { cause: String(cause) })
+      }
+    }
+  }
+}
+
+/** Raw control write through the renderer; falls back to process.stdout when
+ *  the renderer is already gone (the title-stack restore on destroy — at that
+ *  point there is no frame left to tear). */
+function writeRaw(seam: RendererSeam, chunk: string, options?: { evenIfDestroyed?: boolean }): void {
+  try {
+    if (!seam.isDestroyed) {
+      seam.writeOut(chunk)
+      return
+    }
+    if (options?.evenIfDestroyed) process.stdout.write(chunk)
+  } catch (cause) {
+    getLog().warn('chrome', 'control write failed', { cause: String(cause) })
+  }
+}

ui-opentui/src/entry/fakeGateway.ts

@@ -0,0 +1,64 @@
+/**
+ * FakeGateway — the test/dev implementation of GatewayService (spec v4 §2/§5
+ * Layer-3 seam). Provides an emittable event source and a spy `request`, so
+ * store/component tests can drive synthetic streams and assert RPC calls
+ * without spawning Python. Mirrors opencode's injectable fake transport.
+ *
+ * Phase 0 uses it to stream a scripted "hello" so the entry/test renders a
+ * non-empty frame. Phase 1 swaps in `liveGateway.layer` (real `tui_gateway`).
+ */
+import { Effect, Layer } from 'effect'
+
+import { GatewayService, type GatewayServiceShape } from '../boundary/gateway/GatewayService.ts'
+import type { GatewayEvent } from '../boundary/schema/GatewayEvent.ts'
+
+export interface FakeGatewayController {
+  readonly service: GatewayServiceShape
+  /** Emit a decoded event to all subscribers (drives the store in tests). */
+  readonly emit: (event: GatewayEvent) => void
+  /** Recorded (method, params) pairs from `request` calls. */
+  readonly calls: Array<{ method: string; params: unknown }>
+}
+
+/** Build a fresh fake controller (used directly in tests, or wrapped as a Layer). */
+export function makeFakeGateway(initialSessionId = 'fake-session'): FakeGatewayController {
+  const handlers = new Set<(event: GatewayEvent) => void>()
+  const calls: Array<{ method: string; params: unknown }> = []
+
+  const service: GatewayServiceShape = {
+    subscribe: handler =>
+      Effect.sync(() => {
+        handlers.add(handler)
+        return () => {
+          handlers.delete(handler)
+        }
+      }),
+    request: <A>(method: string, params: unknown) =>
+      Effect.sync(() => {
+        calls.push({ method, params })
+        return undefined as A
+      }),
+    sessionId: () => initialSessionId
+  }
+
+  return {
+    service,
+    emit: event => {
+      for (const handler of handlers) handler(event)
+    },
+    calls
+  }
+}
+
+/** A GatewayService layer backed by a fresh FakeGateway. The controller is
+ *  reachable for assertions via the returned tuple in tests; for the dev entry
+ *  use {@link fakeGatewayLayer} and drive it from a scripted effect. */
+export function fakeGatewayLayerWith(controller: FakeGatewayController): Layer.Layer<GatewayService> {
+  return Layer.succeed(GatewayService, controller.service)
+}
+
+/** Convenience: a layer + its controller, for the dev entry's scripted stream. */
+export function makeFakeGatewayLayer(): { layer: Layer.Layer<GatewayService>; controller: FakeGatewayController } {
+  const controller = makeFakeGateway()
+  return { layer: Layer.succeed(GatewayService, controller.service), controller }
+}

ui-opentui/src/entry/main.tsx

@@ -0,0 +1,770 @@
+/**
+ * Entry — the single boundary edge (spec v4 §3.1). This is the ONE place that:
+ *   - acquires the renderer (acquireRelease + Deferred-on-destroy),
+ *   - creates the Solid store,
+ *   - wires GatewayService.subscribe -> store.apply  (Effect->Solid contact #2),
+ *   - does the one-line `render(() => <App/>, renderer)` bridge (contact #1),
+ *   - (live) bootstraps a session and optionally submits an initial prompt,
+ *   - blocks until the renderer is destroyed (user quit),
+ * and at the bottom PROVIDES the layers and runs (`Effect.provide(AppLayer)`).
+ *
+ * Backend selection (import.meta.main):
+ *   - default            → the LIVE `liveGatewayLayer` (spawns the real Python
+ *     `tui_gateway`); after `gateway.ready` it `session.create`s and, if an
+ *     initial prompt is given (HERMES_TUI_PROMPT or argv), `prompt.submit`s it.
+ *     The composer lands in Phase 2 — until then the initial prompt is how a
+ *     streamed reply is driven into the transcript (spec Phase-1 smoke).
+ *   - HERMES_TUI_FAKE=1  → the scripted FakeGateway "hello" (offline dev/CI).
+ *
+ * The body of `run` does not change when the backend swaps — that's the point of
+ * the layer; only `makeAppLayer(...)` differs at the edge.
+ */
+import { createDefaultOpenTuiKeymap } from '@opentui/keymap/opentui'
+import { KeymapProvider } from '@opentui/keymap/solid'
+import { render } from '@opentui/solid'
+import { Deferred, Duration, Effect } from 'effect'
+import { writeFileSync } from 'node:fs'
+
+import { readClipboardImage, writeClipboard } from '../boundary/clipboard.ts'
+import { GatewayService, type GatewayServiceShape } from '../boundary/gateway/GatewayService.ts'
+import { liveGatewayLayer } from '../boundary/gateway/liveGateway.ts'
+import { getLog } from '../boundary/log.ts'
+import { startMemlog } from '../boundary/memlog.ts'
+import { startMemoryMonitor } from '../boundary/memoryMonitor.ts'
+import { startProactiveGc } from '../boundary/proactiveGc.ts'
+import { registerRemoteParsers } from '../boundary/parsers.ts'
+import { acquireRenderer } from '../boundary/renderer.ts'
+import { makeAppLayer } from '../boundary/runtime.ts'
+import { nthAssistantResponse } from '../logic/copy.ts'
+import { performHeapdump } from '../logic/diagnostics.ts'
+import {
+  envFlag,
+  heapdumpOnStart,
+  launchCwd,
+  noConfirmDestructive,
+  resolveMouseEnabled,
+  startupImage,
+  startupPrompt,
+  STARTUP_IMAGE_DEFAULT_PROMPT
+} from '../logic/env.ts'
+import { createPromptHistory, dirHistoryPersister, loadDirHistory } from '../logic/history.ts'
+import { parseProcessList } from '../logic/backgroundActivity.ts'
+import { createPasteStore } from '../logic/pastes.ts'
+import { mapResumeHistory } from '../logic/resume.ts'
+import {
+  classifySubmit,
+  catalogCommandItems,
+  createCompletionGate,
+  dispatchSlash,
+  mapCompletions,
+  mapModelOptions,
+  planCompletion,
+  readReplaceFrom,
+  registerModelPrefetch,
+  type SlashContext
+} from '../logic/slash.ts'
+import { createSessionStore, type SessionStore } from '../logic/store.ts'
+import { App } from '../view/App.tsx'
+import { seedLearnedNames } from '../view/composer.tsx'
+import { TerminalChrome } from '../view/terminalChrome.tsx'
+
+// Syntax-highlighting language expansion: register the remote tree-sitter
+// grammars (python/rust/go/bash/json/c/html/css/yaml/toml) before the first
+// <code>/<markdown> mount initializes the global tree-sitter client. Grammars
+// are fetched from GitHub on first use and cached under HERMES_TUI_PARSER_CACHE.
+registerRemoteParsers()
+import type { SessionPickerOps } from '../view/overlays/sessionPicker.tsx'
+import { ThemeProvider } from '../view/theme.tsx'
+import { makeFakeGatewayLayer, type FakeGatewayController } from './fakeGateway.ts'
+
+export interface TuiInput {
+  /** Mouse tracking on/off. */
+  readonly mouse: boolean
+  /** Skip the live session bootstrap (the fake backend drives the stream itself). */
+  readonly fake: boolean
+  /** Terminal width passed to `session.create` (Ink uses the live cols; 80 is a fine default). */
+  readonly cols: number
+  /** Optional initial prompt submitted once the session is ready — the Phase-1 stand-in for the composer. */
+  readonly initialPrompt?: string
+  /** Optional image PATH attached (image.attach) before the initial prompt — `hermes --tui --image <path>`. */
+  readonly initialImage?: string
+  /** Resume a session instead of creating one: a session id, 'recent'/'last'
+   *  (→ session.most_recent), or 'picker' (bare `--resume` — open the resume
+   *  picker BEFORE any session.create; create stays lazy). */
+  readonly resumeId?: string
+}
+
+const READY_POLL = Duration.millis(100)
+const READY_TIMEOUT_MS = 20_000
+/** Window after a Ctrl+C in which a second Ctrl+C quits the TUI (item 11). */
+const QUIT_WINDOW_MS = 3_000
+
+/** Recursive renderable count under a node (the /mem store-cap diagnostic —
+ *  same walk as scripts/mem-bench.tsx; cheap: one tree pass on demand). */
+function descendantCount(node: { getChildren(): unknown[] }): number {
+  let n = 0
+  for (const child of node.getChildren()) {
+    n += 1
+    if (child && typeof child === 'object' && 'getChildren' in child) {
+      n += descendantCount(child as { getChildren(): unknown[] })
+    }
+  }
+  return n
+}
+
+/**
+ * Resume a session INTO the store: buffer live events across the `session.resume`
+ * RPC, then replace history + replay (gotcha §8 #5 tool rows handled by
+ * mapResumeHistory). Shared by the launch bootstrap and the session switcher.
+ * Timed (rpc_ms / hydrate_ms) for the resume profile.
+ */
+/**
+ * Record the CURRENT session id in `HERMES_TUI_ACTIVE_SESSION_FILE` (item #5).
+ * The launcher reads this on exit to print the right "Resume this session with…"
+ * epilogue (hermes_cli/main.py `_print_tui_exit_summary`). The Ink TUI writes it on
+ * every session change (useSessionLifecycle.writeActiveSessionFile); the native
+ * engine must too, or the launcher falls back to the INITIAL launch session and
+ * shows resume info for the wrong session after a `/session` switch.
+ */
+const writeActiveSession = (sid: string | undefined) => {
+  const file = process.env.HERMES_TUI_ACTIVE_SESSION_FILE
+  if (!file || !sid) return
+  try {
+    writeFileSync(file, JSON.stringify({ session_id: sid }), { mode: 0o600 })
+  } catch (cause) {
+    getLog().warn('bootstrap', 'active-session-file write failed', { cause: String(cause) })
+  }
+}
+
+const resumeInto = (gateway: GatewayServiceShape, store: SessionStore, sid: string, cols: number) =>
+  Effect.gen(function* () {
+    writeActiveSession(sid) // the session we're switching to is now the active one (#5)
+    store.setSessionId(sid)
+    store.beginBuffer()
+    const t0 = Date.now()
+    const resumed = yield* gateway.request<{ messages?: unknown; info?: Record<string, unknown> }>('session.resume', {
+      cols,
+      session_id: sid,
+      // native engine renders tools collapsed → safe to fold each tool's capped
+      // result into the resume snapshot so resumed turns render like live (item 1).
+      with_tool_output: true
+    })
+    const t1 = Date.now()
+    const snapshot = mapResumeHistory(resumed?.messages)
+    store.commitSnapshot(snapshot)
+    if (resumed?.info) store.applyInfo(resumed.info)
+    getLog().info('bootstrap', 'session resumed', {
+      count: snapshot.length,
+      hydrate_ms: Date.now() - t1,
+      rpc_ms: t1 - t0,
+      sid
+    })
+  })
+
+/**
+ * Post-session setup, shared by every way a session comes to exist (create,
+ * boot resume, boot-picker pick): the tools/skills/MCP catalog for the home
+ * panel (item 9 — best-effort), the optional initial prompt, and the `/model`
+ * catalog prefetch (Epic 7 instant open: `model.options` is the slow RPC —
+ * network pricing fetch + Nous tier check — so pay it ONCE in an already-
+ * forked fiber; the promise is STASHED in the slash seam so an early `/model`
+ * awaits THIS request instead of doubling it).
+ */
+const postSessionSetup = (
+  gateway: GatewayServiceShape,
+  store: SessionStore,
+  sid: string,
+  initialPrompt?: string,
+  initialImage?: string
+) =>
+  Effect.gen(function* () {
+    const catalog = yield* gateway
+      .request<unknown>('startup.catalog', { session_id: sid })
+      .pipe(Effect.catchCause(() => Effect.succeed(undefined)))
+    if (catalog) store.setCatalog(catalog)
+
+    // Seed the composer's slash-highlight catalog ONCE at boot (glitch
+    // 2026-06-14): `commands.catalog` returns the full uncapped command + skill
+    // name list ({pairs:[["/name","desc"],…]}); feeding the names through
+    // seedLearnedNames means a cold `/command` highlights on the first keystroke
+    // instead of only after its completion batch was browsed earlier. Best-effort
+    // — a failure just leaves the old lazy-learn behavior.
+    const cmdCatalog = yield* gateway
+      .request<unknown>('commands.catalog', {})
+      .pipe(Effect.catchCause(() => Effect.succeed(undefined)))
+    seedLearnedNames(catalogCommandItems(cmdCatalog))
+
+    // Seeded image (`hermes --tui --image <path>`): attach BEFORE submitting, so
+    // the next prompt.submit picks it up — exact Ink parity (createGatewayEventHandler
+    // scheduleStartupPrompt: image.attach then submit; default prompt when image-only).
+    const image = initialImage?.trim()
+    if (image) {
+      yield* gateway.request('image.attach', { path: image, session_id: sid }).pipe(
+        Effect.catchCause(cause =>
+          Effect.sync(() => {
+            getLog().warn('bootstrap', 'startup image attach failed', { cause: String(cause) })
+            store.pushSystem(`startup image attach failed: ${String(cause)}`)
+          })
+        )
+      )
+    }
+
+    const prompt = initialPrompt?.trim() || (image ? STARTUP_IMAGE_DEFAULT_PROMPT : undefined)
+    if (prompt) {
+      store.pushUser(prompt)
+      yield* gateway.request('prompt.submit', { session_id: sid, text: prompt })
+    }
+
+    const prefetch = Effect.runPromise(
+      gateway
+        .request<unknown>('model.options', { session_id: sid })
+        .pipe(Effect.catchCause(() => Effect.succeed(undefined)))
+    ).then(modelOpts => {
+      const modelItems = mapModelOptions(modelOpts)
+      if (modelItems.length) store.setModelItems(modelItems)
+    })
+    registerModelPrefetch(prefetch)
+    yield* Effect.promise(() => prefetch)
+  })
+
+/** Create a FRESH session + run the post-session setup (the default boot path;
+ *  also the boot-picker's Esc fallback — closing the picker without a pick
+ *  must still leave a usable session behind). */
+const createFreshSession = (gateway: GatewayServiceShape, store: SessionStore, input: TuiInput) =>
+  Effect.gen(function* () {
+    const created = yield* gateway.request<{ session_id?: string; info?: Record<string, unknown> }>('session.create', {
+      cols: input.cols,
+      // The launch directory IS the workspace choice in a terminal (you cd'd
+      // here) — passing it makes the gateway treat it as explicit, so the
+      // session row gets a persisted cwd on first message, the chrome bar shows
+      // the right dir, and /sessions groups this directory's sessions first.
+      // NOT process.cwd(): the hermes launcher runs this engine with cwd set to
+      // its own package dir (ui-opentui), so process.cwd() would be the engine
+      // dir. The launcher exports the REAL launch dir as HERMES_CWD / the
+      // gateway's TERMINAL_CWD; prefer those, falling back to process.cwd()
+      // only when launched standalone (smokes/dev). (Desktop omits cwd — its
+      // launch dir is meaningless; see _ensure_session_db_row.)
+      cwd: launchCwd()
+    })
+    const sid = created?.session_id ?? gateway.sessionId()
+    if (!sid) {
+      getLog().warn('bootstrap', 'session.create returned no session_id')
+      return
+    }
+    if (created?.info) store.applyInfo(created.info)
+    writeActiveSession(sid) // record the new session for the launcher's exit epilogue (#5)
+    store.setSessionId(sid)
+    getLog().info('bootstrap', 'session created', { sid })
+    yield* postSessionSetup(gateway, store, sid, input.initialPrompt, input.initialImage)
+  })
+
+/**
+ * Live session bootstrap: wait for the unsolicited `gateway.ready` handshake,
+ * then either RESUME a session (hydrate its transcript — incl. tool rows — via
+ * the snapshot, buffering live events across the RPC), open the resume PICKER
+ * (`resumeId === 'picker'` — bare `--resume`: no session is created until the
+ * user picks or closes; create is lazy), or CREATE a fresh one, and (if given)
+ * submit the initial prompt. Forked into the entry scope so it runs
+ * concurrently with the render + the quit-await. Any failure is logged and
+ * swallowed — a bootstrap hiccup must never tear down the rendered UI.
+ */
+const bootstrapSession = (gateway: GatewayServiceShape, store: SessionStore, input: TuiInput) =>
+  Effect.gen(function* () {
+    const log = getLog()
+    let waited = 0
+    while (!store.state.ready && waited < READY_TIMEOUT_MS) {
+      yield* Effect.sleep(READY_POLL)
+      waited += 100
+    }
+    if (!store.state.ready) {
+      log.warn('bootstrap', 'no gateway.ready within timeout', { waited })
+      return
+    }
+
+    if (input.resumeId === 'picker') {
+      // Boot picker (design doc §A): opens BEFORE any session.create. The pick
+      // resumes via onResume (which then runs postSessionSetup); a close
+      // without a pick falls back to createFreshSession (onSessionPickerClosed).
+      store.openSessionPicker('recent')
+      return
+    }
+
+    if (input.resumeId) {
+      let sid: string | undefined = input.resumeId
+      if (sid === 'recent' || sid === 'last') {
+        const recent = yield* gateway.request<{ session_id?: string }>('session.most_recent', {})
+        sid = recent?.session_id
+      }
+      if (!sid) {
+        log.warn('bootstrap', 'no session to resume', { resumeId: input.resumeId })
+        return
+      }
+      yield* resumeInto(gateway, store, sid, input.cols)
+      yield* postSessionSetup(gateway, store, sid, input.initialPrompt, input.initialImage)
+      return
+    }
+
+    yield* createFreshSession(gateway, store, input)
+  }).pipe(Effect.catchCause(cause => Effect.sync(() => getLog().warn('bootstrap', 'failed', { cause: String(cause) }))))
+
+/** The entry Effect. Mirrors opencode `app.tsx:177` `run = Effect.fn("Tui.run")`. */
+export const run = Effect.fn('Tui.run')(function* (input: TuiInput) {
+  yield* Effect.scoped(
+    Effect.gen(function* () {
+      // Solid side: the store + reducer. Created here, lives in Solid-land.
+      const store = createSessionStore()
+
+      // Prompt history (item 6): scoped to the launch directory so prior prompts
+      // from the same project dir are recallable (Up/Down), without bleeding
+      // across different dirs. process.cwd() is the user's launch dir under the
+      // real launcher.
+      const historyCwd = process.cwd()
+      const history = createPromptHistory({
+        initial: loadDirHistory(historyCwd),
+        persist: dirHistoryPersister(historyCwd)
+      })
+
+      // Pasted-text store — created ONCE here so it survives the composer
+      // remounting (overlay open/close); a per-composer store would lose a
+      // pending `[Pasted text #N]` mid-compose and submit would send it literally.
+      const pasteStore = createPasteStore()
+
+      // Contact point #2: boundary pushes decoded events into the Solid store.
+      // The callback ALSO drives auto-heal re-resume: a post-crash gateway.ready
+      // (i.e. one that follows a gateway.exited, so `recoverSid` is set) re-resumes
+      // the session so the transcript continues. The INITIAL gateway.ready has
+      // `recoverSid === undefined`, so the normal bootstrap path is untouched.
+      const gateway = yield* GatewayService
+      let recoverSid: string | undefined
+      yield* gateway.subscribe(event => {
+        store.apply(event)
+        if (event.type === 'gateway.exited') {
+          recoverSid = gateway.sessionId() ?? recoverSid
+        } else if (event.type === 'gateway.ready' && recoverSid !== undefined) {
+          const sid = recoverSid
+          recoverSid = undefined
+          Effect.runFork(
+            resumeInto(gateway, store, sid, input.cols).pipe(
+              Effect.catchCause(cause =>
+                Effect.sync(() => getLog().warn('recover', 'resume failed', { cause: String(cause) }))
+              )
+            )
+          )
+        }
+      })
+
+      // ── Ctrl+C state machine (item 11) ──────────────────────────────────
+      // While a turn runs, the first Ctrl+C STOPS the agent (session.interrupt);
+      // a second Ctrl+C within QUIT_WINDOW_MS (or when idle) KILLS the TUI. The
+      // debounce stops a stray Ctrl+C from nuking the session (opencode's
+      // double-press model; the user's preferred behaviour).
+      let quitArmed = false
+      let quitTimer: ReturnType<typeof setTimeout> | undefined
+      let doQuit = () => {} // assigned once the renderer exists
+      const disarmQuit = () => {
+        quitArmed = false
+        if (quitTimer) clearTimeout(quitTimer)
+        quitTimer = undefined
+        store.setHint(undefined)
+      }
+      const armQuit = (message: string) => {
+        quitArmed = true
+        store.setHint(message)
+        if (quitTimer) clearTimeout(quitTimer)
+        quitTimer = setTimeout(disarmQuit, QUIT_WINDOW_MS)
+      }
+      const interruptTurn = () => {
+        const sid = gateway.sessionId()
+        if (!sid) return
+        Effect.runFork(
+          gateway
+            .request('session.interrupt', { session_id: sid })
+            .pipe(
+              Effect.catchCause(cause =>
+                Effect.sync(() => getLog().warn('interrupt', 'failed', { cause: String(cause) }))
+              )
+            )
+        )
+      }
+      const onCtrlC = () => {
+        if (quitArmed) {
+          disarmQuit()
+          doQuit()
+          return
+        }
+        if (store.state.info.running) {
+          interruptTurn()
+          armQuit('⏹ stopped — Ctrl+C again to quit')
+        } else {
+          armQuit('Ctrl+C again to quit')
+        }
+      }
+
+      // Transient hint that auto-clears (used by copy/image-paste feedback).
+      const flashHint = (message: string, ms = 1500) => {
+        store.setHint(message)
+        setTimeout(() => {
+          if (store.state.hint === message) store.setHint(undefined)
+        }, ms)
+      }
+
+      // Copy a mouse selection to the clipboard (item 1) — OSC 52 + native command.
+      // Copies exactly the rendered text the user highlighted (markers are concealed
+      // in the pretty render; the `/copy` command copies a full response's source).
+      const onCopySelection = (text: string) => {
+        void writeClipboard(text)
+        flashHint('Copied selection')
+      }
+
+      // Paste an IMAGE (item 1): read the clipboard image and attach it to the
+      // session (image.attach_bytes); the next prompt.submit picks it up.
+      const onImagePaste = () => {
+        void (async () => {
+          const img = await readClipboardImage()
+          if (!img) {
+            flashHint('No image in clipboard', 2000)
+            return
+          }
+          const sid = gateway.sessionId()
+          if (!sid) {
+            flashHint('No session for image', 2000)
+            return
+          }
+          try {
+            await Effect.runPromise(
+              gateway.request('image.attach_bytes', {
+                content_base64: img.data,
+                filename: 'pasted.png',
+                session_id: sid
+              })
+            )
+            flashHint('🖼 image attached — type a message and send', 3000)
+          } catch {
+            flashHint('Image attach failed', 2000)
+          }
+        })()
+      }
+
+      // A blocking prompt owns Ctrl+C (→ cancel); otherwise the state machine above runs.
+      const { renderer, shutdown } = yield* acquireRenderer({
+        mouse: input.mouse,
+        isBlocked: () => store.state.prompt !== undefined,
+        onCtrlC,
+        onCopySelection
+      })
+      // Fleet memory self-sampling (HERMES_TUI_MEMLOG / diagnostics master
+      // switch — boundary/memlog.ts). Scoped acquire→release like the renderer.
+      const stopMemlog = startMemlog()
+      yield* Effect.addFinalizer(() => Effect.sync(stopMemlog))
+      // Proactive idle GC (W2) — opt-in via a low HERMES_TUI_HEAP_MB (no-op on
+      // the default path). Idle-gated on the store's streaming flag so it never
+      // collects mid-reply. Scoped release like memlog.
+      const proactiveGc = startProactiveGc(() => store.state.info.running === true)
+      yield* Effect.addFinalizer(() => Effect.sync(proactiveGc.stop))
+      // Memory early-warning (#34095 parity) — surfaces a transcript system line
+      // when heap climbs abnormally fast below the OOM ceiling (the silent-death
+      // regime). ON by default: a KB user-facing safety heads-up, not a
+      // diagnostic dump. No auto heap-snapshot (memlog is the diagnosis path).
+      const stopMemoryMonitor = startMemoryMonitor(line => store.pushSystem(line))
+      yield* Effect.addFinalizer(() => Effect.sync(stopMemoryMonitor))
+      // HERMES_HEAPDUMP_ON_START (Ink parity): a deliberate baseline snapshot at
+      // boot. Bypasses the diagnostics master switch (you set it on purpose).
+      // Best-effort + synchronous (writeHeapSnapshot blocks V8) — a failure must
+      // never block launch.
+      if (heapdumpOnStart()) {
+        try {
+          const dump = performHeapdump()
+          store.pushSystem(`heap snapshot written: ${dump.path}`)
+        } catch (cause) {
+          getLog().warn('bootstrap', 'heapdump-on-start failed', { cause: String(cause) })
+        }
+      }
+      doQuit = () => {
+        if (!renderer.isDestroyed) renderer.destroy()
+      }
+
+      // Native keymap host (Phase 3): one keymap bound to this renderer, provided
+      // to the whole Solid tree via <KeymapProvider>. Overlays/prompts register
+      // close (and confirm) layers against it through useCloseLayer/useBindings.
+      const keymap = createDefaultOpenTuiKeymap(renderer)
+
+      // Submit a user turn: the service value is in hand, so `gateway.request(...)`
+      // is Effect<…, never> — fire it detached with runFork; failures are logged.
+      const submitPrompt = (text: string) => {
+        store.pushUser(text)
+        const sid = gateway.sessionId()
+        if (!sid) {
+          getLog().warn('submit', 'no session yet — dropping prompt', { text })
+          return
+        }
+        Effect.runFork(
+          gateway
+            .request('prompt.submit', { session_id: sid, text })
+            .pipe(
+              Effect.catchCause(cause => Effect.sync(() => getLog().warn('submit', 'failed', { cause: String(cause) })))
+            )
+        )
+      }
+
+      // `!cmd` — run a shell command directly (Ink/free-code parity: F9). The
+      // gateway's `shell.exec` runs it (30s timeout, dangerous/hardline guards)
+      // and returns {stdout, stderr, code}; we echo the invocation as a user line
+      // and the combined output (or the error / non-zero exit) as a system line.
+      // No model turn — this never hits prompt.submit. Detached like submitPrompt.
+      const runShell = (cmd: string) => {
+        if (!cmd) return
+        store.pushUser(`!${cmd}`)
+        Effect.runFork(
+          gateway.request<{ stdout?: string; stderr?: string; code?: number }>('shell.exec', { command: cmd }).pipe(
+            Effect.tap(r =>
+              Effect.sync(() => {
+                const out = [r.stdout, r.stderr].filter(Boolean).join('\n').trimEnd()
+                if (out) store.pushSystem(out)
+                if ((r.code ?? 0) !== 0 || !out) store.pushSystem(`exit ${r.code ?? 0}`)
+              })
+            ),
+            Effect.catchCause(cause =>
+              Effect.sync(() => {
+                getLog().warn('shell', 'failed', { cause: String(cause) })
+                store.pushSystem(`error: ${String(cause)}`)
+              })
+            )
+          )
+        )
+      }
+
+      // Resume a chosen session (resume picker pick or `/resume <id>` direct
+      // path) — the same hydrate path as launch. When the picker was the BOOT
+      // surface (bare `--resume`), no create ever ran, so the post-session
+      // setup (catalog, /model prefetch) runs here exactly once.
+      const onResume = (resumeSid: string) => {
+        Effect.runFork(
+          Effect.gen(function* () {
+            yield* resumeInto(gateway, store, resumeSid, input.cols)
+            if (!store.state.catalog) yield* postSessionSetup(gateway, store, resumeSid)
+          }).pipe(
+            Effect.catchCause(cause => Effect.sync(() => getLog().warn('resume', 'failed', { cause: String(cause) })))
+          )
+        )
+      }
+
+      // The resume picker's gateway calls (view/overlays/sessionPicker.tsx).
+      // `rename` goes through `session.title` — the existing title RPC (it
+      // reaches only LIVE gateway sessions; the picker surfaces rejections).
+      const sessionOps: SessionPickerOps = {
+        list: params => Effect.runPromise(gateway.request('session.list', params)),
+        peek: sessionId => Effect.runPromise(gateway.request('session.peek', { session_id: sessionId })),
+        rename: (sessionId, title) =>
+          Effect.runPromise(gateway.request('session.title', { session_id: sessionId, title })).then(() => undefined)
+      }
+
+      // The background-process panel's gateway calls (view/overlays/backgroundPanel.tsx):
+      // `agents.list` lists the OS process registry; `process.stop` kills ALL of them
+      // (the gateway exposes kill-all only — no per-process RPC, hence no per-row kill).
+      const backgroundOps = {
+        list: () => Effect.runPromise(gateway.request('agents.list', {})).then(parseProcessList),
+        stopAll: () => Effect.runPromise(gateway.request('process.stop', {})).then(() => undefined)
+      }
+
+      // Boot-picker Esc fallback: the picker closed without a pick and no
+      // session exists yet (bare `--resume` launch) — create a fresh one so
+      // the composer has somewhere to send prompts.
+      const onSessionPickerClosed = () => {
+        if (gateway.sessionId()) return
+        Effect.runFork(
+          createFreshSession(gateway, store, input).pipe(
+            Effect.catchCause(cause =>
+              Effect.sync(() => getLog().warn('bootstrap', 'post-picker create failed', { cause: String(cause) }))
+            )
+          )
+        )
+      }
+
+      // Slash dispatch context (Solid logic; the boundary just hands it a
+      // Promise-returning `request` + the host capabilities it needs).
+      const slashCtx: SlashContext = {
+        clearTranscript: () => store.clearTranscript(),
+        compact: () => store.state.compact,
+        setCompact: on => store.setCompact(on),
+        details: () => store.state.details,
+        setDetails: mode => store.setDetails(mode),
+        renderableCount: () => {
+          try {
+            return descendantCount(renderer.root)
+          } catch {
+            return undefined
+          }
+        },
+        // HERMES_TUI_NO_CONFIRM (Ink parity): skip the destructive-action confirm
+        // step and run the action immediately. Read per call so a wrapper that
+        // mutates env before launch sees the live value.
+        confirm: (message, onConfirm) => (noConfirmDestructive() ? onConfirm() : store.setConfirm(message, onConfirm)),
+        copyResponse: n => {
+          const text = nthAssistantResponse(store.state.messages, n)
+          if (!text) return false
+          void writeClipboard(text)
+          flashHint(n > 1 ? `Copied response #${n} to clipboard` : 'Copied response to clipboard')
+          return true
+        },
+        modelItems: () => store.state.modelItems,
+        setModelItems: items => store.setModelItems(items),
+        logTail: () =>
+          getLog()
+            .tail(200)
+            .map(e => `${e.scope}: ${e.msg}`),
+        openDashboard: () => store.openDashboard(),
+        openBackgroundPanel: () => store.openBackgroundPanel(),
+        addBgTask: id => store.addBgTask(id),
+        openPager: (title, text) => store.openPager(title, text),
+        openPicker: picker => store.openPicker(picker),
+        openSessionPicker: tab => store.openSessionPicker(tab),
+        resumeSession: onResume,
+        pushSystem: text => store.pushSystem(text),
+        quit: () => {
+          if (!renderer.isDestroyed) renderer.destroy()
+        },
+        request: (method, params) => Effect.runPromise(gateway.request(method, params)),
+        sessionId: () => gateway.sessionId(),
+        submit: submitPrompt
+      }
+
+      // The composer's submit: `!cmd` runs a shell command (F9), `/command`
+      // routes through the slash ladder, else a prompt turn.
+      const submit = (text: string) => {
+        const route = classifySubmit(text)
+        if (route.kind === 'shell') runShell(route.payload)
+        else if (route.kind === 'slash') void dispatchSlash(route.payload, slashCtx)
+        else submitPrompt(route.payload)
+      }
+
+      // Live completions (items 5 + 13): a `/command [args]` line queries
+      // `complete.slash` (the gateway completes names AND args); a trailing
+      // path-like word queries `complete.path` (file/@-mention tagging). The
+      // accepted item replaces from the gateway's `replace_from` (or the token
+      // start), so only the relevant token is spliced — not the whole line.
+      // Fired per keystroke (a debounce is a polish item).
+      //
+      // Out-of-order guard (glitch 2026-06-14): the gateway transport does NOT
+      // guarantee in-order response delivery, and these RPCs fire per keystroke
+      // with no debounce — a slow earlier `complete.slash` could resolve AFTER a
+      // later `@`-mention `complete.path` and clobber the store, blanking the
+      // `@` dropdown ("a leading /path message breaks @-mentions afterward").
+      // The completion gate (claimed on EVERY call, before the clear branch, so
+      // an intermediate keystroke that fires no RPC still invalidates the older
+      // in-flight one) drops any response a newer keystroke has superseded.
+      const completionGate = createCompletionGate()
+      const onType = (text: string, cursor: number = text.length) => {
+        const token = completionGate.claim()
+        const plan = planCompletion(text, cursor)
+        if (!plan) {
+          store.clearCompletions()
+          return
+        }
+        Effect.runPromise(gateway.request(plan.method, plan.params))
+          .then(result => {
+            if (!completionGate.isCurrent(token)) return // a newer keystroke superseded this query
+            store.setCompletions(mapCompletions(result), readReplaceFrom(result, plan.from))
+          })
+          .catch(() => {
+            if (!completionGate.isCurrent(token)) return
+            store.clearCompletions()
+          })
+      }
+
+      // Blocking-prompt replies (clarify/approval/sudo/secret `*.respond`). Same
+      // detached-runFork pattern; failures logged, never thrown into the view.
+      const respond = (method: string, params: Record<string, unknown>) => {
+        Effect.runFork(
+          gateway
+            .request(method, params)
+            .pipe(
+              Effect.catchCause(cause =>
+                Effect.sync(() => getLog().warn('respond', 'failed', { cause: String(cause), method }))
+              )
+            )
+        )
+      }
+
+      // Live backend: drive a session (create + optional initial prompt) concurrently.
+      if (!input.fake) yield* Effect.forkScoped(bootstrapSession(gateway, store, input))
+
+      // (No ambient OS-process poll: the `bg:` badge now counts in-flight
+      // background-PROMPT tasks from the event stream, and the /processes panel
+      // fetches `agents.list` on open. Nothing to poll for.)
+
+      // Contact point #1: the single render bridge. After this, the screen is Solid's.
+      // The theme is sourced reactively from the store (skin events update it).
+      yield* Effect.promise(() =>
+        render(
+          () => (
+            <KeymapProvider keymap={keymap}>
+              <ThemeProvider theme={() => store.state.theme}>
+                <TerminalChrome store={store} />
+                <App
+                  store={store}
+                  onSubmit={submit}
+                  onType={onType}
+                  onRespond={respond}
+                  onResume={onResume}
+                  sessionOps={sessionOps}
+                  onSessionPickerClosed={onSessionPickerClosed}
+                  sessionId={() => gateway.sessionId()}
+                  history={history}
+                  onImagePaste={onImagePaste}
+                  pasteStore={pasteStore}
+                  backgroundOps={backgroundOps}
+                />
+              </ThemeProvider>
+            </KeymapProvider>
+          ),
+          renderer
+        )
+      )
+
+      // Block until the renderer is destroyed (Ctrl+C / quit); finalizers then run.
+      yield* Deferred.await(shutdown)
+    })
+  )
+})
+
+/** Scripted "hello" stream so the fake backend paints a non-empty frame offline. */
+function streamHello(controller: FakeGatewayController): void {
+  controller.emit({ type: 'gateway.ready' })
+  controller.emit({ type: 'message.start' })
+  for (const chunk of ['Hi ', 'there, ', 'glitch!']) {
+    controller.emit({ type: 'message.delta', payload: { text: chunk } })
+  }
+  controller.emit({ type: 'message.complete' })
+}
+
+if (import.meta.main) {
+  const fake = envFlag(process.env.HERMES_TUI_FAKE, false)
+  const cols = process.stdout.columns || 80
+  // `hermes --tui "prompt"` / `--image` seed: the launcher sets HERMES_TUI_QUERY
+  // (+ HERMES_TUI_IMAGE); we also honor HERMES_TUI_PROMPT (OpenTUI alias) and a
+  // bare argv tail (standalone dev). See logic/env.ts startupPrompt/startupImage.
+  const initialPrompt = startupPrompt()
+  const initialImage = startupImage()
+  const resumeId = process.env.HERMES_TUI_RESUME?.trim()
+  // Mouse on by default. Defers to Ink's env surface (HERMES_TUI_MOUSE_TRACKING >
+  // HERMES_TUI_DISABLE_MOUSE > HERMES_TUI_MOUSE alias > default on). See env.ts.
+  const mouse = resolveMouseEnabled()
+  const base = { mouse, fake, cols }
+  const withPrompt = initialPrompt ? { ...base, initialPrompt } : base
+  const withImage = initialImage ? { ...withPrompt, initialImage } : withPrompt
+  const input: TuiInput = resumeId ? { ...withImage, resumeId } : withImage
+
+  const onFatal = (error: unknown) => {
+    getLog().error('entry', 'fatal', { error: String(error) })
+    process.exitCode = 1
+  }
+
+  if (fake) {
+    const { layer, controller } = makeFakeGatewayLayer()
+    // Drive the fake stream shortly after mount so the subscription is live.
+    setTimeout(() => streamHello(controller), 50)
+    Effect.runPromise(run(input).pipe(Effect.provide(makeAppLayer(layer)))).catch(onFatal)
+  } else {
+    Effect.runPromise(run(input).pipe(Effect.provide(makeAppLayer(liveGatewayLayer)))).catch(onFatal)
+  }
+}

ui-opentui/src/logic/backgroundActivity.ts

@@ -0,0 +1,129 @@
+/**
+ * Background-activity logic — pure parsers + derive helpers for the "ambient
+ * activity" feature (notifications, long-running processes, background runs).
+ * No state container here: the store owns the arrays; these functions parse
+ * loose wire payloads (everything off the gateway is `unknown`) and compute
+ * derived values over immutable arrays. Mirrors the defensive loose-read style
+ * of `logic/slash.ts` (`readStr`) and the snake_case→camel mapping the wire
+ * needs.
+ *
+ * Wire shapes (see boundary/schema/GatewayEvent.ts ~134):
+ *   notification.show  payload {text, level, kind, ttl_ms, key, id}  (loose Record)
+ *   notification.clear payload {key}
+ *   agents.list result {processes:[{session_id, command, status, uptime_seconds}]}
+ */
+
+export interface ActivityNotification {
+  id: string
+  key?: string
+  text: string
+  level: 'info' | 'warn' | 'error' | 'success'
+  kind: string
+  ttlMs?: number
+}
+
+export interface BackgroundProcess {
+  sessionId: string
+  command: string
+  status: string
+  uptimeSeconds: number
+}
+
+/** Loose-read a string field off an `unknown` object (slash.ts `readStr` style). */
+function readStr(value: unknown, key: string): string | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'string' ? v : undefined
+}
+
+/** Loose-read a finite number off an `unknown` object. */
+function readNum(value: unknown, key: string): number | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'number' && Number.isFinite(v) ? v : undefined
+}
+
+/** Coerce any wire `level` to the closed union; anything that isn't a known
+ *  level (absent, garbage, wrong-typed) falls back to 'info'. */
+function coerceLevel(value: unknown): ActivityNotification['level'] {
+  return value === 'warn' || value === 'error' || value === 'success' ? value : 'info'
+}
+
+/** A chrome notice (status-bar banner with lifecycle), distinguished from an
+ *  inline card by its lifecycle kind. Credits/usage notices set kind sticky|ttl;
+ *  process/background cards use label kinds (process.complete, etc.). */
+export function isChromeNotice(n: ActivityNotification): boolean {
+  return n.kind === 'sticky' || n.kind === 'ttl'
+}
+
+/**
+ * Parse a `notification.show` payload (unknown) → ActivityNotification, or null
+ * when there's no usable text (text is the load-bearing field — without it the
+ * card has nothing to show). Maps snake_case `ttl_ms` → `ttlMs`, coerces a
+ * garbage/missing `level` to 'info', and defaults `kind` to ''.
+ *
+ * id resolution: prefer the wire `id`, then fall back to `key`, else synthesize
+ * `id = `n:${text}`` (a stable, text-derived id rather than a random one). The
+ * original `key` (if any) is preserved separately so notification.clear by key
+ * still targets the right cards.
+ */
+export function parseNotification(payload: unknown): ActivityNotification | null {
+  const text = readStr(payload, 'text')
+  if (!text) return null
+  const key = readStr(payload, 'key')
+  const id = readStr(payload, 'id') ?? key ?? `n:${text}`
+  const out: ActivityNotification = {
+    id,
+    kind: readStr(payload, 'kind') ?? '',
+    level: coerceLevel((payload as { level?: unknown } | null | undefined)?.level),
+    text
+  }
+  if (key !== undefined) out.key = key
+  const ttlMs = readNum(payload, 'ttl_ms')
+  if (ttlMs !== undefined) out.ttlMs = ttlMs
+  return out
+}
+
+/** Parse an `agents.list` result ({processes:[...]}) → BackgroundProcess[],
+ *  skipping malformed rows (a row missing session_id/command is dropped, not
+ *  defaulted). snake_case `session_id`/`uptime_seconds` → camelCase; a missing
+ *  uptime defaults to 0, a missing status to ''. */
+export function parseProcessList(result: unknown): BackgroundProcess[] {
+  if (!result || typeof result !== 'object') return []
+  const processes = (result as { processes?: unknown }).processes
+  if (!Array.isArray(processes)) return []
+  const out: BackgroundProcess[] = []
+  for (const row of processes) {
+    const sessionId = readStr(row, 'session_id')
+    const command = readStr(row, 'command')
+    if (!sessionId || !command) continue
+    out.push({
+      command,
+      sessionId,
+      status: readStr(row, 'status') ?? '',
+      uptimeSeconds: readNum(row, 'uptime_seconds') ?? 0
+    })
+  }
+  return out
+}
+
+/** Terminal (no-longer-running) process statuses. A process whose status is
+ *  NOT one of these is treated as running — leniently, because the gateway's
+ *  status vocabulary is open-ended and we'd rather over-count the ambient badge
+ *  than silently hide a still-live process under an unfamiliar status string.
+ *  Matched case-insensitively after trimming. */
+/** Terminal (no-longer-running) process statuses — exported as the single
+ *  source of truth (the panel imports `procIsRunning` rather than re-declaring). */
+export const DONE_STATUSES = new Set(['exited', 'failed', 'complete', 'done', 'killed'])
+
+/** Whether a process status is "running-ish": NOT one of DONE_STATUSES. Lenient
+ *  by design — the gateway's status vocabulary is open-ended, so we over-count
+ *  rather than hide a live process under an unfamiliar status. Case-insensitive. */
+export function procIsRunning(status: string): boolean {
+  return !DONE_STATUSES.has(status.trim().toLowerCase())
+}
+
+/** Count of running background processes (the ambient `bg:` badge). */
+export function runningCount(procs: readonly BackgroundProcess[]): number {
+  return procs.filter(p => procIsRunning(p.status)).length
+}

ui-opentui/src/logic/completionMenu.ts

@@ -0,0 +1,66 @@
+/**
+ * Completion-menu key routing (Epic 8) — the pure decision table for the
+ * composer's completions dropdown, kept out of the view so the precedence
+ * rules are unit-testable.
+ *
+ * Precedence (the hard part):
+ *   - Tab accepts the highlighted item and Esc dismisses whenever ANY menu is
+ *     open (slash-command OR path/@-mention) — the pre-Epic-8 semantics.
+ *   - Up/Down move the highlight (wrapping) and Enter accepts it ONLY for the
+ *     SLASH menu (the composer's first token starts with `/`). On a path menu
+ *     — or with a Ctrl/Alt-modified key — they `pass`, keeping their existing
+ *     meanings (prompt history, cursor moves, textarea submit).
+ *   - A closed menu (`count === 0`) always passes.
+ *
+ * The caller owns the side effects: `move` updates the selection signal,
+ * `accept` splices the item into the composer (then arg-completion continues
+ * as before), `dismiss` clears the candidates, `pass` falls through to the
+ * history/cursor handling.
+ */
+
+/** Max dropdown rows shown (the view slices candidates to this). */
+export const MENU_MAX = 8
+
+export interface MenuKeyContext {
+  /** Number of VISIBLE candidates (already capped at MENU_MAX). */
+  count: number
+  /** The currently highlighted row. */
+  selected: number
+  /** Whether this is the slash-command menu (composer text starts with `/`). */
+  slashMenu: boolean
+}
+
+export type MenuKeyAction =
+  | { kind: 'move'; selected: number }
+  | { kind: 'accept'; index: number }
+  | { kind: 'dismiss' }
+  | { kind: 'pass' }
+
+const PASS: MenuKeyAction = { kind: 'pass' }
+
+/** Clamp the selection into the visible range (a shrunken list can strand it). */
+function clampSelected(ctx: MenuKeyContext): number {
+  return Math.min(Math.max(0, ctx.selected), ctx.count - 1)
+}
+
+/**
+ * Route one key press against the open menu. `modified` is Ctrl/Alt/Option —
+ * modified arrows/Enter never belong to the menu (Tab/Esc keep their
+ * pre-existing modifier-blind accept/dismiss semantics).
+ *
+ * ANY open menu owns plain arrows/Enter (glitch, 2026-06-10): @-path and
+ * arg menus navigate exactly like the slash menu — standard editor-
+ * autocomplete behavior; Esc dismisses to hand the cursor keys back.
+ * (`ctx.slashMenu` still feeds the hint text + suggestion rows.)
+ */
+export function routeMenuKey(name: string, modified: boolean, ctx: MenuKeyContext): MenuKeyAction {
+  if (ctx.count <= 0) return PASS
+  if (name === 'tab') return { index: clampSelected(ctx), kind: 'accept' }
+  if (name === 'escape') return { kind: 'dismiss' }
+  if (modified) return PASS
+  const sel = clampSelected(ctx)
+  if (name === 'up') return { kind: 'move', selected: (sel - 1 + ctx.count) % ctx.count }
+  if (name === 'down') return { kind: 'move', selected: (sel + 1) % ctx.count }
+  if (name === 'return') return { index: sel, kind: 'accept' }
+  return PASS
+}

ui-opentui/src/logic/copy.ts

@@ -0,0 +1,41 @@
+/**
+ * Assistant-text extraction (the `/copy [n]` command's pure logic). An assistant
+ * turn's answer lives in `parts` (the `type:'text'` fragments, concatenated) while
+ * live, OR in `.text` once settled/resumed. We copy the ANSWER only — reasoning and
+ * tool parts are excluded. `nthAssistantResponse` indexes newest-first (1-based).
+ *
+ * NB: mouse-selection copies the RENDERED text verbatim (native OpenTUI selection,
+ * `selection.getSelectedText()`), not markdown source — markers are concealed in the
+ * pretty render and can't be recovered from a partial selection (user's choice). The
+ * source-bearing path is this `/copy` command, which copies a whole response's source.
+ */
+import type { Message } from './store.ts'
+
+/** The answer text of one message: concat the `text` parts (trimmed) when live, else `.text`. */
+export function messageText(m: Message): string {
+  if (m.parts && m.parts.length) {
+    return m.parts
+      .filter(p => p.type === 'text')
+      .map(p => p.text)
+      .join('')
+      .trim()
+  }
+  return m.text
+}
+
+/** Newest-first list of the non-empty answer text for every assistant message. */
+export function assistantResponses(messages: Message[]): string[] {
+  const out: string[] = []
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const m = messages[i]
+    if (!m || m.role !== 'assistant') continue
+    const text = messageText(m)
+    if (text) out.push(text)
+  }
+  return out
+}
+
+/** The n-th newest assistant response (1-based; n=1 → last). `undefined` if out of range. */
+export function nthAssistantResponse(messages: Message[], n: number): string | undefined {
+  return assistantResponses(messages)[n - 1]
+}

ui-opentui/src/logic/defer.ts

@@ -0,0 +1,12 @@
+/**
+ * deferClose — defer an overlay/prompt close by one tick.
+ *
+ * Overlays REPLACE the composer (a `<Switch>`), so when one closes the composer
+ * remounts + refocuses. Running the close on the NEXT tick lets the current
+ * key/close event (Esc/q/Enter/y/select) finish dispatching first, so the
+ * keystroke that triggered the close can't leak into the freshly-focused
+ * composer (e.g. `/clear`→y once left a stray "y" in the input).
+ */
+export function deferClose(fn: () => void): void {
+  setTimeout(fn, 0)
+}

ui-opentui/src/logic/details.ts

@@ -0,0 +1,84 @@
+/**
+ * Global detail-mode logic (/details — Epic 3 utility-command port; mirrors Ink
+ * `domain/details.ts`, GLOBAL mode only — per-section overrides are explicitly
+ * deferred). The mode drives how the transcript treats tool + reasoning rows:
+ *
+ *   - `collapsed` (default): today's behaviour — headers with click-to-expand.
+ *   - `expanded`: tool bodies + settled reasoning previews default-OPEN.
+ *   - `hidden`: tool/reasoning runs reduce to ONE muted `⚡ N tools hidden`-style
+ *     line per run (never silently dropped — flipping the mode back restores,
+ *     since the parts stay in the store untouched).
+ *
+ * Pure data + functions; the store carries the flag, messageLine/toolPart/
+ * reasoningPart read it via the display context.
+ */
+import type { Part } from './store.ts'
+
+export type DetailsMode = 'hidden' | 'collapsed' | 'expanded'
+
+/** Cycle order (Ink parity: hidden → collapsed → expanded → hidden). */
+export const DETAILS_MODES = ['hidden', 'collapsed', 'expanded'] as const
+
+/** Gateway `complete.slash` suggests these per-section names after `/details ` —
+ *  recognized so picking one yields an honest "not supported yet" notice instead
+ *  of the generic usage line (per-section overrides are deferred). */
+export const DETAILS_SECTIONS = ['thinking', 'tools', 'subagents', 'activity'] as const
+
+export const DETAILS_USAGE = 'usage: /details [hidden|collapsed|expanded|cycle]'
+
+/** Parse a mode word; null for anything unrecognized (non-strings included). */
+export function parseDetailsMode(v: unknown): DetailsMode | null {
+  if (typeof v !== 'string') return null
+  const norm = v.trim().toLowerCase()
+  return DETAILS_MODES.find(m => m === norm) ?? null
+}
+
+/** The next mode in the cycle (`/details cycle`). */
+export function nextDetailsMode(m: DetailsMode): DetailsMode {
+  return DETAILS_MODES[(DETAILS_MODES.indexOf(m) + 1) % DETAILS_MODES.length] ?? 'collapsed'
+}
+
+/** One collapsed RUN of consecutive tool/reasoning parts (hidden mode). */
+export interface HiddenRun {
+  type: 'hiddenRun'
+  /** Stable-ish key: the first hidden part's id. */
+  id: string
+  tools: number
+  thoughts: number
+}
+
+/** What the transcript renders per part slot: a real part, or a hidden-run marker. */
+export type DisplayPart = Part | HiddenRun
+
+/**
+ * Hidden mode: keep text parts, fold each consecutive run of tool/reasoning
+ * parts into ONE HiddenRun marker (so a 5-tool fan-out reads as a single muted
+ * line, not 5 of them). Pure — the source parts are untouched, so switching
+ * the mode back restores everything.
+ */
+export function collapseHiddenParts(parts: readonly Part[]): DisplayPart[] {
+  const out: DisplayPart[] = []
+  let run: HiddenRun | undefined
+  for (const part of parts) {
+    if (part.type === 'text') {
+      run = undefined
+      out.push(part)
+      continue
+    }
+    if (!run) {
+      run = { id: `hidden-${part.id}`, thoughts: 0, tools: 0, type: 'hiddenRun' }
+      out.push(run)
+    }
+    if (part.type === 'tool') run.tools += 1
+    else run.thoughts += 1
+  }
+  return out
+}
+
+/** Muted one-liner for a hidden run: `2 tools · 1 thought hidden`. */
+export function hiddenRunLabel(run: HiddenRun): string {
+  const segs: string[] = []
+  if (run.tools) segs.push(`${run.tools} tool${run.tools === 1 ? '' : 's'}`)
+  if (run.thoughts) segs.push(`${run.thoughts} thought${run.thoughts === 1 ? '' : 's'}`)
+  return `${segs.join(' · ')} hidden — /details collapsed to show`
+}

ui-opentui/src/logic/diagnostics.ts

@@ -0,0 +1,82 @@
+/**
+ * Process diagnostics for the /mem + /heapdump utility commands (Epic 3 port;
+ * Ink ref `app/slash/commands/debug.ts` + `lib/memory.ts`). Pure formatters plus
+ * the one impure seam (`performHeapdump` → `v8.writeHeapSnapshot`), kept out of
+ * slash.ts so the dispatcher stays plain and tests can mock `node:v8`.
+ */
+import { mkdirSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { dirname, join } from 'node:path'
+import { writeHeapSnapshot } from 'node:v8'
+
+/** `123456789` → `117.7 MB` (binary units, one decimal above bytes). */
+export function formatBytes(n: number): string {
+  if (!Number.isFinite(n) || n < 0) return '0 B'
+  if (n < 1024) return `${Math.round(n)} B`
+  const units = ['KB', 'MB', 'GB', 'TB'] as const
+  let v = n
+  let i = -1
+  do {
+    v /= 1024
+    i += 1
+  } while (v >= 1024 && i < units.length - 1)
+  return `${v.toFixed(1)} ${units[i]}`
+}
+
+/** Where heap snapshots land: `$HERMES_HOME`/`~/.hermes` + `logs/opentui-heap-<ts>.heapsnapshot`. */
+export function heapSnapshotPath(now = new Date()): string {
+  const home = process.env.HERMES_HOME?.trim() || join(homedir(), '.hermes')
+  const ts = now.toISOString().replace(/[:.]/g, '-')
+  return join(home, 'logs', `opentui-heap-${ts}.heapsnapshot`)
+}
+
+export interface HeapdumpResult {
+  path: string
+  before: { heapUsed: number; rss: number }
+  after: { heapUsed: number; rss: number }
+}
+
+/**
+ * Write a V8 heap snapshot (SYNCHRONOUS — blocks the event loop while V8 walks
+ * the heap; that's inherent to writeHeapSnapshot) and report heap/rss before
+ * vs after. Throws on I/O failure — the caller renders the error.
+ */
+export function performHeapdump(): HeapdumpResult {
+  const before = process.memoryUsage()
+  const path = heapSnapshotPath()
+  mkdirSync(dirname(path), { recursive: true })
+  const written = writeHeapSnapshot(path)
+  const after = process.memoryUsage()
+  return {
+    after: { heapUsed: after.heapUsed, rss: after.rss },
+    before: { heapUsed: before.heapUsed, rss: before.rss },
+    path: written
+  }
+}
+
+export interface MemSnapshot {
+  heapUsed: number
+  heapTotal: number
+  external: number
+  arrayBuffers: number
+  rss: number
+}
+
+/**
+ * The /mem system-line body (Ink's Memory panel as aligned rows). `renderables`
+ * is the mounted-renderable count under the live renderer root (the store-cap
+ * diagnostic) — omitted when unavailable (e.g. no renderer in tests).
+ */
+export function memReport(usage: MemSnapshot, uptimeSeconds: number, renderables?: number): string {
+  const rows: Array<[string, string]> = [
+    ['heap used', formatBytes(usage.heapUsed)],
+    ['heap total', formatBytes(usage.heapTotal)],
+    ['external', formatBytes(usage.external)],
+    ['array buffers', formatBytes(usage.arrayBuffers)],
+    ['rss', formatBytes(usage.rss)],
+    ['uptime', `${Math.round(uptimeSeconds)}s`]
+  ]
+  if (renderables !== undefined) rows.push(['renderables', String(renderables)])
+  const pad = Math.max(...rows.map(([k]) => k.length))
+  return ['memory', ...rows.map(([k, v]) => `  ${k.padEnd(pad)}  ${v}`)].join('\n')
+}

ui-opentui/src/logic/diff.ts

@@ -0,0 +1,87 @@
+/**
+ * Pure unified-diff helpers for the file-tool renderer (Epic 2.3). No
+ * OpenTUI/Solid imports — just string work, trivially unit-testable (like
+ * `toolOutput.ts`). The gateway ships the FULL raw unified diff on file-edit
+ * `tool.complete` (`diff_unified`); these helpers turn it into the collapsed
+ * `+N −M` summary and per-file sections for the native `<diff>` renderable
+ * (which parses only the FIRST file of a multi-file diff — so we split).
+ */
+
+/** Added/removed line counts for the collapsed header summary (`+N −M`). */
+export interface DiffStats {
+  added: number
+  removed: number
+}
+
+/** Count changed lines in a unified diff, excluding the `+++`/`---` file headers. */
+export function diffStats(diff: string): DiffStats {
+  let added = 0
+  let removed = 0
+  for (const line of diff.split('\n')) {
+    if (line.startsWith('+++') || line.startsWith('---')) continue
+    if (line.startsWith('+')) added++
+    else if (line.startsWith('-')) removed++
+  }
+  return { added, removed }
+}
+
+/**
+ * Path relative to the session cwd: exact prefix strip only (no `~` for home —
+ * deliberately simple). Paths outside cwd come back unchanged; the cwd itself
+ * becomes `.`. A trailing slash on cwd is tolerated.
+ */
+export function relativizePath(path: string, cwd?: string): string {
+  if (!path || !cwd) return path
+  const base = cwd.endsWith('/') && cwd !== '/' ? cwd.slice(0, -1) : cwd
+  if (path === base) return '.'
+  const prefix = base === '/' ? '/' : base + '/'
+  if (path.startsWith(prefix)) return path.slice(prefix.length) || '.'
+  return path
+}
+
+/** One file's section of a (possibly multi-file) unified diff. */
+export interface DiffFileSection {
+  /** Target path from the `+++ b/…` header (or `--- a/…` for deletions); '' if unknown. */
+  path: string
+  /** The section's unified diff text, parseable on its own. */
+  diff: string
+}
+
+/** Extract the path from a `--- a/x` / `+++ b/x` header line ('' for /dev/null). */
+function headerPath(line: string): string {
+  let p = line.slice(4).trim()
+  const tab = p.indexOf('\t') // difflib may append a date after a tab
+  if (tab !== -1) p = p.slice(0, tab)
+  if (!p || p === '/dev/null') return ''
+  if (p.startsWith('a/') || p.startsWith('b/')) p = p.slice(2)
+  return p
+}
+
+function sectionPath(lines: string[]): string {
+  const to = lines.find(l => l.startsWith('+++ '))
+  const from = lines.find(l => l.startsWith('--- '))
+  return (to ? headerPath(to) : '') || (from ? headerPath(from) : '')
+}
+
+/**
+ * Split a unified diff into per-file sections (the gateway concatenates one
+ * difflib diff per edited file; `patch`-mode diffs can also be multi-file). A
+ * new section starts at a `--- ` header — required to be FOLLOWED by `+++ `
+ * and to come after the current section's hunks, so removed lines that merely
+ * start with `--` can't split a file in half.
+ */
+export function splitUnifiedDiff(diff: string): DiffFileSection[] {
+  const lines = diff.replace(/\n$/, '').split('\n')
+  const sections: string[][] = []
+  let current: string[] = []
+  for (let i = 0; i < lines.length; i++) {
+    const line = lines[i] ?? ''
+    if (current.some(l => l.startsWith('@@')) && line.startsWith('--- ') && (lines[i + 1] ?? '').startsWith('+++ ')) {
+      sections.push(current)
+      current = []
+    }
+    current.push(line)
+  }
+  if (current.length > 0) sections.push(current)
+  return sections.filter(s => s.some(l => l.startsWith('@@'))).map(s => ({ diff: s.join('\n'), path: sectionPath(s) }))
+}

ui-opentui/src/logic/env.ts

@@ -0,0 +1,222 @@
+/**
+ * env — shared boolean env-flag parsing (one source for the TRUE/FALSE regexes).
+ *
+ * Recognized truthy values: 1/true/yes/on; falsy: 0/false/no/off (case-insensitive,
+ * surrounding whitespace trimmed). Anything else (incl. unset) is "unrecognized".
+ */
+export const TRUE_RE = /^(?:1|true|yes|on)$/i
+export const FALSE_RE = /^(?:0|false|no|off)$/i
+
+/** Parse a boolean env var; returns `fallback` when unset/unrecognized. */
+export function envFlag(value: string | undefined, fallback: boolean): boolean {
+  const v = value?.trim() ?? ''
+  if (TRUE_RE.test(v)) return true
+  if (FALSE_RE.test(v)) return false
+  return fallback
+}
+
+/**
+ * Tri-state toggle parse: `true`/`false` for a recognized value, `null` when
+ * unset/unrecognized (so a caller can fall through to the next precedence rung).
+ * Mirrors Ink's `parseToggle` (`ui-tui/src/config/env.ts`).
+ */
+export function envToggle(value: string | undefined): boolean | null {
+  const v = value?.trim() ?? ''
+  if (TRUE_RE.test(v)) return true
+  if (FALSE_RE.test(v)) return false
+  return null
+}
+
+/**
+ * Resolve whether mouse tracking is ON at boot, deferring to Ink's env surface
+ * (`ui-tui/src/config/env.ts`) so muscle memory + docs + support scripts carry
+ * over. Precedence (highest first):
+ *   1. `HERMES_TUI_MOUSE_TRACKING` (toggle) — the explicit force knob; beats all.
+ *   2. `HERMES_TUI_DISABLE_MOUSE=1` — the legacy Ink kill switch (off).
+ *   3. `HERMES_TUI_MOUSE` (toggle) — the OpenTUI-native alias (kept, rule 2);
+ *      it's also what the launcher sets, so it stays a first-class boot knob.
+ *   4. default ON (opencode parity: wheel-scroll, drag-scrollbar, click-to-expand,
+ *      text-aware selection).
+ * OpenTUI's renderer mouse is a single boolean, so Ink's granular off|wheel|
+ * buttons|all collapses to on/off here (any non-off tracking mode → on).
+ */
+export function resolveMouseEnabled(env: { readonly [k: string]: string | undefined } = process.env): boolean {
+  const trackingOverride = envToggle(env.HERMES_TUI_MOUSE_TRACKING)
+  if (trackingOverride !== null) return trackingOverride
+  if (envFlag(env.HERMES_TUI_DISABLE_MOUSE, false)) return false
+  const mouseAlias = envToggle(env.HERMES_TUI_MOUSE)
+  if (mouseAlias !== null) return mouseAlias
+  return true
+}
+
+/**
+ * The seeded initial prompt for `hermes --tui "prompt"` / `--image`.
+ *
+ * The launcher (`hermes_cli/main.py`) sets `HERMES_TUI_QUERY` (the established
+ * cross-engine contract Ink reads via `STARTUP_QUERY`); the OpenTUI engine also
+ * accepts `HERMES_TUI_PROMPT` as its own alias and a bare argv tail for
+ * standalone dev launches. QUERY wins (it's the launcher contract); PROMPT and
+ * argv are fallbacks. Empty → undefined.
+ */
+export function startupPrompt(
+  env: { readonly [k: string]: string | undefined } = process.env,
+  argv: readonly string[] = process.argv.slice(2)
+): string | undefined {
+  const query = env.HERMES_TUI_QUERY?.trim()
+  if (query) return query
+  const prompt = env.HERMES_TUI_PROMPT?.trim()
+  if (prompt) return prompt
+  const tail = argv.join(' ').trim()
+  return tail || undefined
+}
+
+/**
+ * The seeded image PATH for `hermes --tui --image <path>`. The launcher sets
+ * `HERMES_TUI_IMAGE` (Ink reads it as `STARTUP_IMAGE` and `image.attach`es the
+ * path before submitting the query). Empty → undefined.
+ */
+export function startupImage(env: { readonly [k: string]: string | undefined } = process.env): string | undefined {
+  const image = env.HERMES_TUI_IMAGE?.trim()
+  return image || undefined
+}
+
+/** Ink's default prompt when an image is seeded with no query (`STARTUP_QUERY`). */
+export const STARTUP_IMAGE_DEFAULT_PROMPT = 'What do you see in this image?'
+
+/**
+ * `HERMES_TUI_NO_CONFIRM` — skip destructive-action confirm prompts (Ink parity,
+ * `ui-tui/src/config/env.ts` `NO_CONFIRM_DESTRUCTIVE`). When truthy, the `/clear`
+ * and `/new` confirm step is bypassed and the action runs immediately. Default
+ * off (confirm). Same name, same truthy parsing as Ink.
+ */
+export function noConfirmDestructive(env: { readonly [k: string]: string | undefined } = process.env): boolean {
+  return envFlag(env.HERMES_TUI_NO_CONFIRM, false)
+}
+
+/**
+ * `HERMES_HEAPDUMP_ON_START` — write a manual heap snapshot at boot (Ink parity).
+ * A diagnostic escape hatch that BYPASSES the diagnostics master switch (you set
+ * it deliberately to capture a baseline). Default off.
+ */
+export function heapdumpOnStart(env: { readonly [k: string]: string | undefined } = process.env): boolean {
+  return envFlag(env.HERMES_HEAPDUMP_ON_START, false)
+}
+
+/**
+ * `HERMES_TUI_SCROLL_SPEED` (or `CLAUDE_CODE_SCROLL_SPEED` for portability) —
+ * the wheel-scroll speed multiplier (Ink parity, `lib/wheelAccel.ts`
+ * `readScrollSpeedBase`). Default 1 (the engine's native scroll behavior is
+ * untouched), clamped to (0, 20]. Returns `null` when UNSET/garbage so the
+ * caller leaves OpenTUI's native scroll acceleration alone — only an explicit,
+ * in-range value installs a constant-multiplier override.
+ */
+export function scrollSpeedMultiplier(env: { readonly [k: string]: string | undefined } = process.env): number | null {
+  const raw = (env.HERMES_TUI_SCROLL_SPEED ?? env.CLAUDE_CODE_SCROLL_SPEED ?? '').trim()
+  if (!raw) return null
+  const n = Number.parseFloat(raw)
+  if (!Number.isFinite(n) || n <= 0) return null
+  return Math.min(n, 20)
+}
+
+/**
+ * The diagnostics master switch — `HERMES_TUI_DIAGNOSTICS` (default OFF).
+ *
+ * Gates the developer/profiling surface a regular user should never trip
+ * over: the diagnostic slash commands (`/mem`, `/heapdump`) and the default
+ * for `HERMES_TUI_WINDOW_STATS` (which can still be set individually). It is
+ * an enable switch, not a secret: anyone CAN set it (support flows say
+ * "relaunch with HERMES_TUI_DIAGNOSTICS=1"), it just keeps the day-to-day
+ * surface clean. Read per call so tests (and long-lived processes whose
+ * wrapper mutates env before launch) see the live value.
+ */
+export function diagnosticsEnabled(): boolean {
+  return envFlag(process.env.HERMES_TUI_DIAGNOSTICS, false)
+}
+
+/**
+ * Whether rich tool-call OUTPUTS are kept — `HERMES_TUI_TOOL_OUTPUTS` (default
+ * ON). OpenTUI's rich tool cards (full result body + raw result/args dicts) are
+ * its differentiator vs Ink, so they stay on for real users. Setting `=off`
+ * drops both the RENDER and the STORE of those bodies (exact Ink parity: Ink
+ * keeps only a short context line and discards the result/args dicts), which is
+ * the biggest memory lever — used by the bench (D8: a fair Ink-vs-OpenTUI
+ * engine-overhead comparison) and the low-mem mode. The redaction-safe
+ * `argsPreview` one-liner, name/duration/error, and file-edit diffs are KEPT
+ * either way (a diff is a high-value surface, not generic "output"). Read per
+ * call so a wrapper that mutates env before launch sees the live value.
+ */
+export function toolOutputsEnabled(): boolean {
+  return envFlag(process.env.HERMES_TUI_TOOL_OUTPUTS, true)
+}
+
+/**
+ * Parse `HERMES_TUI_TOOL_OUTPUT_LINES` (a TUI-only env var — deliberately NOT
+ * a config.yaml knob): how many output lines an expanded tool body shows.
+ * UNSET → Infinity (UNLIMITED — expanded tool output is uncapped by default;
+ * setting the var is how you RESTORE a cap, e.g. `=200`). A positive integer
+ * → that cap. `0` → Infinity too (back-compat: it was the old opt-in
+ * "unlimited" value). Garbage → Infinity (unrecognized ≙ no cap asked for —
+ * the semantic is "cap only when the user asked for one").
+ */
+export function envOutputLines(value: string | undefined): number {
+  const v = value?.trim() ?? ''
+  if (!/^\d+$/.test(v)) return Number.POSITIVE_INFINITY
+  const n = Number.parseInt(v, 10)
+  return n === 0 ? Number.POSITIVE_INFINITY : n
+}
+
+/**
+ * Default visible-height cap for the composer textarea, in rows (Ink composer
+ * parity — 8 lines, ref feature request #10418). Beyond this the textarea
+ * scrolls INTERNALLY (the native edit buffer keeps the cursor in view).
+ */
+export const COMPOSER_MAX_ROWS = 8
+
+/**
+ * Parse `HERMES_TUI_COMPOSER_ROWS` (a TUI-only env var — deliberately NOT a
+ * config.yaml knob): the composer's visible-height cap before internal scroll
+ * kicks in. A positive integer → that cap; unset / `0` / garbage → the
+ * COMPOSER_MAX_ROWS default.
+ */
+export function envComposerRows(value: string | undefined): number {
+  const v = value?.trim() ?? ''
+  if (!/^\d+$/.test(v)) return COMPOSER_MAX_ROWS
+  const n = Number.parseInt(v, 10)
+  return n > 0 ? n : COMPOSER_MAX_ROWS
+}
+
+/**
+ * Whether NO line cap applies (unset / `0` / unparseable). When unlimited,
+ * the store prefers the always-full raw `result` over a gateway tail-capped
+ * `result_text` — an "unlimited" view of a tail would still be missing its
+ * head — see store.ts tool.complete. With an explicit finite cap the gateway
+ * tail (+ honest omitted note) is kept: the user asked for a bounded view.
+ */
+export function envOutputUnlimited(value: string | undefined): boolean {
+  return envOutputLines(value) === Number.POSITIVE_INFINITY
+}
+
+/**
+ * The session's launch directory for `session.create`'s `cwd` param.
+ *
+ * The hermes launcher runs the OpenTUI engine with its process cwd set to the
+ * engine's own package dir, so `process.cwd()` is NOT where the user ran
+ * hermes. The launcher exports the real launch dir as `HERMES_CWD` (and the
+ * gateway's `TERMINAL_CWD`); prefer those. Falls back to `process.cwd()` only
+ * for standalone launches (smokes/dev) where no launcher set them, and returns
+ * `undefined` when even that is empty so the gateway resolves its own default.
+ */
+export function launchCwd(env: { readonly [k: string]: string | undefined } = process.env): string | undefined {
+  // First NON-BLANK of the launcher's vars (?? would keep a blank HERMES_CWD
+  // and never reach TERMINAL_CWD).
+  for (const value of [env.HERMES_CWD, env.TERMINAL_CWD]) {
+    const trimmed = (value ?? '').trim()
+    if (trimmed) return trimmed
+  }
+  try {
+    const cwd = process.cwd().trim()
+    return cwd || undefined
+  } catch {
+    return undefined
+  }
+}

ui-opentui/src/logic/fuzzy.ts

@@ -0,0 +1,140 @@
+/**
+ * fuzzy.ts — pure fuzzy filtering + grouped presentation for picker overlays
+ * (Epic 7 model picker v2; resume-session picker; skills hub). Matching/ranking
+ * is delegated to `fuzzysort` (the library opencode uses in production, see its
+ * dialog-select.tsx) through a thin adapter that preserves this module's API:
+ * call sites pass weighted `FuzzyField[]` haystacks and get back a ranked list.
+ *
+ * Adapter semantics on top of fuzzysort:
+ * - Multi-key scoring à la opencode: each field is a fuzzysort key; the final
+ *   score is the weight-multiplied SUM of per-key scores (label conventionally
+ *   ×2, opencode's `r[0].score * 2 + r[1].score`), so label hits outrank
+ *   equal-quality group/slug hits.
+ * - Multi-term AND (a feature of the old hand-rolled scorer that fuzzysort
+ *   lacks natively): the query is whitespace-split and fuzzysort runs once per
+ *   term over the progressively-filtered pool — every term must match at least
+ *   one field; per-term scores accumulate. Chosen over a joined single needle
+ *   because it keeps `anthropic son` / `copilot son` matching ACROSS fields.
+ * - Empty/blank query → all items in catalog order (fuzzysort returns nothing
+ *   for an empty needle; the old all-rows behavior is preserved here).
+ * - Equal final scores keep catalog order (fuzzysort's sort is not stable; the
+ *   adapter re-sorts with the original index as tie-break).
+ */
+import fuzzysort from 'fuzzysort'
+
+/** One searchable field of an item (e.g. model id ×2, provider slug, lab name). */
+export interface FuzzyField {
+  text: string
+  /** Score multiplier (default 1). The primary label is conventionally 2. */
+  weight?: number
+}
+
+/** Pool entry: the item plus its precomputed fields, catalog position and the
+ *  per-term accumulated score. */
+interface Entry<T> {
+  item: T
+  at: number
+  fields: FuzzyField[]
+  total: number
+}
+
+/**
+ * Filter + rank items by query. Empty query → the items in catalog order;
+ * otherwise matches sorted by score (descending), ties keeping catalog order.
+ * Every whitespace-split term must fuzzy-match at least one field.
+ */
+export function fuzzyFilter<T>(query: string, items: readonly T[], fieldsOf: (item: T) => FuzzyField[]): T[] {
+  const terms = query.trim().split(/\s+/).filter(Boolean)
+  if (!terms.length) return [...items]
+
+  let pool: Entry<T>[] = items.map((item, at) => ({ at, fields: fieldsOf(item), item, total: 0 }))
+  // Items may carry different field counts (description/haystacks optional):
+  // one key per field slot, missing slots read as '' (never match).
+  const keyCount = pool.reduce((max, e) => Math.max(max, e.fields.length), 0)
+  const keys = Array.from({ length: keyCount }, (_, i) => (e: Entry<T>) => e.fields[i]?.text ?? '')
+
+  for (const term of terms) {
+    const results = fuzzysort.go(term, pool, {
+      keys,
+      // Weighted sum of per-key scores (unmatched keys score 0). Inclusion is
+      // decided by fuzzysort (≥1 key must match); this only ranks.
+      scoreFn: r => {
+        let sum = 0
+        for (let i = 0; i < r.length; i++) sum += (r[i]?.score ?? 0) * (r.obj.fields[i]?.weight ?? 1)
+        return sum
+      }
+    })
+    if (!results.length) return []
+    pool = results.map(r => {
+      r.obj.total += r.score
+      return r.obj
+    })
+  }
+
+  pool.sort((a, b) => b.total - a.total || a.at - b.at)
+  return pool.map(e => e.item)
+}
+
+/** A render row of a grouped picker: a non-selectable group header or an item.
+ *  `index` is the item's position in the flat ARROW-TRAVERSAL order; `-1` marks
+ *  a non-selectable item row (rendered dimmed, skipped by traversal). */
+export type PickerRow<T> = { kind: 'header'; label: string } | { kind: 'item'; item: T; index: number }
+
+/**
+ * Group items for display (group order = first appearance, so a score-sorted
+ * input puts the best group first). Returns the header+item render rows and
+ * the flat selectable list in traversal order — arrows walk `flat` and thus
+ * cross group boundaries seamlessly; headers are never selectable. Items
+ * without a group render headerless (e.g. the skills picker). Items failing
+ * `selectableOf` (picker v2.1: unconfigured-provider hint rows) still RENDER
+ * (index `-1`) but never enter `flat`, so ↑↓ traversal skips them.
+ */
+export function buildPickerRows<T>(
+  items: readonly T[],
+  groupOf: (item: T) => string | undefined,
+  selectableOf: (item: T) => boolean = () => true
+): { rows: PickerRow<T>[]; flat: T[] } {
+  const order: string[] = []
+  const buckets = new Map<string, T[]>()
+  for (const item of items) {
+    const group = groupOf(item) ?? ''
+    let bucket = buckets.get(group)
+    if (!bucket) {
+      bucket = []
+      buckets.set(group, bucket)
+      order.push(group)
+    }
+    bucket.push(item)
+  }
+  const rows: PickerRow<T>[] = []
+  const flat: T[] = []
+  for (const group of order) {
+    if (group) rows.push({ kind: 'header', label: group })
+    for (const item of buckets.get(group) ?? []) {
+      if (selectableOf(item)) {
+        rows.push({ index: flat.length, item, kind: 'item' })
+        flat.push(item)
+      } else {
+        rows.push({ index: -1, item, kind: 'item' })
+      }
+    }
+  }
+  return { flat, rows }
+}
+
+/**
+ * Slice rows to a visible window of at most `cap` rows that keeps the selected
+ * item in view (centered when possible). `above`/`below` are the hidden row
+ * counts for the ↑/↓ "more" indicators.
+ */
+export function visibleRows<T>(
+  rows: readonly PickerRow<T>[],
+  selected: number,
+  cap: number
+): { rows: PickerRow<T>[]; above: number; below: number } {
+  if (rows.length <= cap) return { above: 0, below: 0, rows: [...rows] }
+  const selRow = rows.findIndex(r => r.kind === 'item' && r.index === selected)
+  const anchor = selRow === -1 ? 0 : selRow
+  const start = Math.max(0, Math.min(anchor - Math.floor(cap / 2), rows.length - cap))
+  return { above: start, below: rows.length - (start + cap), rows: rows.slice(start, start + cap) }
+}

ui-opentui/src/logic/gatewayRecovery.ts

@@ -0,0 +1,52 @@
+/**
+ * Pure recovery-budget policy for the gateway exit handler (LOGIC side — no
+ * Effect, no refs, no UI). Ported from Ink's `ui-tui/src/app/gatewayRecovery.ts`
+ * and EXTENDED with opencode-style exponential backoff.
+ *
+ * A gateway that crash-loops on startup must not let the TUI spawn-storm, so
+ * respawn+resume attempts are capped to GATEWAY_RECOVERY_LIMIT within a sliding
+ * GATEWAY_RECOVERY_WINDOW_MS; past the budget the app falls back to the inert
+ * "gateway exited" state. Kept pure (no refs/UI) so the bound — including the
+ * crash-loop case — is unit-testable.
+ */
+export const GATEWAY_RECOVERY_LIMIT = 3
+export const GATEWAY_RECOVERY_WINDOW_MS = 60_000
+
+export interface RecoveryPlan {
+  /** Attempt timestamps to persist (the pruned window, plus `now` iff recovering). */
+  attempts: number[]
+  recover: boolean
+  /**
+   * Session to resume — the live sid, or the not-yet-consumed recovery target
+   * when the live sid was already cleared by a prior exit.
+   */
+  sid: null | string
+}
+
+/**
+ * Decide whether to respawn+resume after a gateway death. `liveSid` is the
+ * current session (nulled on the first exit); `recoverSid` is a pending
+ * recovery target carried across a respawn that died before gateway.ready —
+ * so a startup crash-loop keeps retrying the same session up to the budget
+ * instead of stranding it after one attempt.
+ */
+export function planGatewayRecovery(
+  liveSid: null | string,
+  recoverSid: null | string,
+  attempts: number[],
+  now: number
+): RecoveryPlan {
+  const sid = liveSid ?? recoverSid
+  const recent = attempts.filter(t => now - t < GATEWAY_RECOVERY_WINDOW_MS)
+  const recover = Boolean(sid) && recent.length < GATEWAY_RECOVERY_LIMIT
+
+  return { attempts: recover ? [...recent, now] : recent, recover, sid }
+}
+
+/**
+ * Exponential backoff between respawn attempts (opencode-style): 1s, 2s, 4s, …
+ * capped at 30s. `attempt` is 1-based (the first respawn waits 1s).
+ */
+export function backoffMs(attempt: number): number {
+  return Math.min(1000 * 2 ** Math.max(0, attempt - 1), 30_000)
+}

ui-opentui/src/logic/history.ts

@@ -0,0 +1,122 @@
+/**
+ * Prompt history (item 6) — the SOLID side, plain TS. Up/Down cycle through the
+ * prompts you've sent, scoped PER DIRECTORY: launching Hermes again in the same
+ * project dir reuses that dir's prior prompts (the "bleed for the same dir" the
+ * user asked for), while a session in a different dir keeps its own list.
+ *
+ * `createPromptHistory` is pure + injectable (initial entries + a `persist`
+ * sink) so the cursor logic is unit-tested with no filesystem. The real wiring
+ * uses `loadDirHistory(cwd)` / `dirHistoryPersister(cwd)` to read/append a
+ * per-dir JSONL file under `$HERMES_HOME/tui-history/<hash>.jsonl` (one
+ * JSON-encoded prompt per line, multiline-safe; opencode's prompt-history.jsonl
+ * model, Ink's ~/.hermes/.hermes_history idea, scoped by dir).
+ */
+import { appendFileSync, mkdirSync, readFileSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { createHash } from 'node:crypto'
+import { dirname, join } from 'node:path'
+
+const DEFAULT_MAX = 200
+
+export interface PromptHistoryOptions {
+  /** Entries already on disk for this dir (oldest → newest). */
+  initial?: string[]
+  /** Persist a newly pushed prompt (real use: append to the per-dir file). */
+  persist?: (text: string) => void
+  /** Cap on retained entries (oldest dropped). */
+  max?: number
+}
+
+export interface PromptHistory {
+  /** All cycleable entries (oldest → newest) — loaded prev-session + this session. */
+  entries: () => string[]
+  /** Record a submitted prompt (skips a consecutive duplicate) and reset the cursor. */
+  push: (text: string) => void
+  /** Cycle to the OLDER entry (Up). Stashes `currentInput` as the draft on the first step. */
+  prev: (currentInput: string) => string | null
+  /** Cycle to the NEWER entry (Down); returns the stashed draft at the bottom. */
+  next: () => string | null
+  /** Reset the cursor to the live draft (call on any edit). */
+  reset: () => void
+}
+
+export function createPromptHistory(opts: PromptHistoryOptions = {}): PromptHistory {
+  const entries = [...(opts.initial ?? [])]
+  const max = opts.max ?? DEFAULT_MAX
+  // `idx === entries.length` means "at the live draft" (past the newest entry).
+  let idx = entries.length
+  let draft = ''
+
+  return {
+    entries: () => entries.slice(),
+    push(text) {
+      if (!text.trim()) return
+      if (entries[entries.length - 1] !== text) {
+        entries.push(text)
+        if (entries.length > max) entries.shift()
+        opts.persist?.(text)
+      }
+      idx = entries.length
+      draft = ''
+    },
+    prev(currentInput) {
+      if (entries.length === 0) return null
+      if (idx === entries.length) draft = currentInput // leaving the bottom — stash the draft
+      if (idx > 0) idx--
+      return entries[idx] ?? null
+    },
+    next() {
+      if (idx >= entries.length) return null
+      idx++
+      return idx === entries.length ? draft : (entries[idx] ?? null)
+    },
+    reset() {
+      idx = entries.length
+    }
+  }
+}
+
+// ── per-directory file persistence (best-effort; never throws) ──────────
+
+function hermesHome(): string {
+  return process.env.HERMES_HOME?.trim() || join(homedir(), '.hermes')
+}
+
+/** The history file for a given working directory (keyed by a hash of the abs path). */
+function dirHistoryPath(cwd: string): string {
+  const key = createHash('sha1').update(cwd).digest('hex').slice(0, 16)
+  return join(hermesHome(), 'tui-history', `${key}.jsonl`)
+}
+
+/** Load a directory's prior prompts (oldest → newest); [] if none / unreadable. */
+export function loadDirHistory(cwd: string, max = DEFAULT_MAX): string[] {
+  try {
+    const raw = readFileSync(dirHistoryPath(cwd), 'utf8')
+    const out: string[] = []
+    for (const line of raw.split('\n')) {
+      if (!line.trim()) continue
+      try {
+        const v: unknown = JSON.parse(line)
+        if (typeof v === 'string') out.push(v)
+      } catch {
+        // skip a corrupt line — never let it break loading
+      }
+    }
+    return out.length > max ? out.slice(out.length - max) : out
+  } catch {
+    return []
+  }
+}
+
+/** A persister that appends each pushed prompt to the dir's JSONL file (best-effort). */
+export function dirHistoryPersister(cwd: string): (text: string) => void {
+  const path = dirHistoryPath(cwd)
+  return text => {
+    try {
+      mkdirSync(dirname(path), { recursive: true })
+      appendFileSync(path, JSON.stringify(text) + '\n', 'utf8')
+    } catch {
+      // history persistence is non-essential — a write failure must not disrupt the turn
+    }
+  }
+}

ui-opentui/src/logic/mathPreprocess.ts

@@ -0,0 +1,279 @@
+/**
+ * Fence-aware LaTeX→Unicode span converter. Runs on the raw markdown string
+ * BEFORE it reaches the native `<markdown>` renderable (the one seam in
+ * view/markdown.tsx), so the native parser only ever sees already-converted
+ * unicode text. Tier-A: text-only — no styled spans, no accent color on math
+ * (that needs renderNode hooks into MarkdownRenderable; deferred).
+ *
+ * Span detection ports the Ink tokenizer's EXACT rules (ui-tui/src/components/
+ * markdown.tsx — keep in sync):
+ *
+ *   • inline `$…$` — INLINE_RE group 17:
+ *       (?<!\$)\$([^\s$](?:[^$\n]*?[^\s$])?)\$(?!\$)
+ *     content starts AND ends with a non-space-non-`$`, contains no `$` or
+ *     newline. This is the currency guard: in `I paid $5 and $10` the closing
+ *     `$` is preceded by a space, so nothing matches and the prose survives.
+ *   • inline `\(…\)` — INLINE_RE group 18: `\\\(([^\n]+?)\\\)` (single line).
+ *   • display `$$…$$` / `\[…\]` — MATH_BLOCK_OPEN_RE: opener only at the start
+ *     of a (whitespace-trimmed) line; closes on the same line (`$$x$$`) or on a
+ *     later line ENDING with the closer. No closer anywhere → the line passes
+ *     through verbatim (Ink renders it as a plain paragraph). That rule is what
+ *     makes streaming safe for free: an unclosed `$$`/`$` mid-stream stays
+ *     verbatim and converts exactly once, when the closing delimiter arrives
+ *     (the whole text re-feeds per delta).
+ *
+ * Because the markdown parser hasn't run yet, fence / inline-code state is
+ * tracked here:
+ *   • fenced blocks: ``` or ~~~ runs (3+, any info string) open; a line that is
+ *     only a run of the SAME character, at least as long, closes (CommonMark).
+ *     Everything inside, including the fence lines, passes through untouched.
+ *   • inline code: per-line backtick scan — a run of N backticks opens a span
+ *     closed by the next run of EXACTLY N backticks on the same line
+ *     (CommonMark rule); unmatched runs are literal text. Multi-line inline
+ *     code spans are NOT supported (the Ink tokenizer was per-line too).
+ *
+ * Known, documented deviations from full markdown awareness (both rare, both
+ * shared with or narrower than the Ink renderer's behavior):
+ *   • a paired `$…$` inside a link destination (`[x](http://a$b$c)`) converts;
+ *     Ink's tokenizer matched the link first.
+ *   • 4-space-indented code blocks are not tracked (fences only).
+ *
+ * `\boxed{…}` sentinels (U+0001/U+0002 from texToUnicode) are STRIPPED to the
+ * inner text — injecting a styled span into the native renderable needs
+ * renderer hooks; deferred with the rest of tier-B.
+ *
+ * Perf: this runs over the FULL text on every streaming delta. Early-exit
+ * fast path returns the same string reference when no `$` / `\(` / `\[`
+ * appears at all, and when a scan converts nothing the original reference is
+ * returned too (so the renderable's content prop stays identity-stable).
+ */
+import { BOX_RE, texToUnicode } from './mathUnicode.ts'
+
+const FENCE_OPEN_RE = /^\s*(`{3,}|~{3,})/
+const FENCE_CLOSE_RE = /^\s*(`{3,}|~{3,})\s*$/
+
+// Display math openers/closers — ported verbatim from Ink's markdown.tsx.
+const MATH_BLOCK_OPEN_RE = /^\s*(\$\$|\\\[)(.*)$/
+const MATH_BLOCK_CLOSE_DOLLAR_RE = /^(.*?)\$\$\s*$/
+const MATH_BLOCK_CLOSE_BRACKET_RE = /^(.*?)\\\]\s*$/
+
+// Ink INLINE_RE group 17 / 18, anchored (sticky). The `(?<!\$)` lookbehind is
+// checked by the caller (prev char), everything else is byte-for-byte Ink's.
+const INLINE_DOLLAR_RE = /\$([^\s$](?:[^$\n]*?[^\s$])?)\$(?!\$)/y
+const INLINE_PAREN_RE = /\\\(([^\n]+?)\\\)/y
+
+/** texToUnicode + strip the \boxed highlight sentinels down to plain text. */
+const toUnicode = (tex: string): string => texToUnicode(tex).replace(BOX_RE, '$1')
+
+/** Index of the next run of EXACTLY `len` backticks at/after `from`, or -1. */
+const findBacktickClose = (line: string, from: number, len: number): number => {
+  let i = from
+  while (i < line.length) {
+    if (line[i] !== '`') {
+      i++
+      continue
+    }
+
+    let j = i + 1
+
+    while (j < line.length && line[j] === '`') j++
+
+    if (j - i === len) {
+      return i
+    }
+
+    i = j
+  }
+
+  return -1
+}
+
+// Convert inline `$…$` / `\(…\)` spans in one prose line, skipping inline
+// code spans. Returns the SAME string reference when nothing converted.
+const convertInline = (line: string): string => {
+  let out = ''
+  let i = 0
+  let changed = false
+
+  while (i < line.length) {
+    const ch = line[i]
+
+    if (ch === '`') {
+      let j = i + 1
+
+      while (j < line.length && line[j] === '`') j++
+
+      const close = findBacktickClose(line, j, j - i)
+
+      if (close >= 0) {
+        out += line.slice(i, close + (j - i))
+        i = close + (j - i)
+      } else {
+        out += line.slice(i, j)
+        i = j
+      }
+
+      continue
+    }
+
+    if (ch === '$' && line[i - 1] !== '$') {
+      INLINE_DOLLAR_RE.lastIndex = i
+      const m = INLINE_DOLLAR_RE.exec(line)
+
+      if (m) {
+        out += toUnicode(m[1] ?? '')
+        i = INLINE_DOLLAR_RE.lastIndex
+        changed = true
+
+        continue
+      }
+    }
+
+    if (ch === '\\' && line[i + 1] === '(') {
+      INLINE_PAREN_RE.lastIndex = i
+      const m = INLINE_PAREN_RE.exec(line)
+
+      if (m) {
+        out += toUnicode(m[1] ?? '')
+        i = INLINE_PAREN_RE.lastIndex
+        changed = true
+
+        continue
+      }
+    }
+
+    out += ch
+    i++
+  }
+
+  return changed ? out : line
+}
+
+export function preprocessMath(markdown: string, _opts?: { streaming?: boolean | undefined }): string {
+  // Fast path — REQUIRED, this runs on every streaming delta. No math trigger
+  // characters anywhere → hand back the exact same string (identity).
+  if (!markdown.includes('$') && !markdown.includes('\\(') && !markdown.includes('\\[')) {
+    return markdown
+  }
+
+  const lines = markdown.split('\n')
+  const out: string[] = []
+  let changed = false
+  let fence: { char: string; len: number } | null = null
+  let i = 0
+
+  // Emit a converted display block as its own paragraph: blank-line separated
+  // from surrounding prose (only where a separator is actually missing).
+  const pushDisplay = (block: string[], nextIdx: number) => {
+    if (out.length > 0 && out[out.length - 1]?.trim()) {
+      out.push('')
+    }
+
+    out.push(...block)
+
+    if (nextIdx < lines.length && lines[nextIdx]?.trim()) {
+      out.push('')
+    }
+
+    changed = true
+  }
+
+  while (i < lines.length) {
+    const line = lines[i] ?? ''
+
+    if (fence) {
+      out.push(line)
+      const close = line.match(FENCE_CLOSE_RE)?.[1]
+
+      if (close && close.charAt(0) === fence.char && close.length >= fence.len) {
+        fence = null
+      }
+
+      i++
+
+      continue
+    }
+
+    const open = line.match(FENCE_OPEN_RE)?.[1]
+
+    if (open) {
+      fence = { char: open.charAt(0), len: open.length }
+      out.push(line)
+      i++
+
+      continue
+    }
+
+    const mathOpen = line.match(MATH_BLOCK_OPEN_RE)
+
+    if (mathOpen) {
+      const closeRe = mathOpen[1] === '$$' ? MATH_BLOCK_CLOSE_DOLLAR_RE : MATH_BLOCK_CLOSE_BRACKET_RE
+      const headRest = mathOpen[2] ?? ''
+
+      // Single-line block: `$$x + y = z$$` or `\[x\]`.
+      const sameLineClose = headRest.match(closeRe)
+
+      if (sameLineClose) {
+        const inner = (sameLineClose[1] ?? '').trim()
+        pushDisplay(inner ? [toUnicode(inner)] : [], i + 1)
+        i++
+
+        continue
+      }
+
+      // Multi-line block: scan ahead for a real closer before committing. If
+      // none exists in the rest of the (possibly still-streaming) document,
+      // the line stays verbatim — Ink's paragraph fallback.
+      let closeIdx = -1
+      let closeTail = ''
+
+      for (let j = i + 1; j < lines.length; j++) {
+        const m = (lines[j] ?? '').match(closeRe)
+
+        if (m) {
+          closeIdx = j
+          closeTail = m[1] ?? ''
+
+          break
+        }
+      }
+
+      if (closeIdx >= 0) {
+        const block: string[] = []
+
+        if (headRest.trim()) {
+          block.push(headRest)
+        }
+
+        for (let j = i + 1; j < closeIdx; j++) {
+          block.push(lines[j] ?? '')
+        }
+
+        const tail = closeTail.trimEnd()
+
+        if (tail.trim()) {
+          block.push(tail)
+        }
+
+        pushDisplay(
+          block.map(l => toUnicode(l)),
+          closeIdx + 1
+        )
+        i = closeIdx + 1
+
+        continue
+      }
+    }
+
+    const converted = convertInline(line)
+
+    if (converted !== line) {
+      changed = true
+    }
+
+    out.push(converted)
+    i++
+  }
+
+  return changed ? out.join('\n') : markdown
+}

ui-opentui/src/logic/mathUnicode.ts

@@ -0,0 +1,773 @@
+// Best-effort LaTeX → Unicode for inline / display math (ported verbatim from
+// ui-tui/src/lib/mathUnicode.ts — keep the two in sync). The terminal can't
+// typeset LaTeX, but Unicode covers
+// most of what models actually emit: Greek letters, blackboard / fraktur /
+// calligraphic capitals, set theory + logic operators, common arrows,
+// sub/superscripts, and `\frac{a}{b}` collapsed to `a/b`.
+//
+// Design rules:
+//   • Pure regex pipeline. Anything we don't recognise is preserved
+//     verbatim (so a `\foo{bar}` we've never heard of still survives).
+//     A real LaTeX parser would be more correct but throws on partial
+//     input — terminal users would rather see the raw command than a
+//     parse-error placeholder.
+//   • Longest-match-first ordering on commands so `\le` doesn't shadow
+//     `\leq`, `\sub` doesn't shadow `\subseteq`, etc.
+//   • Word-boundary lookahead `(?![A-Za-z])` after each command so
+//     `\pix` (made-up command) doesn't get partially substituted as `π`.
+//   • `\mathbb{X}`, `\mathcal{X}`, `\mathfrak{X}` only handle a single
+//     letter argument — multi-letter `\mathbb{NN}` is rare and would
+//     need a real parser to do correctly.
+//   • Sub/super scripts only convert if EVERY character has a Unicode
+//     equivalent. Mixed content like `^{n+1}` falls back to the raw
+//     LaTeX so we don't emit `ⁿ+¹` (which has no `+` superscript glyph
+//     in some fonts and reads worse than the source).
+
+const SYMBOLS: Record<string, string> = {
+  // Greek lowercase
+  '\\alpha': 'α',
+  '\\beta': 'β',
+  '\\gamma': 'γ',
+  '\\delta': 'δ',
+  '\\epsilon': 'ε',
+  '\\varepsilon': 'ε',
+  '\\zeta': 'ζ',
+  '\\eta': 'η',
+  '\\theta': 'θ',
+  '\\vartheta': 'ϑ',
+  '\\iota': 'ι',
+  '\\kappa': 'κ',
+  '\\lambda': 'λ',
+  '\\mu': 'μ',
+  '\\nu': 'ν',
+  '\\xi': 'ξ',
+  '\\pi': 'π',
+  '\\varpi': 'ϖ',
+  '\\rho': 'ρ',
+  '\\varrho': 'ϱ',
+  '\\sigma': 'σ',
+  '\\varsigma': 'ς',
+  '\\tau': 'τ',
+  '\\upsilon': 'υ',
+  '\\phi': 'φ',
+  '\\varphi': 'φ',
+  '\\chi': 'χ',
+  '\\psi': 'ψ',
+  '\\omega': 'ω',
+
+  // Greek uppercase
+  '\\Gamma': 'Γ',
+  '\\Delta': 'Δ',
+  '\\Theta': 'Θ',
+  '\\Lambda': 'Λ',
+  '\\Xi': 'Ξ',
+  '\\Pi': 'Π',
+  '\\Sigma': 'Σ',
+  '\\Upsilon': 'Υ',
+  '\\Phi': 'Φ',
+  '\\Psi': 'Ψ',
+  '\\Omega': 'Ω',
+
+  // Big operators
+  '\\sum': '∑',
+  '\\prod': '∏',
+  '\\coprod': '∐',
+  '\\int': '∫',
+  '\\iint': '∬',
+  '\\iiint': '∭',
+  '\\oint': '∮',
+  '\\bigcup': '⋃',
+  '\\bigcap': '⋂',
+  '\\bigvee': '⋁',
+  '\\bigwedge': '⋀',
+  '\\bigoplus': '⨁',
+  '\\bigotimes': '⨂',
+
+  // Calculus
+  '\\partial': '∂',
+  '\\nabla': '∇',
+  '\\sqrt': '√',
+
+  // Sets
+  '\\emptyset': '∅',
+  '\\varnothing': '∅',
+  '\\infty': '∞',
+  '\\in': '∈',
+  '\\notin': '∉',
+  '\\ni': '∋',
+  '\\subset': '⊂',
+  '\\supset': '⊃',
+  '\\subseteq': '⊆',
+  '\\supseteq': '⊇',
+  '\\subsetneq': '⊊',
+  '\\supsetneq': '⊋',
+  '\\cup': '∪',
+  '\\cap': '∩',
+  '\\setminus': '∖',
+  '\\complement': '∁',
+
+  // Logic
+  '\\forall': '∀',
+  '\\exists': '∃',
+  '\\nexists': '∄',
+  '\\land': '∧',
+  '\\lor': '∨',
+  '\\lnot': '¬',
+  '\\neg': '¬',
+  '\\therefore': '∴',
+  '\\because': '∵',
+
+  // Relations
+  '\\le': '≤',
+  '\\leq': '≤',
+  '\\ge': '≥',
+  '\\geq': '≥',
+  '\\ne': '≠',
+  '\\neq': '≠',
+  '\\ll': '≪',
+  '\\gg': '≫',
+  '\\approx': '≈',
+  '\\equiv': '≡',
+  '\\cong': '≅',
+  '\\sim': '∼',
+  '\\simeq': '≃',
+  '\\propto': '∝',
+  '\\perp': '⊥',
+  '\\parallel': '∥',
+  '\\models': '⊨',
+  '\\vdash': '⊢',
+  '\\mid': '∣',
+  '\\nmid': '∤',
+  '\\divides': '∣',
+
+  // Common standalone glyphs
+  '\\blacksquare': '■',
+  '\\square': '□',
+  '\\Box': '□',
+  '\\qed': '∎',
+  '\\bigstar': '★',
+
+  // Modular arithmetic — the `\pmod{p}` form (with arg) is handled below;
+  // the bare `\bmod` / `\mod` commands are simple text substitutions.
+  '\\bmod': 'mod',
+  '\\mod': 'mod',
+
+  // Brackets / fences (named delimiter commands; the `\left\X` / `\right\X`
+  // unwrapping below leaves these behind for the symbol pass to resolve).
+  '\\langle': '⟨',
+  '\\rangle': '⟩',
+  '\\lceil': '⌈',
+  '\\rceil': '⌉',
+  '\\lfloor': '⌊',
+  '\\rfloor': '⌋',
+  '\\|': '‖',
+
+  // Arrows
+  '\\to': '→',
+  '\\rightarrow': '→',
+  '\\leftarrow': '←',
+  '\\leftrightarrow': '↔',
+  '\\Rightarrow': '⇒',
+  '\\Leftarrow': '⇐',
+  '\\Leftrightarrow': '⇔',
+  '\\implies': '⟹',
+  '\\impliedby': '⟸',
+  '\\iff': '⟺',
+  '\\mapsto': '↦',
+  '\\hookrightarrow': '↪',
+  '\\hookleftarrow': '↩',
+  '\\uparrow': '↑',
+  '\\downarrow': '↓',
+  '\\updownarrow': '↕',
+
+  // Binary operators
+  '\\cdot': '⋅',
+  '\\cdots': '⋯',
+  '\\ldots': '…',
+  '\\dots': '…',
+  '\\dotsb': '…',
+  '\\dotsc': '…',
+  '\\vdots': '⋮',
+  '\\ddots': '⋱',
+  '\\times': '×',
+  '\\div': '÷',
+  '\\pm': '±',
+  '\\mp': '∓',
+  '\\circ': '∘',
+  '\\bullet': '•',
+  '\\star': '⋆',
+  '\\ast': '∗',
+  '\\oplus': '⊕',
+  '\\ominus': '⊖',
+  '\\otimes': '⊗',
+  '\\odot': '⊙',
+  '\\diamond': '⋄',
+  '\\angle': '∠',
+  '\\triangle': '△',
+
+  // Spacing — collapse to varying widths of regular space
+  '\\,': ' ',
+  '\\;': ' ',
+  '\\:': ' ',
+  '\\!': '',
+  '\\ ': ' ',
+  '\\quad': '  ',
+  '\\qquad': '    ',
+
+  // Functions (LaTeX renders these in roman; we just keep the name)
+  '\\sin': 'sin',
+  '\\cos': 'cos',
+  '\\tan': 'tan',
+  '\\cot': 'cot',
+  '\\sec': 'sec',
+  '\\csc': 'csc',
+  '\\arcsin': 'arcsin',
+  '\\arccos': 'arccos',
+  '\\arctan': 'arctan',
+  '\\sinh': 'sinh',
+  '\\cosh': 'cosh',
+  '\\tanh': 'tanh',
+  '\\log': 'log',
+  '\\ln': 'ln',
+  '\\exp': 'exp',
+  '\\det': 'det',
+  '\\dim': 'dim',
+  '\\ker': 'ker',
+  '\\lim': 'lim',
+  '\\liminf': 'liminf',
+  '\\limsup': 'limsup',
+  '\\sup': 'sup',
+  '\\inf': 'inf',
+  '\\max': 'max',
+  '\\min': 'min',
+  '\\arg': 'arg',
+  '\\gcd': 'gcd',
+
+  // Escaped literals — model occasionally emits these for display
+  '\\&': '&',
+  '\\%': '%',
+  '\\$': '$',
+  '\\#': '#',
+  '\\_': '_',
+  '\\{': '{',
+  '\\}': '}'
+}
+
+const BB: Record<string, string> = {
+  A: '𝔸',
+  B: '𝔹',
+  C: 'ℂ',
+  D: '𝔻',
+  E: '𝔼',
+  F: '𝔽',
+  G: '𝔾',
+  H: 'ℍ',
+  I: '𝕀',
+  J: '𝕁',
+  K: '𝕂',
+  L: '𝕃',
+  M: '𝕄',
+  N: 'ℕ',
+  O: '𝕆',
+  P: 'ℙ',
+  Q: 'ℚ',
+  R: 'ℝ',
+  S: '𝕊',
+  T: '𝕋',
+  U: '𝕌',
+  V: '𝕍',
+  W: '𝕎',
+  X: '𝕏',
+  Y: '𝕐',
+  Z: 'ℤ'
+}
+
+const CAL: Record<string, string> = {
+  A: '𝒜',
+  B: 'ℬ',
+  C: '𝒞',
+  D: '𝒟',
+  E: 'ℰ',
+  F: 'ℱ',
+  G: '𝒢',
+  H: 'ℋ',
+  I: 'ℐ',
+  J: '𝒥',
+  K: '𝒦',
+  L: 'ℒ',
+  M: 'ℳ',
+  N: '𝒩',
+  O: '𝒪',
+  P: '𝒫',
+  Q: '𝒬',
+  R: 'ℛ',
+  S: '𝒮',
+  T: '𝒯',
+  U: '𝒰',
+  V: '𝒱',
+  W: '𝒲',
+  X: '𝒳',
+  Y: '𝒴',
+  Z: '𝒵'
+}
+
+const FRAK: Record<string, string> = {
+  A: '𝔄',
+  B: '𝔅',
+  C: 'ℭ',
+  D: '𝔇',
+  E: '𝔈',
+  F: '𝔉',
+  G: '𝔊',
+  H: 'ℌ',
+  I: 'ℑ',
+  J: '𝔍',
+  K: '𝔎',
+  L: '𝔏',
+  M: '𝔐',
+  N: '𝔑',
+  O: '𝔒',
+  P: '𝔓',
+  Q: '𝔔',
+  R: 'ℜ',
+  S: '𝔖',
+  T: '𝔗',
+  U: '𝔘',
+  V: '𝔙',
+  W: '𝔚',
+  X: '𝔛',
+  Y: '𝔜',
+  Z: 'ℨ'
+}
+
+const SUPERSCRIPT: Record<string, string> = {
+  '0': '⁰',
+  '1': '¹',
+  '2': '²',
+  '3': '³',
+  '4': '⁴',
+  '5': '⁵',
+  '6': '⁶',
+  '7': '⁷',
+  '8': '⁸',
+  '9': '⁹',
+  '+': '⁺',
+  '-': '⁻',
+  '=': '⁼',
+  '(': '⁽',
+  ')': '⁾',
+  a: 'ᵃ',
+  b: 'ᵇ',
+  c: 'ᶜ',
+  d: 'ᵈ',
+  e: 'ᵉ',
+  f: 'ᶠ',
+  g: 'ᵍ',
+  h: 'ʰ',
+  i: 'ⁱ',
+  j: 'ʲ',
+  k: 'ᵏ',
+  l: 'ˡ',
+  m: 'ᵐ',
+  n: 'ⁿ',
+  o: 'ᵒ',
+  p: 'ᵖ',
+  r: 'ʳ',
+  s: 'ˢ',
+  t: 'ᵗ',
+  u: 'ᵘ',
+  v: 'ᵛ',
+  w: 'ʷ',
+  x: 'ˣ',
+  y: 'ʸ',
+  z: 'ᶻ'
+}
+
+const SUBSCRIPT: Record<string, string> = {
+  '0': '₀',
+  '1': '₁',
+  '2': '₂',
+  '3': '₃',
+  '4': '₄',
+  '5': '₅',
+  '6': '₆',
+  '7': '₇',
+  '8': '₈',
+  '9': '₉',
+  '+': '₊',
+  '-': '₋',
+  '=': '₌',
+  '(': '₍',
+  ')': '₎',
+  a: 'ₐ',
+  e: 'ₑ',
+  h: 'ₕ',
+  i: 'ᵢ',
+  j: 'ⱼ',
+  k: 'ₖ',
+  l: 'ₗ',
+  m: 'ₘ',
+  n: 'ₙ',
+  o: 'ₒ',
+  p: 'ₚ',
+  r: 'ᵣ',
+  s: 'ₛ',
+  t: 'ₜ',
+  u: 'ᵤ',
+  v: 'ᵥ',
+  x: 'ₓ'
+}
+
+// Sentinel control characters used to mark `\boxed` / `\fbox` regions in
+// the converted output. The renderer splits on these to apply a highlight
+// style; consumers that don't want highlighting can strip them with the
+// exported `BOX_RE` below.
+export const BOX_OPEN = '\u0001'
+export const BOX_CLOSE = '\u0002'
+// eslint-disable-next-line no-control-regex
+export const BOX_RE = /\u0001([^\u0001\u0002]*)\u0002/g
+
+const escapeRe = (s: string) => s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+
+// Pre-compile two symbol regexes: one for letter-ending commands (`\pi`,
+// `\sum`) which need a `(?![A-Za-z])` lookahead so they don't partially
+// match `\pix` or `\summa`, and one for punctuation-ending commands
+// (`\{`, `\,`, `\|`) which must NOT have the lookahead — otherwise
+// `\{p` would refuse to substitute because `p` is a letter.
+//
+// Longest commands first inside each group so `\leq` beats `\le`.
+const splitByEnding = (keys: string[]) => {
+  const letter: string[] = []
+  const punct: string[] = []
+
+  for (const k of keys) {
+    if (/[A-Za-z]$/.test(k)) {
+      letter.push(k)
+    } else {
+      punct.push(k)
+    }
+  }
+
+  return { letter, punct }
+}
+
+const buildAlt = (cmds: string[]) =>
+  cmds
+    .sort((a, b) => b.length - a.length)
+    .map(escapeRe)
+    .join('|')
+
+const { letter: LETTER_CMDS, punct: PUNCT_CMDS } = splitByEnding(Object.keys(SYMBOLS))
+
+const SYMBOL_LETTER_RE = new RegExp('(?:' + buildAlt(LETTER_CMDS) + ')(?![A-Za-z])', 'g')
+const SYMBOL_PUNCT_RE = new RegExp('(?:' + buildAlt(PUNCT_CMDS) + ')', 'g')
+
+const convertScript = (input: string, table: Record<string, string>, sigil: '^' | '_'): string => {
+  let out = ''
+  let allMapped = true
+
+  for (const ch of input) {
+    const mapped = table[ch]
+
+    if (!mapped) {
+      allMapped = false
+
+      break
+    }
+
+    out += mapped
+  }
+
+  if (allMapped) {
+    return out
+  }
+
+  // Fallback: if the body is a single visible character (e.g. `∞` after
+  // earlier symbol substitution), render it without braces — `^∞` reads
+  // far better than `^{∞}` in a terminal. Multi-char bodies that don't
+  // fully convert use parens (`e^(iπ)`) instead of braces (`e^{iπ}`)
+  // because parens are normal punctuation while braces look like
+  // unrendered LaTeX.
+  const trimmed = input.trim()
+
+  if ([...trimmed].length === 1) {
+    return `${sigil}${trimmed}`
+  }
+
+  return `${sigil}(${trimmed})`
+}
+
+// Walk the string and parse `{...}` honouring nested braces. Unlike a
+// `\{[^{}]*\}` regex this survives `\frac{|t|^{p-1}|P(t)|^p}{...}` where
+// the numerator contains its own braces from a superscript. Returns the
+// inner content (without the outer braces) and the offset just past the
+// closing `}`. Returns null if there is no balanced brace at `start`.
+const readBraced = (s: string, start: number): { content: string; end: number } | null => {
+  if (s[start] !== '{') {
+    return null
+  }
+
+  let depth = 1
+  let i = start + 1
+
+  while (i < s.length && depth > 0) {
+    const c = s[i]
+
+    // Skip escapes — `\{` and `\}` inside a body are literal braces and
+    // should not change the brace counter.
+    if (c === '\\' && i + 1 < s.length) {
+      i += 2
+      continue
+    }
+
+    if (c === '{') {
+      depth++
+    } else if (c === '}') {
+      depth--
+    }
+
+    if (depth > 0) {
+      i++
+    }
+  }
+
+  if (depth !== 0) {
+    return null
+  }
+
+  return { content: s.slice(start + 1, i), end: i + 1 }
+}
+
+// Replace every occurrence of `\command{arg}` using balanced-brace parsing
+// (so `\boxed{x^{n+1}}` works where a `[^{}]*` regex would fail). The
+// `render` callback receives the inner content already recursed-into, so
+// `\boxed{\boxed{x}}` resolves outside-in cleanly. Unmatched `\command`
+// (no following `{...}`) is preserved verbatim.
+const replaceBracedCommand = (input: string, command: string, render: (content: string) => string): string => {
+  const cmdLen = command.length
+  let out = ''
+  let i = 0
+
+  while (i < input.length) {
+    const idx = input.indexOf(command, i)
+
+    if (idx < 0) {
+      out += input.slice(i)
+
+      return out
+    }
+
+    const after = input[idx + cmdLen]
+
+    if (after && /[A-Za-z]/.test(after)) {
+      out += input.slice(i, idx + cmdLen)
+      i = idx + cmdLen
+      continue
+    }
+
+    out += input.slice(i, idx)
+
+    let p = idx + cmdLen
+
+    while (input[p] === ' ' || input[p] === '\t') p++
+
+    const arg = readBraced(input, p)
+
+    if (!arg) {
+      out += input.slice(idx, p + 1)
+      i = p + 1
+      continue
+    }
+
+    out += render(replaceBracedCommand(arg.content, command, render))
+    i = arg.end
+  }
+
+  return out
+}
+
+// Replace every `\frac{num}{den}` with `num/den` (parens around either
+// side when its precedence demands it). The recursion handles nested
+// fractions naturally: `\frac{1}{\frac{1}{x}}` collapses to `1/(1/x)`
+// because we recurse into `den` before deciding whether to parenthesise.
+const replaceFracs = (input: string): string => {
+  let out = ''
+  let i = 0
+
+  while (i < input.length) {
+    const idx = input.indexOf('\\frac', i)
+
+    if (idx < 0) {
+      out += input.slice(i)
+
+      return out
+    }
+
+    const after = input[idx + 5]
+
+    // `(?![A-Za-z])` — protect hypothetical commands like `\fraction`.
+    if (after && /[A-Za-z]/.test(after)) {
+      out += input.slice(i, idx + 5)
+      i = idx + 5
+      continue
+    }
+
+    out += input.slice(i, idx)
+
+    let p = idx + 5
+
+    while (input[p] === ' ' || input[p] === '\t') p++
+
+    const num = readBraced(input, p)
+
+    if (!num) {
+      out += input.slice(idx, p + 1)
+      i = p + 1
+      continue
+    }
+
+    p = num.end
+
+    while (input[p] === ' ' || input[p] === '\t') p++
+
+    const den = readBraced(input, p)
+
+    if (!den) {
+      out += input.slice(idx, p + 1)
+      i = p + 1
+      continue
+    }
+
+    out += `${wrapForFrac(replaceFracs(num.content))}/${wrapForFrac(replaceFracs(den.content))}`
+    i = den.end
+  }
+
+  return out
+}
+
+// Wrap multi-token expressions in parens so `\frac{a+b}{c}` becomes
+// `(a+b)/c` rather than `a+b/c`. We wrap whenever inline `/` would
+// change the meaning — that's any binary operator (`+`, `-`, `*`, `/`)
+// or whitespace separating tokens. `*` and `/` matter because nested
+// fractions and products like `\frac{a*b}{c}` and `\frac{1/x}{y}` would
+// otherwise read as `a*b/c` (right-associative ambiguity) and `1/x/y`.
+// Atomic factors like `n!`, `x^2`, `\sin x` don't trigger any of these
+// and stay un-parenthesised — wrapping them just clutters the output.
+const wrapForFrac = (expr: string) => {
+  const trimmed = expr.trim()
+
+  if (!trimmed) {
+    return trimmed
+  }
+
+  if (/^\(.*\)$/.test(trimmed)) {
+    return trimmed
+  }
+
+  if (/[+\-/*]|\s/.test(trimmed)) {
+    return `(${trimmed})`
+  }
+
+  return trimmed
+}
+
+export function texToUnicode(input: string): string {
+  let s = input
+
+  s = s.replace(/\\mathbb\s*\{([A-Za-z])\}/g, (raw, c: string) => BB[c] ?? raw)
+  s = s.replace(/\\mathcal\s*\{([A-Za-z])\}/g, (raw, c: string) => CAL[c] ?? raw)
+  s = s.replace(/\\mathfrak\s*\{([A-Za-z])\}/g, (raw, c: string) => FRAK[c] ?? raw)
+  s = s.replace(/\\mathbf\s*\{([^{}]+)\}/g, (_, c: string) => c)
+  s = s.replace(/\\mathit\s*\{([^{}]+)\}/g, (_, c: string) => c)
+  s = s.replace(/\\mathrm\s*\{([^{}]+)\}/g, (_, c: string) => c)
+  s = s.replace(/\\text\s*\{([^{}]+)\}/g, (_, c: string) => c)
+  s = s.replace(/\\operatorname\s*\{([^{}]+)\}/g, (_, c: string) => c)
+
+  s = s.replace(/\\overline\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0305`)
+  s = s.replace(/\\hat\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0302`)
+  s = s.replace(/\\bar\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0304`)
+  s = s.replace(/\\tilde\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0303`)
+  s = s.replace(/\\vec\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u20D7`)
+  s = s.replace(/\\dot\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0307`)
+  s = s.replace(/\\ddot\s*\{([^{}]+)\}/g, (_, c: string) => `${c}\u0308`)
+
+  s = replaceFracs(s)
+
+  // `\boxed{X}` / `\fbox{X}` highlight a final answer. Terminals can't
+  // draw a real box, so we wrap the content in U+0001 / U+0002 control
+  // characters — non-printable, never present in real text — so a renderer
+  // with styled-span hooks can apply a highlight style (inverse video) to
+  // the bracketed region. The OpenTUI engine's preprocessor currently
+  // strips them via BOX_RE (styled-span injection into the native
+  // `<markdown>` renderable is deferred); `texToUnicode` stays pure-string.
+  // Argument is parsed with balanced braces so nested `{...}` from
+  // superscripts / fractions inside the box survive.
+  s = replaceBracedCommand(s, '\\boxed', body => `${BOX_OPEN}${body.trim()}${BOX_CLOSE}`)
+  s = replaceBracedCommand(s, '\\fbox', body => `${BOX_OPEN}${body.trim()}${BOX_CLOSE}`)
+
+  // `\xrightarrow{label}` / `\xleftarrow{label}` collapse to an arrow with
+  // the label inline. LaTeX renders the label above the arrow; in monospace
+  // we put it adjacent — `─label→` is the closest readable approximation.
+  // Run before the symbol pass so the label can still pick up Greek and
+  // operator substitutions afterwards.
+  s = s.replace(/\\xrightarrow\s*\{([^{}]*)\}/g, (_, label: string) => `─${label.trim()}→`)
+  s = s.replace(/\\xleftarrow\s*\{([^{}]*)\}/g, (_, label: string) => `←${label.trim()}─`)
+  s = s.replace(/\\Longrightarrow/g, '⟹')
+  s = s.replace(/\\Longleftarrow/g, '⟸')
+  s = s.replace(/\\Longleftrightarrow/g, '⟺')
+
+  // `\pmod{p}` → ` (mod p)` (LaTeX adds parens automatically); `\pod{p}`
+  // is a paren-less variant; `\tag{n}` is the equation-number annotation
+  // shown to the right of an equation. Collapse to a single-space-prefixed
+  // bracketed form. The leading `\s*` in the pattern absorbs any whitespace
+  // already in the source so we don't end up with `b  (mod p)` (double
+  // space) when the user wrote `b \pmod{p}`.
+  s = s.replace(/\s*\\pmod\s*\{([^{}]*)\}/g, (_, p: string) => ` (mod ${p.trim()})`)
+  s = s.replace(/\s*\\pod\s*\{([^{}]*)\}/g, (_, p: string) => ` (${p.trim()})`)
+  s = s.replace(/\s*\\tag\s*\{([^{}]*)\}/g, (_, n: string) => ` (${n.trim()})`)
+
+  // `\big`, `\Big`, `\bigg`, `\Bigg` (with optional `l`/`r`/`m` suffix)
+  // are sizing wrappers analogous to `\left`/`\right` but without the
+  // automatic-pairing semantics. Strip them and leave whatever delimiter
+  // follows. The trailing `(?![A-Za-z])` protects `\bigtriangleup` and
+  // any other letter-continuation command from being shaved.
+  s = s.replace(/\\(?:Bigg|bigg|Big|big)[lrm]?(?![A-Za-z])/g, '')
+
+  // Style / size hints that don't typeset any glyph and only affect how
+  // things would be sized in a real LaTeX engine. In a terminal every
+  // glyph is one monospace cell, so there's nothing to do — drop them
+  // (with any trailing whitespace) so they don't leak through as raw
+  // `\displaystyle` in the output.
+  s = s.replace(/\\(?:scriptscriptstyle|displaystyle|scriptstyle|textstyle|nolimits|limits)(?![A-Za-z])\s*/g, '')
+
+  // `\left` and `\right` are sizing wrappers around any delimiter — bare
+  // (`\left(`), escaped (`\left\{`), or named (`\left\langle`). Strip the
+  // wrapper unconditionally and let the rest of the pipeline (or the
+  // upcoming symbol pass) handle whatever delimiter follows. The optional
+  // `.?` consumes `\left.` / `\right.` which mean "no delimiter".
+  // Lookahead `(?![A-Za-z])` keeps `\leftarrow` / `\leftrightarrow` safe.
+  s = s.replace(/\\left(?![A-Za-z])\.?/g, '')
+  s = s.replace(/\\right(?![A-Za-z])\.?/g, '')
+
+  // Run symbol substitution BEFORE scripts so a body like `^{\infty}`
+  // becomes `^{∞}` first; convertScript can then either map ∞ to a
+  // superscript (it can't — Unicode lacks one) or fall back to `^∞`
+  // by stripping braces around the now-single-character body.
+  //
+  // Punctuation pass first — these can be followed by letters (`\{p`
+  // is "open-brace then p"), so the letter pass's `(?![A-Za-z])` rule
+  // would wrongly block them.
+  s = s.replace(SYMBOL_PUNCT_RE, m => SYMBOLS[m] ?? m)
+  s = s.replace(SYMBOL_LETTER_RE, m => SYMBOLS[m] ?? m)
+
+  // Bare `^c` / `_c` handles ONLY alphanumerics and `+`/`-`/`=`. Parens
+  // are intentionally excluded because the braced-fallback above can
+  // emit `(...)` and we don't want a second pass to greedily convert
+  // its opening paren into `⁽` and orphan the closing one.
+  s = s.replace(/\^\s*\{([^{}]+)\}/g, (_, body: string) => convertScript(body, SUPERSCRIPT, '^'))
+  s = s.replace(/\^([A-Za-z0-9+\-=])/g, (raw, ch: string) => SUPERSCRIPT[ch] ?? raw)
+  s = s.replace(/_\s*\{([^{}]+)\}/g, (_, body: string) => convertScript(body, SUBSCRIPT, '_'))
+  s = s.replace(/_([A-Za-z0-9+\-=])/g, (raw, ch: string) => SUBSCRIPT[ch] ?? raw)
+
+  return s
+}

ui-opentui/src/logic/memoryMonitor.ts

@@ -0,0 +1,88 @@
+/**
+ * Memory-monitor LOGIC (pure, no node:v8/process/file imports — testable).
+ *
+ * Ports the high-value SMART part of Ink's memory monitor
+ * (`ui-tui/src/lib/memoryMonitor.ts`): the #34095 silent-death EARLY-WARNING.
+ * It deliberately does NOT port Ink's auto heap-snapshot capture — the OpenTUI
+ * engine's always-on `memlog` NDJSON trace (boundary/memlog.ts) is the
+ * diagnosis path, and the rss-vs-heap divergence it records is the better
+ * diagnostic for the native-RSS leak class (#15141) that a V8 heap snapshot
+ * captures poorly anyway. So we skip the #41948 disk-fill bug class entirely.
+ *
+ * The early-warning regime is BELOW the OOM ceiling: Node can OOM from a render-
+ * tree / store blowup at a few hundred MB, well under any "critical" exit
+ * watermark, so a plain level machine never sees it and the death looks silent
+ * (#34095 showed up only as a bare gateway `stdin EOF`). We fire ONCE when heap
+ * both crosses a modest absolute floor AND is climbing steeply (≥150MB between
+ * ticks) — the render-tree-blowup signature — and re-arm only after heap falls
+ * back below the floor. The boundary turns the fire into a visible transcript
+ * system line so the user gets a heads-up before the process dies.
+ */
+
+const MB = 1024 ** 2
+
+/** Heap floor below which we never warn (a small heap climbing is normal). */
+export const WARN_FLOOR_BYTES = 600 * MB
+/** Per-tick growth that, combined with crossing the floor, signals a blowup. */
+export const WARN_GROWTH_STEP_BYTES = 150 * MB
+
+/** Mutable arm/disarm state for the early-warning detector. */
+export interface WarnState {
+  /** Previous heapUsed sample; `-1` until the first sample is seen. */
+  lastHeap: number
+  /** Whether we've already fired since the last re-arm (one-shot until reset). */
+  warned: boolean
+}
+
+/** A fresh, un-seeded warn state (lastHeap < 0 ⇒ first sample can't "grow"). */
+export function createWarnState(): WarnState {
+  return { lastHeap: -1, warned: false }
+}
+
+export interface WarnEvaluation {
+  /** True exactly on the tick the warning should fire (one-shot). */
+  readonly fire: boolean
+  /** The growth since the previous sample (bytes; 0 on the first sample). */
+  readonly growthBytes: number
+}
+
+/**
+ * Advance the early-warning state machine by one sample. MUTATES `state`
+ * (lastHeap + warned) and returns whether to fire this tick.
+ *
+ * Fires once when, while below any OOM ceiling: heap ≥ floor AND grew
+ * ≥ step since the previous sample AND we haven't already fired. Re-arms
+ * (warned ← false) once heap drops back below the floor. The first
+ * (un-seeded) sample only seeds lastHeap and never fires.
+ */
+export function evaluateWarn(
+  state: WarnState,
+  heapUsed: number,
+  floorBytes: number = WARN_FLOOR_BYTES,
+  stepBytes: number = WARN_GROWTH_STEP_BYTES
+): WarnEvaluation {
+  const seeded = state.lastHeap >= 0
+  const growthBytes = seeded ? heapUsed - state.lastHeap : 0
+  let fire = false
+
+  if (seeded) {
+    if (!state.warned && heapUsed >= floorBytes && growthBytes >= stepBytes) {
+      state.warned = true
+      fire = true
+    } else if (heapUsed < floorBytes) {
+      state.warned = false
+    }
+  }
+
+  state.lastHeap = heapUsed
+  return { fire, growthBytes }
+}
+
+/** Render the user-facing early-warning line (KB system line, no disk cost). */
+export function warnLine(heapUsed: number, rss: number, growthBytes: number): string {
+  const mb = (n: number) => Math.round(n / MB)
+  return (
+    `⚠ memory climbing fast — heap ${mb(heapUsed)}MB (+${mb(growthBytes)}MB), rss ${mb(rss)}MB. ` +
+    `If the TUI dies, this is why; relaunch with HERMES_TUI_DIAGNOSTICS=1 for a trace.`
+  )
+}

ui-opentui/src/logic/multiClick.ts

@@ -0,0 +1,168 @@
+/**
+ * Multi-click selection logic — double-click selects the word, triple-click the
+ * line, and a drag after either extends word-by-word / line-by-line while the
+ * originally clicked span stays selected (native macOS / VS Code behavior).
+ * Ported from the Ink fork's `hermes-ink/src/ink/selection.ts` (wordBoundsAt /
+ * selectLineAt / extendSelection) onto OpenTUI's screen model: the rendered
+ * frame is a flat grid of codepoints (`OptimizedBuffer.buffers.char`), so word
+ * scanning reads the frame the user actually sees — concealed markdown, tool
+ * chrome and all.
+ *
+ * Pure string/number work, no OpenTUI imports — the boundary shim
+ * (`boundary/multiClickSelect.ts`) adapts the live buffer to `ScreenText`.
+ */
+
+/** Screen-buffer cell coordinates (0-indexed col/row). */
+export interface Point {
+  readonly x: number
+  readonly y: number
+}
+
+/** Inclusive span from `lo` to `hi` in reading order (row-major). */
+export interface Span {
+  readonly lo: Point
+  readonly hi: Point
+}
+
+/** The multi-clicked span a drag extends from. */
+export interface AnchorSpan extends Span {
+  readonly kind: 'word' | 'line'
+}
+
+/** Read-only view of the rendered frame's character grid. */
+export interface ScreenText {
+  readonly width: number
+  readonly height: number
+  /** Unicode codepoint at cell (x,y); 0 marks a wide-char continuation cell. */
+  readonly codepointAt: (x: number, y: number) => number
+}
+
+/** -1 if a < b, 1 if a > b, 0 if equal (reading order: row then col). */
+export function comparePoints(a: Point, b: Point): number {
+  if (a.y !== b.y) return a.y < b.y ? -1 : 1
+  if (a.x !== b.x) return a.x < b.x ? -1 : 1
+  return 0
+}
+
+// Unicode-aware word character matcher: letters (any script), digits, and the
+// punctuation set iTerm2 treats as word-part by default (`/-+\~_.`). Matching
+// iTerm2's default means double-clicking a path like `src/logic/multiClick.ts`
+// selects the whole path — the muscle memory terminal users have.
+const WORD_CHAR = /[\p{L}\p{N}_/.\-+~\\]/u
+
+/**
+ * Character class for double-click word-expansion: 0 = whitespace/empty,
+ * 1 = word char, 2 = other punctuation. Cells with the same class as the
+ * clicked cell are one run; a class change is a boundary — so double-click on
+ * `foo` selects `foo`, on `->` selects `->`, on spaces the whitespace run.
+ */
+function charClass(cp: number): 0 | 1 | 2 {
+  if (cp === 0 || cp === 32) return 0
+  if (WORD_CHAR.test(String.fromCodePoint(cp))) return 1
+  return 2
+}
+
+/**
+ * Bounds of the same-class character run at (x, y), or null when the click is
+ * out of bounds. Wide-char continuation cells (codepoint 0) belong to the head
+ * glyph at their left: a click on one resolves to the head, the left scan
+ * steps over them to the head's class, and the right scan includes them in the
+ * span so the highlight covers the full glyph.
+ */
+export function wordSpanAt(screen: ScreenText, x: number, y: number): Span | null {
+  if (y < 0 || y >= screen.height || x < 0 || x >= screen.width) return null
+
+  // Land on a continuation cell → step back to the wide-char head.
+  let c = x
+  while (c > 0 && screen.codepointAt(c, y) === 0) c -= 1
+
+  const cls = charClass(screen.codepointAt(c, y))
+
+  let lo = c
+  while (lo > 0) {
+    let prev = lo - 1
+    while (prev > 0 && screen.codepointAt(prev, y) === 0) prev -= 1
+    if (charClass(screen.codepointAt(prev, y)) !== cls) break
+    lo = prev
+  }
+
+  let hi = c
+  while (hi < screen.width - 1) {
+    const cp = screen.codepointAt(hi + 1, y)
+    // A continuation cell after a run member is the tail of the run's last
+    // wide glyph — include it and keep scanning.
+    if (cp !== 0 && charClass(cp) !== cls) break
+    hi += 1
+  }
+
+  return { lo: { x: lo, y }, hi: { x: hi, y } }
+}
+
+/** The full row as a span (triple-click). Null when the row is out of bounds —
+ *  per-renderable `getSelectedText` trims what shouldn't copy, matching the
+ *  Ink fork where line-select spans the visual row. */
+export function lineSpanAt(screen: ScreenText, y: number): Span | null {
+  if (y < 0 || y >= screen.height || screen.width <= 0) return null
+  return { lo: { x: 0, y }, hi: { x: screen.width - 1, y } }
+}
+
+/**
+ * Where a drag at (x, y) puts the selection while an anchor span is held:
+ * the span under the mouse (word at the pointer, or its row in line mode;
+ * raw cell fallback when the pointer is out of bounds) is merged with the
+ * anchor span so the original word/line always stays selected.
+ */
+export function extendedSelection(
+  span: AnchorSpan,
+  screen: ScreenText,
+  x: number,
+  y: number
+): { anchor: Point; focus: Point } {
+  let mouseLo: Point
+  let mouseHi: Point
+
+  if (span.kind === 'word') {
+    const b = wordSpanAt(screen, x, y)
+    mouseLo = b ? b.lo : { x, y }
+    mouseHi = b ? b.hi : { x, y }
+  } else {
+    const row = Math.max(0, Math.min(y, screen.height - 1))
+    mouseLo = { x: 0, y: row }
+    mouseHi = { x: screen.width - 1, y: row }
+  }
+
+  // Mouse target entirely before the anchor span → grow backward from its end;
+  // entirely after → grow forward from its start; overlapping → just the span.
+  if (comparePoints(mouseHi, span.lo) < 0) return { anchor: span.hi, focus: mouseLo }
+  if (comparePoints(mouseLo, span.hi) > 0) return { anchor: span.lo, focus: mouseHi }
+  return { anchor: span.lo, focus: span.hi }
+}
+
+/** Same chain window the Ink fork uses (`App.tsx` MULTI_CLICK_*). */
+export const MULTI_CLICK_TIMEOUT_MS = 500
+export const MULTI_CLICK_DISTANCE = 1
+
+/**
+ * Click-chain counter: a press within MULTI_CLICK_TIMEOUT_MS and
+ * MULTI_CLICK_DISTANCE of the previous press continues the chain, otherwise
+ * the count resets to 1. The returned count is capped at 3 — quadruple+
+ * clicks stay line-select, like every terminal/editor.
+ */
+export function createClickCounter(): (x: number, y: number, now: number) => 1 | 2 | 3 {
+  let lastTime = 0
+  let lastX = -1
+  let lastY = -1
+  let count = 0
+
+  return (x, y, now) => {
+    const chained =
+      now - lastTime <= MULTI_CLICK_TIMEOUT_MS &&
+      Math.abs(x - lastX) <= MULTI_CLICK_DISTANCE &&
+      Math.abs(y - lastY) <= MULTI_CLICK_DISTANCE
+    count = chained ? count + 1 : 1
+    lastTime = now
+    lastX = x
+    lastY = y
+    return count >= 3 ? 3 : (count as 1 | 2)
+  }
+}

ui-opentui/src/logic/notificationDispatcher.ts

@@ -0,0 +1,29 @@
+/**
+ * Notification → desktop-OSC decision. EVERY notification renders an inline
+ * transcript card (so there's nothing to decide there); this only decides whether
+ * a notification is important enough to ALSO fire a desktop/terminal OSC ping
+ * (to pull the user back). The OSC payload is termChrome's `TermNotification`;
+ * the boundary (terminalChrome) owns the actual escape-sequence write.
+ */
+import type { ActivityNotification } from './backgroundActivity.ts'
+import type { TermNotification } from './termChrome.ts'
+
+/** Kind substrings that mark a "the work finished, look here" notification —
+ *  matched case-insensitively anywhere in the kind. */
+const COMPLETION_KIND_HINTS = ['complete', 'done', 'finish']
+
+function isImportant(n: ActivityNotification): boolean {
+  if (n.level === 'error' || n.level === 'warn') return true
+  const kind = n.kind.toLowerCase()
+  return COMPLETION_KIND_HINTS.some(hint => kind.includes(hint))
+}
+
+/**
+ * The desktop OSC notification for `n`, or `undefined` when it's not important
+ * enough to interrupt — level 'error'/'warn', or a kind containing
+ * 'complete'/'done'/'finish' (case-insensitive). Title is always 'Hermes' with
+ * the notification text as the body.
+ */
+export function notificationOsc(n: ActivityNotification): TermNotification | undefined {
+  return isImportant(n) ? { body: n.text, title: 'Hermes' } : undefined
+}

ui-opentui/src/logic/notify.ts

@@ -0,0 +1,35 @@
+/**
+ * Transient-notice seam (per-block copy feedback, Epic: design pass piece 2).
+ * Deep view nodes (e.g. the per-block `⧉` copy affordance in messageLine) need
+ * to flash a short notice ("Copied") on the EXISTING hint line (StatusLine —
+ * the same surface the entry's flashHint uses for /copy and selection-copy),
+ * but they don't hold the store. The store registers its `setHint` here at
+ * creation (one live store per app; the latest registration wins, which is
+ * also what headless tests want), and `flashNotice` mirrors the entry's
+ * flashHint contract: set, then auto-clear after `ms` unless something newer
+ * replaced it. No-op when nothing is registered (bare component tests).
+ */
+
+type NotifySink = (text: string | undefined) => void
+
+let sink: NotifySink | undefined
+let timer: ReturnType<typeof setTimeout> | undefined
+let current: string | undefined
+
+/** Register (or clear) the app-wide notice sink — the store's `setHint`. */
+export function registerNotifier(fn: NotifySink | undefined): void {
+  sink = fn
+}
+
+/** Flash a transient notice on the hint line; auto-clears after `ms`. */
+export function flashNotice(text: string, ms = 1500): void {
+  sink?.(text)
+  current = text
+  if (timer) clearTimeout(timer)
+  timer = setTimeout(() => {
+    if (current === text) {
+      sink?.(undefined)
+      current = undefined
+    }
+  }, ms)
+}

ui-opentui/src/logic/pastes.ts

@@ -0,0 +1,50 @@
+/**
+ * Pasted-text placeholders (free-code's model). A large paste isn't dumped raw
+ * into the composer — instead a compact `[Pasted text #N +M lines]` chip is shown
+ * and the real content is held in a Map, then expanded back on submit. Pure + no
+ * OpenTUI imports → trivially unit-testable.
+ *
+ * The store is created ONCE per session (entry) and passed to the Composer, so it
+ * survives the composer remounting when overlays open/close (a per-composer store
+ * would lose a pending paste mid-compose).
+ */
+
+export interface PasteStore {
+  /** Register a pasted block; returns the placeholder to insert into the input. */
+  add(text: string): string
+  /** Replace every `[Pasted text #N …]` placeholder with its stored content. */
+  expand(input: string): string
+  /** Drop all stored pastes (call after a successful submit). */
+  clear(): void
+}
+
+// Matches `[Pasted text #12]` and `[Pasted text #12 +34 lines]`. The id is the key.
+const REF = /\[Pasted text #(\d+)(?: \+\d+ lines)?\]/g
+
+export function createPasteStore(): PasteStore {
+  const map = new Map<number, string>()
+  let seq = 0
+  return {
+    add(text) {
+      const id = ++seq
+      map.set(id, text)
+      const lines = text.split('\n').length
+      return lines > 1 ? `[Pasted text #${id} +${lines} lines]` : `[Pasted text #${id}]`
+    },
+    // String.replace(/g) is a SINGLE left-to-right pass over the ORIGINAL string,
+    // so content inserted for one ref is never re-scanned for another ref —
+    // a pasted block that itself contains `[Pasted text #k]` is safe.
+    expand(input) {
+      return (input ?? '').replace(REF, (m, id: string) => map.get(Number(id)) ?? m)
+    },
+    clear() {
+      map.clear()
+      seq = 0
+    }
+  }
+}
+
+/** A paste big enough to placeholder rather than inline (conservative thresholds). */
+export function shouldPlaceholder(text: string): boolean {
+  return text.split('\n').length >= 4 || text.length > 400
+}

ui-opentui/src/logic/promptHistory.ts

@@ -0,0 +1,118 @@
+/**
+ * promptHistory — pure logic for the Esc+Esc session prompt viewer (Epic 5).
+ *
+ * Model: free-code's rewind dialog (`useDoublePress.ts`, `MessageSelector.tsx`)
+ * — 800ms double-press window, only-when-input-empty trigger, 7 visible rows
+ * newest-first with a centered window, Enter → confirm step.
+ *
+ * Semantics (spec Epic 5, RESOLVED block):
+ *   - Entries are THIS session's user prompts from the store transcript
+ *     (NOT the per-dir JSONL composer history), newest first. Empty → no modal.
+ *   - Undo = conversation layer (`/undo` removes the LAST user/assistant
+ *     exchange; files kept) → offered ONLY for the most recent prompt. We never
+ *     fake arbitrary-depth conversation rewind.
+ *   - Rollback = filesystem layer (`/rollback` checkpoints; conversation kept).
+ *     Prompt→checkpoint mapping isn't feasible client-side (neither store
+ *     messages nor `session.history` carry timestamps to correlate against
+ *     `rollback.list`'s checkpoint timestamps), so the honest action is plain
+ *     `/rollback`: the gateway's own checkpoint list lands in the transcript
+ *     and the user picks `/rollback <n>` from real data.
+ */
+
+/** Double-press window (free-code `DOUBLE_PRESS_TIMEOUT_MS`). */
+export const DOUBLE_PRESS_WINDOW_MS = 800
+
+/** Max visible prompt rows before the list windows (free-code `MAX_VISIBLE_MESSAGES`). */
+export const MAX_VISIBLE = 7
+
+/**
+ * Double-press detector (pure state machine; the free-code hook without React).
+ * `press(now)` returns true on the SECOND press within the window — and then
+ * disarms, so a third press starts a fresh cycle. `reset()` disarms (call it on
+ * any intervening key, and never call `press` for an Esc something else
+ * consumed — that's what keeps a dropdown-dismiss Esc from arming).
+ */
+export interface DoublePress {
+  press(now?: number): boolean
+  reset(): void
+}
+
+export function createDoublePress(windowMs: number = DOUBLE_PRESS_WINDOW_MS): DoublePress {
+  let armedAt: number | undefined
+  return {
+    // performance.now() is monotonic (Node) — an NTP/wall-clock jump between
+    // two presses can't break or spuriously satisfy the window (review finding).
+    press(now: number = performance.now()): boolean {
+      if (armedAt !== undefined && now - armedAt <= windowMs) {
+        armedAt = undefined
+        return true
+      }
+      armedAt = now
+      return false
+    },
+    reset(): void {
+      armedAt = undefined
+    }
+  }
+}
+
+/** One viewer row: a user prompt of THIS session. `index` is its position in
+ *  the source transcript (stable identity across renders). */
+export interface PromptEntry {
+  readonly index: number
+  readonly text: string
+}
+
+/**
+ * Source the viewer entries from the store transcript: USER prompts only,
+ * non-empty, NEWEST FIRST. Session-only by construction (the store holds only
+ * this session's messages). Empty session → [] (the trigger shows nothing).
+ */
+export function promptHistoryEntries(messages: ReadonlyArray<{ readonly role: string; text: string }>): PromptEntry[] {
+  const entries: PromptEntry[] = []
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const m = messages[i]
+    if (m && m.role === 'user' && m.text.trim() !== '') entries.push({ index: i, text: m.text })
+  }
+  return entries
+}
+
+/** A confirm-step action. */
+export type HistoryAction = 'undo' | 'rollback'
+
+export interface ConfirmOption {
+  readonly action: HistoryAction
+  readonly label: string
+}
+
+/** The exact signed-off confirm labels (spec Epic 5). */
+export const UNDO_LABEL = 'Undo — rewind the conversation (files kept)'
+export const ROLLBACK_LABEL = 'Rollback — restore files from checkpoint (conversation kept)'
+
+/**
+ * The confirm-step options for a selected entry. `/undo` only removes the LAST
+ * exchange, so Undo is offered ONLY for the most recent prompt (`isLatest`) —
+ * an option the gateway can't honor is hidden, never a dead button. Rollback
+ * (filesystem checkpoints) applies regardless of the selected depth.
+ */
+export function confirmOptions(isLatest: boolean): ConfirmOption[] {
+  const options: ConfirmOption[] = []
+  if (isLatest) options.push({ action: 'undo', label: UNDO_LABEL })
+  options.push({ action: 'rollback', label: ROLLBACK_LABEL })
+  return options
+}
+
+/** The slash command an action dispatches — through the SAME command path the
+ *  composer uses (`dispatchSlash` → `slash.exec`/`command.dispatch`). */
+export function actionCommand(action: HistoryAction): string {
+  return action === 'undo' ? '/undo' : '/rollback'
+}
+
+/**
+ * First visible row index for a list window: keep the selection centered until
+ * the window hits either end (free-code `firstVisibleIndex`). Total ≤ visible
+ * → 0 (everything shows).
+ */
+export function windowStart(selected: number, total: number, visible: number = MAX_VISIBLE): number {
+  return Math.max(0, Math.min(selected - Math.floor(visible / 2), total - visible))
+}

ui-opentui/src/logic/replay.ts

@@ -0,0 +1,126 @@
+/**
+ * /replay — spawn-tree inspector logic (Epic 3 port; Ink ref
+ * `app/slash/commands/ops.ts` /replay + `spawnHistoryStore.ts`). The gateway
+ * archives each completed delegation fan-out as a JSON snapshot
+ * (`spawn_tree.save`); these helpers read `spawn_tree.list` / `spawn_tree.load`
+ * payloads and format them as PAGER TEXT — the native engine renders replays
+ * through the existing pager overlay instead of Ink's agents overlay.
+ *
+ * All readers are defensive (wire JSON is loose, snapshots cross versions).
+ */
+
+export interface SpawnTreeEntry {
+  path: string
+  label: string
+  count: number
+  /** Epoch SECONDS (gateway convention). */
+  finishedAt?: number
+  sessionId?: string
+}
+
+function str(v: unknown): string | undefined {
+  return typeof v === 'string' && v ? v : undefined
+}
+
+function num(v: unknown): number | undefined {
+  return typeof v === 'number' && Number.isFinite(v) ? v : undefined
+}
+
+/** Map a `spawn_tree.list` result ({entries:[…]}) into typed rows (pathless rows dropped). */
+export function readSpawnTreeEntries(result: unknown): SpawnTreeEntry[] {
+  if (!result || typeof result !== 'object') return []
+  const entries = (result as { entries?: unknown }).entries
+  if (!Array.isArray(entries)) return []
+  const out: SpawnTreeEntry[] = []
+  for (const e of entries) {
+    if (!e || typeof e !== 'object') continue
+    const o = e as { [k: string]: unknown }
+    const path = str(o['path'])
+    if (!path) continue
+    const entry: SpawnTreeEntry = {
+      count: num(o['count']) ?? 0,
+      label: str(o['label']) ?? '',
+      path
+    }
+    const finishedAt = num(o['finished_at'])
+    if (finishedAt !== undefined) entry.finishedAt = finishedAt
+    const sessionId = str(o['session_id'])
+    if (sessionId !== undefined) entry.sessionId = sessionId
+    out.push(entry)
+  }
+  return out
+}
+
+function fmtWhen(epochSeconds: number | undefined): string {
+  if (epochSeconds === undefined) return '?'
+  try {
+    return new Date(epochSeconds * 1000).toLocaleString()
+  } catch {
+    return '?'
+  }
+}
+
+/** The bare `/replay` listing: indexed rows the user replays by number. */
+export function formatSpawnTreeList(entries: readonly SpawnTreeEntry[]): string {
+  const lines: string[] = ['Archived spawn trees — /replay <n> to view, /replay <path> for any snapshot', '']
+  entries.forEach((e, i) => {
+    const label = e.label || `${e.count} subagent${e.count === 1 ? '' : 's'}`
+    lines.push(`${String(i + 1).padStart(3)}. ${fmtWhen(e.finishedAt)} · ${e.count}× — ${label}`)
+    lines.push(`     ${e.path}`)
+  })
+  return lines.join('\n')
+}
+
+/** Status glyph for an archived subagent row. */
+function statusGlyph(status: string): string {
+  if (status === 'completed') return '✓'
+  if (status === 'error' || status === 'failed' || status === 'timeout') return '✗'
+  if (status === 'interrupted') return '⏹'
+  return '●'
+}
+
+/** One archived subagent → its pager lines (indented by spawn depth). */
+function subagentLines(raw: unknown, index: number): string[] {
+  const o = (raw && typeof raw === 'object' ? raw : {}) as { [k: string]: unknown }
+  const depth = num(o['depth']) ?? 0
+  const pad = '  '.repeat(Math.max(0, depth))
+  const status = str(o['status']) ?? 'completed'
+  const goal = str(o['goal']) ?? 'subagent'
+  const lines = [`${pad}${statusGlyph(status)} [${index + 1}] ${goal}`]
+  const meta: string[] = [status]
+  const model = str(o['model'])
+  if (model) meta.push(model)
+  const duration = num(o['durationSeconds'])
+  if (duration !== undefined) meta.push(`${Math.round(duration)}s`)
+  const tools = num(o['toolCount'])
+  if (tools) meta.push(`${tools} tool${tools === 1 ? '' : 's'}`)
+  const tokIn = num(o['inputTokens'])
+  const tokOut = num(o['outputTokens'])
+  if (tokIn !== undefined || tokOut !== undefined) meta.push(`${tokIn ?? 0} in / ${tokOut ?? 0} out tok`)
+  lines.push(`${pad}    ${meta.join(' · ')}`)
+  const summary = str(o['summary'])
+  if (summary) for (const s of summary.split('\n')) lines.push(`${pad}    ${s}`)
+  const notes = o['notes']
+  if (Array.isArray(notes)) {
+    for (const note of notes) if (typeof note === 'string' && note) lines.push(`${pad}    · ${note}`)
+  }
+  return lines
+}
+
+/** A loaded snapshot (`spawn_tree.load` payload) → the full pager text. */
+export function formatSpawnTree(payload: unknown): string {
+  const o = (payload && typeof payload === 'object' ? payload : {}) as { [k: string]: unknown }
+  const subagents = Array.isArray(o['subagents']) ? (o['subagents'] as unknown[]) : []
+  const header: string[] = []
+  const label = str(o['label'])
+  header.push(label ?? 'spawn tree')
+  const meta: string[] = []
+  const sessionId = str(o['session_id'])
+  if (sessionId) meta.push(`session ${sessionId}`)
+  meta.push(`finished ${fmtWhen(num(o['finished_at']))}`)
+  meta.push(`${subagents.length} subagent${subagents.length === 1 ? '' : 's'}`)
+  header.push(meta.join(' · '))
+  if (!subagents.length) return [...header, '', '(snapshot empty or unreadable)'].join('\n')
+  const body = subagents.flatMap((s, i) => ['', ...subagentLines(s, i)])
+  return [...header, ...body].join('\n')
+}

ui-opentui/src/logic/resume.ts

@@ -0,0 +1,101 @@
+/**
+ * Resume snapshot mapper (spec §1 lifecycle; gotcha §8 #5). Maps the
+ * `session.resume` response `messages` (tui_gateway `_history_to_messages`) into
+ * the store's `Message[]`. Each history entry is either `{role, text}` (user/
+ * assistant/system) or `{role:'tool', name, context}` (NO text — render it).
+ *
+ * Tool rows are folded into the PRECEDING assistant turn's ordered `parts[]`
+ * (state:'complete', summary=context) so a resumed transcript renders inline like
+ * a live one. Resumed assistant text is given a single text part so it renders
+ * through the native markdown path. IDs are `r*` (distinct from live `p*`).
+ */
+import type { Message, Part, SessionItem, ToolPartState } from './store.ts'
+import { stripOmittedNote, stripToolEnvelope } from './toolOutput.ts'
+
+function readStr(value: unknown, key: string): string | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'string' ? v : undefined
+}
+
+function readNum(value: unknown, key: string): number {
+  if (!value || typeof value !== 'object') return 0
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'number' ? v : 0
+}
+
+/** Map a `session.list` result into switcher rows (loose-typed read). */
+export function mapSessionList(result: unknown): SessionItem[] {
+  if (!result || typeof result !== 'object') return []
+  const sessions = (result as { sessions?: unknown }).sessions
+  if (!Array.isArray(sessions)) return []
+  const out: SessionItem[] = []
+  for (const s of sessions) {
+    const id = readStr(s, 'id')
+    if (!id) continue
+    out.push({
+      id,
+      messageCount: readNum(s, 'message_count'),
+      preview: readStr(s, 'preview') ?? '',
+      title: readStr(s, 'title') ?? ''
+    })
+  }
+  return out
+}
+
+export function mapResumeHistory(history: unknown): Message[] {
+  if (!Array.isArray(history)) return []
+  const out: Message[] = []
+  let seq = 0
+  const id = () => `r${++seq}`
+  let currentAssistant: Message | undefined
+
+  for (const raw of history) {
+    const role = readStr(raw, 'role')
+
+    if (role === 'tool') {
+      const name = readStr(raw, 'name') ?? 'tool'
+      const context = readStr(raw, 'context')
+      const tool: ToolPartState = { type: 'tool', id: id(), name, state: 'complete' }
+      // Match the live tool part exactly (item 1): primary-arg preview in the
+      // header, plus the (capped) output so resumed tools are collapsible too.
+      if (context) tool.argsPreview = context
+      const rawResult = readStr(raw, 'result_text')
+      if (rawResult) {
+        const { body, omittedNote } = stripOmittedNote(rawResult)
+        const resultText = stripToolEnvelope(body)
+        if (resultText) {
+          tool.resultText = resultText
+          tool.lineCount = resultText.replace(/\s+$/, '').split('\n').length
+        }
+        if (omittedNote) tool.omittedNote = omittedNote
+      }
+      const args = (raw as { args?: unknown }).args
+      if (args && typeof args === 'object') {
+        try {
+          tool.argsText = JSON.stringify(args, null, 2)
+        } catch {
+          /* unstringifiable — leave unset */
+        }
+      }
+      if (!currentAssistant) {
+        currentAssistant = { role: 'assistant', text: '', parts: [] }
+        out.push(currentAssistant)
+      }
+      ;(currentAssistant.parts ??= []).push(tool)
+      continue
+    }
+
+    const text = readStr(raw, 'text') ?? ''
+    if (role === 'assistant') {
+      const parts: Part[] = text ? [{ type: 'text', id: id(), text }] : []
+      currentAssistant = { role: 'assistant', text, parts }
+      out.push(currentAssistant)
+    } else if (role === 'user' || role === 'system') {
+      out.push({ role, text })
+      currentAssistant = undefined
+    }
+  }
+
+  return out
+}

ui-opentui/src/logic/sessionPicker.ts

@@ -0,0 +1,332 @@
+/**
+ * sessionPicker.ts — pure logic for the tabbed resume picker (design doc
+ * docs/plans/opentui-resume-picker.md §A/§B item 5; supersedes the flat
+ * SessionSwitcher). Everything here is view-free and vitest-covered:
+ *
+ *   - tab definitions + source→tab classification (Recent = interactive
+ *     cli/tui/acp + unknown/custom; Cron = cron; Gateways = the known platform
+ *     sources; All = everything minus the deny-listed `tool`),
+ *   - the `session.list` params each tab queries with (`sources` allow-list —
+ *     note the one honest gap: unknown/custom sources CLASSIFY as Recent but
+ *     can't be expressed in an allow-list, so they surface under All),
+ *   - the client-side search filter chain over title/preview/cwd/id (reuses
+ *     fuzzy.ts — same scorer as the model picker),
+ *   - the key-routing decision table (pattern: completionMenu's routeMenuKey),
+ *   - the relative-time formatter + row-meta composer (time · source · N msgs
+ *     · tail-truncated cwd),
+ *   - `/sessions <tab>` arg parsing and the `/resume <id|name>` resolver.
+ */
+import { fuzzyFilter, type FuzzyField } from './fuzzy.ts'
+
+// ── tabs + classification ─────────────────────────────────────────────────
+
+export type SessionTabId = 'recent' | 'cron' | 'gateways' | 'all'
+
+/** Tab strip order + labels (design doc §A — Recent is the default). */
+export const SESSION_TABS: ReadonlyArray<{ id: SessionTabId; label: string }> = [
+  { id: 'recent', label: 'Recent' },
+  { id: 'cron', label: 'Cron' },
+  { id: 'gateways', label: 'Gateways' },
+  { id: 'all', label: 'All' }
+]
+
+/** Interactive sources — the Recent tab's allow-list. */
+export const INTERACTIVE_SOURCES: readonly string[] = ['cli', 'tui', 'acp']
+
+/** Known platform/gateway sources (the Gateways tab's allow-list). The gateway
+ *  itself deny-lists only `tool`, so this list is the picker's working set of
+ *  "messaging platform" tags; new platforms join here (or show under All). */
+export const PLATFORM_SOURCES: readonly string[] = [
+  'telegram',
+  'discord',
+  'slack',
+  'whatsapp',
+  'signal',
+  'imessage',
+  'matrix',
+  'teams',
+  'email',
+  'webhook',
+  'x',
+  'twitter',
+  'mastodon',
+  'irc',
+  'mattermost'
+]
+
+/** Classify a session `source` tag into its home tab (`tool` = deny-listed —
+ *  never shown). Unknown/custom sources (incl. empty) default to Recent per
+ *  the design table: they're assumed interactive `HERMES_SESSION_SOURCE`s. */
+export function classifySource(source: string | undefined): 'recent' | 'cron' | 'gateways' | 'tool' {
+  const s = (source ?? '').trim().toLowerCase()
+  if (s === 'tool') return 'tool'
+  if (s === 'cron') return 'cron'
+  if (PLATFORM_SOURCES.includes(s)) return 'gateways'
+  return 'recent'
+}
+
+/** Whether a row with this source belongs on the given tab. */
+export function tabAccepts(tab: SessionTabId, source: string | undefined): boolean {
+  const cls = classifySource(source)
+  if (cls === 'tool') return false
+  return tab === 'all' || cls === tab
+}
+
+/**
+ * The `session.list` params a tab queries with. Cron/Gateways push an exact
+ * `sources` allow-list to the gateway; All omits it (the gateway deny-lists
+ * `tool` itself). Recent sends the interactive allow-list — the one honest gap
+ * vs `classifySource` (unknown/custom sources can't be allow-listed, so they
+ * appear under All only); fetching everything and filtering client-side would
+ * make Recent unusable in cron-heavy DBs (1500+ cron rows drown the page).
+ */
+export function listParamsFor(tab: SessionTabId, offset: number, limit: number): Record<string, unknown> {
+  const base: Record<string, unknown> = { limit, offset }
+  if (tab === 'recent') return { ...base, sources: [...INTERACTIVE_SOURCES] }
+  if (tab === 'cron') return { ...base, sources: ['cron'] }
+  if (tab === 'gateways') return { ...base, sources: [...PLATFORM_SOURCES] }
+  return base
+}
+
+// ── session.list row mapping ──────────────────────────────────────────────
+
+/** One picker row — the widened `session.list` projection (gateway 529d8084b). */
+export interface SessionRow {
+  id: string
+  title: string
+  preview: string
+  source: string
+  messageCount: number
+  startedAt: number
+  lastActive: number
+  endedAt?: number
+  model?: string
+  cwd?: string
+}
+
+function readStr(value: unknown, key: string): string | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'string' ? v : undefined
+}
+
+function readNum(value: unknown, key: string): number {
+  if (!value || typeof value !== 'object') return 0
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'number' ? v : 0
+}
+
+/** Map a widened `session.list` result into rows + the honesty flag. */
+export function mapSessionRows(result: unknown): { rows: SessionRow[]; truncated: boolean } {
+  if (!result || typeof result !== 'object') return { rows: [], truncated: false }
+  const sessions = (result as { sessions?: unknown }).sessions
+  const truncated = (result as { truncated?: unknown }).truncated === true
+  if (!Array.isArray(sessions)) return { rows: [], truncated }
+  const rows: SessionRow[] = []
+  for (const s of sessions) {
+    const id = readStr(s, 'id')
+    if (!id) continue
+    const row: SessionRow = {
+      id,
+      lastActive: readNum(s, 'last_active') || readNum(s, 'started_at'),
+      messageCount: readNum(s, 'message_count'),
+      preview: readStr(s, 'preview') ?? '',
+      source: readStr(s, 'source') ?? '',
+      startedAt: readNum(s, 'started_at'),
+      title: readStr(s, 'title') ?? ''
+    }
+    const endedAt = readNum(s, 'ended_at')
+    if (endedAt) row.endedAt = endedAt
+    const model = readStr(s, 'model')
+    if (model) row.model = model
+    const cwd = readStr(s, 'cwd')
+    if (cwd) row.cwd = cwd
+    rows.push(row)
+  }
+  return { rows, truncated }
+}
+
+// ── search filter chain (client-side, within the active tab) ──────────────
+
+/** Fuzzy haystacks of a row: title ×2 (primary), preview, cwd, id. */
+export function sessionFields(row: SessionRow): FuzzyField[] {
+  const fields: FuzzyField[] = []
+  if (row.title) fields.push({ text: row.title, weight: 2 })
+  if (row.preview) fields.push({ text: row.preview })
+  if (row.cwd) fields.push({ text: row.cwd })
+  fields.push({ text: row.id })
+  return fields
+}
+
+/** Filter + rank rows by the search query (empty → all rows, fetch order). */
+export function filterSessions(query: string, rows: readonly SessionRow[]): SessionRow[] {
+  return fuzzyFilter(query, rows, sessionFields)
+}
+
+// ── this-directory grouping ───────────────────────────────────────────────
+
+/** Path equality for cwd grouping: trim + drop trailing slashes. Pure string
+ *  work (no fs) — rows carry the gateway's already-absolute paths. */
+export function normalizeCwd(path: string | undefined): string {
+  return (path ?? '').trim().replace(/\/+$/, '')
+}
+
+/** Display order with sessions started in the CURRENT directory first.
+ *
+ * Browse mode only: while a search query is active the fuzzy score owns the
+ * order (relevance beats locality), so `hereCount` is 0 and rows pass through.
+ * Stable within both groups (each keeps the gateway's recency order). The
+ * view renders section captions off `hereCount`; selection math is untouched
+ * because this just reorders the one flat list.
+ */
+export function orderRowsForCwd(
+  rows: SessionRow[],
+  currentCwd: string | undefined,
+  query: string
+): { rows: SessionRow[]; hereCount: number } {
+  const here = normalizeCwd(currentCwd)
+  if (!here || query.trim()) return { hereCount: 0, rows }
+  const local: SessionRow[] = []
+  const elsewhere: SessionRow[] = []
+  for (const row of rows) (normalizeCwd(row.cwd) === here ? local : elsewhere).push(row)
+  if (!local.length) return { hereCount: 0, rows }
+  return { hereCount: local.length, rows: [...local, ...elsewhere] }
+}
+
+// ── key routing (pattern: completionMenu.ts routeMenuKey) ────────────────
+
+export interface SessionPickerKeyContext {
+  /** Whether the inline Ctrl+R rename is active (it owns Enter/Esc). */
+  renaming: boolean
+  /** Whether the search query is empty (←/→ only cycle tabs when it is). */
+  queryEmpty: boolean
+}
+
+export type SessionPickerAction =
+  | { kind: 'close' }
+  | { kind: 'resume' }
+  | { kind: 'move'; dir: 1 | -1 }
+  | { kind: 'cycle-tab'; dir: 1 | -1 }
+  | { kind: 'preview' }
+  | { kind: 'rename' }
+  | { kind: 'commit-rename' }
+  | { kind: 'cancel-rename' }
+  | { kind: 'pass' }
+
+const PASS: SessionPickerAction = { kind: 'pass' }
+
+/**
+ * Route one key press. While RENAMING, the rename input owns every key except
+ * Enter (commit) and Esc/Ctrl+C (cancel rename — NOT close). Otherwise:
+ * Esc/Ctrl+C close, Enter resumes, ↑↓ (or Ctrl+P/N) move, Tab/Shift+Tab cycle
+ * tabs, ←/→ cycle only on an empty query (with text they stay cursor moves),
+ * Space toggles the preview (it never types — fuzzy terms don't need literal
+ * spaces), Ctrl+R starts the inline rename. Everything else belongs to the
+ * focused search input.
+ */
+export function routeSessionPickerKey(
+  name: string,
+  mods: { ctrl?: boolean; shift?: boolean },
+  ctx: SessionPickerKeyContext
+): SessionPickerAction {
+  if (ctx.renaming) {
+    if (name === 'return') return { kind: 'commit-rename' }
+    if (name === 'escape' || (mods.ctrl && name === 'c')) return { kind: 'cancel-rename' }
+    return PASS
+  }
+  if (name === 'escape' || (mods.ctrl && name === 'c')) return { kind: 'close' }
+  if (name === 'return') return { kind: 'resume' }
+  if (name === 'up' || (mods.ctrl && name === 'p')) return { kind: 'move', dir: -1 }
+  if (name === 'down' || (mods.ctrl && name === 'n')) return { kind: 'move', dir: 1 }
+  if (name === 'tab') return { kind: 'cycle-tab', dir: mods.shift ? -1 : 1 }
+  if ((name === 'left' || name === 'right') && ctx.queryEmpty) {
+    return { kind: 'cycle-tab', dir: name === 'left' ? -1 : 1 }
+  }
+  if (name === 'space') return { kind: 'preview' }
+  if (mods.ctrl && name === 'r') return { kind: 'rename' }
+  return PASS
+}
+
+// ── relative time + row meta ──────────────────────────────────────────────
+
+/** Epoch seconds OR milliseconds → ms (DB rows are seconds; be lenient). */
+function toMs(epoch: number): number {
+  return epoch >= 1e12 ? epoch : epoch * 1000
+}
+
+const TIME_STEPS: ReadonlyArray<{ ms: number; unit: string }> = [
+  { ms: 60_000, unit: 'minute' },
+  { ms: 3_600_000, unit: 'hour' },
+  { ms: 86_400_000, unit: 'day' },
+  { ms: 604_800_000, unit: 'week' },
+  { ms: 2_629_800_000, unit: 'month' },
+  { ms: 31_557_600_000, unit: 'year' }
+]
+
+/** "just now" / "1 minute ago" / "5 hours ago" / "2 weeks ago" … */
+export function relativeTime(epoch: number | undefined, nowMs: number): string {
+  if (!epoch) return 'unknown'
+  const delta = nowMs - toMs(epoch)
+  if (delta < 60_000) return 'just now'
+  for (let i = TIME_STEPS.length - 1; i >= 0; i--) {
+    const step = TIME_STEPS[i]
+    if (step && delta >= step.ms) {
+      const n = Math.floor(delta / step.ms)
+      return `${n} ${step.unit}${n === 1 ? '' : 's'} ago`
+    }
+  }
+  return 'just now'
+}
+
+/** Tail-truncate a path-ish string to `max` chars (`…tail/of/path`). */
+export function tailTruncate(text: string, max: number): string {
+  if (text.length <= max) return text
+  return `…${text.slice(text.length - (max - 1))}`
+}
+
+/** Max cwd tail shown in a row's meta line. */
+const META_CWD_MAX = 40
+
+/** Row meta line: relative time · source · N msgs · cwd (when present). */
+export function rowMeta(row: SessionRow, nowMs: number): string {
+  const parts = [
+    relativeTime(row.lastActive || row.startedAt, nowMs),
+    row.source || 'unknown',
+    `${row.messageCount} msgs`
+  ]
+  if (row.cwd) parts.push(tailTruncate(row.cwd, META_CWD_MAX))
+  return parts.join(' · ')
+}
+
+// ── slash entry points ────────────────────────────────────────────────────
+
+/** Parse a `/sessions <tab>` argument (case-insensitive, singular tolerated;
+ *  bare/empty → the default Recent tab). Garbage → undefined (usage notice). */
+export function parseSessionTabArg(arg: string): SessionTabId | undefined {
+  const a = arg.trim().toLowerCase()
+  if (!a || a === 'recent') return 'recent'
+  if (a === 'cron') return 'cron'
+  if (a === 'gateway' || a === 'gateways') return 'gateways'
+  if (a === 'all') return 'all'
+  return undefined
+}
+
+/**
+ * Resolve a `/resume <id|name>` argument against listed rows (the direct
+ * path): exact id → unique id prefix → exact title (case-insensitive) →
+ * unique case-insensitive title substring. Ambiguous/missing → undefined.
+ */
+export function resolveSessionArg(rows: readonly SessionRow[], arg: string): SessionRow | undefined {
+  const needle = arg.trim()
+  if (!needle) return undefined
+  const exactId = rows.find(r => r.id === needle)
+  if (exactId) return exactId
+  const idPrefix = rows.filter(r => r.id.startsWith(needle))
+  if (idPrefix.length === 1) return idPrefix[0]
+  const lower = needle.toLowerCase()
+  const exactTitle = rows.filter(r => r.title.toLowerCase() === lower)
+  if (exactTitle.length === 1) return exactTitle[0]
+  const sub = rows.filter(r => r.title.toLowerCase().includes(lower))
+  if (sub.length === 1) return sub[0]
+  return undefined
+}

ui-opentui/src/logic/skillMatch.ts

@@ -0,0 +1,161 @@
+/**
+ * Slash-token matching for the composer (Epic 6) — pure tokenizer + matcher,
+ * no deps, fully table-testable. The composer uses this to:
+ *
+ *   1. HIGHLIGHT a `/name` token whose name exactly matches a valid
+ *      command/skill name (native textarea highlight ranges),
+ *   2. SUGGEST an autocorrect when the message IS a bare `/name` token at the
+ *      very start and the name is exactly one edit away (Damerau-Levenshtein /
+ *      OSA distance 1) from exactly ONE valid name — surfaced through the
+ *      existing completion dropdown, never auto-applied.
+ *
+ * Anti-jank rule (the whole point): a `/` in the middle of prose must NOT
+ * trigger completion or autocorrect. Mid-prose tokens get highlight-only when
+ * they exactly match a valid name; otherwise nothing happens. Path-looking
+ * tokens (`a/b`, `/usr/bin`, `./x`) are never tokens at all.
+ *
+ * The catalog of valid names is supplied by the caller (the composer LEARNS it
+ * from the slash-completion batches the gateway already sends — the completion
+ * flow is the source of truth; nothing is hardcoded here).
+ */
+
+/** A standalone `/name` token found in the composer text. */
+export interface SlashToken {
+  /** The name WITHOUT the leading `/`. */
+  name: string
+  /** Char offset of the leading `/` in the text. */
+  start: number
+  /** Char offset one past the last name char. */
+  end: number
+  /** Whether the token sits at the very start of the message (offset 0). */
+  lead: boolean
+}
+
+/** An autocorrect suggestion for the lead token (`/comit` → `commit`). */
+export interface SlashSuggestion {
+  /** The corrected name (no slash). */
+  name: string
+  /** Char offset the accepted suggestion replaces from (just past the `/`). */
+  from: number
+}
+
+export interface SlashAnalysis {
+  /** Tokens whose name EXACTLY matches a valid name — highlight these. */
+  highlights: SlashToken[]
+  /** The one-edit autocorrect for a bare lead token; null when none applies. */
+  suggestion: SlashSuggestion | null
+}
+
+/** Command/skill name charset: starts alphanumeric, then word chars / `.` / `-`.
+ *  Notably EXCLUDES `/` — `/usr/bin` is a path, never a command token. */
+const NAME_RE = /^[A-Za-z0-9][\w.-]*$/
+
+const isSpace = (ch: string | undefined): boolean => ch === ' ' || ch === '\t' || ch === '\n' || ch === '\r'
+
+/**
+ * Extract every standalone `/name` token. Boundary rules:
+ *   - the `/` must be at the start of the text or preceded by whitespace
+ *     (`a/b` and `path/to` are not tokens),
+ *   - the name runs to the next whitespace (or end) and must match NAME_RE
+ *     (`/usr/bin` has a `/` in the body → not a token; bare `/` is nothing).
+ */
+export function slashTokens(text: string): SlashToken[] {
+  const tokens: SlashToken[] = []
+  for (let i = 0; i < text.length; i++) {
+    if (text[i] !== '/') continue
+    if (i > 0 && !isSpace(text[i - 1])) continue
+    let j = i + 1
+    while (j < text.length && !isSpace(text[j])) j++
+    const name = text.slice(i + 1, j)
+    if (NAME_RE.test(name)) tokens.push({ end: j, lead: i === 0, name, start: i })
+    i = j
+  }
+  return tokens
+}
+
+/**
+ * Whether `a` and `b` are EXACTLY one edit apart under Damerau-Levenshtein
+ * (OSA): one substitution, one insertion/deletion, or one adjacent
+ * transposition. Equal strings are zero edits → false.
+ */
+export function isOneEdit(a: string, b: string): boolean {
+  if (a === b) return false
+  const la = a.length
+  const lb = b.length
+  if (Math.abs(la - lb) > 1) return false
+  if (la === lb) {
+    // one substitution, or one adjacent transposition
+    let i = 0
+    while (i < la && a[i] === b[i]) i++
+    if (i === la) return false // identical (handled above, defensive)
+    // try substitution: rest after i must match
+    if (a.slice(i + 1) === b.slice(i + 1)) return true
+    // try transposition of i,i+1
+    return i + 1 < la && a[i] === b[i + 1] && a[i + 1] === b[i] && a.slice(i + 2) === b.slice(i + 2)
+  }
+  // one insertion/deletion: align the longer against the shorter
+  const [short, long] = la < lb ? [a, b] : [b, a]
+  let i = 0
+  while (i < short.length && short[i] === long[i]) i++
+  return short.slice(i) === long.slice(i + 1)
+}
+
+/**
+ * Analyze the composer text against the valid-name catalog.
+ * Matching is CASE-SENSITIVE per the catalog (commands are stored lowercase;
+ * `/Help` is not exact — though it IS one edit from `help`, so it suggests).
+ *
+ * Suggestion rules (anti-jank):
+ *   - only for the LEAD token, and only while the message is EXACTLY the bare
+ *     token (`/comit` — not `/comit args`, never mid-prose),
+ *   - the token must not already be exact,
+ *   - exactly ONE catalog name within one edit; ambiguity → nothing.
+ */
+export function analyzeSlash(text: string, names: ReadonlySet<string>): SlashAnalysis {
+  const tokens = slashTokens(text)
+  const highlights = tokens.filter(t => names.has(t.name))
+  let suggestion: SlashSuggestion | null = null
+  const lead = tokens[0]
+  if (lead && lead.lead && text === `/${lead.name}` && !names.has(lead.name)) {
+    let candidate: string | undefined
+    let count = 0
+    for (const n of names) {
+      if (isOneEdit(lead.name, n)) {
+        candidate = n
+        if (++count > 1) break
+      }
+    }
+    if (count === 1 && candidate !== undefined) suggestion = { from: 1, name: candidate }
+  }
+  return { highlights, suggestion }
+}
+
+/**
+ * Names learnable from a slash-completion batch. Only when the composer text is
+ * a bare lead token (`/…` with no space) are the candidates command/skill NAMES
+ * — after a space the gateway completes ARGS (`/details thinking`), which must
+ * not pollute the catalog. Item text arrives as `name `, `name`, or `/name`
+ * (the gateway's extras carry the slash); all normalize to the bare name.
+ */
+export function learnableNames(text: string, items: ReadonlyArray<{ text: string }>): string[] {
+  if (!/^\/\S*$/.test(text)) return []
+  const out: string[] = []
+  for (const item of items) {
+    let name = item.text.trim()
+    if (name.startsWith('/')) name = name.slice(1)
+    if (NAME_RE.test(name)) out.push(name)
+  }
+  return out
+}
+
+/**
+ * Convert a JS string offset into the native highlight char offset: the native
+ * char-range counter skips newlines (mirror of ExtmarksController.
+ * offsetExcludingNewlines in @opentui/core for plain-width text).
+ */
+export function nativeCharOffset(text: string, offset: number): number {
+  let newlines = 0
+  const max = Math.min(offset, text.length)
+  for (let i = 0; i < max; i++) if (text[i] === '\n') newlines++
+  return offset - newlines
+}

ui-opentui/src/logic/slash.ts

@@ -0,0 +1,894 @@
+/**
+ * Slash command system — the SOLID side (spec §1; mirrors Ink
+ * `app/createSlashHandler.ts` + `domain/slash.ts`). Plain functions/data, NOT
+ * Effect; the boundary injects a Promise-returning `request` so dispatch can call
+ * `slash.exec` / `command.dispatch` / `commands.catalog`.
+ *
+ * Dispatch ladder (Ink parity):
+ *   1. client-local command (the TUI-only set — handled in-process)
+ *   2. `slash.exec {command, session_id}` → `{output, warning?}` → system line
+ *   3. on reject → `command.dispatch {arg, name, session_id}` → typed action
+ *      (exec/plugin → system · alias → re-dispatch · skill/send → submit a turn ·
+ *       prefill → notice). Long output routes to the pager (Phase 5a).
+ */
+import { diagnosticsEnabled } from './env.ts'
+import { DETAILS_SECTIONS, DETAILS_USAGE, type DetailsMode, nextDetailsMode, parseDetailsMode } from './details.ts'
+import { formatBytes, memReport, performHeapdump } from './diagnostics.ts'
+import { formatSpawnTree, formatSpawnTreeList, readSpawnTreeEntries } from './replay.ts'
+import { mapSessionRows, parseSessionTabArg, resolveSessionArg, type SessionTabId } from './sessionPicker.ts'
+import type { CompletionItem, PickerItem, PickerState } from './store.ts'
+
+export interface ParsedSlash {
+  name: string
+  arg: string
+}
+
+/** Parse `/name rest…` → {name, arg}; null if not a slash command. */
+export function parseSlash(input: string): ParsedSlash | null {
+  if (!input.startsWith('/')) return null
+  const body = input.slice(1).trimStart()
+  if (!body) return null
+  const sp = body.indexOf(' ')
+  return sp === -1 ? { arg: '', name: body } : { arg: body.slice(sp + 1).trim(), name: body.slice(0, sp) }
+}
+
+/** How a submitted composer line is routed (F9 + slash ladder): a `!cmd` runs a
+ *  shell command, a `/command` goes through the slash dispatcher, everything else
+ *  is a prompt turn. `payload` is the command (shell) with the lead `!` stripped
+ *  and trimmed, or the original text (slash/prompt). */
+export type SubmitRoute =
+  | { kind: 'shell'; payload: string }
+  | { kind: 'slash'; payload: string }
+  | { kind: 'prompt'; payload: string }
+
+export function classifySubmit(text: string): SubmitRoute {
+  if (text.startsWith('!')) return { kind: 'shell', payload: text.slice(1).trim() }
+  if (text.startsWith('/')) return { kind: 'slash', payload: text }
+  return { kind: 'prompt', payload: text }
+}
+
+/** The host capabilities the dispatcher needs (wired by the entry boundary). */
+export interface SlashContext {
+  /** Server RPC (resolves with the result, rejects on GatewayError). */
+  readonly request: (method: string, params: Record<string, unknown>) => Promise<unknown>
+  readonly sessionId: () => string | undefined
+  readonly pushSystem: (text: string) => void
+  /** Open the full-screen pager (long output: /status, /logs, …). */
+  readonly openPager: (title: string, text: string) => void
+  /** Submit a user turn (skill/send dispatch results). */
+  readonly submit: (text: string) => void
+  /** Open a local Y/N confirm; `onConfirm` runs on Yes. */
+  readonly confirm: (message: string, onConfirm: () => void) => void
+  readonly clearTranscript: () => void
+  /** Copy the n-th newest assistant response to the clipboard; returns whether something was copied. */
+  readonly copyResponse: (n: number) => boolean
+  readonly quit: () => void
+  /** Recent log lines for `/logs` (the ring buffer). */
+  readonly logTail: () => string[]
+  /** Open the tabbed resume picker on the given tab (/sessions, bare /resume). */
+  readonly openSessionPicker: (tab: SessionTabId) => void
+  /** Resume a session directly by id (`/resume <id|name>` — no picker). */
+  readonly resumeSession: (sessionId: string) => void
+  /** Open a generic picker (model picker, skills hub). */
+  readonly openPicker: (picker: PickerState) => void
+  /** Open the agents dashboard (/agents, /tasks). */
+  readonly openDashboard: () => void
+  /** Open the OS background-process panel (/processes). */
+  readonly openBackgroundPanel: () => void
+  /** Track an in-flight background-prompt task id (`/bg` → prompt.background). */
+  readonly addBgTask: (id: string) => void
+  /** Cached `/model` picker rows (Epic 7 instant open); undefined until prefetched. */
+  readonly modelItems: () => PickerItem[] | undefined
+  /** Update the cached `/model` picker rows. */
+  readonly setModelItems: (items: PickerItem[]) => void
+  /** Read / set the compact-transcript display flag (/compact — Epic 3). */
+  readonly compact: () => boolean
+  readonly setCompact: (on: boolean) => void
+  /** Read / set the global tool/reasoning detail mode (/details — Epic 3). */
+  readonly details: () => DetailsMode
+  readonly setDetails: (mode: DetailsMode) => void
+  /** Mounted-renderable count under the live renderer root (a /mem diagnostic);
+   *  undefined when no renderer is reachable (tests). */
+  readonly renderableCount: () => number | undefined
+}
+
+function readStr(value: unknown, key: string): string | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const v = (value as { [k: string]: unknown })[key]
+  return typeof v === 'string' ? v : undefined
+}
+
+const titleCase = (name: string) => name.charAt(0).toUpperCase() + name.slice(1)
+
+/** A planned completion query (item 5/13): which RPC + params, and where an
+ *  accepted item replaces from if the RPC omits its own `replace_from`. */
+export interface CompletionPlan {
+  method: 'complete.slash' | 'complete.path'
+  params: Record<string, unknown>
+  from: number
+}
+
+/** The command-name grammar for the lead `/token` (mirrors skillMatch NAME_RE):
+ *  starts alphanumeric, then word chars / `.` / `-`. Notably EXCLUDES `/`, so a
+ *  path like `/usr/bin` is NEVER a slash command (F2). */
+const SLASH_NAME_RE = /^[A-Za-z0-9][\w.-]*$/
+
+/** `@`-mention is the ONLY file/dir completion trigger now (F8b — glitch
+ *  2026-06-13: drop `~`/`./`/`/`/bare-path as triggers; the gateway's
+ *  complete.path still understands `@file:`/`@folder:`/fuzzy basename). */
+function isPathLike(word: string): boolean {
+  return word.startsWith('@')
+}
+
+/**
+ * Decide what to complete for the composer text + cursor offset:
+ *   - the text is a slash command — `/` at the very start → `complete.slash
+ *     {text}`. A bare `/` opens the full command list immediately (glitch
+ *     2026-06-13); `/m`, `/model foo` narrow it. A `/abs/path` whose first token
+ *     isn't a valid name (F2) → no slash menu.
+ *   - the WORD under the cursor is an `@`-mention → `complete.path {word}` for
+ *     file/dir tagging (F8b).
+ *   - otherwise nothing.
+ *
+ * Cursor-aware (F7/F8): completion is computed from the line/token at the cursor,
+ * so it keeps working on later lines after Shift+Enter (the old whole-buffer
+ * `includes('\n')` bail killed it on every multi-line buffer). `cursor` defaults
+ * to the end of `text`. Slash commands stay first-line-only (a `/` mid-buffer is
+ * prose, never a command).
+ * Returns null when there's no completion to run (so the dropdown clears).
+ */
+export function planCompletion(text: string, cursor: number = text.length): CompletionPlan | null {
+  // Slash command: only when the WHOLE buffer's lead token is a command. A `/`
+  // after a newline is prose, so a slash command never spans lines.
+  if (text.startsWith('/') && !text.includes('\n')) {
+    const body = text.slice(1)
+    const space = body.search(/\s/)
+    const name = space === -1 ? body : body.slice(0, space)
+    // Hydrate on a BARE `/` (body === '', glitch 2026-06-13 — open the full
+    // command list on the first slash) or a valid command name. A `/abs/path`
+    // (the lead token contains a `/`) is never a command (F2), and a `/ ` with a
+    // trailing space past an empty name is not arg-completion on nothing.
+    if (body === '' || SLASH_NAME_RE.test(name)) {
+      return { from: 0, method: 'complete.slash', params: { text } }
+    }
+    return null
+  }
+  // @-mention: the whitespace-delimited token the cursor sits in/just after.
+  const pos = Math.max(0, Math.min(cursor, text.length))
+  const head = text.slice(0, pos)
+  const tokenStart = head.search(/\S+$/)
+  if (tokenStart === -1) return null
+  const word = head.slice(tokenStart)
+  if (isPathLike(word)) {
+    return { from: tokenStart, method: 'complete.path', params: { word } }
+  }
+  return null
+}
+
+/** Read a `replace_from` offset off a completion result, falling back to `fallback`. */
+export function readReplaceFrom(result: unknown, fallback: number): number {
+  if (result && typeof result === 'object') {
+    const rf = (result as { replace_from?: unknown }).replace_from
+    if (typeof rf === 'number') return rf
+  }
+  return fallback
+}
+
+/** Map a `complete.slash`/`complete.path` result ({items:[{text,display,meta}]}) into candidates. */
+export function mapCompletions(result: unknown): CompletionItem[] {
+  if (!result || typeof result !== 'object') return []
+  const items = (result as { items?: unknown }).items
+  if (!Array.isArray(items)) return []
+  const out: CompletionItem[] = []
+  for (const it of items) {
+    const text = readStr(it, 'text')
+    if (!text) continue
+    out.push({ display: readStr(it, 'display') ?? text, meta: readStr(it, 'meta') ?? '', text })
+  }
+  return out
+}
+
+/** Extract `{text}` items from a `commands.catalog` result ({pairs:[["/name",
+ *  "desc"],…]}) for seeding the composer's slash-highlight catalog at boot
+ *  (glitch 2026-06-14). Each pair's first element is the `/name`; non-string or
+ *  empty entries are skipped. Shape-defensive — any junk → []. */
+export function catalogCommandItems(result: unknown): { text: string }[] {
+  if (!result || typeof result !== 'object') return []
+  const pairs = (result as { pairs?: unknown }).pairs
+  if (!Array.isArray(pairs)) return []
+  const out: { text: string }[] = []
+  for (const pair of pairs as unknown[]) {
+    const name = Array.isArray(pair) ? (pair as unknown[])[0] : undefined
+    if (typeof name === 'string' && name) out.push({ text: name })
+  }
+  return out
+}
+
+/**
+ * A monotonic gate for the per-keystroke completion RPCs (glitch 2026-06-14).
+ * The gateway transport does NOT guarantee in-order response delivery and
+ * `onType` fires an RPC per keystroke with no debounce, so a slow earlier query
+ * (the first bare-`/` `complete.slash`) can resolve AFTER a newer one (an
+ * `@`-mention `complete.path`) and clobber the store with stale results — which
+ * is what made "a leading /path message breaks @-mentions afterward."
+ *
+ * `claim()` is called once per keystroke (BEFORE the early-return clear branch,
+ * so an intermediate keystroke that fires no RPC still invalidates the older
+ * in-flight one) and returns a token; `isCurrent(token)` is true only for the
+ * most recently claimed token, so a resolving response applies ONLY when no
+ * newer keystroke has superseded it.
+ */
+export interface CompletionGate {
+  claim: () => number
+  isCurrent: (token: number) => boolean
+}
+
+export function createCompletionGate(): CompletionGate {
+  let seq = 0
+  return {
+    claim: () => ++seq,
+    isCurrent: (token: number) => token === seq
+  }
+}
+
+/** Long output → the pager; short → a system line (Ink: >180 chars or >2 lines). */
+function present(ctx: SlashContext, title: string, text: string): void {
+  const long = text.length > 180 || text.split('\n').filter(Boolean).length > 2
+  if (long) ctx.openPager(title, text)
+  else ctx.pushSystem(text)
+}
+
+/** Process-diagnostic commands — hidden behind `HERMES_TUI_DIAGNOSTICS`
+ *  (logic/env.ts). Regular users never see them; support flows enable them
+ *  with one env var. Keep this set in sync with the `(diag)` lines below.
+ *  DESIGN ASSUMPTION (review 2026-06-12): these stay CLIENT-ONLY. Completion
+ *  is gateway-driven and hides them only because the gateway doesn't know
+ *  them — adding a server command with one of these names requires gating it
+ *  gateway-side too (the early return below would shadow, not hide, it). */
+const DIAGNOSTIC_COMMANDS = new Set(['mem', 'heapdump'])
+
+const CLIENT_HELP_LINES = [
+  '/help — list commands',
+  '/model [name] — switch model (picker if bare)',
+  '/copy [n] — copy the last (or n-th) response',
+  '/skills — browse skills',
+  '/skin [name] — switch theme skin (live)',
+  '/sessions [cron|gateways|all] — browse/resume sessions (tabbed picker)',
+  '/resume [id|name] — resume directly, or open the picker',
+  '/clear, /new — clear the transcript (confirm)',
+  '/compact [on|off|toggle] — compact transcript spacing',
+  '/details [hidden|collapsed|expanded|cycle] — tool/reasoning detail',
+  '/bg <prompt> — launch a background prompt',
+  '/processes — OS background processes (list + stop all)',
+  '/replay [n|path] — inspect an archived spawn tree',
+  '/mem — live memory stats (diag)',
+  '/heapdump — write a V8 heap snapshot (diag)',
+  '/logs — recent engine log lines',
+  '/quit, /exit — quit',
+  '(other /commands run on the gateway)'
+]
+
+function clientHelp(): string {
+  const lines = diagnosticsEnabled() ? CLIENT_HELP_LINES : CLIENT_HELP_LINES.filter(l => !l.includes('(diag)'))
+  return lines.join('\n')
+}
+
+type ClientHandler = (arg: string, ctx: SlashContext) => void | Promise<void>
+
+/** `/sessions [recent|cron|gateways|all]` — open the tabbed resume picker,
+ *  pre-selecting the named tab (shared by /sessions, /switch, /session). */
+const sessionsCmd: ClientHandler = (arg, ctx) => {
+  const tab = parseSessionTabArg(arg)
+  if (!tab) {
+    ctx.pushSystem('usage: /sessions [recent|cron|gateways|all]')
+    return
+  }
+  ctx.openSessionPicker(tab)
+}
+
+/** `/resume` — bare opens the picker; `/resume <id|name>` keeps the DIRECT
+ *  path: resolve the arg against `session.list` (exact id → unique id prefix
+ *  → exact/unique title) and hydrate without the overlay. */
+const resumeCmd: ClientHandler = async (arg, ctx) => {
+  const needle = arg.trim()
+  if (!needle) {
+    ctx.openSessionPicker('recent')
+    return
+  }
+  try {
+    // One bounded page over ALL sources (the gateway deny-lists `tool`) — the
+    // direct path targets a known session, not a browse.
+    const { rows } = mapSessionRows(await ctx.request('session.list', { limit: 200 }))
+    const hit = resolveSessionArg(rows, needle)
+    if (!hit) {
+      ctx.pushSystem(`/resume: no session matching “${needle}” — try /sessions`)
+      return
+    }
+    ctx.resumeSession(hit.id)
+  } catch (error) {
+    ctx.pushSystem(`/resume: ${error instanceof Error ? error.message : 'session.list failed'}`)
+  }
+}
+
+/**
+ * Flatten `model.options` into grouped picker rows (Epic 7; v2.1 availability):
+ * group = the provider's display ("lab") name, haystacks = slug + lab name (so
+ * `oai`/`copilot`/`anthropic` fuzzy-match the whole group), value = the FULL
+ * switch arg `<model> --provider <slug>` so picking a model under a different
+ * provider actually switches provider+model (the gateway's
+ * `_apply_model_switch` parses `--provider` via parse_model_flags). The current
+ * model is flagged, not baked into the label, so the fuzzy scorer never matches
+ * the ✓.
+ *
+ * UNCONFIGURED providers (`authenticated: false` skeleton rows — the gateway
+ * sends them via `build_models_payload(include_unconfigured=True,
+ * picker_hints=True)`, with `key_env`/`warning` setup hints) become one
+ * `unavailable` hint row each (`no API key — set <ENV_VAR>`): hidden by
+ * default, revealed dimmed + non-selectable by the picker's Ctrl+U toggle.
+ */
+export function mapModelOptions(opts: unknown): PickerItem[] {
+  if (!opts || typeof opts !== 'object') return []
+  const providers = (opts as { providers?: unknown }).providers
+  if (!Array.isArray(providers)) return []
+  const current = readStr(opts, 'model')
+  const currentProvider = readStr(opts, 'provider')
+  const items: PickerItem[] = []
+  for (const p of providers) {
+    if (!p || typeof p !== 'object') continue
+    const slug = readStr(p, 'slug') ?? readStr(p, 'name') ?? ''
+    const lab = readStr(p, 'name') ?? slug
+    if ((p as { authenticated?: unknown }).authenticated === false) {
+      // Unconfigured provider → one dimmed hint row under its own group header.
+      // Identity (slug + display name) is the haystack so a provider-name query
+      // still narrows to the group; the hint text itself is not searched.
+      const keyEnv = readStr(p, 'key_env')
+      const item: PickerItem = {
+        group: lab || slug,
+        label: keyEnv ? `no API key — set ${keyEnv}` : (readStr(p, 'warning') ?? 'not configured'),
+        unavailable: true,
+        value: slug || lab
+      }
+      const hay = [slug, lab].filter(Boolean)
+      if (hay.length) item.haystacks = hay
+      items.push(item)
+      continue
+    }
+    if ((p as { authenticated?: unknown }).authenticated !== true) continue
+    // The gateway's own normalized "this row is the active provider" flag —
+    // more reliable than comparing `provider` to `slug` (the agent's provider
+    // string can be the API dialect, e.g. an openai-compatible base_url).
+    const rowCurrent = (p as { is_current?: unknown }).is_current === true
+    const models = (p as { models?: unknown }).models
+    if (!Array.isArray(models)) continue
+    for (const m of models) {
+      if (typeof m !== 'string') continue
+      const item: PickerItem = { label: m, value: slug ? `${m} --provider ${slug}` : m }
+      // current = same model id under the active provider (row flag first,
+      // then the slug comparison, then "no provider known at all").
+      if (m === current && (rowCurrent || currentProvider === slug || !currentProvider)) item.current = true
+      if (lab) item.group = lab
+      const haystacks = [slug, lab].filter(Boolean)
+      if (haystacks.length) item.haystacks = haystacks
+      items.push(item)
+    }
+  }
+  // Provider matching failed entirely (string-normalization drift) but the
+  // model id is known → flag the first id match so the ✓ never just vanishes.
+  if (current && !items.some(i => i.current)) {
+    const fallback = items.find(i => i.label === current)
+    if (fallback) fallback.current = true
+  }
+  return items
+}
+
+/**
+ * Provider tab order for the model picker's chip strip (picker v2.2): each
+ * CONFIGURED provider's group (= lab display name) in catalog order, with
+ * Nous-identified groups (slug or lab name containing `nous`) hoisted to the
+ * front. Unconfigured providers (`unavailable` hint rows) get NO tab — they
+ * stay reachable via Ctrl+U under the picker's trailing `All` tab (which the
+ * picker appends itself; it is not part of this list).
+ */
+export function buildModelTabs(items: readonly PickerItem[]): string[] {
+  const seen = new Set<string>()
+  const nous: string[] = []
+  const rest: string[] = []
+  for (const it of items) {
+    if (it.unavailable || !it.group || seen.has(it.group)) continue
+    seen.add(it.group)
+    const identity = [it.group, ...(it.haystacks ?? [])].join(' ').toLowerCase()
+    ;(identity.includes('nous') ? nous : rest).push(it.group)
+  }
+  return [...nous, ...rest]
+}
+
+/** Flatten `skills.manage {action:'list'}` ({skills: Record<category, names[]>}) into
+ *  grouped picker rows (category = group header; also a fuzzy haystack). */
+function mapSkills(result: unknown): PickerItem[] {
+  if (!result || typeof result !== 'object') return []
+  const skills = (result as { skills?: unknown }).skills
+  if (!skills || typeof skills !== 'object') return []
+  const items: PickerItem[] = []
+  for (const [category, names] of Object.entries(skills as { [k: string]: unknown })) {
+    if (!Array.isArray(names)) continue
+    for (const n of names) if (typeof n === 'string') items.push({ group: category, label: n, value: n })
+  }
+  return items
+}
+
+/** Re-fetch `model.options` and update the cached picker rows. Resolves with
+ *  the fresh rows (the open picker swaps them in live — Ctrl+R, picker v2.1);
+ *  rejections are the CALLER's to handle (background callers fire-and-forget). */
+function refreshModelItems(ctx: SlashContext): Promise<PickerItem[]> {
+  return ctx.request('model.options', { session_id: ctx.sessionId() }).then(opts => {
+    const items = mapModelOptions(opts)
+    if (items.length) ctx.setModelItems(items)
+    return items
+  })
+}
+
+/**
+ * The open picker's manual-refresh seam (picker v2.1 Ctrl+R). Whoever opens a
+ * picker registers (or clears) the catalog re-fetch here; the mounted Picker
+ * triggers it via `runPickerRefresh` and swaps in the resolved rows live. A
+ * module slot rather than a Picker prop because the App→Picker prop plumbing
+ * carries only the PickerState basics; the seam keeps the overlay generic for
+ * the upcoming resume-session picker (register a `session.list` re-fetch).
+ */
+let activePickerRefresh: (() => Promise<PickerItem[]>) | undefined
+
+/** Register (or clear, with `undefined`) the open picker's catalog re-fetch. */
+export function registerPickerRefresh(fn: (() => Promise<PickerItem[]>) | undefined): void {
+  activePickerRefresh = fn
+}
+
+/** Whether a refresh is registered (the picker's footer hint is gated on it). */
+export function canRefreshPicker(): boolean {
+  return activePickerRefresh !== undefined
+}
+
+/** Run the registered catalog re-fetch; undefined when none is registered. */
+export function runPickerRefresh(): Promise<PickerItem[]> | undefined {
+  return activePickerRefresh?.()
+}
+
+/**
+ * The open picker's tab-strip seam (picker v2.2 provider tabs) — same pattern
+ * as the refresh seam above: whoever opens a picker registers (or clears) a
+ * tab DERIVATION over the picker's live rows; the mounted Picker re-derives
+ * through it whenever the rows swap (Ctrl+R), so fresh providers grow chips
+ * without re-opening. `/model` registers `buildModelTabs`; pickers without
+ * tabs (skills) clear it and render the classic stripless view.
+ */
+let activePickerTabs: ((items: readonly PickerItem[]) => string[]) | undefined
+
+/** Register (or clear, with `undefined`) the open picker's tab derivation. */
+export function registerPickerTabs(fn: ((items: readonly PickerItem[]) => string[]) | undefined): void {
+  activePickerTabs = fn
+}
+
+/** Derive the open picker's tabs from its rows; [] when no tabs are registered. */
+export function pickerTabs(items: readonly PickerItem[]): string[] {
+  return activePickerTabs?.(items) ?? []
+}
+
+/**
+ * The bootstrap `model.options` prefetch seam (perf: prefetch dedupe). The
+ * entry stashes its in-flight prefetch promise here; a bare `/model` that
+ * finds the cache empty AWAITS it (bounded by `waitMs`) and re-checks the
+ * cache instead of issuing a second concurrent `model.options` RPC. A hung
+ * prefetch only delays the picker by the bound — `/model` then opens via its
+ * own fetch as before.
+ */
+let modelPrefetch: { promise: Promise<unknown>; waitMs: number } | undefined
+
+/** Register (or clear, with `undefined`) the in-flight bootstrap prefetch. */
+export function registerModelPrefetch(promise: Promise<unknown> | undefined, waitMs = 2000): void {
+  modelPrefetch = promise ? { promise, waitMs } : undefined
+}
+
+/** Await the registered prefetch (bounded); resolves immediately when none. */
+function awaitModelPrefetch(): Promise<void> {
+  const pending = modelPrefetch
+  if (!pending) return Promise.resolve()
+  return Promise.race([pending.promise, new Promise(resolve => setTimeout(resolve, pending.waitMs))]).then(
+    () => undefined
+  )
+}
+
+/** Switch the model via the server (shared by `/model <name>` and the picker pick).
+ *  A successful switch refreshes the cached rows in the background (fresh ✓). */
+async function switchModel(ctx: SlashContext, name: string): Promise<void> {
+  try {
+    const r = await ctx.request('slash.exec', { command: `model ${name}`, session_id: ctx.sessionId() })
+    ctx.pushSystem(readStr(r, 'output') || `→ ${name}`)
+    void refreshModelItems(ctx).catch(() => {})
+  } catch (error) {
+    ctx.pushSystem(`/model ${name}: ${error instanceof Error ? error.message : 'switch failed'}`)
+  }
+}
+
+/** `/model` — bare opens the model picker; `/model <name>` switches directly.
+ *  Opens from the CACHED catalog when present — zero RPCs, same-frame paint
+ *  (Epic 7; the catalog is prefetched at bootstrap and refreshed on switch).
+ *  An empty cache first awaits the in-flight bootstrap prefetch (bounded) so
+ *  an early `/model` never doubles the slow `model.options` RPC. */
+const modelCmd: ClientHandler = async (arg, ctx) => {
+  if (arg.trim()) {
+    await switchModel(ctx, arg.trim())
+    return
+  }
+  const open = (items: PickerItem[]) => {
+    // Ctrl+R in the open picker re-fetches the catalog (and re-syncs the cache).
+    registerPickerRefresh(() => refreshModelItems(ctx))
+    // Provider chip strip (picker v2.2): Nous-first configured-provider tabs.
+    registerPickerTabs(buildModelTabs)
+    ctx.openPicker({ items, onPick: name => void switchModel(ctx, name), title: 'Switch model' })
+  }
+  const cached = ctx.modelItems()
+  if (cached?.length) {
+    open(cached)
+    return
+  }
+  // Cache empty but the bootstrap prefetch may be in flight — await it
+  // (bounded) and re-check instead of racing a SECOND model.options RPC.
+  await awaitModelPrefetch()
+  const prefetched = ctx.modelItems()
+  if (prefetched?.length) {
+    open(prefetched)
+    return
+  }
+  const items = mapModelOptions(await ctx.request('model.options', { session_id: ctx.sessionId() }))
+  // Unavailable hint rows alone are not a usable catalog — keep the notice.
+  if (!items.some(i => !i.unavailable)) {
+    ctx.pushSystem('No models available (no authenticated providers).')
+    return
+  }
+  ctx.setModelItems(items)
+  open(items)
+}
+
+/** `/skills` — open the skills hub; picking a skill shows its info in the pager. */
+const skillsCmd: ClientHandler = async (_arg, ctx) => {
+  const items = mapSkills(await ctx.request('skills.manage', { action: 'list' }))
+  if (!items.length) {
+    ctx.pushSystem('No skills found.')
+    return
+  }
+  registerPickerRefresh(undefined) // no Ctrl+R catalog re-fetch for skills (yet)
+  registerPickerTabs(undefined) // no tab strip for skills — classic grouped view
+  ctx.openPicker({
+    items,
+    onPick: name =>
+      void ctx
+        .request('skills.manage', { action: 'inspect', query: name })
+        .then(info => ctx.openPager(`Skill: ${name}`, readStr(info, 'info') || JSON.stringify(info, null, 2)))
+        .catch(() => ctx.pushSystem(`/skills: could not inspect ${name}`)),
+    title: 'Skills'
+  })
+}
+
+/** `on`/`off`/`toggle`/bare → the next flag value; null on garbage (Ink flagFromArg). */
+function flagFromArg(arg: string, current: boolean): boolean | null {
+  const mode = arg.trim().toLowerCase()
+  if (!mode || mode === 'toggle') return !current
+  if (mode === 'on') return true
+  if (mode === 'off') return false
+  return null
+}
+
+/** `/compact [on|off|toggle]` — compact transcript spacing. The flag flips locally
+ *  (the store drives the render); persistence mirrors Ink: a fire-and-forget
+ *  `config.set {key:'compact'}` so the Ink TUI + future launches share the pref
+ *  (the gateway does NOT send the persisted value to this TUI, so each launch
+ *  starts off — see store.ts `compact`). */
+const compactCmd: ClientHandler = (arg, ctx) => {
+  const next = flagFromArg(arg, ctx.compact())
+  if (next === null) {
+    ctx.pushSystem('usage: /compact [on|off|toggle]')
+    return
+  }
+  ctx.setCompact(next)
+  void ctx.request('config.set', { key: 'compact', value: next ? 'on' : 'off' }).catch(() => {})
+  ctx.pushSystem(`compact ${next ? 'on' : 'off'}`)
+}
+
+/**
+ * `/details [hidden|collapsed|expanded|cycle]` — GLOBAL detail mode (per-section
+ * overrides deferred; the gateway's arg completion also suggests section names,
+ * so those get an honest "not supported yet" notice). Bare `/details` reports the
+ * persisted mode (`config.get details_mode`) and syncs the local flag to it; a
+ * mode set persists via `config.set` (fire-and-forget, Ink parity).
+ */
+const detailsCmd: ClientHandler = async (arg, ctx) => {
+  const first = arg.trim().toLowerCase().split(/\s+/)[0] ?? ''
+  if (!first) {
+    try {
+      const r = await ctx.request('config.get', { key: 'details_mode' })
+      const mode = parseDetailsMode(readStr(r, 'value')) ?? ctx.details()
+      ctx.setDetails(mode)
+      ctx.pushSystem(`details: ${mode}`)
+    } catch {
+      ctx.pushSystem(`details: ${ctx.details()}`)
+    }
+    return
+  }
+  if ((DETAILS_SECTIONS as readonly string[]).includes(first)) {
+    ctx.pushSystem(`per-section detail overrides are not supported in the native engine yet — ${DETAILS_USAGE}`)
+    return
+  }
+  const next = first === 'cycle' || first === 'toggle' ? nextDetailsMode(ctx.details()) : parseDetailsMode(first)
+  if (!next) {
+    ctx.pushSystem(DETAILS_USAGE)
+    return
+  }
+  ctx.setDetails(next)
+  void ctx.request('config.set', { key: 'details_mode', value: next }).catch(() => {})
+  ctx.pushSystem(`details: ${next}`)
+}
+
+/** `/skin [name]` — switch the active theme skin (Ink parity:
+ *  ui-tui/src/app/slash/commands/session.ts). Bare `/skin` reports the persisted
+ *  skin (`config.get skin`); `/skin <name>` persists via `config.set` which makes
+ *  the gateway emit `skin.changed` → the store re-themes the running UI LIVE (no
+ *  relaunch). Skin-name arg completion comes from the gateway's `complete.slash`
+ *  for free. Fire-and-forget with a guarded notice, matching compact/details. */
+const skinCmd: ClientHandler = async (arg, ctx) => {
+  const name = arg.trim()
+  if (!name) {
+    try {
+      const r = await ctx.request('config.get', { key: 'skin' })
+      ctx.pushSystem(`skin: ${readStr(r, 'value') || 'default'}`)
+    } catch {
+      ctx.pushSystem('skin: default')
+    }
+    return
+  }
+  try {
+    const r = await ctx.request('config.set', { key: 'skin', value: name })
+    ctx.pushSystem(`skin → ${readStr(r, 'value') || name}`)
+  } catch (error) {
+    ctx.pushSystem(`/skin: ${error instanceof Error ? error.message : 'config.set failed'}`)
+  }
+}
+
+/** Fetch + map the session's archived spawn trees (`spawn_tree.list`). */
+async function listSpawnTrees(ctx: SlashContext) {
+  const r = await ctx.request('spawn_tree.list', { limit: 30, session_id: ctx.sessionId() ?? 'default' })
+  return readSpawnTreeEntries(r)
+}
+
+/**
+ * `/replay [n|path]` — spawn-tree inspector through the pager (Ink renders these
+ * in its agents overlay; the flow + RPCs are the same): bare lists the archived
+ * trees with indices, `<n>` loads the n-th listed tree, anything else is treated
+ * as a snapshot path on disk (`load <path>` accepted for Ink muscle memory).
+ */
+const replayCmd: ClientHandler = async (arg, ctx) => {
+  const raw = arg.trim()
+  const lower = raw.toLowerCase()
+  try {
+    if (!raw || lower === 'list' || lower === 'ls') {
+      const entries = await listSpawnTrees(ctx)
+      if (!entries.length) {
+        ctx.pushSystem('no archived spawn trees for this session — completed delegations are archived automatically')
+        return
+      }
+      ctx.openPager('Spawn trees', formatSpawnTreeList(entries))
+      return
+    }
+    if (/^\d+$/.test(raw)) {
+      const n = Number.parseInt(raw, 10)
+      const entries = await listSpawnTrees(ctx)
+      const entry = entries[n - 1]
+      if (!entry) {
+        ctx.pushSystem(
+          entries.length
+            ? `replay: index out of range 1..${entries.length} — /replay to list`
+            : 'no archived spawn trees for this session'
+        )
+        return
+      }
+      const tree = await ctx.request('spawn_tree.load', { path: entry.path })
+      ctx.openPager(`Replay ${n}`, formatSpawnTree(tree))
+      return
+    }
+    const path = lower.startsWith('load ') ? raw.slice(5).trim() : raw
+    const tree = await ctx.request('spawn_tree.load', { path })
+    ctx.openPager('Replay', formatSpawnTree(tree))
+  } catch (error) {
+    ctx.pushSystem(`/replay: ${error instanceof Error ? error.message : 'failed'}`)
+  }
+}
+
+/** `/heapdump` — write a V8 heap snapshot to `$HERMES_HOME|~/.hermes/logs/` and
+ *  report the path + heap/rss before vs after (Ink ref debug.ts /heapdump). */
+const heapdumpCmd: ClientHandler = (_arg, ctx) => {
+  const pre = process.memoryUsage()
+  ctx.pushSystem(`writing heap dump (heap ${formatBytes(pre.heapUsed)} · rss ${formatBytes(pre.rss)})…`)
+  try {
+    const { after, before, path } = performHeapdump()
+    ctx.pushSystem(
+      `heapdump: ${path}\n` +
+        `heap ${formatBytes(before.heapUsed)} → ${formatBytes(after.heapUsed)} · ` +
+        `rss ${formatBytes(before.rss)} → ${formatBytes(after.rss)}`
+    )
+  } catch (error) {
+    ctx.pushSystem(`heapdump failed: ${error instanceof Error ? error.message : String(error)}`)
+  }
+}
+
+/** `/mem` — live V8 heap/rss numbers + uptime + the mounted-renderable count
+ *  (the store-cap diagnostic) as one system block (Ink ref debug.ts /mem). */
+const memCmd: ClientHandler = (_arg, ctx) => {
+  ctx.pushSystem(memReport(process.memoryUsage(), process.uptime(), ctx.renderableCount()))
+}
+
+/** `/tools` — fetch the tool roster from the gateway and show it in the pager (navigable). */
+const toolsCmd: ClientHandler = async (arg, ctx) => {
+  const command = arg.trim() ? `tools ${arg.trim()}` : 'tools'
+  try {
+    const r = await ctx.request('slash.exec', { command, session_id: ctx.sessionId() })
+    ctx.openPager('Tools', readStr(r, 'output') || '(no tool info)')
+  } catch (error) {
+    ctx.pushSystem(`/tools: ${error instanceof Error ? error.message : 'failed'}`)
+  }
+}
+
+/** `/bg <prompt>` (aliases /background, /btw) — launch a background PROMPT via
+ *  `prompt.background` (Ink parity): echo "bg <id> started" and track the task so
+ *  the `bg: N` badge counts it until `background.complete` clears it. NOT the OS
+ *  process panel (that's /processes). */
+const backgroundCmd: ClientHandler = async (arg, ctx) => {
+  const text = arg.trim()
+  if (!text) {
+    ctx.pushSystem('/bg <prompt> — launch a background prompt')
+    return
+  }
+  try {
+    const r = await ctx.request('prompt.background', { session_id: ctx.sessionId(), text })
+    const taskId = readStr(r, 'task_id')
+    if (taskId) {
+      ctx.addBgTask(taskId)
+      ctx.pushSystem(`bg ${taskId} started`)
+    } else {
+      ctx.pushSystem('/bg: no task id returned')
+    }
+  } catch (error) {
+    ctx.pushSystem(`/bg: ${error instanceof Error ? error.message : 'failed'}`)
+  }
+}
+
+/** The TUI-only client commands (run in-process, never hit the gateway). */
+const CLIENT: Record<string, ClientHandler> = {
+  agents: (_arg, ctx) => ctx.openDashboard(),
+  background: backgroundCmd,
+  bg: backgroundCmd,
+  btw: backgroundCmd,
+  clear: (_arg, ctx) => ctx.confirm('Clear the transcript?', ctx.clearTranscript),
+  compact: compactCmd,
+  copy: (arg, ctx) => {
+    const n = Math.max(1, Number.parseInt(arg, 10) || 1)
+    if (!ctx.copyResponse(n)) ctx.pushSystem('Nothing to copy yet.')
+  },
+  detail: detailsCmd,
+  details: detailsCmd,
+  exit: (_arg, ctx) => ctx.quit(),
+  heapdump: heapdumpCmd,
+  mem: memCmd,
+  processes: (_arg, ctx) => ctx.openBackgroundPanel(),
+  procs: (_arg, ctx) => ctx.openBackgroundPanel(),
+  model: modelCmd,
+  replay: replayCmd,
+  resume: resumeCmd,
+  session: sessionsCmd,
+  sessions: sessionsCmd,
+  skills: skillsCmd,
+  skin: skinCmd,
+  switch: sessionsCmd,
+  tasks: (_arg, ctx) => ctx.openDashboard(),
+  tools: toolsCmd,
+  help: async (_arg, ctx) => {
+    // Prefer the live catalog; fall back to the client list if it's unavailable.
+    try {
+      const cat = await ctx.request('commands.catalog', {})
+      ctx.pushSystem(renderCatalog(cat) || clientHelp())
+    } catch {
+      ctx.pushSystem(clientHelp())
+    }
+  },
+  logs: (_arg, ctx) => ctx.openPager('Logs', ctx.logTail().join('\n') || '(log empty)'),
+  new: (_arg, ctx) => ctx.confirm('Start fresh? (clears the transcript)', ctx.clearTranscript),
+  quit: (_arg, ctx) => ctx.quit()
+}
+
+/** The registered client-command names (catalog introspection — tests/menus). */
+export function clientCommandNames(): string[] {
+  const names = Object.keys(CLIENT)
+  return (diagnosticsEnabled() ? names : names.filter(n => !DIAGNOSTIC_COMMANDS.has(n))).sort()
+}
+
+/** Render the gateway `commands.catalog` into a help block (loose-typed read).
+ *  The TUI catalog shape is `{ pairs: [["/name","desc"], …], canon, categories }`
+ *  (tui_gateway/server.py `commands.catalog`). */
+function renderCatalog(cat: unknown): string {
+  if (!cat || typeof cat !== 'object') return ''
+  const pairs = (cat as { pairs?: unknown }).pairs
+  if (!Array.isArray(pairs)) return ''
+  const lines = pairs
+    .map(pair => {
+      if (!Array.isArray(pair) || typeof pair[0] !== 'string') return null
+      const desc = typeof pair[1] === 'string' ? pair[1] : ''
+      return desc ? `${pair[0]} — ${desc}` : pair[0]
+    })
+    .filter((l): l is string => l !== null)
+  return lines.length ? lines.join('\n') : ''
+}
+
+function handleDispatchResult(parsed: ParsedSlash, raw: unknown, ctx: SlashContext): void {
+  const type = readStr(raw, 'type')
+  const argTail = parsed.arg ? ` ${parsed.arg}` : ''
+  switch (type) {
+    case 'exec':
+    case 'plugin':
+      ctx.pushSystem(readStr(raw, 'output') || '(no output)')
+      return
+    case 'alias': {
+      const target = readStr(raw, 'target')
+      if (target) void dispatchSlash(`/${target}${argTail}`, ctx)
+      return
+    }
+    case 'skill':
+    case 'send': {
+      const notice = readStr(raw, 'notice')
+      if (notice) ctx.pushSystem(notice)
+      const message = readStr(raw, 'message')
+      if (message?.trim()) ctx.submit(message)
+      else ctx.pushSystem(`/${parsed.name}: empty message`)
+      return
+    }
+    case 'prefill': {
+      // /undo etc. — composer prefill lands with the composer-ref plumbing; show it for now.
+      const message = readStr(raw, 'message')
+      ctx.pushSystem(message ? `(edit & resubmit) ${message}` : `/${parsed.name}: nothing to prefill`)
+      return
+    }
+    default:
+      ctx.pushSystem(`error: invalid response: command.dispatch`)
+  }
+}
+
+/** Dispatch a `/command` through the ladder. Returns once the (async) work settles. */
+export async function dispatchSlash(input: string, ctx: SlashContext): Promise<void> {
+  const parsed = parseSlash(input)
+  if (!parsed) return
+
+  if (DIAGNOSTIC_COMMANDS.has(parsed.name) && !diagnosticsEnabled()) {
+    // Not a secret — an enable switch. Tell the user exactly how to get it.
+    ctx.pushSystem(`/${parsed.name} is a diagnostic command — relaunch with HERMES_TUI_DIAGNOSTICS=1 to enable it.`)
+    return
+  }
+
+  const client = CLIENT[parsed.name]
+  if (client) {
+    await client(parsed.arg, ctx)
+    return
+  }
+
+  const sid = ctx.sessionId()
+  try {
+    const result = await ctx.request('slash.exec', { command: input.slice(1), session_id: sid })
+    const output = readStr(result, 'output') || `/${parsed.name}: no output`
+    const warning = readStr(result, 'warning')
+    const text = warning ? `warning: ${warning}\n${output}` : output
+    // Long output → pager (Ink: >180 chars or >2 non-empty lines), else a system line.
+    present(ctx, titleCase(parsed.name), text)
+  } catch {
+    try {
+      const raw = await ctx.request('command.dispatch', { arg: parsed.arg, name: parsed.name, session_id: sid })
+      handleDispatchResult(parsed, raw, ctx)
+    } catch (error) {
+      ctx.pushSystem(`error: ${error instanceof Error ? error.message : String(error)}`)
+    }
+  }
+}