hermes-agent/tui_gateway/server.py

import atexit
import concurrent.futures
import contextvars
import copy
import inspect
import json
import logging
import os
import queue
import subprocess
import sys
import threading
import time
import uuid
from datetime import datetime
from pathlib import Path
from typing import Any, Optional

from hermes_constants import get_hermes_home
from hermes_cli.env_loader import load_hermes_dotenv
from utils import is_truthy_value
from tui_gateway.transport import (
    StdioTransport,
    Transport,
    bind_transport,
    current_transport,
    reset_transport,
)

logger = logging.getLogger(__name__)

_hermes_home = get_hermes_home()
load_hermes_dotenv(
    hermes_home=_hermes_home, project_env=Path(__file__).parent.parent / ".env"
)


# ── Panic logger ─────────────────────────────────────────────────────
# Gateway crashes in a TUI session leave no forensics: stdout is the
# JSON-RPC pipe (TUI side parses it, doesn't log raw), the root logger
# only catches handled warnings, and the subprocess exits before stderr
# flushes through the stderr->gateway.stderr event pump. This hook
# appends every unhandled exception to ~/.hermes/logs/tui_gateway_crash.log
# AND re-emits a one-line summary to stderr so the TUI can surface it in
# Activity — exactly what was missing when the voice-mode turns started
# exiting the gateway mid-TTS.
_CRASH_LOG = os.path.join(_hermes_home, "logs", "tui_gateway_crash.log")


def _panic_hook(exc_type, exc_value, exc_tb):
    import traceback

    trace = "".join(traceback.format_exception(exc_type, exc_value, exc_tb))
    try:
        os.makedirs(os.path.dirname(_CRASH_LOG), exist_ok=True)
        with open(_CRASH_LOG, "a", encoding="utf-8") as f:
            f.write(
                f"\n=== unhandled exception · {time.strftime('%Y-%m-%d %H:%M:%S')} ===\n"
            )
            f.write(trace)
    except Exception:
        pass
    # Stderr goes through to the TUI as a gateway.stderr Activity line —
    # the first line here is what the user will see without opening any
    # log files.  Rest of the stack is still in the log for full context.
    first = (
        str(exc_value).strip().splitlines()[0]
        if str(exc_value).strip()
        else exc_type.__name__
    )
    print(f"[gateway-crash] {exc_type.__name__}: {first}", file=sys.stderr, flush=True)
    # Chain to the default hook so the process still terminates normally.
    sys.__excepthook__(exc_type, exc_value, exc_tb)


sys.excepthook = _panic_hook


def _thread_panic_hook(args):
    # threading.excepthook signature: SimpleNamespace(exc_type, exc_value, exc_traceback, thread)
    import traceback

    trace = "".join(
        traceback.format_exception(args.exc_type, args.exc_value, args.exc_traceback)
    )
    try:
        os.makedirs(os.path.dirname(_CRASH_LOG), exist_ok=True)
        with open(_CRASH_LOG, "a", encoding="utf-8") as f:
            f.write(
                f"\n=== thread exception · {time.strftime('%Y-%m-%d %H:%M:%S')} "
                f"· thread={args.thread.name} ===\n"
            )
            f.write(trace)
    except Exception:
        pass
    first_line = (
        str(args.exc_value).strip().splitlines()[0]
        if str(args.exc_value).strip()
        else args.exc_type.__name__
    )
    print(
        f"[gateway-crash] thread {args.thread.name} raised {args.exc_type.__name__}: {first_line}",
        file=sys.stderr,
        flush=True,
    )


threading.excepthook = _thread_panic_hook

try:
    from hermes_cli.banner import prefetch_update_check

    prefetch_update_check()
except Exception:
    pass

from tui_gateway.render import make_stream_renderer, render_diff, render_message

_sessions: dict[str, dict] = {}
_methods: dict[str, callable] = {}
_pending: dict[str, tuple[str, threading.Event]] = {}
_pending_prompt_payloads: dict[str, tuple[str, dict]] = {}
_answers: dict[str, str] = {}
_db = None
_db_error: str | None = None
_stdout_lock = threading.Lock()
_cfg_lock = threading.Lock()
_cfg_cache: dict | None = None
_cfg_mtime: float | None = None
_cfg_path = None
try:
    _slash_timeout = float(os.environ.get("HERMES_TUI_SLASH_TIMEOUT_S") or "45")
except (ValueError, TypeError):
    _slash_timeout = 45.0
_SLASH_WORKER_TIMEOUT_S = max(5.0, _slash_timeout)
_DETAIL_SECTION_NAMES = ("thinking", "tools", "subagents", "activity")
_DETAIL_MODES = frozenset({"hidden", "collapsed", "expanded"})

# ── Async RPC dispatch (#12546) ──────────────────────────────────────
# A handful of handlers block the dispatcher loop in entry.py for seconds
# to minutes (slash.exec, cli.exec, shell.exec, session.resume,
# session.branch, session.compress, skills.manage).  While they're running, inbound RPCs —
# notably approval.respond and session.interrupt — sit unread in the
# stdin pipe.  We route only those slow handlers onto a small thread pool;
# everything else stays on the main thread so ordering stays sane for the
# fast path.  write_json is already _stdout_lock-guarded, so concurrent
# response writes are safe.
_LONG_HANDLERS = frozenset(
    {
        "browser.manage",
        "cli.exec",
        "session.branch",
        "session.compress",
        "session.resume",
        "shell.exec",
        "skills.manage",
        "slash.exec",
    }
)

try:
    _rpc_pool_workers = max(
        2, int(os.environ.get("HERMES_TUI_RPC_POOL_WORKERS") or "4")
    )
except (ValueError, TypeError):
    _rpc_pool_workers = 4
_pool = concurrent.futures.ThreadPoolExecutor(
    max_workers=_rpc_pool_workers,
    thread_name_prefix="tui-rpc",
)
atexit.register(lambda: _pool.shutdown(wait=False, cancel_futures=True))

# Reserve real stdout for JSON-RPC only; redirect Python's stdout to stderr
# so stray print() from libraries/tools becomes harmless gateway.stderr instead
# of corrupting the JSON protocol.
_real_stdout = sys.stdout
sys.stdout = sys.stderr

# Module-level stdio transport — fallback sink when no transport is bound via
# contextvar or session. Stream resolved through a lambda so runtime monkey-
# patches of `_real_stdout` (used extensively in tests) still land correctly.
_stdio_transport = StdioTransport(lambda: _real_stdout, _stdout_lock)


class _SlashWorker:
    """Persistent HermesCLI subprocess for slash commands."""

    def __init__(self, session_key: str, model: str):
        self._lock = threading.Lock()
        self._seq = 0
        self.stderr_tail: list[str] = []
        self.stdout_queue: queue.Queue[dict | None] = queue.Queue()

        argv = [
            sys.executable,
            "-m",
            "tui_gateway.slash_worker",
            "--session-key",
            session_key,
        ]
        if model:
            argv += ["--model", model]

        self.proc = subprocess.Popen(
            argv,
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
            bufsize=1,
            cwd=os.getcwd(),
            env=os.environ.copy(),
        )
        threading.Thread(target=self._drain_stdout, daemon=True).start()
        threading.Thread(target=self._drain_stderr, daemon=True).start()

    def _drain_stdout(self):
        for line in self.proc.stdout or []:
            try:
                self.stdout_queue.put(json.loads(line))
            except json.JSONDecodeError:
                continue
        self.stdout_queue.put(None)

    def _drain_stderr(self):
        for line in self.proc.stderr or []:
            if text := line.rstrip("\n"):
                self.stderr_tail = (self.stderr_tail + [text])[-80:]

    def run(self, command: str) -> str:
        if self.proc.poll() is not None:
            raise RuntimeError("slash worker exited")

        with self._lock:
            self._seq += 1
            rid = self._seq
            self.proc.stdin.write(json.dumps({"id": rid, "command": command}) + "\n")
            self.proc.stdin.flush()

            while True:
                try:
                    msg = self.stdout_queue.get(timeout=_SLASH_WORKER_TIMEOUT_S)
                except queue.Empty:
                    raise RuntimeError("slash worker timed out")
                if msg is None:
                    break
                if msg.get("id") != rid:
                    continue
                if not msg.get("ok"):
                    raise RuntimeError(msg.get("error", "slash worker failed"))
                return str(msg.get("output", "")).rstrip()

            raise RuntimeError(
                f"slash worker closed pipe{': ' + chr(10).join(self.stderr_tail[-8:]) if self.stderr_tail else ''}"
            )

    def close(self):
        try:
            if self.proc.poll() is None:
                self.proc.terminate()
                self.proc.wait(timeout=1)
        except Exception:
            try:
                self.proc.kill()
            except Exception:
                pass


def _load_busy_input_mode() -> str:
    display = _load_cfg().get("display")
    if not isinstance(display, dict):
        display = {}
    raw = str(display.get("busy_input_mode", "") or "").strip().lower()
    return raw if raw in {"queue", "steer", "interrupt"} else "interrupt"


def _notify_session_boundary(event_type: str, session_id: str | None) -> None:
    """Fire session lifecycle hooks with CLI parity."""
    try:
        from hermes_cli.plugins import invoke_hook as _invoke_hook

        _invoke_hook(event_type, session_id=session_id, platform="tui")
    except Exception:
        pass


def _finalize_session(session: dict | None, end_reason: str = "tui_close") -> None:
    """Best-effort finalize hook + memory commit for a session."""
    if not session or session.get("_finalized"):
        return
    session["_finalized"] = True
    stop_event = session.get("_notif_stop")
    if stop_event is not None:
        stop_event.set()

    agent = session.get("agent")
    lock = session.get("history_lock")
    if lock is not None:
        with lock:
            history = list(session.get("history", []))
    else:
        history = list(session.get("history", []))
    if agent is not None and history and hasattr(agent, "commit_memory_session"):
        try:
            agent.commit_memory_session(history)
        except Exception:
            pass

    session_key = session.get("session_key")
    session_id = getattr(agent, "session_id", None) or session_key
    _notify_session_boundary("on_session_finalize", session_id)

    # Mark session ended in DB so it doesn't linger as a ghost row in /resume.
    # Use session_id (from agent.session_id) not session_key — after compression,
    # session_key may be stale (the ended parent) while session_id is the live
    # continuation. Fix for #20001.
    if session_id:
        try:
            db = _get_db()
            if db is not None:
                db.end_session(session_id, end_reason)
        except Exception:
            pass


def _shutdown_sessions() -> None:
    for session in list(_sessions.values()):
        _finalize_session(session, end_reason="tui_shutdown")
        try:
            worker = session.get("slash_worker")
            if worker:
                worker.close()
        except Exception:
            pass


atexit.register(_shutdown_sessions)


# ── Plumbing ──────────────────────────────────────────────────────────


def _get_db():
    global _db, _db_error
    if _db is None:
        from hermes_state import SessionDB

        try:
            _db = SessionDB()
            _db_error = None
        except Exception as exc:
            _db_error = str(exc)
            logger.warning(
                "TUI session store unavailable — continuing without state.db features: %s",
                exc,
            )
            return None
    return _db


def _db_unavailable_error(rid, *, code: int):
    detail = _db_error or "state.db unavailable"
    return _err(rid, code, f"state.db unavailable: {detail}")


def write_json(obj: dict) -> bool:
    """Emit one JSON frame. Routes via the most-specific transport available.

    Precedence:

    1. Event frames with a session id → the transport stored on that session,
       so async events land with the client that owns the session even if
       the emitting thread has no contextvar binding.
    2. Otherwise the transport bound on the current context (set by
       :func:`dispatch` for the lifetime of a request).
    3. Otherwise the module-level stdio transport, matching the historical
       behaviour and keeping tests that monkey-patch ``_real_stdout`` green.
    """
    if obj.get("method") == "event":
        sid = ((obj.get("params") or {}).get("session_id")) or ""
        if sid and (t := (_sessions.get(sid) or {}).get("transport")) is not None:
            return t.write(obj)

    return (current_transport() or _stdio_transport).write(obj)


def _emit(event: str, sid: str, payload: dict | None = None):
    params = {"type": event, "session_id": sid}
    if payload is not None:
        params["payload"] = payload
    write_json({"jsonrpc": "2.0", "method": "event", "params": params})


def _status_update(sid: str, kind: str, text: str | None = None):
    body = (text if text is not None else kind).strip()
    if not body:
        return
    _emit(
        "status.update",
        sid,
        {"kind": kind if text is not None else "status", "text": body},
    )


def _estimate_image_tokens(width: int, height: int) -> int:
    """Very rough UI estimate for image prompt cost.

    Uses 512px tiles at ~85 tokens/tile as a lightweight cross-provider hint.
    This is intentionally approximate and only used for attachment display.
    """
    if width <= 0 or height <= 0:
        return 0
    return max(1, (width + 511) // 512) * max(1, (height + 511) // 512) * 85


def _image_meta(path: Path) -> dict:
    meta = {"name": path.name}
    try:
        from PIL import Image

        with Image.open(path) as img:
            width, height = img.size
        meta["width"] = int(width)
        meta["height"] = int(height)
        meta["token_estimate"] = _estimate_image_tokens(int(width), int(height))
    except Exception:
        pass
    return meta


def _ok(rid, result: dict) -> dict:
    return {"jsonrpc": "2.0", "id": rid, "result": result}


def _err(rid, code: int, msg: str) -> dict:
    return {"jsonrpc": "2.0", "id": rid, "error": {"code": code, "message": msg}}


def method(name: str):
    def dec(fn):
        _methods[name] = fn
        return fn

    return dec


def _normalize_request(req: Any) -> tuple[Any, str, dict] | dict:
    """Validate a JSON-RPC request enough for safe local dispatch."""
    if not isinstance(req, dict):
        return _err(None, -32600, "invalid request: expected an object")

    rid = req.get("id")
    method = req.get("method")
    if not isinstance(method, str) or not method:
        return _err(rid, -32600, "invalid request: method must be a non-empty string")

    params = req.get("params", {})
    if params is None:
        params = {}
    elif not isinstance(params, dict):
        return _err(rid, -32602, "invalid params: expected an object")

    return rid, method, params


def handle_request(req: dict) -> dict | None:
    normalized = _normalize_request(req)
    if isinstance(normalized, dict):
        return normalized

    rid, method, params = normalized
    fn = _methods.get(method)
    if not fn:
        return _err(rid, -32601, f"unknown method: {method}")
    return fn(rid, params)


def dispatch(req: dict, transport: Optional[Transport] = None) -> dict | None:
    """Route inbound RPCs — long handlers to the pool, everything else inline.

    Returns a response dict when handled inline. Returns None when the
    handler was scheduled on the pool; the worker writes its own response
    via the bound transport when done.

    *transport* (optional): pins every write produced by this request —
    including any events emitted by the handler — to the given transport.
    Omitting it falls back to the module-level stdio transport, preserving
    the original behaviour for ``tui_gateway.entry``.
    """
    t = transport or _stdio_transport
    token = bind_transport(t)
    try:
        normalized = _normalize_request(req)
        if isinstance(normalized, dict):
            return normalized

        _rid, method, _params = normalized
        if method not in _LONG_HANDLERS:
            return handle_request(req)

        # Snapshot the context so the pool worker sees the bound transport.
        ctx = contextvars.copy_context()

        def run():
            try:
                resp = handle_request(req)
            except Exception as exc:
                resp = _err(req.get("id"), -32000, f"handler error: {exc}")
            if resp is not None:
                t.write(resp)

        _pool.submit(lambda: ctx.run(run))

        return None
    finally:
        reset_transport(token)


def _wait_agent(session: dict, rid: str, timeout: float = 30.0) -> dict | None:
    ready = session.get("agent_ready")
    if ready is not None and not ready.wait(timeout=timeout):
        return _err(rid, 5032, "agent initialization timed out")
    err = session.get("agent_error")
    return _err(rid, 5032, err) if err else None


def _start_agent_build(sid: str, session: dict) -> None:
    """Start building the real AIAgent for a TUI session, once.

    Classic `hermes` shows the prompt before constructing AIAgent; the TUI used
    to eagerly build it during session.create, making startup feel blocked on
    tool discovery/model metadata even though the composer was visible.  Keep
    the shell responsive by deferring this work until the first prompt (or any
    command that actually needs the agent), while retaining the same ready/error
    event contract for the frontend.
    """
    ready = session.get("agent_ready")
    if ready is None:
        return
    lock = session.setdefault("agent_build_lock", threading.Lock())
    with lock:
        if ready.is_set() or session.get("agent_build_started"):
            return
        session["agent_build_started"] = True
    key = session["session_key"]

    def _build() -> None:
        current = _sessions.get(sid)
        if current is None:
            ready.set()
            return

        worker = None
        notify_registered = False
        try:
            tokens = _set_session_context(key)
            try:
                agent = _make_agent(sid, key)
            finally:
                _clear_session_context(tokens)

            # Session DB row deferred to first run_conversation() call.
            # pending_title applied post-first-message (see cli.exec handler).
            current["agent"] = agent

            try:
                worker = _SlashWorker(key, getattr(agent, "model", _resolve_model()))
                current["slash_worker"] = worker
            except Exception:
                pass

            try:
                from tools.approval import (
                    register_gateway_notify,
                    load_permanent_allowlist,
                )

                register_gateway_notify(
                    key, lambda data: _emit("approval.request", sid, data)
                )
                notify_registered = True
                load_permanent_allowlist()
            except Exception:
                pass

            _wire_callbacks(sid)
            _sessions[sid]["_notif_stop"] = _start_notification_poller(sid, _sessions[sid])
            _notify_session_boundary("on_session_reset", key)

            info = _session_info(agent, current)
            cfg_warn = _probe_config_health(_load_cfg())
            if cfg_warn:
                info["config_warning"] = cfg_warn
                logger.warning(cfg_warn)
            _emit("session.info", sid, info)
        except Exception as e:
            current["agent_error"] = str(e)
            _emit("error", sid, {"message": f"agent init failed: {e}"})
        finally:
            if _sessions.get(sid) is not current:
                if worker is not None:
                    try:
                        worker.close()
                    except Exception:
                        pass
                if notify_registered:
                    try:
                        from tools.approval import unregister_gateway_notify

                        unregister_gateway_notify(key)
                    except Exception:
                        pass
            ready.set()

    threading.Thread(target=_build, daemon=True).start()


def _sess_nowait(params, rid):
    s = _sessions.get(params.get("session_id") or "")
    return (s, None) if s else (None, _err(rid, 4001, "session not found"))


def _sess(params, rid):
    s, err = _sess_nowait(params, rid)
    if err:
        return (None, err)
    _start_agent_build(params.get("session_id") or "", s)
    return (s, _wait_agent(s, rid))


def _normalize_completion_path(path_part: str) -> str:
    expanded = os.path.expanduser(path_part)
    if os.name != "nt":
        normalized = expanded.replace("\\", "/")
        if (
            len(normalized) >= 3
            and normalized[1] == ":"
            and normalized[2] == "/"
            and normalized[0].isalpha()
        ):
            return f"/mnt/{normalized[0].lower()}/{normalized[3:]}"
    return expanded


def _completion_cwd(params: dict | None = None) -> str:
    raw = (
        (params or {}).get("cwd")
        or _sessions.get((params or {}).get("session_id") or "", {}).get("cwd")
        or os.environ.get("TERMINAL_CWD")
        or os.getcwd()
    )
    try:
        resolved = os.path.abspath(os.path.expanduser(str(raw)))
        if os.path.isdir(resolved):
            return resolved
    except Exception:
        pass
    return os.getcwd()


def _git_branch_for_cwd(cwd: str) -> str:
    try:
        result = subprocess.run(
            ["git", "-C", cwd, "branch", "--show-current"],
            capture_output=True,
            text=True,
            timeout=1.5,
            check=False,
        )
        if result.returncode == 0:
            branch = result.stdout.strip()
            if branch:
                return branch
        head = subprocess.run(
            ["git", "-C", cwd, "rev-parse", "--short", "HEAD"],
            capture_output=True,
            text=True,
            timeout=1.5,
            check=False,
        )
        return head.stdout.strip() if head.returncode == 0 else ""
    except Exception:
        return ""


def _session_cwd(session: dict | None) -> str:
    if session and session.get("cwd"):
        return str(session["cwd"])
    return _completion_cwd()


def _register_session_cwd(session: dict | None) -> None:
    if not session:
        return
    try:
        from tools.terminal_tool import register_task_env_overrides

        register_task_env_overrides(
            session["session_key"], {"cwd": _session_cwd(session)}
        )
    except Exception:
        pass


def _set_session_cwd(session: dict, cwd: str) -> str:
    resolved = os.path.abspath(os.path.expanduser(str(cwd)))
    if not os.path.isdir(resolved):
        raise ValueError(f"working directory does not exist: {cwd}")
    session["cwd"] = resolved
    _register_session_cwd(session)
    db = _get_db()
    if db is not None:
        try:
            db.update_session_cwd(session.get("session_key", ""), resolved)
        except Exception:
            logger.debug("failed to persist session cwd", exc_info=True)
    try:
        from tools.terminal_tool import cleanup_vm

        cleanup_vm(session["session_key"])
    except Exception:
        pass
    return resolved


# ── Config I/O ────────────────────────────────────────────────────────


# Keep aligned with `INDICATOR_STYLES` / `DEFAULT_INDICATOR_STYLE` in
# ``ui-tui/src/app/interfaces.ts`` — both ends validate against the
# same shape so `config.get indicator` and the live TUI render agree.
_INDICATOR_STYLES: tuple[str, ...] = ("ascii", "emoji", "kaomoji", "unicode")
_INDICATOR_DEFAULT = "kaomoji"


def _load_cfg() -> dict:
    global _cfg_cache, _cfg_mtime, _cfg_path
    try:
        import yaml

        p = _hermes_home / "config.yaml"
        mtime = p.stat().st_mtime if p.exists() else None
        with _cfg_lock:
            if _cfg_cache is not None and _cfg_mtime == mtime and _cfg_path == p:
                return copy.deepcopy(_cfg_cache)
        if p.exists():
            with open(p, encoding="utf-8") as f:
                data = yaml.safe_load(f) or {}
        else:
            data = {}
        with _cfg_lock:
            _cfg_cache = copy.deepcopy(data)
            _cfg_mtime = mtime
            _cfg_path = p
        return data
    except Exception:
        pass
    return {}


def _save_cfg(cfg: dict):
    global _cfg_cache, _cfg_mtime, _cfg_path
    import yaml

    path = _hermes_home / "config.yaml"
    with open(path, "w", encoding="utf-8") as f:
        yaml.safe_dump(cfg, f)
    with _cfg_lock:
        _cfg_cache = copy.deepcopy(cfg)
        _cfg_path = path
        try:
            _cfg_mtime = path.stat().st_mtime
        except Exception:
            _cfg_mtime = None


def _set_session_context(session_key: str) -> list:
    try:
        from gateway.session_context import set_session_vars

        return set_session_vars(session_key=session_key)
    except Exception:
        return []


def _clear_session_context(tokens: list) -> None:
    if not tokens:
        return
    try:
        from gateway.session_context import clear_session_vars

        clear_session_vars(tokens)
    except Exception:
        pass


def _enable_gateway_prompts() -> None:
    """Route approvals through gateway callbacks instead of CLI input()."""
    os.environ["HERMES_GATEWAY_SESSION"] = "1"
    os.environ["HERMES_EXEC_ASK"] = "1"
    os.environ["HERMES_INTERACTIVE"] = "1"


# ── Blocking prompt factory ──────────────────────────────────────────


def _block(event: str, sid: str, payload: dict, timeout: int = 300) -> str:
    rid = uuid.uuid4().hex[:8]
    ev = threading.Event()
    _pending[rid] = (sid, ev)
    payload["request_id"] = rid
    _pending_prompt_payloads[rid] = (event, dict(payload))
    try:
        _emit(event, sid, payload)
        ev.wait(timeout=timeout)
    finally:
        _pending.pop(rid, None)
        _pending_prompt_payloads.pop(rid, None)
    return _answers.pop(rid, "")


def _clear_pending(sid: str | None = None) -> None:
    """Release pending prompts with an empty answer.

    When *sid* is provided, only prompts owned by that session are
    released — critical for session.interrupt, which must not
    collaterally cancel clarify/sudo/secret prompts on unrelated
    sessions sharing the same tui_gateway process.  When *sid* is
    None, every pending prompt is released (used during shutdown).
    """
    for rid, (owner_sid, ev) in list(_pending.items()):
        if sid is None or owner_sid == sid:
            _answers[rid] = ""
            ev.set()


# ── Agent factory ────────────────────────────────────────────────────


def resolve_skin() -> dict:
    try:
        from hermes_cli.skin_engine import init_skin_from_config, get_active_skin

        init_skin_from_config(_load_cfg())
        skin = get_active_skin()
        return {
            "name": skin.name,
            "colors": skin.colors,
            "branding": skin.branding,
            "banner_logo": skin.banner_logo,
            "banner_hero": skin.banner_hero,
            "tool_prefix": skin.tool_prefix,
            "help_header": (skin.branding or {}).get("help_header", ""),
        }
    except Exception:
        return {}


def _resolve_model() -> str:
    env = (
        os.environ.get("HERMES_MODEL", "")
        or os.environ.get("HERMES_INFERENCE_MODEL", "")
    ).strip()
    if env:
        return env
    m = _load_cfg().get("model", "")
    if isinstance(m, dict):
        return str(m.get("default", "") or "").strip()
    if isinstance(m, str) and m:
        return m.strip()
    return "anthropic/claude-sonnet-4"


def _resolve_startup_runtime() -> tuple[str, str | None]:
    model = _resolve_model()
    explicit_provider = os.environ.get("HERMES_TUI_PROVIDER", "").strip()
    if explicit_provider:
        return model, explicit_provider

    explicit_model = (
        os.environ.get("HERMES_MODEL", "")
        or os.environ.get("HERMES_INFERENCE_MODEL", "")
    ).strip()
    if not explicit_model:
        return model, None

    try:
        from hermes_cli.models import detect_static_provider_for_model

        cfg = _load_cfg().get("model") or {}
        current_provider = (
            (
                str(cfg.get("provider") or "").strip().lower()
                if isinstance(cfg, dict)
                else ""
            )
            or os.environ.get("HERMES_INFERENCE_PROVIDER", "").strip().lower()
            or "auto"
        )
        detected = detect_static_provider_for_model(explicit_model, current_provider)
        if detected:
            provider, detected_model = detected
            return detected_model, provider
    except Exception:
        pass
    return model, None


def _write_config_key(key_path: str, value):
    cfg = _load_cfg()
    current = cfg
    keys = key_path.split(".")
    for key in keys[:-1]:
        if key not in current or not isinstance(current.get(key), dict):
            current[key] = {}
        current = current[key]
    current[keys[-1]] = value
    _save_cfg(cfg)


_STATUSBAR_MODES = frozenset({"off", "top", "bottom"})


def _coerce_statusbar(raw) -> str:
    if raw is False:
        return "off"
    if isinstance(raw, str) and (s := raw.strip().lower()) in _STATUSBAR_MODES:
        return s
    return "top"


_MOUSE_TRACKING_ALIASES = {
    "0": "off",
    "1": "all",
    "all": "all",
    "any": "all",
    "button": "buttons",
    "buttons": "buttons",
    "click": "buttons",
    "false": "off",
    "full": "all",
    "no": "off",
    "off": "off",
    "on": "all",
    "scroll": "wheel",
    "true": "all",
    "wheel": "wheel",
    "yes": "all",
}


def _display_mouse_tracking(display: dict) -> str:
    """Resolve display.mouse_tracking to one of ``off|wheel|buttons|all``.

    Boolean values keep their legacy meaning (``True`` → ``all``, ``False`` →
    ``off``). The ``wheel`` preset (DEC 1000+1006) is the tmux-friendly
    subset — wheel + click only, no hover events to trigger prompt-row
    clipboard probes. Legacy ``tui_mouse`` is honored only when
    ``mouse_tracking`` is absent.
    """
    if not isinstance(display, dict):
        return "all"
    if "mouse_tracking" in display:
        raw = display.get("mouse_tracking")
    else:
        raw = display.get("tui_mouse", True)
    if raw is False or raw == 0:
        return "off"
    if raw is True or raw is None:
        return "all"
    if isinstance(raw, (int, float)):
        return "all"
    if isinstance(raw, str):
        return _MOUSE_TRACKING_ALIASES.get(raw.strip().lower(), "all")
    return "all"


def _load_reasoning_config() -> dict | None:
    from hermes_constants import parse_reasoning_effort

    effort = str(
        (_load_cfg().get("agent") or {}).get("reasoning_effort", "") or ""
    ).strip()
    return parse_reasoning_effort(effort)


def _load_service_tier() -> str | None:
    raw = (
        str((_load_cfg().get("agent") or {}).get("service_tier", "") or "")
        .strip()
        .lower()
    )
    if not raw or raw in {"normal", "default", "standard", "off", "none"}:
        return None
    if raw in {"fast", "priority", "on"}:
        return "priority"
    return None


def _load_show_reasoning() -> bool:
    return bool((_load_cfg().get("display") or {}).get("show_reasoning", False))


def _load_tool_progress_mode() -> str:
    env = os.environ.get("HERMES_TUI_TOOL_PROGRESS", "").strip().lower()
    if env in {"off", "new", "all", "verbose"}:
        return env
    raw = (_load_cfg().get("display") or {}).get("tool_progress", "all")
    if raw is False:
        return "off"
    if raw is True:
        return "all"
    mode = str(raw or "all").strip().lower()
    return mode if mode in {"off", "new", "all", "verbose"} else "all"


def _load_enabled_toolsets() -> list[str] | None:
    explicit = [
        item.strip()
        for item in os.environ.get("HERMES_TUI_TOOLSETS", "").split(",")
        if item.strip()
    ]
    cfg = None
    fallback_notice = None

    try:
        from toolsets import validate_toolset
    except Exception:
        validate_toolset = None

    if explicit and validate_toolset is not None:
        built_in = [name for name in explicit if validate_toolset(name)]
        unresolved = [name for name in explicit if name not in built_in]

        if unresolved:
            try:
                from hermes_cli.plugins import discover_plugins

                discover_plugins()
                plugin_valid = [name for name in unresolved if validate_toolset(name)]
            except Exception:
                plugin_valid = []

            if plugin_valid:
                built_in.extend(plugin_valid)
                unresolved = [name for name in unresolved if name not in plugin_valid]

        if any(name in {"all", "*"} for name in built_in):
            ignored = [name for name in explicit if name not in {"all", "*"}]
            if ignored:
                print(
                    "[tui] HERMES_TUI_TOOLSETS=all enables every toolset; "
                    f"ignoring additional entries: {', '.join(ignored)}",
                    file=sys.stderr,
                    flush=True,
                )
            return None

        if not unresolved:
            return built_in

        mcp_names: set[str] = set()
        mcp_disabled: set[str] = set()
        try:
            from hermes_cli.config import read_raw_config
            from hermes_cli.tools_config import _parse_enabled_flag

            raw_cfg = read_raw_config()
            mcp_servers = (
                raw_cfg.get("mcp_servers")
                if isinstance(raw_cfg.get("mcp_servers"), dict)
                else {}
            )
            for name, server_cfg in mcp_servers.items():
                if not isinstance(server_cfg, dict):
                    continue
                if _parse_enabled_flag(server_cfg.get("enabled", True), default=True):
                    mcp_names.add(str(name))
                else:
                    mcp_disabled.add(str(name))
        except Exception:
            mcp_names = set()
            mcp_disabled = set()

        mcp_valid = [name for name in unresolved if name in mcp_names]
        disabled = [name for name in unresolved if name in mcp_disabled]
        unknown = [
            name
            for name in unresolved
            if name not in mcp_names and name not in mcp_disabled
        ]
        valid = built_in + mcp_valid

        if unknown:
            print(
                f"[tui] ignoring unknown HERMES_TUI_TOOLSETS entries: {', '.join(unknown)}",
                file=sys.stderr,
                flush=True,
            )
        if disabled:
            print(
                "[tui] ignoring disabled MCP servers in HERMES_TUI_TOOLSETS "
                "(set enabled: true in config.yaml to use): "
                f"{', '.join(disabled)}",
                file=sys.stderr,
                flush=True,
            )

        if valid:
            return valid

        fallback_notice = (
            "[tui] no valid HERMES_TUI_TOOLSETS entries; using configured CLI toolsets"
        )

    try:
        from hermes_cli.config import load_config
        from hermes_cli.tools_config import _get_platform_tools

        cfg = cfg if cfg is not None else load_config()

        # Runtime toolset resolution must include default MCP servers so the
        # agent can actually call them. Passing ``False`` here is the
        # config-editing variant — used when we need to persist a toolset
        # list without baking in implicit MCP defaults. Using the wrong
        # variant at agent creation time makes MCP tools silently missing
        # from the TUI. See PR #3252 for the original design split.
        enabled = sorted(
            _get_platform_tools(cfg, "cli", include_default_mcp_servers=True)
        )
        if fallback_notice is not None:
            print(fallback_notice, file=sys.stderr, flush=True)
        return enabled or None
    except Exception:
        if fallback_notice is not None:
            print(
                "[tui] no valid HERMES_TUI_TOOLSETS entries and configured CLI toolsets could not be loaded; enabling all toolsets",
                file=sys.stderr,
                flush=True,
            )
        return None


def _session_tool_progress_mode(sid: str) -> str:
    return str(_sessions.get(sid, {}).get("tool_progress_mode", "all") or "all")


def _session_verbose(sid: str) -> bool:
    return _session_tool_progress_mode(sid) == "verbose"


def _tool_progress_enabled(sid: str) -> bool:
    return _session_tool_progress_mode(sid) != "off"


def _restart_slash_worker(session: dict):
    worker = session.get("slash_worker")
    if worker:
        try:
            worker.close()
        except Exception:
            pass
    try:
        session["slash_worker"] = _SlashWorker(
            session["session_key"],
            getattr(session.get("agent"), "model", _resolve_model()),
        )
    except Exception:
        session["slash_worker"] = None


def _persist_model_switch(result) -> None:
    from hermes_cli.config import save_config

    cfg = _load_cfg()
    model_cfg = cfg.get("model")
    if not isinstance(model_cfg, dict):
        model_cfg = {}
        cfg["model"] = model_cfg

    model_cfg["default"] = result.new_model
    model_cfg["provider"] = result.target_provider
    if result.base_url:
        model_cfg["base_url"] = result.base_url
    else:
        model_cfg.pop("base_url", None)
    save_config(cfg)


def _apply_model_switch(sid: str, session: dict, raw_input: str) -> dict:
    from hermes_cli.model_switch import parse_model_flags, switch_model
    from hermes_cli.runtime_provider import resolve_runtime_provider

    model_input, explicit_provider, persist_global, _force_refresh = parse_model_flags(raw_input)
    if not model_input:
        raise ValueError("model value required")

    agent = session.get("agent")
    if agent:
        current_provider = getattr(agent, "provider", "") or ""
        current_model = getattr(agent, "model", "") or ""
        current_base_url = getattr(agent, "base_url", "") or ""
        current_api_key = getattr(agent, "api_key", "") or ""
    else:
        runtime = resolve_runtime_provider(requested=None)
        current_provider = str(runtime.get("provider", "") or "")
        current_model = _resolve_model()
        current_base_url = str(runtime.get("base_url", "") or "")
        # Preserve a callable api_key (Azure Foundry Entra ID bearer
        # provider) unchanged — ``str(...)`` would produce
        # ``"<function ...>"`` and poison downstream switch_model
        # validation. Match the agent-present branch's behavior at the
        # top of this block.
        _runtime_key = runtime.get("api_key", "")
        if callable(_runtime_key) and not isinstance(_runtime_key, str):
            current_api_key = _runtime_key
        else:
            current_api_key = str(_runtime_key or "")

    # Load user-defined providers so switch_model can resolve named custom
    # endpoints (e.g. "ollama-launch") and validate against saved model lists.
    user_provs = None
    custom_provs = None
    try:
        from hermes_cli.config import get_compatible_custom_providers, load_config

        cfg = load_config()
        user_provs = cfg.get("providers")
        custom_provs = get_compatible_custom_providers(cfg)
    except Exception:
        pass

    result = switch_model(
        raw_input=model_input,
        current_provider=current_provider,
        current_model=current_model,
        current_base_url=current_base_url,
        current_api_key=current_api_key,
        is_global=persist_global,
        explicit_provider=explicit_provider,
        user_providers=user_provs,
        custom_providers=custom_provs,
    )
    if not result.success:
        raise ValueError(result.error_message or "model switch failed")

    if agent:
        agent.switch_model(
            new_model=result.new_model,
            new_provider=result.target_provider,
            api_key=result.api_key,
            base_url=result.base_url,
            api_mode=result.api_mode,
        )
        _restart_slash_worker(session)
        _emit("session.info", sid, _session_info(agent, session))

    os.environ["HERMES_MODEL"] = result.new_model
    os.environ["HERMES_INFERENCE_MODEL"] = result.new_model
    # Keep the process-level provider env vars in sync with the user's
    # explicit choice so any ambient re-resolution (credential pool refresh,
    # compressor rebuild, aux clients) and startup re-resolution on /new
    # both pick up the new provider instead of the original one persisted
    # in config or env.
    #
    # HERMES_TUI_PROVIDER is the canonical "explicit-this-process" carrier
    # consumed by _resolve_startup_runtime() — set it unconditionally on
    # /model so /new can't fall through to static-catalog detection and
    # pick a coincidentally-matching native provider (fixes #16857).
    if result.target_provider:
        os.environ["HERMES_INFERENCE_PROVIDER"] = result.target_provider
        os.environ["HERMES_TUI_PROVIDER"] = result.target_provider
    if persist_global:
        _persist_model_switch(result)
    return {"value": result.new_model, "warning": result.warning_message or ""}


def _compress_session_history(
    session: dict,
    focus_topic: str | None = None,
    approx_tokens: int | None = None,
    before_messages: list | None = None,
    history_version: int | None = None,
) -> tuple[int, dict]:
    from agent.model_metadata import estimate_request_tokens_rough

    agent = session["agent"]
    # Snapshot history under the lock so the LLM-bound compression call
    # below does NOT hold history_lock for the duration of the request —
    # otherwise other handlers acquiring the lock (prompt.submit etc.)
    # block on the dispatcher loop while compaction runs.
    if before_messages is None or history_version is None:
        with session["history_lock"]:
            before_messages = list(session.get("history", []))
            history_version = int(session.get("history_version", 0))
    history = before_messages
    if len(history) < 4:
        usage = _get_usage(agent)
        return 0, usage
    if approx_tokens is None:
        # Include system prompt + tool schemas so the figure reflects real
        # request pressure, not a transcript-only underestimate (#6217).
        _sys_prompt = getattr(agent, "_cached_system_prompt", "") or ""
        _tools = getattr(agent, "tools", None) or None
        approx_tokens = estimate_request_tokens_rough(
            history, system_prompt=_sys_prompt, tools=_tools
        )
    # Pass system_message=None so AIAgent._compress_context rebuilds the
    # system prompt cleanly via _build_system_prompt(None). Passing the
    # cached prompt (which already contains the agent identity block)
    # makes the rebuild append the identity a second time. Mirrors the
    # CLI's _manual_compress fix for issue #15281.
    compressed, _ = agent._compress_context(
        history,
        None,
        approx_tokens=approx_tokens,
        focus_topic=focus_topic or None,
    )
    with session["history_lock"]:
        if int(session.get("history_version", 0)) != history_version:
            # External mutation during compaction — drop the compressed
            # result so we don't clobber concurrent edits.
            usage = _get_usage(agent)
            return 0, usage
        session["history"] = compressed
        session["history_version"] = history_version + 1
    usage = _get_usage(agent)
    return len(history) - len(compressed), usage


def _sync_session_key_after_compress(
    sid: str,
    session: dict,
    *,
    clear_pending_title: bool = True,
    restart_slash_worker: bool = True,
) -> None:
    """Re-anchor session_key when AIAgent._compress_context rotates session_id.

    AIAgent._compress_context ends the current SessionDB session and creates
    a new continuation session, rotating ``agent.session_id``.  The TUI
    gateway keeps the gateway-side ``session_key`` separate (used for
    approval routing, slash worker init, DB title/history lookups, yolo
    state).  Without this sync, those operations would target the ended
    parent session while the agent writes to the new continuation session.

    Policy flags:
        clear_pending_title: True for manual /compress (title belongs to old
            session). False for post-turn auto-compression (preserve user
            intent so pending_title can be applied to the continuation).
        restart_slash_worker: True for manual /compress and post-turn
            auto-compression (worker holds stale session key). False only
            if the caller manages the worker lifecycle separately.
    """
    agent = session.get("agent")
    new_session_id = getattr(agent, "session_id", None) or ""
    old_key = session.get("session_key", "") or ""
    if not new_session_id or new_session_id == old_key:
        return

    try:
        from tools.approval import (
            disable_session_yolo,
            enable_session_yolo,
            is_session_yolo_enabled,
            register_gateway_notify,
            unregister_gateway_notify,
        )

        try:
            unregister_gateway_notify(old_key)
        except Exception:
            pass
        session["session_key"] = new_session_id
        try:
            yolo_was_on = is_session_yolo_enabled(old_key)
        except Exception:
            yolo_was_on = False
        if yolo_was_on:
            try:
                enable_session_yolo(new_session_id)
                disable_session_yolo(old_key)
            except Exception:
                pass
        try:
            register_gateway_notify(
                new_session_id,
                lambda data: _emit("approval.request", sid, data),
            )
        except Exception:
            pass
    except Exception:
        # Even if the approval module fails to import, still anchor the
        # session_key on the new continuation id so downstream lookups
        # don't keep targeting the ended row.
        session["session_key"] = new_session_id

    if clear_pending_title:
        session["pending_title"] = None
    if restart_slash_worker:
        try:
            _restart_slash_worker(session)
        except Exception:
            pass


def _get_usage(agent) -> dict:
    g = lambda k, fb=None: getattr(agent, k, 0) or (getattr(agent, fb, 0) if fb else 0)
    usage = {
        "model": getattr(agent, "model", "") or "",
        "input": g("session_input_tokens", "session_prompt_tokens"),
        "output": g("session_output_tokens", "session_completion_tokens"),
        "cache_read": g("session_cache_read_tokens"),
        "cache_write": g("session_cache_write_tokens"),
        "reasoning": g("session_reasoning_tokens"),
        "prompt": g("session_prompt_tokens"),
        "completion": g("session_completion_tokens"),
        "total": g("session_total_tokens"),
        "calls": g("session_api_calls"),
    }
    comp = getattr(agent, "context_compressor", None)
    if comp:
        ctx_used = getattr(comp, "last_prompt_tokens", 0) or usage["total"] or 0
        ctx_max = getattr(comp, "context_length", 0) or 0
        if ctx_max:
            usage["context_used"] = ctx_used
            usage["context_max"] = ctx_max
            usage["context_percent"] = max(0, min(100, round(ctx_used / ctx_max * 100)))
        usage["compressions"] = getattr(comp, "compression_count", 0) or 0
    try:
        from agent.usage_pricing import CanonicalUsage, estimate_usage_cost

        cost = estimate_usage_cost(
            usage["model"],
            CanonicalUsage(
                input_tokens=usage["input"],
                output_tokens=usage["output"],
                cache_read_tokens=usage["cache_read"],
                cache_write_tokens=usage["cache_write"],
            ),
            provider=getattr(agent, "provider", None),
            base_url=getattr(agent, "base_url", None),
        )
        usage["cost_status"] = cost.status
        if cost.amount_usd is not None:
            usage["cost_usd"] = float(cost.amount_usd)
    except Exception:
        pass
    return usage


def _probe_credentials(agent) -> str:
    """Light credential check at session creation — returns warning or ''."""
    try:
        key = getattr(agent, "api_key", "") or ""
        provider = getattr(agent, "provider", "") or ""
        if not key or key == "no-key-required":
            return f"No API key configured for provider '{provider}'. First message will fail."
    except Exception:
        pass
    return ""


def _probe_config_health(cfg: dict) -> str:
    """Flag bare YAML keys (`agent:` with no value → None) that silently
    drop nested settings. Returns warning or ''."""
    if not isinstance(cfg, dict):
        return ""
    warnings: list[str] = []
    null_keys = sorted(k for k, v in cfg.items() if v is None)
    if not null_keys:
        pass
    else:
        keys = ", ".join(f"`{k}`" for k in null_keys)
        warnings.append(
            f"config.yaml has empty section(s): {keys}. "
            f"Remove the line(s) or set them to `{{}}` — "
            f"empty sections silently drop nested settings."
        )
    display_cfg = cfg.get("display")
    agent_cfg = cfg.get("agent")
    if isinstance(display_cfg, dict):
        personality = str(display_cfg.get("personality", "") or "").strip().lower()
        if (
            personality
            and personality not in {"default", "none", "neutral"}
            and isinstance(agent_cfg, dict)
            and agent_cfg.get("personalities") is None
        ):
            warnings.append(
                "`display.personality` is set but `agent.personalities` is empty/null; "
                "personality overlay will be skipped."
            )
    return " ".join(warnings).strip()


def _current_profile_name() -> str:
    try:
        from hermes_cli.profiles import get_active_profile_name

        return get_active_profile_name() or "default"
    except Exception:
        return "default"


# Monotonic GUI<->backend contract version. The desktop app refuses to drive a
# backend reporting less than its required value (or none at all — a pre-GUI
# checkout), surfacing a one-click "update to align" prompt instead of failing
# cryptically downstream. Bump whenever the desktop's backend contract changes.
DESKTOP_BACKEND_CONTRACT = 1


def _session_info(agent, session: dict | None = None) -> dict:
    if session is None:
        for candidate in _sessions.values():
            if candidate.get("agent") is agent:
                session = candidate
                break
    cwd = _session_cwd(session)
    cfg_personality = ((_load_cfg().get("display") or {}).get("personality") or "")
    personality = (session or {}).get("personality", cfg_personality)
    reasoning_config = getattr(agent, "reasoning_config", None)
    reasoning_effort = ""
    if (
        isinstance(reasoning_config, dict)
        and reasoning_config.get("enabled") is not False
    ):
        reasoning_effort = str(reasoning_config.get("effort", "") or "")
    service_tier = getattr(agent, "service_tier", None) or ""
    info: dict = {
        "model": getattr(agent, "model", ""),
        "reasoning_effort": reasoning_effort,
        "service_tier": service_tier,
        "fast": service_tier == "priority",
        "tools": {},
        "skills": {},
        "cwd": cwd,
        "branch": _git_branch_for_cwd(cwd),
        "personality": str(personality or ""),
        "running": bool((session or {}).get("running")),
        "desktop_contract": DESKTOP_BACKEND_CONTRACT,
        "version": "",
        "release_date": "",
        "update_behind": None,
        "update_command": "",
        "usage": _get_usage(agent),
        "profile_name": _current_profile_name(),
    }
    try:
        from hermes_cli import __version__, __release_date__

        info["version"] = __version__
        info["release_date"] = __release_date__
    except Exception:
        pass
    try:
        from model_tools import get_toolset_for_tool

        for t in getattr(agent, "tools", []) or []:
            name = t["function"]["name"]
            info["tools"].setdefault(get_toolset_for_tool(name) or "other", []).append(
                name
            )
    except Exception:
        pass
    try:
        from hermes_cli.banner import get_available_skills

        info["skills"] = get_available_skills()
    except Exception:
        pass
    try:
        from tools.mcp_tool import get_mcp_status

        info["mcp_servers"] = get_mcp_status()
    except Exception:
        info["mcp_servers"] = []
    try:
        info["system_prompt"] = getattr(agent, "_cached_system_prompt", "") or ""
    except Exception:
        pass
    try:
        from hermes_cli.banner import get_update_result
        from hermes_cli.config import recommended_update_command

        info["update_behind"] = get_update_result(timeout=0.5)
        info["update_command"] = recommended_update_command()
    except Exception:
        pass
    warn = _probe_credentials(agent)
    if warn:
        info["credential_warning"] = warn
    return info


def _tool_ctx(name: str, args: dict) -> str:
    try:
        from agent.display import build_tool_preview

        return build_tool_preview(name, args, max_len=80) or ""
    except Exception:
        return ""


_TUI_VERBOSE_TEXT_MAX_CHARS = 16_000
_TUI_VERBOSE_TEXT_MAX_LINES = 240


def _cap_tui_verbose_text(text: str) -> str:
    if (
        len(text) <= _TUI_VERBOSE_TEXT_MAX_CHARS
        and text.count("\n") < _TUI_VERBOSE_TEXT_MAX_LINES
    ):
        return text

    idx = len(text)
    start = 0
    for _ in range(_TUI_VERBOSE_TEXT_MAX_LINES):
        idx = text.rfind("\n", 0, idx)
        if idx < 0:
            start = 0
            break
        start = idx + 1

    line_start = start
    start = max(line_start, len(text) - _TUI_VERBOSE_TEXT_MAX_CHARS)
    if start > line_start:
        next_break = text.find("\n", start)
        if 0 <= next_break < len(text) - 1:
            start = next_break + 1

    tail = text[start:].lstrip()
    omitted_chars = max(0, len(text) - len(tail))
    omitted_lines = text[:start].count("\n")
    if omitted_lines:
        label = (
            "[showing verbose tail; omitted "
            f"{omitted_lines} lines / {omitted_chars} chars]\n"
        )
    else:
        label = f"[showing verbose tail; omitted {omitted_chars} chars]\n"
    return f"{label}{tail}"


def _redact_tui_verbose_text(text: str) -> str:
    try:
        from agent.redact import redact_sensitive_text

        redacted = redact_sensitive_text(str(text), force=True)
    except Exception:
        return ""
    return _cap_tui_verbose_text(redacted)


def _tool_args_text(args: dict) -> str:
    try:
        raw = json.dumps(args or {}, indent=2, ensure_ascii=False, default=str)
    except Exception:
        raw = str(args or {})
    return _redact_tui_verbose_text(raw)


def _tool_result_text(result: object) -> str:
    try:
        from agent.tool_dispatch_helpers import _multimodal_text_summary

        raw = _multimodal_text_summary(result)
    except Exception:
        raw = str(result)
    return _redact_tui_verbose_text(raw)


def _fmt_tool_duration(seconds: float | None) -> str:
    if seconds is None:
        return ""
    if seconds < 10:
        return f"{seconds:.1f}s"
    if seconds < 60:
        return f"{round(seconds)}s"
    mins, secs = divmod(int(round(seconds)), 60)
    return f"{mins}m {secs}s" if secs else f"{mins}m"


def _count_list(obj: object, *path: str) -> int | None:
    cur = obj
    for key in path:
        if not isinstance(cur, dict):
            return None
        cur = cur.get(key)
    return len(cur) if isinstance(cur, list) else None


def _tool_summary(name: str, result: str, duration_s: float | None) -> str | None:
    try:
        data = json.loads(result)
    except Exception:
        data = None

    dur = _fmt_tool_duration(duration_s)
    suffix = f" in {dur}" if dur else ""
    text = None

    if name == "web_search" and isinstance(data, dict):
        n = _count_list(data, "data", "web")
        if n is not None:
            text = f"Did {n} {'search' if n == 1 else 'searches'}"

    elif name == "web_extract" and isinstance(data, dict):
        n = _count_list(data, "results") or _count_list(data, "data", "results")
        if n is not None:
            text = f"Extracted {n} {'page' if n == 1 else 'pages'}"

    if isinstance(data, dict) and data.get("fallback_warning"):
        warning = str(data.get("fallback_warning") or "").strip()
        if warning:
            return f"{warning}{suffix}"

    return f"{text}{suffix}" if text else None


def _on_tool_start(sid: str, tool_call_id: str, name: str, args: dict):
    session = _sessions.get(sid)
    if session is not None:
        try:
            from agent.display import capture_local_edit_snapshot

            snapshot = capture_local_edit_snapshot(name, args)
            if snapshot is not None:
                session.setdefault("edit_snapshots", {})[tool_call_id] = snapshot
        except Exception:
            pass
        session.setdefault("tool_started_at", {})[tool_call_id] = time.time()
    if _tool_progress_enabled(sid):
        payload = {
            "tool_id": tool_call_id,
            "name": name,
            "context": _tool_ctx(name, args),
        }
        if _session_verbose(sid):
            args_text = _tool_args_text(args)
            if args_text:
                payload["args_text"] = args_text
        # tool.complete is the source of truth for todos (full list from the
        # tool result). args.todos here may be a partial merge update.
        _emit("tool.start", sid, payload)


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}
    session = _sessions.get(sid)
    snapshot = None
    started_at = None
    if session is not None:
        snapshot = session.setdefault("edit_snapshots", {}).pop(tool_call_id, None)
        started_at = session.setdefault("tool_started_at", {}).pop(tool_call_id, None)
    duration_s = time.time() - started_at if started_at else None
    if duration_s is not None:
        payload["duration_s"] = duration_s
    try:
        payload["result"] = json.loads(result)
    except Exception:
        payload["result"] = 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)
            if isinstance(data, dict) and isinstance(data.get("todos"), list):
                payload["todos"] = data.get("todos")
        except Exception:
            pass
    try:
        from agent.display import render_edit_diff_with_delta

        rendered: list[str] = []
        if render_edit_diff_with_delta(
            name,
            result,
            function_args=args,
            snapshot=snapshot,
            print_fn=rendered.append,
        ):
            payload["inline_diff"] = "\n".join(rendered)
    except Exception:
        pass
    if _tool_progress_enabled(sid) or payload.get("inline_diff"):
        _emit("tool.complete", sid, payload)


def _on_tool_progress(
    sid: str,
    event_type: str,
    name: str | None = None,
    preview: str | None = None,
    _args: dict | None = None,
    **_kwargs,
):
    if not _tool_progress_enabled(sid):
        return
    if event_type == "tool.started" and name:
        # `_on_tool_start` already emits the authoritative `tool.start` with
        # the stable tool id and args. Emitting another id-less progress row
        # here makes the desktop live view diverge from hydrated history.
        return
    if event_type == "reasoning.available" and preview:
        payload: dict[str, object] = {"text": str(preview)}
        if _session_verbose(sid):
            payload["verbose"] = True
        _emit("reasoning.available", sid, payload)
        return
    if event_type.startswith("subagent."):
        payload = {
            "goal": str(_kwargs.get("goal") or ""),
            "task_count": int(_kwargs.get("task_count") or 1),
            "task_index": int(_kwargs.get("task_index") or 0),
        }
        # Identity fields for the TUI spawn tree.  All optional — older
        # emitters that omit them fall back to flat rendering client-side.
        if _kwargs.get("subagent_id"):
            payload["subagent_id"] = str(_kwargs["subagent_id"])
        if _kwargs.get("parent_id"):
            payload["parent_id"] = str(_kwargs["parent_id"])
        if _kwargs.get("depth") is not None:
            payload["depth"] = int(_kwargs["depth"])
        if _kwargs.get("model"):
            payload["model"] = str(_kwargs["model"])
        if _kwargs.get("tool_count") is not None:
            payload["tool_count"] = int(_kwargs["tool_count"])
        if _kwargs.get("toolsets"):
            payload["toolsets"] = [str(t) for t in _kwargs["toolsets"]]
        # Per-branch rollups emitted on subagent.complete (features 1+2+4).
        for int_key in (
            "input_tokens",
            "output_tokens",
            "reasoning_tokens",
            "api_calls",
        ):
            val = _kwargs.get(int_key)
            if val is not None:
                try:
                    payload[int_key] = int(val)
                except (TypeError, ValueError):
                    pass
        if _kwargs.get("cost_usd") is not None:
            try:
                payload["cost_usd"] = float(_kwargs["cost_usd"])
            except (TypeError, ValueError):
                pass
        if _kwargs.get("files_read"):
            payload["files_read"] = [str(p) for p in _kwargs["files_read"]]
        if _kwargs.get("files_written"):
            payload["files_written"] = [str(p) for p in _kwargs["files_written"]]
        if _kwargs.get("output_tail"):
            payload["output_tail"] = list(_kwargs["output_tail"])  # list of dicts
        if name:
            payload["tool_name"] = str(name)
        if preview:
            payload["text"] = str(preview)
        if _kwargs.get("status"):
            payload["status"] = str(_kwargs["status"])
        if _kwargs.get("summary"):
            payload["summary"] = str(_kwargs["summary"])
        if _kwargs.get("duration_seconds") is not None:
            payload["duration_seconds"] = float(_kwargs["duration_seconds"])
        if preview and event_type == "subagent.tool":
            payload["tool_preview"] = str(preview)
            payload["text"] = str(preview)
        _emit(event_type, sid, payload)


def _agent_cbs(sid: str) -> dict:
    return {
        "tool_start_callback": lambda tc_id, name, args: _on_tool_start(
            sid, tc_id, name, args
        ),
        "tool_complete_callback": lambda tc_id, name, args, result: _on_tool_complete(
            sid, tc_id, name, args, result
        ),
        "tool_progress_callback": lambda event_type, name=None, preview=None, args=None, **kwargs: _on_tool_progress(
            sid, event_type, name, preview, args, **kwargs
        ),
        "tool_gen_callback": lambda name: _tool_progress_enabled(sid)
        and _emit("tool.generating", sid, {"name": name}),
        "thinking_callback": lambda text: _emit("thinking.delta", sid, {"text": text}),
        "reasoning_callback": lambda text: _emit(
            "reasoning.delta",
            sid,
            {"text": text, **({"verbose": True} if _session_verbose(sid) else {})},
        ),
        "status_callback": lambda kind, text=None: _status_update(
            sid, str(kind), None if text is None else str(text)
        ),
        "clarify_callback": lambda q, c: _block(
            "clarify.request", sid, {"question": q, "choices": c}
        ),
    }


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))

    def secret_cb(env_var, prompt, metadata=None):
        pl = {"prompt": prompt, "env_var": env_var}
        if metadata:
            pl["metadata"] = metadata
        val = _block("secret.request", sid, pl)
        if not val:
            return {
                "success": True,
                "stored_as": env_var,
                "validated": False,
                "skipped": True,
                "message": "skipped",
            }
        from hermes_cli.config import save_env_value_secure

        return {
            **save_env_value_secure(env_var, val),
            "skipped": False,
            "message": "ok",
        }

    set_secret_capture_callback(secret_cb)


def _render_personality_prompt(value) -> str:
    if isinstance(value, dict):
        parts = [value.get("system_prompt", "")]
        if value.get("tone"):
            parts.append(f'Tone: {value["tone"]}')
        if value.get("style"):
            parts.append(f'Style: {value["style"]}')
        return "\n".join(p for p in parts if p)
    return str(value)


def _available_personalities(cfg: dict | None = None) -> dict:
    try:
        from cli import load_cli_config

        return (load_cli_config().get("agent") or {}).get("personalities", {}) or {}
    except Exception:
        try:
            from hermes_cli.config import load_config as _load_full_cfg

            return (_load_full_cfg().get("agent") or {}).get("personalities", {}) or {}
        except Exception:
            cfg = cfg or _load_cfg()
            return (cfg.get("agent") or {}).get("personalities", {}) or {}


def _validate_personality(value: str, cfg: dict | None = None) -> tuple[str, str]:
    raw = str(value or "").strip()
    name = raw.lower()
    if not name or name in {"none", "default", "neutral"}:
        return "", ""

    personalities = _available_personalities(cfg)
    if name not in personalities:
        names = sorted(personalities)
        available = ", ".join(f"`{n}`" for n in names)
        base = f"Unknown personality: `{raw}`."
        if available:
            base += f"\n\nAvailable: `none`, {available}"
        else:
            base += "\n\nNo personalities configured."
        raise ValueError(base)

    return name, _render_personality_prompt(personalities[name])


def _prompt_text(value) -> str:
    """Normalize config prompt values from YAML before handing them to AIAgent."""
    if value is None:
        return ""
    if isinstance(value, str):
        return value.strip()
    if isinstance(value, list):
        return "\n".join(str(item).strip() for item in value if str(item).strip())
    return str(value).strip()


def _apply_personality_to_session(
    sid: str, session: dict, new_prompt: str, personality: str = ""
) -> tuple[bool, dict | None]:
    """Apply a personality change to an existing session without resetting history.

    Updates the agent's ephemeral system prompt in-place so the new personality
    takes effect on the next turn.  The cached base system prompt is left intact
    (ephemeral_system_prompt is appended at API-call time, not baked into the
    cache), which preserves prompt-cache hits.

    Also injects a system-role marker into the conversation history so the model
    knows to pivot its style from this point forward (without this, LLMs tend to
    continue the tone established by earlier messages in the transcript).

    Returns (history_reset, info) — history_reset is always False since we
    preserve the conversation.
    """
    if not session:
        return False, None
    session["personality"] = personality

    agent = session.get("agent")
    if agent:
        agent.ephemeral_system_prompt = new_prompt or None
        # Inject a pivot marker into history so the model sees the change point.
        # This prevents it from pattern-matching its prior style.
        if new_prompt:
            marker = (
                "[System: The user has changed the assistant's personality. "
                "From this point forward, adopt the following persona and respond "
                f"accordingly: {new_prompt}]"
            )
        else:
            marker = (
                "[System: The user has cleared the personality overlay. "
                "From this point forward, respond in your normal default style.]"
            )
        with session["history_lock"]:
            session["history"].append({"role": "user", "content": marker})
            session["history_version"] = int(session.get("history_version", 0)) + 1
        info = _session_info(agent)
        _emit("session.info", sid, info)
        return False, info
    return False, None


def _cfg_max_turns(cfg: dict, default: int) -> int:
    try:
        env_max = int(os.environ.get("HERMES_TUI_MAX_TURNS", "") or 0)
        if env_max > 0:
            return env_max
    except (TypeError, ValueError):
        pass
    agent_cfg = cfg.get("agent") or {}
    return int(agent_cfg.get("max_turns") or cfg.get("max_turns") or default)


def _parse_tui_skills_env() -> list[str]:
    raw = os.environ.get("HERMES_TUI_SKILLS", "")
    skills: list[str] = []
    seen: set[str] = set()
    for part in raw.replace("\n", ",").split(","):
        item = part.strip()
        if item and item not in seen:
            seen.add(item)
            skills.append(item)
    return skills


def _background_agent_kwargs(agent, task_id: str) -> dict:
    cfg = _load_cfg()

    return {
        "base_url": getattr(agent, "base_url", None) or None,
        "api_key": getattr(agent, "api_key", None) or None,
        "provider": getattr(agent, "provider", None) or None,
        "api_mode": getattr(agent, "api_mode", None) or None,
        "acp_command": getattr(agent, "acp_command", None) or None,
        "acp_args": getattr(agent, "acp_args", None) or None,
        "model": getattr(agent, "model", None) or _resolve_model(),
        "max_iterations": _cfg_max_turns(cfg, 25),
        "enabled_toolsets": getattr(agent, "enabled_toolsets", None)
        or _load_enabled_toolsets(),
        "quiet_mode": True,
        "verbose_logging": False,
        "ephemeral_system_prompt": getattr(agent, "ephemeral_system_prompt", None)
        or None,
        "providers_allowed": getattr(agent, "providers_allowed", None),
        "providers_ignored": getattr(agent, "providers_ignored", None),
        "providers_order": getattr(agent, "providers_order", None),
        "provider_sort": getattr(agent, "provider_sort", None),
        "provider_require_parameters": getattr(
            agent, "provider_require_parameters", False
        ),
        "provider_data_collection": getattr(agent, "provider_data_collection", None),
        "openrouter_min_coding_score": getattr(agent, "openrouter_min_coding_score", None),
        "session_id": task_id,
        "reasoning_config": getattr(agent, "reasoning_config", None)
        or _load_reasoning_config(),
        "service_tier": getattr(agent, "service_tier", None) or _load_service_tier(),
        "request_overrides": dict(getattr(agent, "request_overrides", {}) or {}),
        "platform": "tui",
        "session_db": _get_db(),
        "fallback_model": getattr(agent, "_fallback_model", None),
    }


def _ephemeral_preview_agent_kwargs(agent, task_id: str) -> dict:
    kwargs = _background_agent_kwargs(agent, task_id)
    kwargs.update(
        {
            "enabled_toolsets": ["terminal", "file"],
            "session_db": None,
            "skip_memory": True,
        }
    )
    return kwargs


def _preview_restart_history(session: dict, max_messages: int = 24, max_tool_chars: int = 1200) -> list[dict]:
    """Distill the parent session's recent history into a context the
    ephemeral preview-restart agent can actually use.

    The restart agent has no idea what app the user was building, what
    server they ran, what cwd was active, or which port belongs to which
    project. Without this, it would take the bare URL + console logs and
    guess — usually starting the wrong thing.

    We keep the last ``max_messages`` messages from the parent session so
    the restart agent sees recent user prompts, assistant replies, and
    most importantly any terminal/tool calls. Tool result payloads are
    truncated so we don't blow the context window with file dumps.
    """
    try:
        with session["history_lock"]:
            history = list(session.get("history", []) or [])
    except Exception:
        history = list(session.get("history", []) or [])

    if not history:
        return []

    # Anchor on the last user turn so we always include at least the most
    # recent request and the assistant/tool work that followed it. Then
    # extend backwards up to max_messages so we capture the prior context.
    last_user_idx = None
    for idx in range(len(history) - 1, -1, -1):
        if history[idx].get("role") == "user":
            last_user_idx = idx
            break

    start = max(0, len(history) - max_messages)
    if last_user_idx is not None:
        start = min(start, last_user_idx)

    trimmed: list[dict] = []
    for msg in history[start:]:
        if not isinstance(msg, dict):
            continue
        role = msg.get("role")
        if role not in ("user", "assistant", "tool", "system"):
            continue

        copy = {k: v for k, v in msg.items() if k != "reasoning"}
        # Truncate heavy tool outputs so a single 50KB file read doesn't
        # crowd out the rest of the context.
        if role == "tool":
            content = copy.get("content")
            if isinstance(content, str) and len(content) > max_tool_chars:
                copy["content"] = (
                    content[:max_tool_chars]
                    + f"\n... (truncated, original {len(content)} chars)"
                )
        trimmed.append(copy)

    return trimmed


def _preview_tool_result_preview(name: str, result: str) -> str:
    try:
        data = json.loads(result)
    except Exception:
        return ""

    if not isinstance(data, dict):
        return ""

    if name == "terminal":
        output = str(data.get("output") or "").strip()
        exit_code = data.get("exit_code")
        if output:
            return output[-1200:]
        if data.get("session_id"):
            return f"Background process started: {data.get('session_id')}"
        if exit_code is not None:
            return f"terminal exited with code {exit_code}"

    return str(data.get("error") or "").strip()[:1200]


def _preview_restart_callbacks(parent: str, task_id: str) -> dict:
    started_at: dict[str, float] = {}

    def progress(message: str, level: str = "info") -> None:
        text = str(message or "").strip()
        if text:
            _emit("preview.restart.progress", parent, {"task_id": task_id, "level": level, "text": text})

    def tool_start(tool_call_id: str, name: str, args: dict) -> None:
        started_at[tool_call_id] = time.time()
        ctx = _tool_ctx(name, args)
        progress(f"Running {name}{f': {ctx}' if ctx else ''}")

    def tool_complete(tool_call_id: str, name: str, _args: dict, result: str) -> None:
        duration_s = time.time() - started_at.get(tool_call_id, time.time())
        summary = _tool_summary(name, result, duration_s) or f"Finished {name}{f' in {_fmt_tool_duration(duration_s)}' if duration_s else ''}"
        output = _preview_tool_result_preview(name, result)
        progress(summary + (f"\n{output}" if output else ""))

    def tool_progress(event_type: str, name: str | None = None, preview: str | None = None, **_kwargs) -> None:
        if preview:
            progress(str(preview))
        elif name:
            progress(f"{event_type.replace('.', ' ')}: {name}")

    return {
        "tool_start_callback": tool_start,
        "tool_complete_callback": tool_complete,
        "tool_progress_callback": tool_progress,
        "tool_gen_callback": lambda name: progress(f"Preparing {name}"),
        "status_callback": lambda kind, text=None: progress(text if text is not None else kind),
    }


def _reset_session_agent(sid: str, session: dict) -> dict:
    tokens = _set_session_context(session["session_key"])
    try:
        new_agent = _make_agent(
            sid, session["session_key"], session_id=session["session_key"]
        )
    finally:
        _clear_session_context(tokens)
    session["agent"] = new_agent
    session["attached_images"] = []
    session["edit_snapshots"] = {}
    session["image_counter"] = 0
    session["running"] = False
    session["show_reasoning"] = _load_show_reasoning()
    session["tool_progress_mode"] = _load_tool_progress_mode()
    session["tool_started_at"] = {}
    with session["history_lock"]:
        session["history"] = []
        session["history_version"] = int(session.get("history_version", 0)) + 1
    info = _session_info(new_agent, session)
    _emit("session.info", sid, info)
    _restart_slash_worker(session)
    return info


def _make_agent(sid: str, key: str, session_id: str | None = None):
    from run_agent import AIAgent
    from hermes_cli.runtime_provider import resolve_runtime_provider

    # MCP tool discovery runs in a background daemon thread at startup so a
    # dead server can't freeze the shell (see tui_gateway/entry.py).  The agent
    # snapshots its tool list once here and never re-reads it, so briefly wait
    # for in-flight discovery to land before building — bounded, so a slow/dead
    # server still can't block.  No-op once discovery has finished (every build
    # after the first during a slow startup).
    try:
        from tui_gateway.entry import wait_for_mcp_discovery

        wait_for_mcp_discovery()
    except Exception:
        pass

    cfg = _load_cfg()
    agent_cfg = cfg.get("agent") or {}
    system_prompt = _prompt_text(agent_cfg.get("system_prompt", ""))
    startup_skills = _parse_tui_skills_env()
    if startup_skills:
        from agent.skill_commands import build_preloaded_skills_prompt

        skills_prompt, _loaded_skills, missing_skills = build_preloaded_skills_prompt(
            startup_skills,
            task_id=session_id or key,
        )
        if missing_skills:
            raise ValueError(f"Unknown skill(s): {', '.join(missing_skills)}")
        if skills_prompt:
            system_prompt = "\n\n".join(
                part for part in (system_prompt, skills_prompt) if part
            ).strip()
    model, requested_provider = _resolve_startup_runtime()
    runtime = resolve_runtime_provider(
        requested=requested_provider,
        target_model=model or None,
    )
    return AIAgent(
        model=model,
        max_iterations=_cfg_max_turns(cfg, 90),
        provider=runtime.get("provider"),
        base_url=runtime.get("base_url"),
        api_key=runtime.get("api_key"),
        api_mode=runtime.get("api_mode"),
        acp_command=runtime.get("command"),
        acp_args=runtime.get("args"),
        credential_pool=runtime.get("credential_pool"),
        quiet_mode=True,
        # verbose_logging controls DEBUG-level agent logging; it is intentionally
        # independent of tool_progress_mode (which only controls per-tool
        # display detail).  See cli.py PR (decoupling fix) for the matching
        # change on the classic CLI side.
        verbose_logging=False,
        reasoning_config=_load_reasoning_config(),
        service_tier=_load_service_tier(),
        enabled_toolsets=_load_enabled_toolsets(),
        platform="tui",
        session_id=session_id or key,
        session_db=_get_db(),
        ephemeral_system_prompt=system_prompt or None,
        checkpoints_enabled=is_truthy_value(os.environ.get("HERMES_TUI_CHECKPOINTS")),
        pass_session_id=is_truthy_value(os.environ.get("HERMES_TUI_PASS_SESSION_ID")),
        skip_context_files=is_truthy_value(os.environ.get("HERMES_IGNORE_RULES")),
        skip_memory=is_truthy_value(os.environ.get("HERMES_IGNORE_RULES")),
        **_agent_cbs(sid),
    )


def _init_session(sid: str, key: str, agent, history: list, cols: int = 80):
    now = time.time()
    _sessions[sid] = {
        "agent": agent,
        "session_key": key,
        "history": history,
        "history_lock": threading.Lock(),
        "history_version": 0,
        "inflight_turn": None,
        "created_at": now,
        "last_active": now,
        "running": False,
        "attached_images": [],
        "image_counter": 0,
        "cwd": _completion_cwd(),
        "cols": cols,
        "slash_worker": None,
        "show_reasoning": _load_show_reasoning(),
        "tool_progress_mode": _load_tool_progress_mode(),
        "edit_snapshots": {},
        "tool_started_at": {},
        # Pin async event emissions to whichever transport created the
        # session (stdio for Ink, JSON-RPC WS for the dashboard sidebar).
        "transport": current_transport() or _stdio_transport,
    }
    db = _get_db()
    if db is not None:
        row = db.get_session(key)
        if row and row.get("cwd"):
            _sessions[sid]["cwd"] = row["cwd"]
        else:
            try:
                db.update_session_cwd(key, _sessions[sid]["cwd"])
            except Exception:
                logger.debug("failed to persist resumed session cwd", exc_info=True)
    _register_session_cwd(_sessions[sid])
    try:
        _sessions[sid]["slash_worker"] = _SlashWorker(
            key, getattr(agent, "model", _resolve_model())
        )
    except Exception:
        # Defer hard-failure to slash.exec; chat still works without slash worker.
        _sessions[sid]["slash_worker"] = None
    try:
        from tools.approval import register_gateway_notify, load_permanent_allowlist

        register_gateway_notify(key, lambda data: _emit("approval.request", sid, data))
        load_permanent_allowlist()
    except Exception:
        pass
    # Surface the self-improvement background review's "💾 …" summary as a
    # review.summary event so Ink can render it as a persistent system line
    # in the transcript. In the CLI path this message is printed via
    # prompt_toolkit; the TUI has no equivalent print surface, so without
    # this callback the review would write the skill/memory change silently.
    try:
        agent.background_review_callback = lambda message, _sid=sid: _emit(
            "review.summary", _sid, {"text": str(message)}
        )
    except Exception:
        # Bare AIAgents that don't expose the attribute (unlikely, but keep
        # session startup resilient).
        pass
    _wire_callbacks(sid)
    _sessions[sid]["_notif_stop"] = _start_notification_poller(sid, _sessions[sid])
    _notify_session_boundary("on_session_reset", key)
    _emit("session.info", sid, _session_info(agent, _sessions[sid]))


def _new_session_key() -> str:
    return f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:6]}"


def _with_checkpoints(session, fn):
    return fn(session["agent"]._checkpoint_mgr, _session_cwd(session))


def _resolve_checkpoint_hash(mgr, cwd: str, ref: str) -> str:
    try:
        checkpoints = mgr.list_checkpoints(cwd)
        idx = int(ref) - 1
    except ValueError:
        return ref
    if 0 <= idx < len(checkpoints):
        return checkpoints[idx].get("hash", ref)
    raise ValueError(f"Invalid checkpoint number. Use 1-{len(checkpoints)}.")


def _enrich_with_attached_images(user_text: str, image_paths: list[str]) -> str:
    """Pre-analyze attached images via vision and prepend descriptions to user text."""
    import asyncio, json as _json
    from tools.vision_tools import vision_analyze_tool

    prompt = (
        "Describe everything visible in this image in thorough detail. "
        "Include any text, code, data, objects, people, layout, colors, "
        "and any other notable visual information."
    )

    parts: list[str] = []
    for path in image_paths:
        p = Path(path)
        if not p.exists():
            continue
        hint = f"[You can examine it with vision_analyze using image_url: {p}]"
        try:
            r = _json.loads(
                asyncio.run(vision_analyze_tool(image_url=str(p), user_prompt=prompt))
            )
            desc = r.get("analysis", "") if r.get("success") else None
            parts.append(
                f"[The user attached an image:\n{desc}]\n{hint}"
                if desc
                else f"[The user attached an image but analysis failed.]\n{hint}"
            )
        except Exception:
            parts.append(f"[The user attached an image but analysis failed.]\n{hint}")

    text = user_text or ""
    prefix = "\n\n".join(parts)
    if prefix:
        return f"{prefix}\n\n{text}" if text else prefix
    return text or "What do you see in this image?"


def _content_display_text(content: Any) -> str:
    if content is None:
        return ""
    if isinstance(content, str):
        return content
    if isinstance(content, (int, float)):
        return str(content)
    if isinstance(content, list):
        parts = []
        for part in content:
            text = _content_display_text(part).strip()
            if text:
                parts.append(text)
        return "\n".join(parts)
    if isinstance(content, dict):
        kind = content.get("type")
        if kind in {"text", "input_text", "output_text"}:
            return str(content.get("text") or content.get("content") or "")
        if kind in {"image_url", "input_image", "image"}:
            return "[image]"
        if kind in {"input_audio", "audio"}:
            return "[audio]"
        if kind:
            return f"[{kind}]"
        if "text" in content:
            return str(content.get("text") or "")
        return "[structured content]"
    return str(content)


def _coerce_message_text(content: Any) -> str:
    """Render ``message['content']`` as a plain string for transport.

    Provider-side, ``content`` may be a string (most common), a list of
    multimodal parts (e.g. ``[{"type": "text", "text": "..."},
    {"type": "image_url", "image_url": {...}}]``), or a single structured
    dict. Calling ``.strip()`` on a list raises ``'list' object has no
    attribute 'strip'`` and breaks session resume entirely.

    Image parts (``image_url``) are preserved by appending the underlying
    URL (data: or http:) into the text. The desktop renderer pulls these
    back out via ``extractEmbeddedImages`` so the user sees the image
    instead of the URL — and it stops the resume payload from disagreeing
    with the cached message (which would otherwise cause the inline image
    to flash, then disappear when the resume payload overwrites the cache).

    Other structured dict shapes (audio, unknown types) fall back to a
    bracketed placeholder so resume doesn't drop the message entirely.
    """
    if content is None:
        return ""
    if isinstance(content, str):
        return content
    if isinstance(content, (int, float)):
        return str(content)
    if isinstance(content, list):
        chunks: list[str] = []
        for part in content:
            if isinstance(part, str):
                chunks.append(part)
                continue
            if not isinstance(part, dict):
                continue
            text = part.get("text")
            if isinstance(text, str):
                chunks.append(text)
                continue
            kind = part.get("type")
            if kind in {"text", "input_text", "output_text"}:
                t = part.get("text") or part.get("content") or ""
                if t:
                    chunks.append(str(t))
                continue
            if kind in {"image_url", "input_image", "image"}:
                image_url = part.get("image_url")
                url = ""
                if isinstance(image_url, dict):
                    candidate = image_url.get("url")
                    if isinstance(candidate, str):
                        url = candidate
                elif isinstance(image_url, str):
                    url = image_url
                if url:
                    chunks.append(f"\n{url}")
                else:
                    chunks.append("\n[image]")
                continue
            if kind in {"input_audio", "audio"}:
                chunks.append("\n[audio]")
                continue
            if kind:
                chunks.append(f"\n[{kind}]")
        return "".join(chunks)
    if isinstance(content, dict):
        kind = content.get("type")
        if kind in {"text", "input_text", "output_text"}:
            return str(content.get("text") or content.get("content") or "")
        if kind in {"image_url", "input_image", "image"}:
            image_url = content.get("image_url")
            url = ""
            if isinstance(image_url, dict):
                candidate = image_url.get("url")
                if isinstance(candidate, str):
                    url = candidate
            elif isinstance(image_url, str):
                url = image_url
            return url or "[image]"
        if kind in {"input_audio", "audio"}:
            return "[audio]"
        if kind:
            return f"[{kind}]"
        if "text" in content:
            return str(content.get("text") or "")
        return "[structured content]"
    return str(content)


def _history_to_messages(history: list[dict]) -> list[dict]:
    messages = []
    tool_call_args = {}

    for m in history:
        if not isinstance(m, dict):
            continue
        role = m.get("role")
        if role not in {"user", "assistant", "tool", "system"}:
            continue
        content_text = _coerce_message_text(m.get("content"))
        if role == "assistant" and m.get("tool_calls"):
            for tc in m["tool_calls"]:
                fn = tc.get("function", {})
                tc_id = tc.get("id", "")
                if tc_id and fn.get("name"):
                    try:
                        args = json.loads(fn.get("arguments", "{}"))
                    except (json.JSONDecodeError, TypeError):
                        args = {}
                    tool_call_args[tc_id] = (fn["name"], args)
            if not content_text.strip():
                continue
        if role == "tool":
            tc_id = m.get("tool_call_id", "")
            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)}
            )
            continue
        if not content_text.strip():
            continue
        msg = {"role": role, "text": content_text}
        if role == "assistant":
            for key in (
                "reasoning",
                "reasoning_content",
                "reasoning_details",
                "codex_reasoning_items",
            ):
                if key in m and m.get(key) is not None:
                    msg[key] = m.get(key)
        messages.append(msg)

    return messages


def _coerce_seed_history(value: Any) -> list[dict]:
    if not isinstance(value, list):
        return []

    history = []
    for item in value:
        if not isinstance(item, dict):
            continue

        role = item.get("role")
        if role not in ("user", "assistant", "system"):
            continue

        content = item.get("content")
        if content is None:
            content = item.get("text")
        if not isinstance(content, str) or not content.strip():
            continue

        history.append({"role": role, "content": content})

    return history


def _content_display_text(content: Any) -> str:
    if content is None:
        return ""
    if isinstance(content, str):
        return content
    if isinstance(content, (int, float)):
        return str(content)
    if isinstance(content, list):
        parts = []
        for part in content:
            text = _content_display_text(part).strip()
            if text:
                parts.append(text)
        return "\n".join(parts)
    if isinstance(content, dict):
        kind = content.get("type")
        if kind in {"text", "input_text", "output_text"}:
            return str(content.get("text") or content.get("content") or "")
        if kind in {"image_url", "input_image", "image"}:
            return "[image]"
        if kind in {"input_audio", "audio"}:
            return "[audio]"
        if kind:
            return f"[{kind}]"
        if "text" in content:
            return str(content.get("text") or "")
        return "[structured content]"
    return str(content)


def _inflight_text(value: Any) -> str:
    return _content_display_text(value).strip()


def _start_inflight_turn(session: dict, text: Any) -> None:
    now = time.time()
    session["inflight_turn"] = {
        "assistant": "",
        "started_at": now,
        "streaming": True,
        "updated_at": now,
        "user": _inflight_text(text),
    }


def _append_inflight_delta(session: dict, delta: Any) -> None:
    text = "" if delta is None else str(delta)
    if not text:
        return
    turn = session.get("inflight_turn")
    if not isinstance(turn, dict):
        turn = {"assistant": "", "streaming": True, "user": ""}
    turn["assistant"] = f"{turn.get('assistant') or ''}{text}"
    turn["streaming"] = True
    turn["updated_at"] = time.time()
    session["inflight_turn"] = turn


def _clear_inflight_turn(session: dict) -> None:
    session["inflight_turn"] = None


def _inflight_snapshot(session: dict) -> dict | None:
    turn = session.get("inflight_turn")
    if not isinstance(turn, dict):
        return None
    user = str(turn.get("user") or "").strip()
    assistant = str(turn.get("assistant") or "")
    streaming = bool(turn.get("streaming"))
    if not user and not assistant and not streaming:
        return None
    return {
        "assistant": assistant,
        "streaming": streaming,
        "user": user,
    }


# ── Methods: session ─────────────────────────────────────────────────


@method("session.create")
def _(rid, params: dict) -> dict:
    sid = uuid.uuid4().hex[:8]
    key = _new_session_key()
    cols = int(params.get("cols", 80))
    history = _coerce_seed_history(params.get("messages"))
    title = str(params.get("title") or "").strip()
    _enable_gateway_prompts()

    ready = threading.Event()
    now = time.time()

    _sessions[sid] = {
        "agent": None,
        "agent_error": None,
        "agent_ready": ready,
        "attached_images": [],
        "cols": cols,
        "created_at": now,
        "edit_snapshots": {},
        "history": history,
        "history_lock": threading.Lock(),
        "history_version": 0,
        "image_counter": 0,
        "cwd": _completion_cwd(params),
        "inflight_turn": None,
        "last_active": now,
        "pending_title": title or None,
        "running": False,
        "session_key": key,
        "show_reasoning": _load_show_reasoning(),
        "slash_worker": None,
        "tool_progress_mode": _load_tool_progress_mode(),
        "tool_started_at": {},
        "transport": current_transport() or _stdio_transport,
    }
    _register_session_cwd(_sessions[sid])
    db = _get_db()
    if db is not None:
        try:
            db.create_session(
                key,
                source="tui",
                model=_resolve_model(),
                cwd=_sessions[sid]["cwd"],
            )
        except Exception:
            logger.debug("failed to pre-create desktop session row", exc_info=True)

    # Return the lightweight session immediately so Ink can paint the composer
    # + skeleton panel, then build the real AIAgent just after this response is
    # flushed.  This keeps startup responsive while still hydrating tools/skills
    # without requiring the user to submit a first prompt.
    def _deferred_build() -> None:
        session = _sessions.get(sid)
        if session is not None:
            _start_agent_build(sid, session)

    build_timer = threading.Timer(0.05, _deferred_build)
    build_timer.daemon = True
    build_timer.start()

    return _ok(
        rid,
        {
            "session_id": sid,
            "stored_session_id": key,
            "message_count": len(history),
            "messages": _history_to_messages(history),
            "info": {
                "model": _resolve_model(),
                "tools": {},
                "skills": {},
                "cwd": _sessions[sid]["cwd"],
                "branch": _git_branch_for_cwd(_sessions[sid]["cwd"]),
                "lazy": True,
                "profile_name": _current_profile_name(),
            },
        },
    )


@method("session.list")
def _(rid, params: dict) -> dict:
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5006)
    try:
        # Resume picker should surface human conversation sessions from every
        # user-facing surface — CLI, TUI, all gateway platforms (including new
        # ones not enumerated here), ACP adapter clients, webhook sessions,
        # custom `HERMES_SESSION_SOURCE` values, and older installs with
        # different source labels. We deny-list only the noisy internal
        # sources (``tool`` sub-agent runs) rather than allow-listing a
        # fixed set of platform names that goes stale whenever a new
        # 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]
        return _ok(
            rid,
            {
                "sessions": [
                    {
                        "id": s["id"],
                        "title": s.get("title") or "",
                        "preview": s.get("preview") or "",
                        "started_at": s.get("started_at") or 0,
                        "message_count": s.get("message_count") or 0,
                        "source": s.get("source") or "",
                    }
                    for s in rows
                ]
            },
        )
    except Exception as e:
        return _err(rid, 5006, str(e))


@method("session.most_recent")
def _(rid, params: dict) -> dict:
    """Return the most recent human-facing session id, or ``None``.

    Mirrors ``session.list``'s deny-list behaviour (drops ``tool``
    sub-agent rows).  Used by TUI auto-resume when
    ``display.tui_auto_resume_recent`` is on; the field is also handy
    for any CLI tooling that wants "latest session" without paginating
    the full list.

    Contract: a ``{"session_id": null}`` result means "no eligible
    session found right now".  Errors are also folded into that
    null-result shape (and logged) so callers don't have to special-
    case JSON-RPC error envelopes for what is a normal "no answer".
    """
    db = _get_db()
    if db is None:
        return _ok(rid, {"session_id": None})
    try:
        deny = frozenset({"tool"})
        # Over-fetch by a generous bounded amount so heavy sub-agent
        # users (lots of recent ``tool`` rows) don't get a false
        # "no eligible session" answer.  ``session.list`` uses a
        # similar over-fetch strategy.
        rows = db.list_sessions_rich(source=None, limit=200)
        for row in rows:
            src = (row.get("source") or "").strip().lower()
            if src in deny:
                continue
            return _ok(
                rid,
                {
                    "session_id": row.get("id"),
                    "title": row.get("title") or "",
                    "started_at": row.get("started_at") or 0,
                    "source": row.get("source") or "",
                },
            )
        return _ok(rid, {"session_id": None})
    except Exception:
        logger.exception("session.most_recent failed")
        return _ok(rid, {"session_id": None})


@method("session.resume")
def _(rid, params: dict) -> dict:
    target = params.get("session_id", "")
    if not target:
        return _err(rid, 4006, "session_id required")
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5000)
    found = db.get_session(target)
    if not found:
        found = db.get_session_by_title(target)
        if found:
            target = found["id"]
        else:
            return _err(rid, 4007, "session not found")
    sid = uuid.uuid4().hex[:8]
    _enable_gateway_prompts()
    try:
        db.reopen_session(target)
        history = db.get_messages_as_conversation(target)
        display_history = db.get_messages_as_conversation(
            target, include_ancestors=True
        )
        messages = _history_to_messages(display_history)
        tokens = _set_session_context(target)
        try:
            agent = _make_agent(sid, target, session_id=target)
        finally:
            _clear_session_context(tokens)
        _init_session(sid, target, agent, history, cols=int(params.get("cols", 80)))
    except Exception as e:
        return _err(rid, 5000, f"resume failed: {e}")
    return _ok(
        rid,
        {
            "session_id": sid,
            "resumed": target,
            "message_count": len(messages),
            "messages": messages,
            "info": _session_info(agent, _sessions.get(sid)),
        },
    )


@method("session.cwd.set")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    if session.get("running"):
        return _err(rid, 4009, "session busy")
    raw = str(params.get("cwd", "") or "").strip()
    if not raw:
        return _err(rid, 4016, "cwd required")
    try:
        cwd = _set_session_cwd(session, raw)
    except ValueError as e:
        return _err(rid, 4017, str(e))
    agent = session.get("agent")
    info = _session_info(agent, session) if agent is not None else {
        "cwd": cwd,
        "branch": _git_branch_for_cwd(cwd),
        "lazy": True,
    }
    _emit("session.info", params.get("session_id", ""), info)
    return _ok(rid, info)


def _session_pending_kind(sid: str) -> str:
    for rid, (owner_sid, _ev) in list(_pending.items()):
        if owner_sid != sid:
            continue
        event, _payload = _pending_prompt_payloads.get(rid, ("input.request", {}))
        return str(event).removesuffix(".request")
    return ""


def _session_live_status(sid: str, session: dict) -> str:
    if _session_pending_kind(sid):
        return "waiting"
    ready = session.get("agent_ready")
    if ready is not None and not ready.is_set():
        return "starting"
    if session.get("running"):
        return "working"
    return "idle"


def _message_preview(history: list) -> str:
    for msg in reversed(history or []):
        text = _content_display_text(msg.get("content", msg.get("text", ""))).strip()
        if text:
            return " ".join(text.split())[:160]
    return ""


def _session_live_title(session: dict, key: str) -> str:
    title = str(session.get("pending_title") or "").strip()
    db = _get_db()
    if db is not None:
        try:
            title = str(db.get_session_title(key) or title or "").strip()
        except Exception:
            pass
    return title


def _session_live_item(sid: str, session: dict, current_sid: str = "") -> dict:
    key = str(session.get("session_key") or sid)
    agent = session.get("agent")
    history = list(session.get("history") or [])
    status = _session_live_status(sid, session)
    inflight = _inflight_snapshot(session)
    preview = _message_preview(history)
    if inflight:
        preview = inflight.get("assistant") or inflight.get("user") or preview
        preview = " ".join(str(preview).split())[:160]
    now = time.time()
    return {
        "current": sid == current_sid,
        "id": sid,
        "last_active": float(session.get("last_active") or session.get("created_at") or now),
        "message_count": len(history),
        "model": str(getattr(agent, "model", "") or _resolve_model()),
        "preview": preview,
        "session_key": key,
        "started_at": float(session.get("created_at") or now),
        "status": status,
        "title": _session_live_title(session, key),
    }


def _fallback_session_info(session: dict) -> dict:
    agent = session.get("agent")
    if agent is not None:
        return _session_info(agent)
    return {
        "cwd": os.getenv("TERMINAL_CWD", os.getcwd()),
        "lazy": True,
        "model": _resolve_model(),
        "skills": {},
        "tools": {},
    }


@method("session.active_list")
def _(rid, params: dict) -> dict:
    """Return live TUI sessions in this gateway process.

    Unlike ``session.list`` this is not a historical DB browser: it reports only
    sessions with in-memory agents/workers that the current TUI can switch to
    without closing siblings.
    """
    current = str(params.get("current_session_id") or "")
    try:
        snapshot = list(_sessions.items())
    except Exception as e:
        return _err(rid, 5036, f"could not enumerate active sessions: {e}")

    # Keep the natural creation/insertion order from ``_sessions``.  The
    # frontend marks the focused session with ``current``; it should not jump to
    # the top just because the user switched to it.
    rows = [_session_live_item(sid, session, current) for sid, session in snapshot]
    return _ok(rid, {"sessions": rows})


@method("session.activate")
def _(rid, params: dict) -> dict:
    """Attach the frontend to an already-live TUI session.

    This intentionally does not close the previously focused session; it merely
    returns enough state for Ink to redraw around another live session id.
    """
    sid = str(params.get("session_id") or "")
    session, err = _sess_nowait({"session_id": sid}, rid)
    if err:
        return err

    with session["history_lock"]:
        session["last_active"] = time.time()
        history = list(session.get("display_history") or session.get("history") or [])
        inflight = _inflight_snapshot(session)
        running = bool(session.get("running"))
    status = _session_live_status(sid, session)
    payload = {
        "info": _fallback_session_info(session),
        "message_count": len(history),
        "messages": _history_to_messages(history),
        "running": running,
        "session_id": sid,
        "session_key": session.get("session_key") or sid,
        "started_at": float(session.get("created_at") or time.time()),
        "status": status,
    }
    if inflight:
        payload["inflight"] = inflight
    return _ok(
        rid,
        payload,
    )


@method("session.delete")
def _(rid, params: dict) -> dict:
    """Delete a stored session and its on-disk transcript files.

    Used by the TUI resume picker (``d`` key) so users can prune old
    sessions without dropping to the CLI.  Refuses to delete a session
    that is currently active in this gateway process — those rows are
    still being written to and removing them out from under the live
    agent corrupts message ordering and trips FK constraints when the
    next message append flushes.
    """
    target = params.get("session_id", "")
    if not target:
        return _err(rid, 4006, "session_id required")
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5036)
    # Block deletion of any session currently bound to a live TUI session
    # in this process.  The picker hides the active session anyway, but a
    # racing caller could still target it.  Snapshot via ``list(...)``
    # because ``_sessions`` is mutated by concurrent RPCs on the thread
    # pool — iterating the dict directly can raise ``RuntimeError:
    # dictionary changed size during iteration``.  If even the snapshot
    # raises, fail closed (refuse the delete) rather than fail open.
    try:
        snapshot = list(_sessions.values())
    except Exception as e:
        return _err(rid, 5036, f"could not enumerate active sessions: {e}")
    active = {s.get("session_key") for s in snapshot if s.get("session_key")}
    if target in active:
        return _err(rid, 4023, "cannot delete an active session")
    sessions_dir = get_hermes_home() / "sessions"
    try:
        deleted = db.delete_session(target, sessions_dir=sessions_dir)
    except Exception as e:
        return _err(rid, 5036, f"delete failed: {e}")
    if not deleted:
        return _err(rid, 4007, "session not found")
    return _ok(rid, {"deleted": target})


@method("session.title")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5007)
    key = session["session_key"]
    if "title" not in params:
        fallback = session.get("pending_title") or ""
        try:
            resolved_title = db.get_session_title(key) or ""
            if fallback:
                if db.set_session_title(key, fallback):
                    session["pending_title"] = None
                    resolved_title = fallback
                else:
                    existing_row = db.get_session(key)
                    existing_title = ((existing_row or {}).get("title") or "").strip()
                    if existing_title == fallback:
                        session["pending_title"] = None
                        resolved_title = fallback
                    elif not resolved_title:
                        resolved_title = fallback
            elif resolved_title:
                session["pending_title"] = None
        except Exception:
            resolved_title = fallback
        return _ok(
            rid,
            {
                "title": resolved_title,
                "session_key": key,
            },
        )
    title = (params.get("title", "") or "").strip()
    if not title:
        return _err(rid, 4021, "title required")
    try:
        if db.set_session_title(key, title):
            session["pending_title"] = None
            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
            return _ok(
                rid,
                {
                    "pending": False,
                    "title": (existing_row.get("title") or title),
                },
            )
        session["pending_title"] = title
        return _ok(rid, {"pending": True, "title": title})
    except ValueError as e:
        return _err(rid, 4022, str(e))
    except Exception as e:
        return _err(rid, 5007, str(e))


@method("session.usage")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    agent = session.get("agent")
    return _ok(
        rid,
        (
            _get_usage(agent)
            if agent is not None
            else {"calls": 0, "input": 0, "output": 0, "total": 0}
        ),
    )


@method("session.status")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err

    from hermes_constants import display_hermes_home

    key = session.get("session_key") or params.get("session_id") or ""
    agent = session.get("agent")
    meta = {}
    db = _get_db()
    if db and key:
        try:
            meta = db.get_session(key) or {}
        except Exception:
            meta = {}

    def _dt(value, fallback: datetime | None = None) -> datetime:
        if value:
            try:
                return datetime.fromtimestamp(float(value))
            except Exception:
                pass
        return fallback or datetime.now()

    created = _dt(meta.get("started_at"))
    updated = created
    for field in ("updated_at", "last_updated_at", "last_activity_at"):
        if meta.get(field):
            updated = _dt(meta.get(field), created)
            break

    usage = _get_usage(agent) if agent is not None else {}
    provider = getattr(agent, "provider", None) or "unknown"
    model = getattr(agent, "model", None) or "(unknown)"
    lines = [
        "Hermes TUI Status",
        "",
        f"Session ID: {key}",
        f"Path: {display_hermes_home()}",
    ]
    title = (meta.get("title") or "").strip()
    if title:
        lines.append(f"Title: {title}")
    lines.extend(
        [
            f"Model: {model} ({provider})",
            f"Created: {created.strftime('%Y-%m-%d %H:%M')}",
            f"Last Activity: {updated.strftime('%Y-%m-%d %H:%M')}",
            f"Tokens: {int(usage.get('total') or 0):,}",
            f"Agent Running: {'Yes' if session.get('running') else 'No'}",
        ]
    )
    return _ok(rid, {"output": "\n".join(lines)})


@method("session.history")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    history = list(session.get("history", []))
    db = _get_db()
    if db is not None and session.get("session_key"):
        try:
            history = db.get_messages_as_conversation(
                session["session_key"], include_ancestors=True
            )
        except Exception:
            pass
    return _ok(
        rid,
        {
            "count": len(history),
            "messages": _history_to_messages(history),
        },
    )


@method("session.undo")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    # Reject during an in-flight turn.  If we mutated history while
    # the agent thread is running, prompt.submit's post-run history
    # write would either clobber the undo (version matches) or
    # silently drop the agent's output (version mismatch, see below).
    # Neither is what the user wants — make them /interrupt first.
    if session.get("running"):
        return _err(
            rid, 4009, "session busy — /interrupt the current turn before /undo"
        )
    removed = 0
    with session["history_lock"]:
        history = session.get("history", [])
        while history and history[-1].get("role") in {"assistant", "tool"}:
            history.pop()
            removed += 1
        if history and history[-1].get("role") == "user":
            history.pop()
            removed += 1
        if removed:
            session["history_version"] = int(session.get("history_version", 0)) + 1
    return _ok(rid, {"removed": removed})


@method("session.compress")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    if session.get("running"):
        return _err(
            rid, 4009, "session busy — /interrupt the current turn before /compress"
        )
    sid = params.get("session_id", "")
    focus_topic = str(params.get("focus_topic", "") or "").strip()
    try:
        from agent.manual_compression_feedback import summarize_manual_compression
        from agent.model_metadata import estimate_request_tokens_rough

        with session["history_lock"]:
            before_messages = list(session.get("history", []))
            history_version = int(session.get("history_version", 0))
        before_count = len(before_messages)
        _agent = session["agent"]
        _sys_prompt = getattr(_agent, "_cached_system_prompt", "") or ""
        _tools = getattr(_agent, "tools", None) or None
        before_tokens = (
            estimate_request_tokens_rough(
                before_messages, system_prompt=_sys_prompt, tools=_tools
            )
            if before_count
            else 0
        )

        if before_count >= 4:
            focus_suffix = f', focus: "{focus_topic}"' if focus_topic else ""
            _status_update(
                sid,
                "compressing",
                f"⠋ compressing {before_count} messages "
                f"(~{before_tokens:,} tok){focus_suffix}…",
            )

        try:
            removed, usage = _compress_session_history(
                session,
                focus_topic,
                approx_tokens=before_tokens,
                before_messages=before_messages,
                history_version=history_version,
            )
            with session["history_lock"]:
                messages = list(session.get("history", []))
            after_count = len(messages)
            # Re-read system prompt + tools after compression — _compress_context
            # may have rebuilt the system prompt (_cached_system_prompt=None).
            _sys_prompt_after = (
                getattr(_agent, "_cached_system_prompt", "") or _sys_prompt
            )
            _tools_after = getattr(_agent, "tools", None) or _tools
            after_tokens = (
                estimate_request_tokens_rough(
                    messages,
                    system_prompt=_sys_prompt_after,
                    tools=_tools_after,
                )
                if after_count
                else 0
            )
            agent = session["agent"]
            _sync_session_key_after_compress(sid, session)
            summary = summarize_manual_compression(
                before_messages, messages, before_tokens, after_tokens
            )
            info = _session_info(agent, session)
            _emit("session.info", sid, info)
            return _ok(
                rid,
                {
                    "status": "compressed",
                    "removed": removed,
                    "before_messages": before_count,
                    "after_messages": after_count,
                    "before_tokens": before_tokens,
                    "after_tokens": after_tokens,
                    "summary": summary,
                    "usage": usage,
                    "info": info,
                    "messages": messages,
                },
            )
        finally:
            # Always clear the pinned compressing status so the bar
            # reverts to neutral whether compaction succeeded, was a
            # no-op, or raised.
            _status_update(sid, "ready")
    except Exception as e:
        return _err(rid, 5005, str(e))


@method("session.save")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    import time as _time

    filename = os.path.abspath(
        f"hermes_conversation_{_time.strftime('%Y%m%d_%H%M%S')}.json"
    )
    try:
        with open(filename, "w", encoding="utf-8") as f:
            json.dump(
                {
                    "model": getattr(session["agent"], "model", ""),
                    "messages": session.get("history", []),
                },
                f,
                indent=2,
                ensure_ascii=False,
            )
        return _ok(rid, {"file": filename})
    except Exception as e:
        return _err(rid, 5011, str(e))


@method("session.close")
def _(rid, params: dict) -> dict:
    sid = params.get("session_id", "")
    session = _sessions.pop(sid, None)
    if not session:
        return _ok(rid, {"closed": False})
    _finalize_session(session)
    try:
        from tools.approval import unregister_gateway_notify

        unregister_gateway_notify(session["session_key"])
    except Exception:
        pass
    try:
        agent = session.get("agent")
        if agent and hasattr(agent, "close"):
            agent.close()
    except Exception:
        pass
    try:
        worker = session.get("slash_worker")
        if worker:
            worker.close()
    except Exception:
        pass
    return _ok(rid, {"closed": True})


@method("session.branch")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5008)
    old_key = session["session_key"]
    with session["history_lock"]:
        history = [dict(msg) for msg in session.get("history", [])]
    if not history:
        return _err(rid, 4008, "nothing to branch — send a message first")
    new_key = _new_session_key()
    branch_name = params.get("name", "")
    try:
        if branch_name:
            title = branch_name
        else:
            current = db.get_session_title(old_key) or "branch"
            title = (
                db.get_next_title_in_lineage(current)
                if hasattr(db, "get_next_title_in_lineage")
                else f"{current} (branch)"
            )
        db.create_session(
            new_key,
            source="tui",
            model=_resolve_model(),
            parent_session_id=old_key,
            cwd=_session_cwd(session),
        )
        for msg in history:
            db.append_message(
                session_id=new_key,
                role=msg.get("role", "user"),
                content=msg.get("content"),
            )
        db.set_session_title(new_key, title)
    except Exception as e:
        return _err(rid, 5008, f"branch failed: {e}")
    new_sid = uuid.uuid4().hex[:8]
    try:
        tokens = _set_session_context(new_key)
        try:
            agent = _make_agent(new_sid, new_key, session_id=new_key)
        finally:
            _clear_session_context(tokens)
        _init_session(
            new_sid, new_key, agent, list(history), cols=session.get("cols", 80)
        )
    except Exception as e:
        return _err(rid, 5000, f"agent init failed on branch: {e}")
    return _ok(rid, {"session_id": new_sid, "title": title, "parent": old_key})


@method("session.interrupt")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    if hasattr(session["agent"], "interrupt"):
        session["agent"].interrupt()
    # Scope the pending-prompt release to THIS session.  A global
    # _clear_pending() would collaterally cancel clarify/sudo/secret
    # prompts on unrelated sessions sharing the same tui_gateway
    # process, silently resolving them to empty strings.
    _clear_pending(params.get("session_id", ""))
    try:
        from tools.approval import resolve_gateway_approval

        resolve_gateway_approval(session["session_key"], "deny", resolve_all=True)
    except Exception:
        pass
    return _ok(rid, {"status": "interrupted"})


# ── Delegation: subagent tree observability + controls ───────────────
# Powers the TUI's /agents overlay (see ui-tui/src/components/agentsOverlay).
# The registry lives in tools/delegate_tool — these handlers are thin
# translators between JSON-RPC and the Python API.


@method("delegation.status")
def _(rid, params: dict) -> dict:
    from tools.delegate_tool import (
        is_spawn_paused,
        list_active_subagents,
        _get_max_concurrent_children,
        _get_max_spawn_depth,
    )

    return _ok(
        rid,
        {
            "active": list_active_subagents(),
            "paused": is_spawn_paused(),
            "max_spawn_depth": _get_max_spawn_depth(),
            "max_concurrent_children": _get_max_concurrent_children(),
        },
    )


@method("delegation.pause")
def _(rid, params: dict) -> dict:
    from tools.delegate_tool import set_spawn_paused

    paused = bool(params.get("paused", True))
    return _ok(rid, {"paused": set_spawn_paused(paused)})


@method("subagent.interrupt")
def _(rid, params: dict) -> dict:
    from tools.delegate_tool import interrupt_subagent

    subagent_id = str(params.get("subagent_id") or "").strip()
    if not subagent_id:
        return _err(rid, 4000, "subagent_id required")
    ok = interrupt_subagent(subagent_id)
    return _ok(rid, {"found": ok, "subagent_id": subagent_id})


# ── Spawn-tree snapshots: TUI-written, disk-persisted ────────────────
# The TUI is the source of truth for subagent state (it assembles payloads
# from the event stream).  On turn-complete it posts the final tree here;
# /replay and /replay-diff fetch past snapshots by session_id + filename.
#
# Layout:  $HERMES_HOME/spawn-trees/<session_id>/<timestamp>.json
# Each file contains { session_id, started_at, finished_at, subagents: [...] }.


def _spawn_trees_root():
    from hermes_constants import get_hermes_home

    root = get_hermes_home() / "spawn-trees"
    root.mkdir(parents=True, exist_ok=True)
    return root


def _spawn_tree_session_dir(session_id: str):
    safe = (
        "".join(c if c.isalnum() or c in "-_" else "_" for c in session_id) or "unknown"
    )
    d = _spawn_trees_root() / safe
    d.mkdir(parents=True, exist_ok=True)
    return d


# Per-session append-only index of lightweight snapshot metadata.  Read by
# `spawn_tree.list` so scanning doesn't require reading every full snapshot
# file (Copilot review on #14045).  One JSON object per line.
_SPAWN_TREE_INDEX = "_index.jsonl"


def _append_spawn_tree_index(session_dir, entry: dict) -> None:
    try:
        with (session_dir / _SPAWN_TREE_INDEX).open("a", encoding="utf-8") as f:
            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
    except OSError as exc:
        # Index is a cache — losing a line just means list() falls back
        # to a directory scan for that entry.  Never block the save.
        logger.debug("spawn_tree index append failed: %s", exc)


def _read_spawn_tree_index(session_dir) -> list[dict]:
    index_path = session_dir / _SPAWN_TREE_INDEX
    if not index_path.exists():
        return []
    out: list[dict] = []
    try:
        with index_path.open("r", encoding="utf-8") as f:
            for line in f:
                line = line.strip()
                if not line:
                    continue
                try:
                    out.append(json.loads(line))
                except json.JSONDecodeError:
                    continue
    except OSError:
        return []
    return out


@method("spawn_tree.save")
def _(rid, params: dict) -> dict:
    session_id = str(params.get("session_id") or "").strip()
    subagents = params.get("subagents") or []
    if not isinstance(subagents, list) or not subagents:
        return _err(rid, 4000, "subagents list required")

    from datetime import datetime

    started_at = params.get("started_at")
    finished_at = params.get("finished_at") or time.time()
    label = str(params.get("label") or "")
    ts = datetime.utcfromtimestamp(float(finished_at)).strftime("%Y%m%dT%H%M%S")
    fname = f"{ts}.json"
    d = _spawn_tree_session_dir(session_id or "default")
    path = d / fname
    try:
        payload = {
            "session_id": session_id,
            "started_at": float(started_at) if started_at else None,
            "finished_at": float(finished_at),
            "label": label,
            "subagents": subagents,
        }
        path.write_text(json.dumps(payload, ensure_ascii=False), encoding="utf-8")
    except OSError as exc:
        return _err(rid, 5000, f"spawn_tree.save failed: {exc}")

    _append_spawn_tree_index(
        d,
        {
            "path": str(path),
            "session_id": session_id,
            "started_at": payload["started_at"],
            "finished_at": payload["finished_at"],
            "label": label,
            "count": len(subagents),
        },
    )

    return _ok(rid, {"path": str(path), "session_id": session_id})


@method("spawn_tree.list")
def _(rid, params: dict) -> dict:
    session_id = str(params.get("session_id") or "").strip()
    limit = int(params.get("limit") or 50)
    cross_session = bool(params.get("cross_session"))

    if cross_session:
        root = _spawn_trees_root()
        roots = [p for p in root.iterdir() if p.is_dir()]
    else:
        roots = [_spawn_tree_session_dir(session_id or "default")]

    entries: list[dict] = []
    for d in roots:
        indexed = _read_spawn_tree_index(d)
        if indexed:
            # Skip index entries whose snapshot file was manually deleted.
            entries.extend(
                e for e in indexed if (p := e.get("path")) and Path(p).exists()
            )
            continue

        # Fallback for legacy (pre-index) sessions: full scan.  O(N) reads
        # but only runs once per session until the next save writes the index.
        for p in d.glob("*.json"):
            if p.name == _SPAWN_TREE_INDEX:
                continue
            try:
                stat = p.stat()
                try:
                    raw = json.loads(p.read_text(encoding="utf-8"))
                except Exception:
                    raw = {}
                subagents = raw.get("subagents") or []
                entries.append(
                    {
                        "path": str(p),
                        "session_id": raw.get("session_id") or d.name,
                        "finished_at": raw.get("finished_at") or stat.st_mtime,
                        "started_at": raw.get("started_at"),
                        "label": raw.get("label") or "",
                        "count": len(subagents) if isinstance(subagents, list) else 0,
                    }
                )
            except OSError:
                continue

    entries.sort(key=lambda e: e.get("finished_at") or 0, reverse=True)
    return _ok(rid, {"entries": entries[:limit]})


@method("spawn_tree.load")
def _(rid, params: dict) -> dict:
    from pathlib import Path

    raw_path = str(params.get("path") or "").strip()
    if not raw_path:
        return _err(rid, 4000, "path required")

    # Reject paths escaping the spawn-trees root.
    root = _spawn_trees_root().resolve()
    try:
        resolved = Path(raw_path).resolve()
        resolved.relative_to(root)
    except (ValueError, OSError) as exc:
        return _err(rid, 4030, f"path outside spawn-trees root: {exc}")

    try:
        payload = json.loads(resolved.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError) as exc:
        return _err(rid, 5000, f"spawn_tree.load failed: {exc}")

    return _ok(rid, payload)


@method("session.steer")
def _(rid, params: dict) -> dict:
    """Inject a user message into the next tool result without interrupting.

    Mirrors AIAgent.steer(). Safe to call while a turn is running — the text
    lands on the last tool result of the next tool batch and the model sees
    it on its next iteration. No interrupt, no new user turn, no role
    alternation violation.
    """
    text = (params.get("text") or "").strip()
    if not text:
        return _err(rid, 4002, "text is required")
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    agent = session.get("agent")
    if agent is None or not hasattr(agent, "steer"):
        return _err(rid, 4010, "agent does not support steer")
    try:
        accepted = agent.steer(text)
    except Exception as exc:
        return _err(rid, 5000, f"steer failed: {exc}")
    return _ok(rid, {"status": "queued" if accepted else "rejected", "text": text})


@method("terminal.resize")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    session["cols"] = int(params.get("cols", 80))
    return _ok(rid, {"cols": session["cols"]})


# ── Methods: prompt ──────────────────────────────────────────────────


@method("prompt.submit")
def _(rid, params: dict) -> dict:
    sid, text = params.get("session_id", ""), params.get("text", "")
    truncate_user_ordinal = params.get("truncate_before_user_ordinal")
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    # Re-bind to the current client transport for this request. This keeps
    # streaming events on the active websocket even if an earlier disconnect
    # or fallback moved the session transport to stdio.
    if (t := current_transport()) is not None:
        session["transport"] = t
    with session["history_lock"]:
        if session.get("running"):
            return _err(rid, 4009, "session busy")
        if truncate_user_ordinal is not None:
            try:
                ordinal = int(truncate_user_ordinal)
            except (TypeError, ValueError):
                return _err(rid, 4004, "truncate_before_user_ordinal must be an integer")
            history = session.get("history", [])
            user_indices = [i for i, m in enumerate(history) if m.get("role") == "user"]
            if ordinal >= len(user_indices):
                return _err(rid, 4018, "target user message is no longer in session history")
            truncated = history[: user_indices[ordinal]]
            session["history"] = truncated
            session["history_version"] = int(session.get("history_version", 0)) + 1
            if (db := _get_db()) is not None:
                try:
                    db.replace_messages(session["session_key"], truncated)
                except Exception as exc:
                    print(f"[tui_gateway] prompt.submit: replace_messages failed: {exc}", file=sys.stderr)
        session["running"] = True
        session["last_active"] = time.time()
        _start_inflight_turn(session, text)

    _start_agent_build(sid, session)

    def run_after_agent_ready() -> None:
        err = _wait_agent(session, rid)
        if err:
            _emit(
                "error",
                sid,
                {
                    "message": err.get("error", {}).get(
                        "message", "agent initialization failed"
                    )
                },
            )
            with session["history_lock"]:
                session["running"] = False
                _clear_inflight_turn(session)
            return
        _run_prompt_submit(rid, sid, session, text)

    threading.Thread(target=run_after_agent_ready, daemon=True).start()
    return _ok(rid, {"status": "streaming"})


def _notification_poller_loop(
    stop_event: threading.Event, sid: str, session: dict
) -> None:
    """Poll completion_queue and dispatch notifications autonomously.

    Runs in a daemon thread started by _init_session(). Emits a
    status.update (kind=process) for user visibility, then chains an
    agent turn via _run_prompt_submit if the session is idle.

    NOTE: The completion_queue is global (one per process). If multiple
    TUI sessions coexist, whichever poller wakes first grabs the event,
    even if the process was started by a different session. This matches
    CLI/gateway behavior (single session per process).
    """
    from tools.process_registry import process_registry, format_process_notification

    while not stop_event.is_set() and not session.get("_finalized"):
        try:
            evt = process_registry.completion_queue.get(timeout=0.5)
        except Exception:
            continue

        _evt_sid = evt.get("session_id", "")
        if evt.get("type") == "completion" and process_registry.is_completion_consumed(_evt_sid):
            continue

        text = format_process_notification(evt)
        if not text:
            continue

        _emit("status.update", sid, {"kind": "process", "text": text})

        with session["history_lock"]:
            if session.get("running"):
                process_registry.completion_queue.put(evt)
                continue
            session["running"] = True

        rid = f"__notif__{int(time.time() * 1000)}"
        try:
            _emit("message.start", sid)
            _run_prompt_submit(rid, sid, session, text)
        except Exception as exc:
            print(
                f"[tui_gateway] notification poller dispatch failed: "
                f"{type(exc).__name__}: {exc}",
                file=sys.stderr,
            )
            with session["history_lock"]:
                session["running"] = False

    # Drain any remaining events after stop signal (process all pending
    # before exiting so nothing is lost on shutdown).
    while not process_registry.completion_queue.empty():
        try:
            evt = process_registry.completion_queue.get_nowait()
        except Exception:
            break
        _evt_sid = evt.get("session_id", "")
        if evt.get("type") == "completion" and process_registry.is_completion_consumed(_evt_sid):
            continue
        text = format_process_notification(evt)
        if not text:
            continue

        _emit("status.update", sid, {"kind": "process", "text": text})

        with session["history_lock"]:
            if session.get("running"):
                process_registry.completion_queue.put(evt)
                break
            session["running"] = True

        rid = f"__notif__{int(time.time() * 1000)}"
        try:
            _emit("message.start", sid)
            _run_prompt_submit(rid, sid, session, text)
        except Exception as exc:
            print(
                f"[tui_gateway] notification poller dispatch failed: "
                f"{type(exc).__name__}: {exc}",
                file=sys.stderr,
            )
            with session["history_lock"]:
                session["running"] = False


def _start_notification_poller(sid: str, session: dict) -> threading.Event:
    """Start the background notification poller for a TUI session."""
    stop = threading.Event()
    t = threading.Thread(
        target=_notification_poller_loop,
        args=(stop, sid, session),
        daemon=True,
    )
    t.start()
    return stop


def _run_prompt_submit(rid, sid: str, session: dict, text: Any) -> None:
    with session["history_lock"]:
        history = list(session["history"])
        history_version = int(session.get("history_version", 0))
        images = list(session.get("attached_images", []))
        session["attached_images"] = []
        if not isinstance(session.get("inflight_turn"), dict):
            _start_inflight_turn(session, text)
    agent = session["agent"]
    _emit("message.start", sid)

    def run():
        approval_token = None
        session_tokens = []
        goal_followup = None  # set by the post-turn goal hook below
        try:
            from tools.approval import (
                reset_current_session_key,
                set_current_session_key,
            )

            approval_token = set_current_session_key(session["session_key"])
            session_tokens = _set_session_context(session["session_key"])
            cwd = _session_cwd(session)
            _register_session_cwd(session)
            cols = session.get("cols", 80)
            streamer = make_stream_renderer(cols)
            prompt = text

            if isinstance(prompt, str) and "@" in prompt:
                from agent.context_references import preprocess_context_references
                from agent.model_metadata import get_model_context_length

                ctx_len = get_model_context_length(
                    getattr(agent, "model", "") or _resolve_model(),
                    base_url=getattr(agent, "base_url", "") or "",
                    api_key=getattr(agent, "api_key", "") or "",
                    provider=getattr(agent, "provider", "") or "",
                    config_context_length=getattr(
                        agent, "_config_context_length", None
                    ),
                )
                ctx = preprocess_context_references(
                    prompt,
                    cwd=cwd,
                    allowed_root=cwd,
                    context_length=ctx_len,
                )
                if ctx.blocked:
                    _emit(
                        "error",
                        sid,
                        {
                            "message": "\n".join(ctx.warnings)
                            or "Context injection refused."
                        },
                    )
                    return
                prompt = ctx.message

            # Decide image routing per-turn based on active provider/model.
            # "native" → pass pixels to the main model as OpenAI-style content
            # parts (adapters translate for Anthropic/Gemini/Bedrock/etc.).
            # "text"   → pre-analyze with vision_analyze and prepend the text.
            # See agent/image_routing.py for the full decision table.
            run_message: Any = prompt
            if images:
                try:
                    from agent.image_routing import (
                        decide_image_input_mode,
                        build_native_content_parts,
                    )
                    from agent.auxiliary_client import (
                        _read_main_model,
                        _read_main_provider,
                    )
                    from hermes_cli.config import load_config as _tui_load_config

                    _cfg = _tui_load_config()
                    _mode = decide_image_input_mode(
                        _read_main_provider(),
                        _read_main_model(),
                        _cfg,
                    )
                    if getattr(agent, "api_mode", "") == "codex_app_server":
                        _mode = "text"
                except Exception as _img_exc:
                    print(
                        f"[tui_gateway] image_routing decision failed, defaulting to text: {_img_exc}",
                        file=sys.stderr,
                    )
                    _mode = "text"

                if _mode == "native":
                    try:
                        _parts, _skipped = build_native_content_parts(
                            prompt,
                            images,
                        )
                        if _skipped:
                            print(
                                f"[tui_gateway] native image attachment skipped {len(_skipped)} unreadable path(s)",
                                file=sys.stderr,
                            )
                        if any(p.get("type") == "image_url" for p in _parts):
                            run_message = _parts
                        else:
                            run_message = _enrich_with_attached_images(prompt, images)
                    except Exception as _img_exc:
                        print(
                            f"[tui_gateway] native attach failed, falling back to text: {_img_exc}",
                            file=sys.stderr,
                        )
                        run_message = _enrich_with_attached_images(prompt, images)
                else:
                    run_message = _enrich_with_attached_images(prompt, images)

            def _stream(delta):
                with session["history_lock"]:
                    _append_inflight_delta(session, delta)
                payload = {"text": delta}
                if streamer and (r := streamer.feed(delta)) is not None:
                    payload["rendered"] = r
                _emit("message.delta", sid, payload)

            run_kwargs = {
                "conversation_history": list(history),
                "stream_callback": _stream,
            }
            try:
                if "task_id" in inspect.signature(agent.run_conversation).parameters:
                    run_kwargs["task_id"] = session["session_key"]
            except (TypeError, ValueError):
                pass
            result = agent.run_conversation(run_message, **run_kwargs)

            last_reasoning = None
            status_note = None
            if isinstance(result, dict):
                if isinstance(result.get("messages"), list):
                    with session["history_lock"]:
                        current_version = int(session.get("history_version", 0))
                        if current_version == history_version:
                            session["history"] = result["messages"]
                            session["history_version"] = history_version + 1
                        else:
                            # History mutated externally during the turn
                            # (undo/compress/retry/rollback now guard on
                            # session.running, but this is the defensive
                            # backstop for any path that slips past).
                            # Surface the desync rather than silently
                            # dropping the agent's output — the UI can
                            # show the response and warn that it was
                            # not persisted.
                            print(
                                f"[tui_gateway] prompt.submit: history_version mismatch "
                                f"(expected={history_version} current={current_version}) — "
                                f"agent output NOT written to session history",
                                file=sys.stderr,
                            )
                            status_note = (
                                "History changed during this turn — the response above is visible "
                                "but was not saved to session history."
                            )

                # If auto-compression fired inside run_conversation(), agent.session_id
                # may have rotated. Sync session_key before downstream title/goal/finalize
                # handling uses it. Preserve pending_title (user intent) so it can be
                # applied to the continuation. Restart slash worker so subsequent
                # worker-backed commands (/title etc.) target the live session.
                # Fix for #20001.
                _sync_session_key_after_compress(
                    sid, session, clear_pending_title=False, restart_slash_worker=True,
                )

                raw = result.get("final_response", "")
                status = (
                    "interrupted"
                    if result.get("interrupted")
                    else "error" if result.get("error") else "complete"
                )
                # When the backend produced no visible response AND reported a
                # real error (e.g. invalid model slug → provider 4xx), surface
                # that error as the visible text instead of shipping an empty
                # turn to Ink. Mirrors classic CLI behavior at cli.py where
                # (failed|partial) + no final_response → "Error: <detail>".
                # Leaves the None-with-no-error path untouched: an empty
                # successful turn still renders as empty, and the existing
                # "(empty)" sentinel handling stays in its own lane.
                if (not raw) and result.get("error") and (
                    result.get("failed") or result.get("partial")
                ):
                    raw = f"Error: {result.get('error')}"
                lr = result.get("last_reasoning")
                if isinstance(lr, str) and lr.strip():
                    last_reasoning = lr.strip()
            else:
                raw = str(result)
                status = "complete"

            payload = {"text": raw, "usage": _get_usage(agent), "status": status}
            if last_reasoning:
                payload["reasoning"] = last_reasoning
            if status_note:
                payload["warning"] = status_note
            rendered = render_message(raw, cols)
            if rendered:
                payload["rendered"] = rendered
            with session["history_lock"]:
                _clear_inflight_turn(session)
            _emit("message.complete", sid, payload)

            # ── /goal continuation (Ralph-style loop) ─────────────────
            # After every TUI turn, if a /goal is active, ask the judge
            # whether the goal is done and — if not and we're still under
            # budget — queue a continuation prompt to run after this
            # thread releases session["running"]. The verdict message
            # ("✓ Goal achieved" / "⏸ budget exhausted") is surfaced as
            # a system line so the user sees progress regardless of
            # outcome. Mirrors gateway/run._post_turn_goal_continuation.
            if status == "complete" and isinstance(raw, str) and raw.strip():
                try:
                    from hermes_cli.goals import GoalManager

                    sid_key = session.get("session_key") or ""
                    if sid_key:
                        try:
                            goals_cfg = _load_cfg().get("goals") or {}
                            goal_max_turns = int(goals_cfg.get("max_turns", 20) or 20)
                        except Exception:
                            goal_max_turns = 20
                        goal_mgr = GoalManager(
                            session_id=sid_key,
                            default_max_turns=goal_max_turns,
                        )
                        if goal_mgr.is_active():
                            decision = goal_mgr.evaluate_after_turn(
                                raw,
                                user_initiated=True,
                            )
                            verdict_msg = decision.get("message") or ""
                            if verdict_msg:
                                _emit(
                                    "status.update",
                                    sid,
                                    {"kind": "goal", "text": verdict_msg},
                                )
                            if decision.get("should_continue"):
                                cont_prompt = decision.get("continuation_prompt") or ""
                                if cont_prompt:
                                    goal_followup = cont_prompt
                except Exception as _goal_exc:
                    print(
                        f"[tui_gateway] goal continuation hook failed: "
                        f"{type(_goal_exc).__name__}: {_goal_exc}",
                        file=sys.stderr,
                    )

            # Apply pending_title now that the DB row exists.
            _pending = session.get("pending_title")
            if _pending and status == "complete":
                _pdb = _get_db()
                if _pdb:
                    _session_key = session.get("session_key") or sid
                    try:
                        if _pdb.set_session_title(_session_key, _pending):
                            session["pending_title"] = None
                    except ValueError as exc:
                        # Invalid/duplicate title — non-retryable, drop it.
                        # Auto-title will take over. Fix for #19029.
                        session["pending_title"] = None
                        logger.info(
                            "Dropping pending title for session %s: %s",
                            _session_key, exc,
                        )
                    except Exception:
                        # Transient DB failure — keep pending_title for retry.
                        pass

            if (
                status == "complete"
                and isinstance(raw, str)
                and raw.strip()
                and isinstance(text, str)
                and text.strip()
            ):
                try:
                    from agent.title_generator import maybe_auto_title

                    maybe_auto_title(
                        _get_db(),
                        session.get("session_key") or sid,
                        text,
                        raw,
                        session.get("history", []),
                    )
                except Exception:
                    pass

            # 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
            # noisy to read aloud.
            if (
                status == "complete"
                and isinstance(raw, str)
                and raw.strip()
                and _voice_tts_enabled()
            ):
                try:
                    from hermes_cli.voice import speak_text

                    spoken = raw
                    threading.Thread(
                        target=speak_text, args=(spoken,), daemon=True
                    ).start()
                except ImportError:
                    logger.warning("voice TTS skipped: hermes_cli.voice unavailable")
                except Exception as e:
                    logger.warning("voice TTS dispatch failed: %s", e)
        except Exception as e:
            import traceback

            trace = traceback.format_exc()
            try:
                os.makedirs(os.path.dirname(_CRASH_LOG), exist_ok=True)
                with open(_CRASH_LOG, "a", encoding="utf-8") as f:
                    f.write(
                        f"\n=== turn-dispatcher exception · "
                        f"{time.strftime('%Y-%m-%d %H:%M:%S')} · sid={sid} ===\n"
                    )
                    f.write(trace)
            except Exception:
                pass
            print(
                f"[gateway-turn] {type(e).__name__}: {e}", file=sys.stderr, flush=True
            )
            _emit("error", sid, {"message": str(e)})
        finally:
            try:
                if approval_token is not None:
                    reset_current_session_key(approval_token)
            except Exception:
                pass
            _clear_session_context(session_tokens)
            with session["history_lock"]:
                session["running"] = False
                session["last_active"] = time.time()
                _clear_inflight_turn(session)
            _emit("session.info", sid, _session_info(agent, session))

        # Chain a goal-continuation turn if the judge said so. We do
        # this AFTER the finally releases session["running"], so the
        # nested _run_prompt_submit doesn't deadlock on the busy
        # guard. A real user prompt that races us wins because
        # prompt.submit sets running=True under the history_lock and
        # we check that guard before re-firing.
        if goal_followup:
            with session["history_lock"]:
                if session.get("running"):
                    # User already sent something — their turn wins,
                    # the judge will re-run on the next turn anyway.
                    return
                session["running"] = True
            try:
                _emit("message.start", sid)
                _run_prompt_submit(rid, sid, session, goal_followup)
            except Exception as _cont_exc:
                print(
                    f"[tui_gateway] goal continuation dispatch failed: "
                    f"{type(_cont_exc).__name__}: {_cont_exc}",
                    file=sys.stderr,
                )
                with session["history_lock"]:
                    session["running"] = False

        # Drain completion notifications that arrived during this turn.
        # The background poller handles between-turn delivery; this is
        # the safety net for events that arrived mid-turn.
        try:
            from tools.process_registry import process_registry

            for _evt, synth in process_registry.drain_notifications():
                with session["history_lock"]:
                    if session.get("running"):
                        process_registry.completion_queue.put(_evt)
                        break
                    session["running"] = True
                try:
                    _emit("message.start", sid)
                    _run_prompt_submit(rid, sid, session, synth)
                except Exception as _n_exc:
                    print(
                        f"[tui_gateway] completion notification dispatch failed: "
                        f"{type(_n_exc).__name__}: {_n_exc}",
                        file=sys.stderr,
                    )
                    with session["history_lock"]:
                        session["running"] = False
        except Exception as _drain_exc:
            print(
                f"[tui_gateway] completion queue drain failed: "
                f"{type(_drain_exc).__name__}: {_drain_exc}",
                file=sys.stderr,
            )

    threading.Thread(target=run, daemon=True).start()


@method("clipboard.paste")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    try:
        from hermes_cli.clipboard import has_clipboard_image, save_clipboard_image
    except Exception as e:
        return _err(rid, 5027, f"clipboard unavailable: {e}")

    session["image_counter"] = session.get("image_counter", 0) + 1
    img_dir = _hermes_home / "images"
    img_dir.mkdir(parents=True, exist_ok=True)
    img_path = (
        img_dir
        / f"clip_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{session['image_counter']}.png"
    )

    # Save-first: mirrors CLI keybinding path; more robust than has_image() precheck
    if not save_clipboard_image(img_path):
        session["image_counter"] = max(0, session["image_counter"] - 1)
        msg = (
            "Clipboard has image but extraction failed"
            if has_clipboard_image()
            else "No image found in clipboard"
        )
        return _ok(rid, {"attached": False, "message": msg})

    session.setdefault("attached_images", []).append(str(img_path))
    return _ok(
        rid,
        {
            "attached": True,
            "path": str(img_path),
            "count": len(session["attached_images"]),
            **_image_meta(img_path),
        },
    )


@method("image.attach")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    raw = str(params.get("path", "") or "").strip()
    if not raw:
        return _err(rid, 4015, "path required")
    try:
        from cli import (
            _IMAGE_EXTENSIONS,
            _detect_file_drop,
            _resolve_attachment_path,
            _split_path_input,
        )

        dropped = _detect_file_drop(raw)
        if dropped:
            image_path = dropped["path"]
            remainder = dropped["remainder"]
        else:
            path_token, remainder = _split_path_input(raw)
            image_path = _resolve_attachment_path(path_token)
            if image_path is None:
                return _err(rid, 4016, f"image not found: {path_token}")
        if image_path.suffix.lower() not in _IMAGE_EXTENSIONS:
            return _err(rid, 4016, f"unsupported image: {image_path.name}")
        session.setdefault("attached_images", []).append(str(image_path))
        return _ok(
            rid,
            {
                "attached": True,
                "path": str(image_path),
                "count": len(session["attached_images"]),
                "remainder": remainder,
                "text": remainder or f"[User attached image: {image_path.name}]",
                **_image_meta(image_path),
            },
        )
    except Exception as e:
        return _err(rid, 5027, str(e))


@method("image.detach")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    raw = str(params.get("path", "") or "").strip()
    if not raw:
        return _err(rid, 4015, "path required")
    images = session.setdefault("attached_images", [])
    before = len(images)
    session["attached_images"] = [path for path in images if path != raw]
    return _ok(
        rid,
        {
            "detached": len(session["attached_images"]) != before,
            "count": len(session["attached_images"]),
        },
    )


@method("input.detect_drop")
def _(rid, params: dict) -> dict:
    session, err = _sess_nowait(params, rid)
    if err:
        return err
    try:
        from cli import _detect_file_drop

        raw = str(params.get("text", "") or "")
        dropped = _detect_file_drop(raw)
        if not dropped:
            return _ok(rid, {"matched": False})

        drop_path = dropped["path"]
        remainder = dropped["remainder"]
        if dropped["is_image"]:
            session.setdefault("attached_images", []).append(str(drop_path))
            text = remainder or f"[User attached image: {drop_path.name}]"
            return _ok(
                rid,
                {
                    "matched": True,
                    "is_image": True,
                    "path": str(drop_path),
                    "count": len(session["attached_images"]),
                    "text": text,
                    **_image_meta(drop_path),
                },
            )

        text = f"[User attached file: {drop_path}]" + (
            f"\n{remainder}" if remainder else ""
        )
        return _ok(
            rid,
            {
                "matched": True,
                "is_image": False,
                "path": str(drop_path),
                "name": drop_path.name,
                "text": text,
            },
        )
    except Exception as e:
        return _err(rid, 5027, str(e))


@method("prompt.background")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    text, parent = params.get("text", ""), params.get("session_id", "")
    if not text:
        return _err(rid, 4012, "text required")
    task_id = f"bg_{uuid.uuid4().hex[:6]}"

    def run():
        session_tokens = _set_session_context(task_id)
        try:
            from run_agent import AIAgent

            result = AIAgent(
                **_background_agent_kwargs(session["agent"], task_id)
            ).run_conversation(
                user_message=text,
                task_id=task_id,
            )
            _emit(
                "background.complete",
                parent,
                {
                    "task_id": task_id,
                    "text": (
                        result.get("final_response", str(result))
                        if isinstance(result, dict)
                        else str(result)
                    ),
                },
            )
        except Exception as e:
            _emit(
                "background.complete",
                parent,
                {"task_id": task_id, "text": f"error: {e}"},
            )
        finally:
            _clear_session_context(session_tokens)

    threading.Thread(target=run, daemon=True).start()
    return _ok(rid, {"task_id": task_id})


@method("preview.restart")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err

    url = str(params.get("url") or "").strip()
    cwd = str(params.get("cwd") or "").strip()
    context = str(params.get("context") or "").strip()

    if not url:
        return _err(rid, 4012, "url required")

    task_id = f"preview_{uuid.uuid4().hex[:6]}"
    parent = params.get("session_id", "")
    parent_history = _preview_restart_history(session)
    has_history = bool(parent_history)
    prompt = "\n".join(
        line
        for line in [
            "The desktop preview pane cannot load a local server URL.",
            "",
            f"Preview URL: {url}",
            f"Current working directory: {cwd or '(unknown)'}",
            "",
            f"Preview console:\n{context}" if context else "",
            "" if context else "",
            (
                "The conversation history above is from the user's main session — including the commands you (the assistant) previously ran to start servers, edit files, or check ports. Use it to figure out exactly which server should be running at this Preview URL. The user did not start a brand new task; recover what they had working."
                if has_history
                else None
            ),
            "Restart exactly the app intended for the Preview URL, not Hermes Desktop itself.",
            "The Preview URL and port are the target. Preserve that target unless you conclude it is impossible.",
            "If the prior conversation shows a specific command that bound this URL/port, prefer re-running THAT exact command (in the same cwd) over guessing a new one.",
            "First inspect what process, if any, owns the Preview URL port. If a stale server exists, inspect its cwd and prefer that cwd over the Hermes/Desktop process cwd.",
            "The Current working directory is only a hint. Do not assume it is the preview app root when the port owner or files indicate another root.",
            "If the console shows a module-script MIME error for src/main.tsx or similar, a static server is serving source files. Do not restart python -m http.server or any dumb static server for that app.",
            "For module-script MIME failures, inspect package.json/vite config in the candidate app root and start the real dev server/bundler (for example npm/pnpm/yarn dev) so module transforms happen.",
            "Before declaring success, verify the Preview URL responds with the intended app, not Hermes Desktop. If it serves Hermes/Desktop UI or another unrelated app, stop that process and report failure.",
            "Do not modify files. Do not ask the user unless blocked.",
            "Prefer existing project scripts or commands when they are clear.",
            "If a stale process owns the needed port, handle it safely.",
            "Start long-running servers detached/in the background, then return immediately.",
            "Do not run a foreground dev server command that blocks this background task.",
            "Keep the final response short: what command/server was started, or why it could not be restarted.",
        ]
        if line
    )

    def run():
        session_tokens = _set_session_context(task_id)
        try:
            from run_agent import AIAgent
            from tools.terminal_tool import register_task_env_overrides

            if cwd and os.path.isdir(os.path.abspath(os.path.expanduser(cwd))):
                register_task_env_overrides(task_id, {"cwd": os.path.abspath(os.path.expanduser(cwd))})

            history_note = (
                f" (with {len(parent_history)} parent-session messages of context)"
                if parent_history
                else ""
            )
            _emit(
                "preview.restart.progress",
                parent,
                {"task_id": task_id, "text": f"Starting hidden restart agent{history_note}"},
            )
            result = AIAgent(
                **_ephemeral_preview_agent_kwargs(session["agent"], task_id),
                **_preview_restart_callbacks(parent, task_id),
            ).run_conversation(
                user_message=prompt,
                task_id=task_id,
                conversation_history=parent_history or None,
            )
            text = (
                result.get("final_response", str(result))
                if isinstance(result, dict)
                else str(result)
            )
            _emit("preview.restart.complete", parent, {"task_id": task_id, "text": text})
        except Exception as e:
            _emit(
                "preview.restart.complete",
                parent,
                {"task_id": task_id, "text": f"error: {e}"},
            )
        finally:
            try:
                from tools.terminal_tool import clear_task_env_overrides

                clear_task_env_overrides(task_id)
            except Exception:
                pass
            _clear_session_context(session_tokens)

    threading.Thread(target=run, daemon=True).start()
    return _ok(rid, {"task_id": task_id})


# ── Methods: respond ─────────────────────────────────────────────────


def _respond(rid, params, key):
    r = params.get("request_id", "")
    entry = _pending.get(r)
    if not entry:
        return _err(rid, 4009, f"no pending {key} request")
    _, ev = entry
    _answers[r] = params.get(key, "")
    ev.set()
    return _ok(rid, {"status": "ok"})


@method("clarify.respond")
def _(rid, params: dict) -> dict:
    return _respond(rid, params, "answer")


@method("sudo.respond")
def _(rid, params: dict) -> dict:
    return _respond(rid, params, "password")


@method("secret.respond")
def _(rid, params: dict) -> dict:
    return _respond(rid, params, "value")


@method("approval.respond")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    try:
        from tools.approval import resolve_gateway_approval

        return _ok(
            rid,
            {
                "resolved": resolve_gateway_approval(
                    session["session_key"],
                    params.get("choice", "deny"),
                    resolve_all=params.get("all", False),
                )
            },
        )
    except Exception as e:
        return _err(rid, 5004, str(e))


# ── Methods: config ──────────────────────────────────────────────────


@method("config.set")
def _(rid, params: dict) -> dict:
    key, value = params.get("key", ""), params.get("value", "")
    session = _sessions.get(params.get("session_id", ""))

    if key == "model":
        try:
            if not value:
                return _err(rid, 4002, "model value required")
            if session:
                # Reject during an in-flight turn.  agent.switch_model()
                # mutates self.model / self.provider / self.base_url /
                # self.client in place; the worker thread running
                # agent.run_conversation is reading those on every
                # iteration.  A mid-turn swap can send an HTTP request
                # with the new base_url but old model (or vice versa),
                # producing 400/404s the user never asked for.  Parity
                # with the gateway's running-agent /model guard.
                if session.get("running"):
                    return _err(
                        rid,
                        4009,
                        "session busy — /interrupt the current turn before switching models",
                    )
                if session.get("agent") is None:
                    session_id = params.get("session_id", "")
                    _start_agent_build(session_id, session)
                    init_err = _wait_agent(session, rid)
                    if init_err:
                        return init_err
                    if session.get("agent") is None:
                        return _err(rid, 5032, "agent initialization failed")
                result = _apply_model_switch(
                    params.get("session_id", ""), session, value
                )
            else:
                result = _apply_model_switch("", {"agent": None}, value)
            return _ok(
                rid,
                {"key": key, "value": result["value"], "warning": result["warning"]},
            )
        except Exception as e:
            return _err(rid, 5001, str(e))

    if key == "fast":
        raw = str(value or "").strip().lower()
        agent = session.get("agent") if session else None
        if agent is not None:
            current_fast = getattr(agent, "service_tier", None) == "priority"
        else:
            current_fast = _load_service_tier() == "priority"

        if raw in {"status"}:
            return _ok(
                rid,
                {"key": key, "value": "fast" if current_fast else "normal"},
            )

        if raw in {"", "toggle"}:
            nv = "normal" if current_fast else "fast"
        elif raw in {"fast", "on"}:
            nv = "fast"
        elif raw in {"normal", "off"}:
            nv = "normal"
        else:
            return _err(rid, 4002, f"unknown fast mode: {value}")

        overrides = None
        if nv == "fast":
            from hermes_cli.models import resolve_fast_mode_overrides

            target_model = (
                getattr(agent, "model", None) if agent is not None else _resolve_model()
            )
            if not target_model:
                return _err(
                    rid,
                    4002,
                    "fast mode is not available without a selected model",
                )
            overrides = resolve_fast_mode_overrides(target_model)
            if overrides is None:
                return _err(
                    rid,
                    4002,
                    "fast mode is not available for this model",
                )

        _write_config_key("agent.service_tier", nv)
        if agent is not None:
            agent.service_tier = "priority" if nv == "fast" else None
            current_overrides = dict(getattr(agent, "request_overrides", {}) or {})
            current_overrides.pop("service_tier", None)
            current_overrides.pop("speed", None)
            if nv == "fast":
                current_overrides.update(overrides)
            agent.request_overrides = current_overrides
            _emit(
                "session.info",
                params.get("session_id", ""),
                _session_info(agent, session),
            )
        return _ok(rid, {"key": key, "value": nv})

    if key == "busy":
        raw = str(value or "").strip().lower()
        if raw in {"", "status"}:
            return _ok(rid, {"key": key, "value": _load_busy_input_mode()})
        if raw not in {"queue", "steer", "interrupt"}:
            return _err(rid, 4002, f"unknown busy mode: {value}")
        _write_config_key("display.busy_input_mode", raw)
        return _ok(rid, {"key": key, "value": raw})

    if key == "verbose":
        cycle = ["off", "new", "all", "verbose"]
        cur = (
            session.get("tool_progress_mode", _load_tool_progress_mode())
            if session
            else _load_tool_progress_mode()
        )
        if value and value != "cycle":
            nv = str(value).strip().lower()
            if nv not in cycle:
                return _err(rid, 4002, f"unknown verbose mode: {value}")
        else:
            try:
                idx = cycle.index(cur)
            except ValueError:
                idx = 2
            nv = cycle[(idx + 1) % len(cycle)]
        _write_config_key("display.tool_progress", nv)
        if session:
            session["tool_progress_mode"] = nv
            agent = session.get("agent")
            if agent is not None:
                agent.verbose_logging = nv == "verbose"
        return _ok(rid, {"key": key, "value": nv})

    if key == "yolo":
        try:
            if session:
                from tools.approval import (
                    disable_session_yolo,
                    enable_session_yolo,
                    is_session_yolo_enabled,
                )

                current = is_session_yolo_enabled(session["session_key"])
                if current:
                    disable_session_yolo(session["session_key"])
                    nv = "0"
                else:
                    enable_session_yolo(session["session_key"])
                    nv = "1"
            else:
                current = is_truthy_value(os.environ.get("HERMES_YOLO_MODE"))
                if current:
                    os.environ.pop("HERMES_YOLO_MODE", None)
                    nv = "0"
                else:
                    os.environ["HERMES_YOLO_MODE"] = "1"
                    nv = "1"
            return _ok(rid, {"key": key, "value": nv})
        except Exception as e:
            return _err(rid, 5001, str(e))

    if key == "reasoning":
        try:
            from hermes_constants import parse_reasoning_effort

            arg = str(value or "").strip().lower()
            if arg in {"show", "on"}:
                cfg = _load_cfg()
                display = (
                    cfg.get("display") if isinstance(cfg.get("display"), dict) else {}
                )
                sections = (
                    display.get("sections")
                    if isinstance(display.get("sections"), dict)
                    else {}
                )
                display["show_reasoning"] = True
                sections["thinking"] = "expanded"
                display["sections"] = sections
                cfg["display"] = display
                _save_cfg(cfg)
                if session:
                    session["show_reasoning"] = True
                return _ok(rid, {"key": key, "value": "show"})
            if arg in {"hide", "off"}:
                cfg = _load_cfg()
                display = (
                    cfg.get("display") if isinstance(cfg.get("display"), dict) else {}
                )
                sections = (
                    display.get("sections")
                    if isinstance(display.get("sections"), dict)
                    else {}
                )
                display["show_reasoning"] = False
                sections["thinking"] = "hidden"
                display["sections"] = sections
                cfg["display"] = display
                _save_cfg(cfg)
                if session:
                    session["show_reasoning"] = False
                return _ok(rid, {"key": key, "value": "hide"})

            parsed = parse_reasoning_effort(arg)
            if parsed is None:
                return _err(rid, 4002, f"unknown reasoning value: {value}")
            _write_config_key("agent.reasoning_effort", arg)
            if session and session.get("agent") is not None:
                session["agent"].reasoning_config = parsed
            return _ok(rid, {"key": key, "value": arg})
        except Exception as e:
            return _err(rid, 5001, str(e))

    if key == "details_mode":
        nv = str(value or "").strip().lower()
        if nv not in _DETAIL_MODES:
            return _err(rid, 4002, f"unknown details_mode: {value}")
        cfg = _load_cfg()
        display = cfg.get("display") if isinstance(cfg.get("display"), dict) else {}
        sections = (
            display.get("sections") if isinstance(display.get("sections"), dict) else {}
        )
        display["details_mode"] = nv
        for section in _DETAIL_SECTION_NAMES:
            sections[section] = nv
        display["sections"] = sections
        cfg["display"] = display
        _save_cfg(cfg)
        return _ok(rid, {"key": key, "value": nv})

    if key.startswith("details_mode."):
        # Per-section override: `details_mode.<section>` writes to
        # `display.sections.<section>`. Empty value clears the explicit
        # override and lets frontend resolution apply built-in section defaults
        # before the global details_mode.
        section = key.split(".", 1)[1]
        if section not in _DETAIL_SECTION_NAMES:
            return _err(rid, 4002, f"unknown section: {section}")

        cfg = _load_cfg()
        display = cfg.get("display") if isinstance(cfg.get("display"), dict) else {}
        sections_cfg = (
            display.get("sections") if isinstance(display.get("sections"), dict) else {}
        )

        nv = str(value or "").strip().lower()
        if not nv:
            sections_cfg.pop(section, None)
            display["sections"] = sections_cfg
            cfg["display"] = display
            _save_cfg(cfg)
            return _ok(rid, {"key": key, "value": ""})

        if nv not in _DETAIL_MODES:
            return _err(rid, 4002, f"unknown details_mode: {value}")

        sections_cfg[section] = nv
        display["sections"] = sections_cfg
        cfg["display"] = display
        _save_cfg(cfg)
        return _ok(rid, {"key": key, "value": nv})

    if key == "thinking_mode":
        nv = str(value or "").strip().lower()
        allowed_tm = frozenset({"collapsed", "truncated", "full"})
        if nv not in allowed_tm:
            return _err(rid, 4002, f"unknown thinking_mode: {value}")
        _write_config_key("display.thinking_mode", nv)
        # Backward compatibility bridge: keep details_mode aligned.
        _write_config_key(
            "display.details_mode", "expanded" if nv == "full" else "collapsed"
        )
        return _ok(rid, {"key": key, "value": nv})

    if key == "compact":
        raw = str(value or "").strip().lower()
        cfg0 = _load_cfg()
        d0 = cfg0.get("display") if isinstance(cfg0.get("display"), dict) else {}
        cur_b = bool(d0.get("tui_compact", False))
        if raw in {"", "toggle"}:
            nv_b = not cur_b
        elif raw == "on":
            nv_b = True
        elif raw == "off":
            nv_b = False
        else:
            return _err(rid, 4002, f"unknown compact value: {value}")
        _write_config_key("display.tui_compact", nv_b)
        return _ok(rid, {"key": key, "value": "on" if nv_b else "off"})

    if key == "statusbar":
        raw = str(value or "").strip().lower()
        display = _load_cfg().get("display")
        d0 = display if isinstance(display, dict) else {}
        current = _coerce_statusbar(d0.get("tui_statusbar", "top"))

        if raw in {"", "toggle"}:
            nv = "top" if current == "off" else "off"
        elif raw == "on":
            nv = "top"
        elif raw in _STATUSBAR_MODES:
            nv = raw
        else:
            return _err(rid, 4002, f"unknown statusbar value: {value}")

        _write_config_key("display.tui_statusbar", nv)
        return _ok(rid, {"key": key, "value": nv})

    if key == "mouse":
        # Explicit None check rather than `value or ""` so falsy non-string
        # inputs (0, False) reach the alias map as themselves — both map to
        # 'off' via _MOUSE_TRACKING_ALIASES — instead of being collapsed to
        # '' and triggering the toggle path. The slash command always passes
        # a string, but programmatic JSON-RPC callers may send booleans.
        raw = ("" if value is None else str(value)).strip().lower()
        cfg = _load_cfg()
        display = cfg.get("display") if isinstance(cfg.get("display"), dict) else {}
        current = _display_mouse_tracking(display)

        if raw in {"", "toggle"}:
            nv = "all" if current == "off" else "off"
        elif raw in _MOUSE_TRACKING_ALIASES:
            nv = _MOUSE_TRACKING_ALIASES[raw]
        else:
            return _err(rid, 4002, f"unknown mouse value: {value}")

        _write_config_key("display.mouse_tracking", nv)
        return _ok(rid, {"key": key, "value": nv})

    if key == "indicator":
        # Use an explicit None check rather than `value or ""` so falsy
        # non-string inputs (0, False, []) still surface as themselves
        # in the error message instead of looking like a blank value.
        raw = ("" if value is None else str(value)).strip().lower()
        if raw not in _INDICATOR_STYLES:
            return _err(
                rid,
                4002,
                f"unknown indicator: {raw!r}; pick one of {'|'.join(_INDICATOR_STYLES)}",
            )
        _write_config_key("display.tui_status_indicator", raw)
        return _ok(rid, {"key": key, "value": raw})

    if key in {"cwd", "terminal.cwd", "workdir"}:
        raw = str(value or "").strip()
        if not raw:
            return _err(rid, 4002, "cwd required")
        cwd = os.path.abspath(os.path.expanduser(raw))
        if not os.path.isdir(cwd):
            return _err(rid, 4002, f"working directory does not exist: {raw}")
        _write_config_key("terminal.cwd", cwd)
        os.environ["TERMINAL_CWD"] = cwd
        return _ok(
            rid,
            {"key": "terminal.cwd", "value": cwd, "cwd": cwd, "branch": _git_branch_for_cwd(cwd)},
        )

    if key in {"prompt", "personality", "skin"}:
        try:
            cfg = _load_cfg()
            if key == "prompt":
                if value == "clear":
                    cfg.pop("custom_prompt", None)
                    nv = ""
                else:
                    cfg["custom_prompt"] = value
                    nv = value
                _save_cfg(cfg)
            elif key == "personality":
                sid_key = params.get("session_id", "")
                pname, new_prompt = _validate_personality(str(value or ""), cfg)
                _write_config_key("display.personality", pname)
                _write_config_key("agent.system_prompt", new_prompt)
                nv = str(value or "none")
                history_reset, info = _apply_personality_to_session(
                    sid_key, session, new_prompt, pname
                )
            else:
                _write_config_key(f"display.{key}", value)
                nv = value
                if key == "skin":
                    _emit("skin.changed", "", resolve_skin())
            resp = {"key": key, "value": nv}
            if key == "personality":
                resp["history_reset"] = history_reset
                if info is not None:
                    resp["info"] = info
            return _ok(rid, resp)
        except Exception as e:
            return _err(rid, 5001, str(e))

    return _err(rid, 4002, f"unknown config key: {key}")


@method("config.get")
def _(rid, params: dict) -> dict:
    key = params.get("key", "")
    if key == "provider":
        try:
            from hermes_cli.models import list_available_providers, normalize_provider

            model = _resolve_model()
            parts = model.split("/", 1)
            return _ok(
                rid,
                {
                    "model": model,
                    "provider": (
                        normalize_provider(parts[0]) if len(parts) > 1 else "unknown"
                    ),
                    "providers": list_available_providers(),
                },
            )
        except Exception as e:
            return _err(rid, 5013, str(e))
    if key == "profile":
        from hermes_constants import display_hermes_home

        return _ok(rid, {"home": str(_hermes_home), "display": display_hermes_home()})
    if key == "project":
        cfg_terminal = _load_cfg().get("terminal") or {}
        raw = str(params.get("cwd", "") or cfg_terminal.get("cwd", "") or "").strip()
        cwd = _completion_cwd({"cwd": raw} if raw else {})
        return _ok(rid, {"cwd": cwd, "branch": _git_branch_for_cwd(cwd)})
    if key == "full":
        return _ok(rid, {"config": _load_cfg()})
    if key == "prompt":
        return _ok(rid, {"prompt": _load_cfg().get("custom_prompt", "")})
    if key == "skin":
        return _ok(
            rid, {"value": (_load_cfg().get("display") or {}).get("skin", "default")}
        )
    if key == "indicator":
        # Normalize so a hand-edited config.yaml with stray casing or
        # an unknown value reads back the SAME value the TUI actually
        # rendered (frontend's `normalizeIndicatorStyle` falls back to
        # `_INDICATOR_DEFAULT` for the same inputs).  Otherwise
        # `/indicator` would print one thing while the UI shows another.
        raw = (_load_cfg().get("display") or {}).get("tui_status_indicator", "")
        norm = str(raw).strip().lower()
        return _ok(
            rid,
            {"value": norm if norm in _INDICATOR_STYLES else _INDICATOR_DEFAULT},
        )
    if key == "personality":
        return _ok(
            rid,
            {"value": (_load_cfg().get("display") or {}).get("personality") or "none"},
        )
    if key == "reasoning":
        cfg = _load_cfg()
        effort = str(
            (cfg.get("agent") or {}).get("reasoning_effort", "medium") or "medium"
        )
        display = (
            "show"
            if bool((cfg.get("display") or {}).get("show_reasoning", False))
            else "hide"
        )
        return _ok(rid, {"value": effort, "display": display})
    if key == "fast":
        return _ok(
            rid,
            {
                "value": (
                    "fast"
                    if (session := _sessions.get(params.get("session_id", "")))
                    and getattr(session.get("agent"), "service_tier", None)
                    == "priority"
                    else ("fast" if _load_service_tier() == "priority" else "normal")
                ),
            },
        )
    if key == "busy":
        return _ok(rid, {"value": _load_busy_input_mode()})
    if key == "details_mode":
        allowed_dm = frozenset({"hidden", "collapsed", "expanded"})
        raw = (
            str(
                (_load_cfg().get("display") or {}).get("details_mode", "collapsed")
                or "collapsed"
            )
            .strip()
            .lower()
        )
        nv = raw if raw in allowed_dm else "collapsed"
        return _ok(rid, {"value": nv})
    if key == "thinking_mode":
        allowed_tm = frozenset({"collapsed", "truncated", "full"})
        cfg = _load_cfg()
        raw = (
            str((cfg.get("display") or {}).get("thinking_mode", "") or "")
            .strip()
            .lower()
        )
        if raw in allowed_tm:
            nv = raw
        else:
            dm = (
                str(
                    (cfg.get("display") or {}).get("details_mode", "collapsed")
                    or "collapsed"
                )
                .strip()
                .lower()
            )
            nv = "full" if dm == "expanded" else "collapsed"
        return _ok(rid, {"value": nv})
    if key == "compact":
        on = bool((_load_cfg().get("display") or {}).get("tui_compact", False))
        return _ok(rid, {"value": "on" if on else "off"})
    if key == "statusbar":
        display = _load_cfg().get("display")
        raw = (
            display.get("tui_statusbar", "top") if isinstance(display, dict) else "top"
        )
        return _ok(rid, {"value": _coerce_statusbar(raw)})
    if key == "mouse":
        display = _load_cfg().get("display")
        return _ok(rid, {"value": _display_mouse_tracking(display)})
    if key == "mtime":
        cfg_path = _hermes_home / "config.yaml"
        try:
            return _ok(
                rid, {"mtime": cfg_path.stat().st_mtime if cfg_path.exists() else 0}
            )
        except Exception:
            return _ok(rid, {"mtime": 0})
    return _err(rid, 4002, f"unknown config key: {key}")


@method("setup.status")
def _(rid, params: dict) -> dict:
    try:
        from hermes_cli.main import _has_any_provider_configured

        return _ok(rid, {"provider_configured": bool(_has_any_provider_configured())})
    except Exception as e:
        return _err(rid, 5016, str(e))


@method("setup.runtime_check")
def _(rid, params: dict) -> dict:
    """Strict provider check: does the configured/default model actually resolve to a usable runtime?

    Unlike setup.status (which returns True if ANY provider auth state is
    discoverable, including indirect fallbacks like ``gh auth token`` for
    Copilot), this runs the same resolve_runtime_provider() call the agent
    uses on session creation. It returns ok=False with the auth error message
    when the user's configured model cannot actually be served, so UIs can
    surface onboarding before the user submits a doomed prompt.
    """
    try:
        from hermes_cli.runtime_provider import resolve_runtime_provider
        from hermes_cli.auth import has_usable_secret
        from hermes_cli.main import _has_any_provider_configured

        runtime = resolve_runtime_provider(requested=None)
        provider_configured = bool(_has_any_provider_configured())
        provider = runtime.get("provider") or "provider"
        source = str(runtime.get("source") or "")
        if not provider_configured and provider == "bedrock" and source in {
            "iam-role",
            "aws-sdk-default-chain",
        }:
            return _ok(
                rid,
                {
                    "ok": False,
                    "provider": provider,
                    "model": runtime.get("model"),
                    "source": source,
                    "error": "No Hermes provider is configured.",
                },
            )

        api_key = runtime.get("api_key")
        api_key_text = "" if callable(api_key) else str(api_key or "").strip()
        credential_ok = (
            callable(api_key)
            or api_key_text in {"aws-sdk", "no-key-required"}
            or has_usable_secret(api_key_text)
            or bool(runtime.get("command"))
        )

        if not credential_ok:
            return _ok(
                rid,
                {
                    "ok": False,
                    "provider": provider,
                    "model": runtime.get("model"),
                    "source": runtime.get("source"),
                    "error": f"No usable credentials found for {provider}.",
                },
            )

        return _ok(
            rid,
            {
                "ok": True,
                "provider": runtime.get("provider"),
                "model": runtime.get("model"),
                "source": runtime.get("source"),
            },
        )
    except Exception as e:
        return _ok(rid, {"ok": False, "error": str(e)})


# ── Methods: tools & system ──────────────────────────────────────────


@method("process.stop")
def _(rid, params: dict) -> dict:
    try:
        from tools.process_registry import process_registry

        return _ok(rid, {"killed": process_registry.kill_all()})
    except Exception as e:
        return _err(rid, 5010, str(e))


@method("reload.mcp")
def _(rid, params: dict) -> dict:
    session = _sessions.get(params.get("session_id", ""))
    try:
        # Gate: /reload-mcp invalidates the prompt cache for this session.
        # Respect the ``approvals.mcp_reload_confirm`` config toggle — if
        # set (default true) AND the caller did not pass ``confirm=true``
        # in params, surface a warning to the transcript instead of just
        # reloading silently.  Users pass confirm=true either by
        # re-invoking after reading the warning, or by setting the
        # config key to false permanently.
        user_confirm = bool(params.get("confirm", False))
        if not user_confirm:
            try:
                from hermes_cli.config import load_config as _load_config

                _cfg = _load_config()
                _approvals = _cfg.get("approvals") if isinstance(_cfg, dict) else None
                _confirm_required = True
                if isinstance(_approvals, dict):
                    _confirm_required = bool(_approvals.get("mcp_reload_confirm", True))
            except Exception:
                _confirm_required = True
            if _confirm_required:
                # Return a structured response the Ink client can surface
                # as a warning/confirmation without actually reloading yet.
                # Ink's ops.ts reads ``status`` and prints ``message`` to
                # the transcript; a follow-up invocation with confirm=true
                # (or an `always` choice that flips the config) proceeds.
                return _ok(
                    rid,
                    {
                        "status": "confirm_required",
                        "message": (
                            "⚠️  /reload-mcp invalidates the prompt cache (next "
                            "message re-sends full input tokens). Reply `/reload-mcp "
                            "now` to proceed, or `/reload-mcp always` to proceed and "
                            "silence this prompt permanently."
                        ),
                    },
                )

        from tools.mcp_tool import shutdown_mcp_servers, discover_mcp_tools

        shutdown_mcp_servers()
        discover_mcp_tools()
        if session:
            agent = session["agent"]
            # Rebuild the cached agent's tool snapshot so the current session
            # picks up added/removed MCP tools without `/new` (which discards
            # history).  The agent snapshots tools once at build and never
            # re-reads the registry, so an explicit rebuild is required here.
            # The user already consented to the prompt-cache invalidation via
            # the confirm gate above.  Mirrors gateway/run.py::_execute_mcp_reload.
            try:
                from model_tools import get_tool_definitions

                new_defs = get_tool_definitions(
                    enabled_toolsets=_load_enabled_toolsets(),
                    quiet_mode=True,
                )
                agent.tools = new_defs
                agent.valid_tool_names = (
                    {t["function"]["name"] for t in new_defs} if new_defs else set()
                )
            except Exception as _exc:
                logger.warning(
                    "Failed to refresh cached agent tools after /reload-mcp: %s",
                    _exc,
                )
            _emit(
                "session.info",
                params.get("session_id", ""),
                _session_info(agent, session),
            )

        # Honor `always=true` by persisting the opt-out to config.
        if bool(params.get("always", False)):
            try:
                from cli import save_config_value as _save_cfg

                _save_cfg("approvals.mcp_reload_confirm", False)
            except Exception as _exc:
                logger.warning("Failed to persist mcp_reload_confirm=false: %s", _exc)

        return _ok(rid, {"status": "reloaded"})
    except Exception as e:
        return _err(rid, 5015, str(e))


@method("reload.env")
def _(rid, params: dict) -> dict:
    """Re-read ``~/.hermes/.env`` into the gateway process via
    ``hermes_cli.config.reload_env``, matching classic CLI's ``/reload``
    handler.  Newly added API keys take effect on the next agent call
    without restarting the TUI.

    The credential pool / provider routing for any *already-constructed*
    agent does not auto-rebuild — that's the same behaviour as classic
    CLI's ``/reload``.  Users who want a brand-new credential resolution
    should follow with ``/new``.
    """
    try:
        from hermes_cli.config import reload_env

        count = reload_env()
        return _ok(rid, {"updated": int(count)})
    except Exception as e:
        return _err(rid, 5015, str(e))


_TUI_HIDDEN: frozenset[str] = frozenset(
    {
        "sethome",
        "set-home",
        "commands",
        "approve",
        "deny",
    }
)

_TUI_EXTRA: list[tuple[str, str, str]] = [
    ("/compact", "Toggle compact display mode", "TUI"),
    ("/logs", "Show recent gateway log lines", "TUI"),
    (
        "/mouse",
        "Set mouse tracking preset [on|off|toggle|wheel|buttons|all]",
        "TUI",
    ),
    ("/sessions", "Switch between live TUI sessions", "TUI"),
]

# Commands that queue messages onto _pending_input in the CLI.
# In the TUI the slash worker subprocess has no reader for that queue,
# so slash.exec rejects them → TUI falls through to command.dispatch.
_PENDING_INPUT_COMMANDS: frozenset[str] = frozenset(
    {
        "retry",
        "queue",
        "q",
        "steer",
        "plan",
        "goal",
    }
)

_WORKER_BLOCKED_COMMANDS: frozenset[str] = frozenset({"snapshot", "snap"})


@method("commands.catalog")
def _(rid, params: dict) -> dict:
    """Registry-backed slash metadata for the TUI — categorized, no aliases."""
    try:
        from hermes_cli.commands import (
            COMMAND_REGISTRY,
            SUBCOMMANDS,
            _build_description,
        )

        all_pairs: list[list[str]] = []
        canon: dict[str, str] = {}
        categories: list[dict] = []
        cat_map: dict[str, list[list[str]]] = {}
        cat_order: list[str] = []

        for cmd in COMMAND_REGISTRY:
            if cmd.name in _TUI_HIDDEN or cmd.gateway_only:
                continue

            c = f"/{cmd.name}"
            canon[c.lower()] = c
            for a in cmd.aliases:
                canon[f"/{a}".lower()] = c

            desc = _build_description(cmd)
            all_pairs.append([c, desc])

            cat = cmd.category
            if cat not in cat_map:
                cat_map[cat] = []
                cat_order.append(cat)
            cat_map[cat].append([c, desc])

        for name, desc, cat in _TUI_EXTRA:
            all_pairs.append([name, desc])
            if cat not in cat_map:
                cat_map[cat] = []
                cat_order.append(cat)
            cat_map[cat].append([name, desc])

        warning = ""
        try:
            qcmds = _load_cfg().get("quick_commands", {}) or {}
            if isinstance(qcmds, dict) and qcmds:
                bucket = "User commands"
                if bucket not in cat_map:
                    cat_map[bucket] = []
                    cat_order.append(bucket)
                for qname, qc in sorted(qcmds.items()):
                    if not isinstance(qc, dict):
                        continue
                    key = f"/{qname}"
                    canon[key.lower()] = key
                    qtype = qc.get("type", "")
                    if qtype == "exec":
                        default_desc = f"exec: {qc.get('command', '')}"
                    elif qtype == "alias":
                        default_desc = f"alias → {qc.get('target', '')}"
                    else:
                        default_desc = qtype or "quick command"
                    qdesc = str(qc.get("description") or default_desc)
                    qdesc = qdesc[:120] + ("…" if len(qdesc) > 120 else "")
                    all_pairs.append([key, qdesc])
                    cat_map[bucket].append([key, qdesc])
        except Exception as e:
            if not warning:
                warning = f"quick_commands discovery unavailable: {e}"

        skill_count = 0
        try:
            from agent.skill_commands import scan_skill_commands

            for k, info in sorted(scan_skill_commands().items()):
                d = str(info.get("description", "Skill"))
                all_pairs.append([k, d[:120] + ("…" if len(d) > 120 else "")])
                skill_count += 1
        except Exception as e:
            warning = f"skill discovery unavailable: {e}"

        for cat in cat_order:
            categories.append({"name": cat, "pairs": cat_map[cat]})

        sub = {k: v[:] for k, v in SUBCOMMANDS.items()}
        return _ok(
            rid,
            {
                "pairs": all_pairs,
                "sub": sub,
                "canon": canon,
                "categories": categories,
                "skill_count": skill_count,
                "warning": warning,
            },
        )
    except Exception as e:
        return _err(rid, 5020, str(e))


def _cli_exec_blocked(argv: list[str]) -> str | None:
    """Return user hint if this argv must not run headless in the gateway process."""
    if not argv:
        return "bare `hermes` is interactive — use `/hermes chat -q …` or run `hermes` in another terminal"
    a0 = argv[0].lower()
    if a0 == "setup":
        return "`hermes setup` needs a full terminal — run it outside the TUI"
    if a0 == "gateway":
        return "`hermes gateway` is long-running — run it in another terminal"
    if a0 == "sessions" and len(argv) > 1 and argv[1].lower() == "browse":
        return "`hermes sessions browse` is interactive — use /resume here, or run browse in another terminal"
    if a0 == "config" and len(argv) > 1 and argv[1].lower() == "edit":
        return "`hermes config edit` needs $EDITOR in a real terminal"
    return None


@method("cli.exec")
def _(rid, params: dict) -> dict:
    """Run `python -m hermes_cli.main` with argv; capture stdout/stderr (non-interactive only)."""
    argv = params.get("argv", [])
    if not isinstance(argv, list) or not all(isinstance(x, str) for x in argv):
        return _err(rid, 4003, "argv must be list[str]")
    hint = _cli_exec_blocked(argv)
    if hint:
        return _ok(rid, {"blocked": True, "hint": hint, "code": -1, "output": ""})
    try:
        r = subprocess.run(
            [sys.executable, "-m", "hermes_cli.main", *argv],
            capture_output=True,
            text=True,
            timeout=min(int(params.get("timeout", 240)), 600),
            cwd=os.getcwd(),
            env=os.environ.copy(),
        )
        parts = [r.stdout or "", r.stderr or ""]
        out = "\n".join(p for p in parts if p).strip() or "(no output)"
        return _ok(
            rid, {"blocked": False, "code": r.returncode, "output": out[:48_000]}
        )
    except subprocess.TimeoutExpired:
        return _err(rid, 5016, "cli.exec: timeout")
    except Exception as e:
        return _err(rid, 5017, str(e))


@method("command.resolve")
def _(rid, params: dict) -> dict:
    try:
        from hermes_cli.commands import resolve_command

        r = resolve_command(params.get("name", ""))
        if r:
            return _ok(
                rid,
                {
                    "canonical": r.name,
                    "description": r.description,
                    "category": r.category,
                },
            )
        return _err(rid, 4011, f"unknown command: {params.get('name')}")
    except Exception as e:
        return _err(rid, 5012, str(e))


def _resolve_name(name: str) -> str:
    try:
        from hermes_cli.commands import resolve_command

        r = resolve_command(name)
        return r.name if r else name
    except Exception:
        return name


@method("command.dispatch")
def _(rid, params: dict) -> dict:
    name, arg = params.get("name", "").lstrip("/"), params.get("arg", "")
    resolved = _resolve_name(name)
    if resolved != name:
        name = resolved
    session = _sessions.get(params.get("session_id", ""))

    qcmds = _load_cfg().get("quick_commands", {})
    if name in qcmds:
        qc = qcmds[name]
        if qc.get("type") == "exec":
            r = subprocess.run(
                qc.get("command", ""),
                shell=True,
                capture_output=True,
                text=True,
                timeout=30,
            )
            output = (
                (r.stdout or "")
                + ("\n" if r.stdout and r.stderr else "")
                + (r.stderr or "")
            ).strip()[:4000]
            if r.returncode != 0:
                return _err(
                    rid,
                    4018,
                    output or f"quick command failed with exit code {r.returncode}",
                )
            return _ok(rid, {"type": "exec", "output": output})
        if qc.get("type") == "alias":
            return _ok(rid, {"type": "alias", "target": qc.get("target", "")})

    try:
        from hermes_cli.plugins import (
            get_plugin_command_handler,
            resolve_plugin_command_result,
        )

        handler = get_plugin_command_handler(name)
        if handler:
            result = resolve_plugin_command_result(handler(arg))
            return _ok(rid, {"type": "plugin", "output": str(result or "")})
    except Exception:
        pass

    try:
        from agent.skill_commands import (
            scan_skill_commands,
            build_skill_invocation_message,
        )

        cmds = scan_skill_commands()
        key = f"/{name}"
        if key in cmds:
            msg = build_skill_invocation_message(
                key, arg, task_id=session.get("session_key", "") if session else ""
            )
            if msg:
                return _ok(
                    rid,
                    {
                        "type": "skill",
                        "message": msg,
                        "name": cmds[key].get("name", name),
                    },
                )
    except Exception:
        pass

    # ── Commands that queue messages onto _pending_input in the CLI ───
    # In the TUI the slash worker subprocess has no reader for that queue,
    # so we handle them here and return a structured payload.

    if name in {"queue", "q"}:
        if not arg:
            return _err(rid, 4004, "usage: /queue <prompt>")
        return _ok(rid, {"type": "send", "message": arg})

    if name == "retry":
        if not session:
            return _err(rid, 4001, "no active session to retry")
        if session.get("running"):
            return _err(
                rid, 4009, "session busy — /interrupt the current turn before /retry"
            )
        history = session.get("history", [])
        if not history:
            return _err(rid, 4018, "no previous user message to retry")
        # Walk backwards to find the last user message
        last_user_idx = None
        for i in range(len(history) - 1, -1, -1):
            if history[i].get("role") == "user":
                last_user_idx = i
                break
        if last_user_idx is None:
            return _err(rid, 4018, "no previous user message to retry")
        content = history[last_user_idx].get("content", "")
        if isinstance(content, list):
            content = " ".join(
                p.get("text", "")
                for p in content
                if isinstance(p, dict) and p.get("type") == "text"
            )
        if not content:
            return _err(rid, 4018, "last user message is empty")
        # Truncate history: remove everything from the last user message onward
        # (mirrors CLI retry_last() which strips the failed exchange)
        with session["history_lock"]:
            session["history"] = history[:last_user_idx]
            session["history_version"] = int(session.get("history_version", 0)) + 1
        return _ok(rid, {"type": "send", "message": content})

    if name == "steer":
        if not arg:
            return _err(rid, 4004, "usage: /steer <prompt>")
        agent = session.get("agent") if session else None
        if agent and hasattr(agent, "steer"):
            try:
                accepted = agent.steer(arg)
                if accepted:
                    return _ok(
                        rid,
                        {
                            "type": "exec",
                            "output": f"⏩ Steer queued — arrives after the next tool call: {arg[:80]}{'...' if len(arg) > 80 else ''}",
                        },
                    )
            except Exception:
                pass
        # Fallback: no active run, treat as next-turn message
        return _ok(rid, {"type": "send", "message": arg})

    if name == "goal":
        if not session:
            return _err(rid, 4001, "no active session")
        try:
            from hermes_cli.goals import GoalManager
        except Exception as exc:
            return _err(rid, 5030, f"goals unavailable: {exc}")

        sid_key = session.get("session_key") or ""
        if not sid_key:
            return _err(rid, 4001, "no session key")

        try:
            goals_cfg = _load_cfg().get("goals") or {}
            max_turns = int(goals_cfg.get("max_turns", 20) or 20)
        except Exception:
            max_turns = 20
        mgr = GoalManager(session_id=sid_key, default_max_turns=max_turns)

        lower = arg.strip().lower()
        if not arg.strip() or lower == "status":
            return _ok(rid, {"type": "exec", "output": mgr.status_line()})
        if lower == "pause":
            state = mgr.pause(reason="user-paused")
            out = "No goal set." if state is None else f"⏸ Goal paused: {state.goal}"
            return _ok(rid, {"type": "exec", "output": out})
        if lower == "resume":
            state = mgr.resume()
            if state is None:
                return _ok(rid, {"type": "exec", "output": "No goal to resume."})
            return _ok(
                rid,
                {
                    "type": "exec",
                    "output": (
                        f"▶ Goal resumed: {state.goal}\n"
                        "Send any message to continue, or wait — I'll take the next step on the next turn."
                    ),
                },
            )
        if lower in {"clear", "stop", "done"}:
            had = mgr.has_goal()
            mgr.clear()
            return _ok(
                rid,
                {
                    "type": "exec",
                    "output": "✓ Goal cleared." if had else "No active goal.",
                },
            )

        # Otherwise — treat the remaining text as the new goal.
        try:
            state = mgr.set(arg)
        except ValueError as exc:
            return _err(rid, 4004, f"invalid goal: {exc}")

        notice = (
            f"⊙ Goal set ({state.max_turns}-turn budget): {state.goal}\n"
            "I'll keep working until the goal is done, you pause/clear it, or the budget is exhausted.\n"
            "Controls: /goal status · /goal pause · /goal resume · /goal clear"
        )
        # Send the goal text as the kickoff prompt. The TUI client sees
        # {type: send, notice, message} → renders `notice` as a sys line,
        # then submits `message` as a user turn. The post-turn judge
        # wired in _run_prompt_submit takes over from there.
        return _ok(
            rid,
            {"type": "send", "notice": notice, "message": state.goal},
        )

    if name in {"snapshot", "snap"}:
        subcommand = arg.split(maxsplit=1)[0].lower() if arg else ""
        if subcommand in {"restore", "rewind"}:
            return _ok(
                rid,
                {
                    "type": "exec",
                    "output": (
                        "/snapshot restore is blocked in the TUI because it changes "
                        "config/state on disk while the live agent has cached settings. "
                        "Run it in the classic CLI, then restart the TUI."
                    ),
                },
            )

    return _err(rid, 4018, f"not a quick/plugin/skill command: {name}")


# ── Methods: paste ────────────────────────────────────────────────────

_paste_counter = 0


@method("paste.collapse")
def _(rid, params: dict) -> dict:
    global _paste_counter
    text = params.get("text", "")
    if not text:
        return _err(rid, 4004, "empty paste")

    _paste_counter += 1
    line_count = text.count("\n") + 1
    paste_dir = _hermes_home / "pastes"
    paste_dir.mkdir(parents=True, exist_ok=True)

    from datetime import datetime

    paste_file = (
        paste_dir / f"paste_{_paste_counter}_{datetime.now().strftime('%H%M%S')}.txt"
    )
    paste_file.write_text(text, encoding="utf-8")

    placeholder = (
        f"[Pasted text #{_paste_counter}: {line_count} lines \u2192 {paste_file}]"
    )
    return _ok(
        rid, {"placeholder": placeholder, "path": str(paste_file), "lines": line_count}
    )


# ── Methods: complete ─────────────────────────────────────────────────

_FUZZY_CACHE_TTL_S = 5.0
_FUZZY_CACHE_MAX_FILES = 20000
_FUZZY_FALLBACK_EXCLUDES = frozenset(
    {
        ".git",
        ".hg",
        ".svn",
        ".next",
        ".cache",
        ".venv",
        "venv",
        "node_modules",
        "__pycache__",
        "dist",
        "build",
        "target",
        ".mypy_cache",
        ".pytest_cache",
        ".ruff_cache",
    }
)
_fuzzy_cache_lock = threading.Lock()
_fuzzy_cache: dict[str, tuple[float, list[str]]] = {}


def _list_repo_files(root: str) -> list[str]:
    """Return file paths relative to ``root``.

    Uses ``git ls-files`` from the repo top (resolved via
    ``rev-parse --show-toplevel``) so the listing covers tracked + untracked
    files anywhere in the repo, then converts each path back to be relative
    to ``root``. Files outside ``root`` (parent directories of cwd, sibling
    subtrees) are excluded so the picker stays scoped to what's reachable
    from the gateway's cwd. Falls back to a bounded ``os.walk(root)`` when
    ``root`` isn't inside a git repo. Result cached per-root for
    ``_FUZZY_CACHE_TTL_S`` so rapid keystrokes don't respawn git processes.
    """
    now = time.monotonic()
    with _fuzzy_cache_lock:
        cached = _fuzzy_cache.get(root)
        if cached and now - cached[0] < _FUZZY_CACHE_TTL_S:
            return cached[1]

    files: list[str] = []
    try:
        top_result = subprocess.run(
            ["git", "-C", root, "rev-parse", "--show-toplevel"],
            capture_output=True,
            timeout=2.0,
            check=False,
        )
        if top_result.returncode == 0:
            top = top_result.stdout.decode("utf-8", "replace").strip()
            list_result = subprocess.run(
                [
                    "git",
                    "-C",
                    top,
                    "ls-files",
                    "-z",
                    "--cached",
                    "--others",
                    "--exclude-standard",
                ],
                capture_output=True,
                timeout=2.0,
                check=False,
            )
            if list_result.returncode == 0:
                for p in list_result.stdout.decode("utf-8", "replace").split("\0"):
                    if not p:
                        continue
                    rel = os.path.relpath(os.path.join(top, p), root).replace(
                        os.sep, "/"
                    )
                    # Skip parents/siblings of cwd — keep the picker scoped
                    # to root-and-below, matching Cmd-P workspace semantics.
                    if rel.startswith("../"):
                        continue
                    files.append(rel)
                    if len(files) >= _FUZZY_CACHE_MAX_FILES:
                        break
    except (OSError, subprocess.TimeoutExpired):
        pass

    if not files:
        # Fallback walk: skip vendor/build dirs + dot-dirs so the walk stays
        # tractable. Dotfiles themselves survive — the ranker decides based
        # on whether the query starts with `.`.
        try:
            for dirpath, dirnames, filenames in os.walk(root, followlinks=False):
                dirnames[:] = [
                    d
                    for d in dirnames
                    if d not in _FUZZY_FALLBACK_EXCLUDES and not d.startswith(".")
                ]
                rel_dir = os.path.relpath(dirpath, root)
                for f in filenames:
                    rel = f if rel_dir == "." else f"{rel_dir}/{f}"
                    files.append(rel.replace(os.sep, "/"))
                    if len(files) >= _FUZZY_CACHE_MAX_FILES:
                        break
                if len(files) >= _FUZZY_CACHE_MAX_FILES:
                    break
        except OSError:
            pass

    with _fuzzy_cache_lock:
        _fuzzy_cache[root] = (now, files)

    return files


def _fuzzy_basename_rank(name: str, query: str) -> tuple[int, int] | None:
    """Rank ``name`` against ``query``; lower is better. Returns None to reject.

    Tiers (kind):
      0 — exact basename
      1 — basename prefix (e.g. `app` → `appChrome.tsx`)
      2 — word-boundary / camelCase hit (e.g. `chrome` → `appChrome.tsx`)
      3 — substring anywhere in basename
      4 — subsequence match (every query char appears in order)

    Secondary key is `len(name)` so shorter names win ties.
    """
    if not query:
        return (3, len(name))

    nl = name.lower()
    ql = query.lower()

    if nl == ql:
        return (0, len(name))

    if nl.startswith(ql):
        return (1, len(name))

    # Word-boundary split: `foo-bar_baz.qux` → ["foo","bar","baz","qux"].
    # camelCase split: `appChrome` → ["app","Chrome"]. Cheap approximation;
    # falls through to substring/subsequence if it misses.
    parts: list[str] = []
    buf = ""
    for ch in name:
        if ch in "-_." or (ch.isupper() and buf and not buf[-1].isupper()):
            if buf:
                parts.append(buf)
            buf = ch if ch not in "-_." else ""
        else:
            buf += ch
    if buf:
        parts.append(buf)
    for p in parts:
        if p.lower().startswith(ql):
            return (2, len(name))

    if ql in nl:
        return (3, len(name))

    i = 0
    for ch in nl:
        if ch == ql[i]:
            i += 1
            if i == len(ql):
                return (4, len(name))

    return None


@method("complete.path")
def _(rid, params: dict) -> dict:
    word = params.get("word", "")
    if not word:
        return _ok(rid, {"items": []})

    items: list[dict] = []
    try:
        root = _completion_cwd(params)
        is_context = word.startswith("@")
        query = word[1:] if is_context else word

        if is_context and not query:
            items = [
                {"text": "@diff", "display": "@diff", "meta": "git diff"},
                {"text": "@staged", "display": "@staged", "meta": "staged diff"},
                {"text": "@file:", "display": "@file:", "meta": "attach file"},
                {"text": "@folder:", "display": "@folder:", "meta": "attach folder"},
                {"text": "@url:", "display": "@url:", "meta": "fetch url"},
                {"text": "@git:", "display": "@git:", "meta": "git log"},
            ]
            return _ok(rid, {"items": items})

        # Accept both `@folder:path` and the bare `@folder` form so the user
        # sees directory listings as soon as they finish typing the keyword,
        # without first accepting the static `@folder:` hint.
        if is_context and query in {"file", "folder"}:
            prefix_tag, path_part = query, ""
        elif is_context and query.startswith(("file:", "folder:")):
            prefix_tag, _, tail = query.partition(":")
            path_part = tail
        else:
            prefix_tag = ""
            path_part = query if is_context else query

        # Fuzzy basename search across the repo when the user types a bare
        # name with no path separator — `@appChrome` surfaces every file
        # whose basename matches, regardless of directory depth. Matches what
        # editors like Cursor / VS Code do for Cmd-P. Path-ish queries (with
        # `/`, `./`, `~/`, `/abs`) fall through to the directory-listing
        # path so explicit navigation intent is preserved.
        if (
            is_context
            and path_part
            and len(path_part.strip()) >= 2
            and "/" not in path_part
            and prefix_tag != "folder"
        ):
            ranked: list[tuple[tuple[int, int], str, str]] = []
            for rel in _list_repo_files(root):
                basename = os.path.basename(rel)
                if basename.startswith(".") and not path_part.startswith("."):
                    continue
                rank = _fuzzy_basename_rank(basename, path_part)
                if rank is None:
                    continue
                ranked.append((rank, rel, basename))

            ranked.sort(key=lambda r: (r[0], len(r[1]), r[1]))
            tag = prefix_tag or "file"
            for _, rel, basename in ranked[:30]:
                items.append(
                    {
                        "text": f"@{tag}:{rel}",
                        "display": basename,
                        "meta": os.path.dirname(rel),
                    }
                )

            return _ok(rid, {"items": items})

        expanded = _normalize_completion_path(path_part) if path_part else "."
        if expanded == "." or not expanded:
            search_dir, match = ".", ""
        elif expanded.endswith("/"):
            search_dir, match = expanded, ""
        else:
            search_dir = os.path.dirname(expanded) or "."
            match = os.path.basename(expanded)

        search_dir = (
            search_dir if os.path.isabs(search_dir) else os.path.join(root, search_dir)
        )
        if not os.path.isdir(search_dir):
            return _ok(rid, {"items": []})

        want_dir = prefix_tag == "folder"
        match_lower = match.lower()
        for entry in sorted(os.listdir(search_dir)):
            if match and not entry.lower().startswith(match_lower):
                continue
            if is_context and entry in _FUZZY_FALLBACK_EXCLUDES:
                continue
            if is_context and not prefix_tag and entry.startswith("."):
                continue
            full = os.path.join(search_dir, entry)
            is_dir = os.path.isdir(full)
            # Explicit `@folder:` / `@file:` — honour the user's filter.  Skip
            # the opposite kind instead of auto-rewriting the completion tag,
            # which used to defeat the prefix and let `@folder:` list files.
            if prefix_tag and want_dir != is_dir:
                continue
            rel = os.path.relpath(full, root).replace(os.sep, "/")
            suffix = "/" if is_dir else ""

            if is_context and prefix_tag:
                text = f"@{prefix_tag}:{rel}{suffix}"
            elif is_context:
                kind = "folder" if is_dir else "file"
                text = f"@{kind}:{rel}{suffix}"
            elif word.startswith("~"):
                text = "~/" + os.path.relpath(full, os.path.expanduser("~")) + suffix
            elif word.startswith("./"):
                text = "./" + rel + suffix
            else:
                text = rel + suffix

            items.append(
                {
                    "text": text,
                    "display": entry + suffix,
                    "meta": "dir" if is_dir else "",
                }
            )
            if len(items) >= 30:
                break
    except Exception as e:
        return _err(rid, 5021, str(e))

    return _ok(rid, {"items": items})


def _details_completion_item(value: str, meta: str = "") -> dict:
    return {"text": value, "display": value, "meta": meta}


def _details_root_completion_item(
    value: str, meta: str, needs_leading_space: bool
) -> dict:
    return _details_completion_item(
        f" {value}" if needs_leading_space else value,
        meta,
    )


def _details_completions(text: str) -> list[dict] | None:
    if not text.lower().startswith("/details"):
        return None

    stripped = text.strip()
    if stripped and not "/details".startswith(stripped.lower().split()[0]):
        return None

    body = text[len("/details") :]
    if body.startswith(" "):
        body = body[1:]
    parts = body.split()
    has_trailing_space = text.endswith(" ")
    sections = ("thinking", "tools", "subagents", "activity")
    modes = ("hidden", "collapsed", "expanded")

    if not body or (len(parts) == 0 and has_trailing_space):
        return [
            *[
                _details_root_completion_item(
                    mode, "global mode", not has_trailing_space
                )
                for mode in modes
            ],
            _details_root_completion_item(
                "cycle", "cycle global mode", not has_trailing_space
            ),
            *[
                _details_root_completion_item(
                    section, "section override", not has_trailing_space
                )
                for section in sections
            ],
        ]

    if len(parts) == 1 and not has_trailing_space:
        prefix = parts[0].lower()
        candidates = [*modes, "cycle", *sections]
        return [
            _details_completion_item(
                candidate,
                (
                    "section override"
                    if candidate in sections
                    else "cycle global mode" if candidate == "cycle" else "global mode"
                ),
            )
            for candidate in candidates
            if candidate.startswith(prefix) and candidate != prefix
        ]

    if len(parts) == 1 and has_trailing_space and parts[0].lower() in sections:
        return [
            *[
                _details_completion_item(mode, f"set {parts[0].lower()}")
                for mode in modes
            ],
            _details_completion_item("reset", f"clear {parts[0].lower()} override"),
        ]

    if len(parts) == 2 and not has_trailing_space and parts[0].lower() in sections:
        prefix = parts[1].lower()
        return [
            _details_completion_item(
                candidate,
                (
                    f"clear {parts[0].lower()} override"
                    if candidate == "reset"
                    else f"set {parts[0].lower()}"
                ),
            )
            for candidate in (*modes, "reset")
            if candidate.startswith(prefix) and candidate != prefix
        ]

    return []


@method("complete.slash")
def _(rid, params: dict) -> dict:
    text = params.get("text", "")
    if not text.startswith("/"):
        return _ok(rid, {"items": []})

    try:
        from hermes_cli.commands import SlashCommandCompleter
        from prompt_toolkit.document import Document
        from prompt_toolkit.formatted_text import to_plain_text

        from agent.skill_commands import get_skill_commands
        from agent.skill_bundles import get_skill_bundles

        completer = SlashCommandCompleter(
            skill_commands_provider=lambda: get_skill_commands(),
            skill_bundles_provider=lambda: get_skill_bundles(),
        )
        doc = Document(text, len(text))
        items = [
            {
                "text": c.text,
                # prompt_toolkit gives us FormattedText (a list of (style,
                # text) tuples) for display/display_meta. Serialize both as
                # plain strings — the TUI's CompletionItem.display contract
                # is a string, and sending the raw list trips Ink's row
                # layout into 1-char truncation of the next column.
                "display": to_plain_text(c.display) if c.display else c.text,
                "meta": to_plain_text(c.display_meta) if c.display_meta else "",
            }
            for c in completer.get_completions(doc, None)
        ][:30]
        text_lower = text.lower()
        extras = [
            {
                "text": "/compact",
                "display": "/compact",
                "meta": "Toggle compact display mode",
            },
            {
                "text": "/details",
                "display": "/details",
                "meta": "Control agent detail visibility",
            },
            {
                "text": "/logs",
                "display": "/logs",
                "meta": "Show recent gateway log lines",
            },
            {
                "text": "/mouse",
                "display": "/mouse",
                "meta": "Set mouse tracking preset [on|off|toggle|wheel|buttons|all]",
            },
        ]
        for extra in extras:
            if extra["text"].startswith(text_lower) and not any(
                item["text"] == extra["text"] for item in items
            ):
                items.append(extra)

        details_items = _details_completions(text)
        if details_items is not None:
            return _ok(
                rid,
                {
                    "items": details_items,
                    "replace_from": text.rfind(" ") + 1 if " " in text else len(text),
                },
            )

        return _ok(
            rid,
            {"items": items, "replace_from": text.rfind(" ") + 1 if " " in text else 1},
        )
    except Exception as e:
        return _err(rid, 5020, str(e))


@method("model.options")
def _(rid, params: dict) -> dict:
    try:
        from hermes_cli.inventory import build_models_payload, load_picker_context

        session = _sessions.get(params.get("session_id", ""))
        agent = session.get("agent") if session else None
        # Layer agent-session state on top of disk config — once an agent
        # is spawned, IT owns the live provider/model/base_url. Empty
        # agent attributes must NOT clobber disk config (with_overrides
        # is truthy-only).
        ctx = load_picker_context().with_overrides(
            current_provider=getattr(agent, "provider", "") if agent else "",
            current_model=(
                (getattr(agent, "model", "") if agent else "") or _resolve_model()
            ),
            current_base_url=getattr(agent, "base_url", "") if agent else "",
        )
        # picker_hints + canonical_order produce the TUI's required shape:
        # `authenticated`/`auth_type`/`key_env`/`warning` per row, in
        # CANONICAL_PROVIDERS declaration order. include_unconfigured=True
        # so the picker can show the full provider universe (with the
        # setup-hint warning attached) instead of only authed rows.
        # Curated model lists are preserved — list_authenticated_providers
        # populates `models` from the curated catalog, not provider_model_ids
        # (which would pull non-agentic models like TTS/embeddings/etc.).
        payload = build_models_payload(
            ctx,
            include_unconfigured=True,
            picker_hints=True,
            canonical_order=True,
            pricing=True,
            max_models=50,
        )
        return _ok(rid, payload)
    except Exception as e:
        return _err(rid, 5033, str(e))


@method("model.save_key")
def _(rid, params: dict) -> dict:
    """Save an API key for a provider, then return its refreshed model list.

    Params:
        slug: provider slug (e.g. "deepseek", "xai")
        api_key: the key value to save

    Returns the provider dict with models populated (same shape as
    model.options entries) on success.
    """
    try:
        from hermes_cli.auth import PROVIDER_REGISTRY
        from hermes_cli.config import is_managed, save_env_value
        from hermes_cli.inventory import build_models_payload, load_picker_context

        slug = (params.get("slug") or "").strip()
        api_key = (params.get("api_key") or "").strip()
        if not slug or not api_key:
            return _err(rid, 4001, "slug and api_key are required")

        if is_managed():
            return _err(rid, 4006, "managed install — credentials are read-only")

        pconfig = PROVIDER_REGISTRY.get(slug)
        if not pconfig:
            return _err(rid, 4002, f"unknown provider: {slug}")
        if pconfig.auth_type != "api_key":
            return _err(
                rid,
                4003,
                f"{pconfig.name} uses {pconfig.auth_type} auth — "
                f"run `hermes model` to configure",
            )
        if not pconfig.api_key_env_vars:
            return _err(rid, 4004, f"no env var defined for {pconfig.name}")

        # Save the key to ~/.hermes/.env
        env_var = pconfig.api_key_env_vars[0]
        save_env_value(env_var, api_key)
        # Also set in current process so the refreshed inventory sees it.
        import os

        os.environ[env_var] = api_key

        # Refresh provider data via the shared inventory builder so this
        # surface stays in lock-step with model.options + dashboard
        # /api/model/options. picker_hints=True ensures the returned row
        # carries `authenticated` for the TUI frontend.
        session = _sessions.get(params.get("session_id", ""))
        agent = session.get("agent") if session else None
        ctx = load_picker_context().with_overrides(
            current_provider=getattr(agent, "provider", "") if agent else "",
            current_model=(
                (getattr(agent, "model", "") if agent else "") or _resolve_model()
            ),
            current_base_url=getattr(agent, "base_url", "") if agent else "",
        )
        payload = build_models_payload(
            ctx, picker_hints=True, max_models=50,
        )
        provider_data = next(
            (p for p in payload["providers"] if p["slug"] == slug), None
        )
        if provider_data is None:
            # Key was saved but provider didn't appear — still return success.
            provider_data = {
                "slug": slug,
                "name": pconfig.name,
                "is_current": False,
                "models": [],
                "total_models": 0,
                "authenticated": True,
            }
        # picker_hints sets `authenticated` from the row state, but the
        # synthetic fallback above doesn't go through that path.
        provider_data["authenticated"] = True
        return _ok(rid, {"provider": provider_data})
    except Exception as e:
        return _err(rid, 5034, str(e))


@method("model.disconnect")
def _(rid, params: dict) -> dict:
    """Remove credentials for a provider.

    Params:
        slug: provider slug (e.g. "deepseek", "xai")

    Returns success status and the provider's slug.
    """
    try:
        from hermes_cli.auth import PROVIDER_REGISTRY, clear_provider_auth
        from hermes_cli.config import remove_env_value

        slug = (params.get("slug") or "").strip()
        if not slug:
            return _err(rid, 4001, "slug is required")

        pconfig = PROVIDER_REGISTRY.get(slug)
        cleared_env = False
        cleared_auth = False

        # Remove API key env vars from .env and process
        if pconfig and pconfig.api_key_env_vars:
            for ev in pconfig.api_key_env_vars:
                if remove_env_value(ev):
                    cleared_env = True

        # Clear OAuth / credential pool state
        cleared_auth = clear_provider_auth(slug)

        if not cleared_env and not cleared_auth:
            return _err(rid, 4005, f"no credentials found for {slug}")

        provider_name = pconfig.name if pconfig else slug
        return _ok(
            rid,
            {
                "slug": slug,
                "name": provider_name,
                "disconnected": True,
            },
        )
    except Exception as e:
        return _err(rid, 5035, str(e))


# ── Methods: slash.exec ──────────────────────────────────────────────


def _mirror_slash_side_effects(sid: str, session: dict, command: str) -> str:
    """Apply side effects that must also hit the gateway's live agent."""
    parts = command.lstrip("/").split(None, 1)
    if not parts:
        return ""
    name, arg, agent = (
        parts[0],
        (parts[1].strip() if len(parts) > 1 else ""),
        session.get("agent"),
    )

    # Reject agent-mutating commands during an in-flight turn.  These
    # all do read-then-mutate on live agent/session state that the
    # worker thread running agent.run_conversation is using.  Parity
    # with the session.compress / session.undo guards and the gateway
    # runner's running-agent /model guard.
    _MUTATES_WHILE_RUNNING = {"model", "personality", "prompt", "compress"}
    if name in _MUTATES_WHILE_RUNNING and session.get("running"):
        return f"session busy — /interrupt the current turn before running /{name}"

    try:
        if name == "model" and arg and agent:
            result = _apply_model_switch(sid, session, arg)
            return result.get("warning", "")
        elif name == "personality" and arg and agent:
            pname, new_prompt = _validate_personality(arg, _load_cfg())
            _apply_personality_to_session(sid, session, new_prompt, pname)
        elif name == "prompt" and agent:
            cfg = _load_cfg()
            new_prompt = _prompt_text((cfg.get("agent") or {}).get("system_prompt", ""))
            agent.ephemeral_system_prompt = new_prompt or None
            agent._cached_system_prompt = None
        elif name == "compress" and agent:
            _compress_session_history(session, arg)
            _sync_session_key_after_compress(sid, session)
            _emit("session.info", sid, _session_info(agent, session))
        elif name == "fast" and agent:
            mode = arg.lower()
            if mode in {"fast", "on"}:
                agent.service_tier = "priority"
            elif mode in {"normal", "off"}:
                agent.service_tier = None
            _emit("session.info", sid, _session_info(agent, session))
        elif name == "reload-mcp" and agent and hasattr(agent, "reload_mcp_tools"):
            agent.reload_mcp_tools()
        elif name == "stop":
            from tools.process_registry import process_registry

            process_registry.kill_all()
    except Exception as e:
        return f"live session sync failed: {e}"
    return ""


@method("slash.exec")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err

    cmd = params.get("command", "").strip()
    if not cmd:
        return _err(rid, 4004, "empty command")

    # Skill slash commands and _pending_input commands must NOT go through the
    # slash worker — see _PENDING_INPUT_COMMANDS definition above. Plugin
    # commands must also avoid the worker, but unlike skills/pending-input they
    # still return normal slash.exec output so the TUI keeps the pager path.
    _cmd_text = cmd.lstrip("/") if cmd.startswith("/") else cmd
    _cmd_parts = _cmd_text.split(maxsplit=1)
    _cmd_base = (_cmd_parts[0] if _cmd_parts else "").lower()
    _cmd_arg = _cmd_parts[1] if len(_cmd_parts) > 1 else ""

    if _cmd_base in _PENDING_INPUT_COMMANDS:
        return _err(
            rid, 4018, f"pending-input command: use command.dispatch for /{_cmd_base}"
        )

    if _cmd_base in _WORKER_BLOCKED_COMMANDS:
        subcommand = _cmd_arg.split(maxsplit=1)[0].lower() if _cmd_arg else ""
        if subcommand in {"restore", "rewind"}:
            return _err(
                rid,
                4018,
                "snapshot restore mutates live config/state; use command.dispatch for /snapshot restore",
            )

    try:
        from agent.skill_commands import get_skill_commands

        _cmd_key = f"/{_cmd_base}"
        if _cmd_key in get_skill_commands():
            return _err(
                rid, 4018, f"skill command: use command.dispatch for {_cmd_key}"
            )
    except Exception:
        pass

    plugin_handler = None
    resolve_plugin_command_result = None
    if _cmd_base:
        try:
            from hermes_cli.plugins import (
                get_plugin_command_handler,
                resolve_plugin_command_result,
            )

            plugin_handler = get_plugin_command_handler(_cmd_base)
        except Exception:
            plugin_handler = None
            resolve_plugin_command_result = None

    if plugin_handler and resolve_plugin_command_result:
        try:
            result = resolve_plugin_command_result(plugin_handler(_cmd_arg))
            return _ok(rid, {"output": str(result or "(no output)")})
        except Exception as e:
            return _ok(rid, {"output": f"Plugin command error: {e}"})

    worker = session.get("slash_worker")
    if not worker:
        try:
            worker = _SlashWorker(
                session["session_key"],
                getattr(session.get("agent"), "model", _resolve_model()),
            )
            session["slash_worker"] = worker
        except Exception as e:
            return _err(rid, 5030, f"slash worker start failed: {e}")

    try:
        output = worker.run(cmd)
        warning = _mirror_slash_side_effects(params.get("session_id", ""), session, cmd)
        payload = {"output": output or "(no output)"}
        if warning:
            payload["warning"] = warning
        return _ok(rid, payload)
    except Exception as e:
        try:
            worker.close()
        except Exception:
            pass
        session["slash_worker"] = None
        return _err(rid, 5030, str(e))


# ── Methods: voice ───────────────────────────────────────────────────


_voice_sid_lock = threading.Lock()
_voice_event_sid: str = ""


def _voice_emit(event: str, payload: dict | None = None) -> None:
    """Emit a voice event toward the session that most recently turned the
    mode on. Voice is process-global (one microphone), so there's only ever
    one sid to target; the TUI handler treats an empty sid as "active
    session". Kept separate from _emit to make the lack of per-call sid
    argument explicit."""
    with _voice_sid_lock:
        sid = _voice_event_sid
    _emit(event, sid, payload)


def _voice_mode_enabled() -> bool:
    """Current voice-mode flag (runtime-only, CLI parity).

    cli.py initialises ``_voice_mode = False`` at startup and only flips
    it via ``/voice on``; it never reads a persisted enable bit from
    config.yaml.  We match that: no config lookup, env var only.  This
    avoids the TUI auto-starting in REC the next time the user opens it
    just because they happened to enable voice in a prior session.
    """
    return os.environ.get("HERMES_VOICE", "").strip() == "1"


def _voice_tts_enabled() -> bool:
    """Whether agent replies should be spoken back via TTS (runtime only)."""
    return os.environ.get("HERMES_VOICE_TTS", "").strip() == "1"


def _voice_cfg_dict() -> dict:
    """Shape-safe accessor for the ``voice:`` block in config.yaml.

    ``_load_cfg()`` returns raw ``yaml.safe_load()`` output, so both the
    root AND ``voice`` may be any YAML scalar / list / None. A hand-edit
    like ``voice: true`` or a malformed top-level config that parses to
    a scalar would otherwise break ``.get("…")`` and take every
    ``voice.*`` branch down with it (Copilot round-3..7 review on
    #19835). Coerce through ``isinstance`` at every level so malformed
    config falls back to an empty dict instead of crashing /voice.
    """
    cfg = _load_cfg()
    voice_cfg = cfg.get("voice") if isinstance(cfg, dict) else None

    return voice_cfg if isinstance(voice_cfg, dict) else {}


def _voice_record_key() -> str:
    """Current ``voice.record_key`` value, documented default on error."""
    record_key = _voice_cfg_dict().get("record_key")

    return str(record_key) if isinstance(record_key, str) and record_key else "ctrl+b"


@method("voice.toggle")
def _(rid, params: dict) -> dict:
    """CLI parity for the ``/voice`` slash command.

    Subcommands:

    * ``status`` — report mode + TTS flags (default when action is unknown).
    * ``on`` / ``off`` — flip voice *mode* (the umbrella bit). Turning it
      off also tears down any active continuous recording loop. Does NOT
      start recording on its own; recording is driven by ``voice.record``
      (Ctrl+B) after mode is on, matching cli.py's enable/Ctrl+B split.
    * ``tts`` — toggle speech-output of agent replies. Requires mode on
      (mirrors CLI's _toggle_voice_tts guard).
    """
    action = params.get("action", "status")

    if action == "status":
        # Mirror CLI's _show_voice_status: include STT/TTS provider
        # availability so the user can tell at a glance *why* voice mode
        # isn't working ("STT provider: MISSING ..." is the common case).
        # ``record_key`` mirrors the configured ``voice.record_key`` so the
        # TUI can both bind it (frontend ``isVoiceToggleKey``) and display
        # it in /voice status — previously the TUI hardcoded Ctrl+B and
        # ignored the config (#18994).
        payload: dict = {
            "enabled": _voice_mode_enabled(),
            "record_key": _voice_record_key(),
            "tts": _voice_tts_enabled(),
        }
        try:
            from tools.voice_mode import check_voice_requirements

            reqs = check_voice_requirements()
            payload["available"] = bool(reqs.get("available"))
            payload["audio_available"] = bool(reqs.get("audio_available"))
            payload["stt_available"] = bool(reqs.get("stt_available"))
            payload["details"] = reqs.get("details") or ""
        except Exception as e:
            # check_voice_requirements pulls optional transcription deps —
            # swallow so /voice status always returns something useful.
            logger.warning("voice.toggle status: requirements probe failed: %s", e)

        return _ok(rid, payload)

    if action in {"on", "off"}:
        enabled = action == "on"
        # Runtime-only flag (CLI parity) — no _write_config_key, so the
        # next TUI launch starts with voice OFF instead of auto-REC from a
        # persisted stale toggle.
        os.environ["HERMES_VOICE"] = "1" if enabled else "0"

        if not enabled:
            # Disabling the mode must tear the continuous loop down; the
            # loop holds the microphone and would otherwise keep running.
            try:
                from hermes_cli.voice import stop_continuous

                stop_continuous()
            except ImportError:
                pass
            except Exception as e:
                logger.warning("voice: stop_continuous failed during toggle off: %s", e)

            # Clear TTS so it can be toggled independently after voice is off.
            os.environ["HERMES_VOICE_TTS"] = "0"

        return _ok(
            rid,
            {
                "enabled": enabled,
                "record_key": _voice_record_key(),
                "tts": _voice_tts_enabled(),
            },
        )

    if action == "tts":
        if not _voice_mode_enabled():
            return _err(rid, 4014, "enable voice mode first: /voice on")
        new_value = not _voice_tts_enabled()
        # Runtime-only flag (CLI parity) — see voice.toggle on/off above.
        os.environ["HERMES_VOICE_TTS"] = "1" if new_value else "0"
        # Include ``record_key`` on every branch so a /voice tts toggle
        # doesn't reset the TUI's cached shortcut to the default when a
        # user has a custom binding configured (Copilot review, round 2
        # on #19835). Keeps parity with the status/on/off branches above.
        return _ok(
            rid,
            {
                "enabled": True,
                "record_key": _voice_record_key(),
                "tts": new_value,
            },
        )

    return _err(rid, 4013, f"unknown voice action: {action}")


@method("voice.record")
def _(rid, params: dict) -> dict:
    """VAD-bounded push-to-talk capture, CLI-parity.

    ``start`` begins one VAD-bounded capture and emits ``voice.transcript``
    after silence stops the recorder. ``stop`` forces transcription of the
    active buffer, matching classic CLI push-to-talk. The voice wrapper retains
    no-speech counts across single-shot starts, so three consecutive silent
    captures emit ``voice.transcript`` with ``no_speech_limit=True``.
    """
    action = params.get("action", "start")

    if action not in {"start", "stop"}:
        return _err(rid, 4019, f"unknown voice action: {action}")

    try:
        if action == "start":
            if not _voice_mode_enabled():
                return _err(rid, 4015, "voice mode is off — enable with /voice on")

            with _voice_sid_lock:
                global _voice_event_sid
                _voice_event_sid = params.get("session_id") or _voice_event_sid

            from hermes_cli.voice import start_continuous

            # Shape-safe lookups: malformed ``voice:`` YAML (bool/scalar/list)
            # must not crash /voice with a 5025 — fall back to VAD defaults.
            #
            # Exclude ``bool`` from the numeric check since Python's bool is
            # a subclass of int — a hand-edit like ``silence_threshold: true``
            # would otherwise forward as ``1`` instead of falling back to
            # the documented 200 / 3.0 defaults (Copilot round-12 on #19835).
            voice_cfg = _voice_cfg_dict()
            threshold = voice_cfg.get("silence_threshold")
            duration = voice_cfg.get("silence_duration")
            safe_threshold = (
                threshold
                if isinstance(threshold, (int, float))
                and not isinstance(threshold, bool)
                else 200
            )
            safe_duration = (
                duration
                if isinstance(duration, (int, float)) and not isinstance(duration, bool)
                else 3.0
            )
            started = start_continuous(
                on_transcript=lambda t: _voice_emit("voice.transcript", {"text": t}),
                on_status=lambda s: _voice_emit("voice.status", {"state": s}),
                on_silent_limit=lambda: _voice_emit(
                    "voice.transcript", {"no_speech_limit": True}
                ),
                silence_threshold=safe_threshold,
                silence_duration=safe_duration,
                auto_restart=False,
            )
            if started is False:
                return _ok(rid, {"status": "busy"})
            return _ok(rid, {"status": "recording"})

        # action == "stop"
        with _voice_sid_lock:
            _voice_event_sid = params.get("session_id") or _voice_event_sid

        from hermes_cli.voice import stop_continuous

        stop_continuous(force_transcribe=True)
        return _ok(rid, {"status": "stopped"})
    except ImportError:
        return _err(
            rid, 5025, "voice module not available — install audio dependencies"
        )
    except Exception as e:
        return _err(rid, 5025, str(e))


@method("voice.tts")
def _(rid, params: dict) -> dict:
    text = params.get("text", "")
    if not text:
        return _err(rid, 4020, "text required")
    try:
        from hermes_cli.voice import speak_text

        threading.Thread(target=speak_text, args=(text,), daemon=True).start()
        return _ok(rid, {"status": "speaking"})
    except ImportError:
        return _err(rid, 5026, "voice module not available")
    except Exception as e:
        return _err(rid, 5026, str(e))


# ── Methods: insights ────────────────────────────────────────────────


@method("insights.get")
def _(rid, params: dict) -> dict:
    days = params.get("days", 30)
    db = _get_db()
    if db is None:
        return _db_unavailable_error(rid, code=5017)
    try:
        cutoff = time.time() - days * 86400
        rows = [
            s
            for s in db.list_sessions_rich(limit=500)
            if (s.get("started_at") or 0) >= cutoff
        ]
        return _ok(
            rid,
            {
                "days": days,
                "sessions": len(rows),
                "messages": sum(s.get("message_count", 0) for s in rows),
            },
        )
    except Exception as e:
        return _err(rid, 5017, str(e))


# ── Methods: rollback ────────────────────────────────────────────────


@method("rollback.list")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    try:

        def go(mgr, cwd):
            if not mgr.enabled:
                return _ok(rid, {"enabled": False, "checkpoints": []})
            return _ok(
                rid,
                {
                    "enabled": True,
                    "checkpoints": [
                        {
                            "hash": c.get("hash", ""),
                            "timestamp": c.get("timestamp", ""),
                            "message": c.get("message", ""),
                        }
                        for c in mgr.list_checkpoints(cwd)
                    ],
                },
            )

        return _with_checkpoints(session, go)
    except Exception as e:
        return _err(rid, 5020, str(e))


@method("rollback.restore")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    target = params.get("hash", "")
    file_path = params.get("file_path", "")
    if not target:
        return _err(rid, 4014, "hash required")
    # Full-history rollback mutates session history.  Rejecting during
    # an in-flight turn prevents prompt.submit from silently dropping
    # the agent's output (version mismatch path) or clobbering the
    # rollback (version-matches path).  A file-scoped rollback only
    # touches disk, so we allow it.
    if not file_path and session.get("running"):
        return _err(
            rid,
            4009,
            "session busy — /interrupt the current turn before full rollback.restore",
        )
    try:

        def go(mgr, cwd):
            resolved = _resolve_checkpoint_hash(mgr, cwd, target)
            result = mgr.restore(cwd, resolved, file_path=file_path or None)
            if result.get("success") and not file_path:
                removed = 0
                with session["history_lock"]:
                    history = session.get("history", [])
                    while history and history[-1].get("role") in {"assistant", "tool"}:
                        history.pop()
                        removed += 1
                    if history and history[-1].get("role") == "user":
                        history.pop()
                        removed += 1
                    if removed:
                        session["history_version"] = (
                            int(session.get("history_version", 0)) + 1
                        )
                result["history_removed"] = removed
            return result

        return _ok(rid, _with_checkpoints(session, go))
    except Exception as e:
        return _err(rid, 5021, str(e))


@method("rollback.diff")
def _(rid, params: dict) -> dict:
    session, err = _sess(params, rid)
    if err:
        return err
    target = params.get("hash", "")
    if not target:
        return _err(rid, 4014, "hash required")
    try:
        r = _with_checkpoints(
            session,
            lambda mgr, cwd: mgr.diff(cwd, _resolve_checkpoint_hash(mgr, cwd, target)),
        )
        raw = r.get("diff", "")[:4000]
        payload = {"stat": r.get("stat", ""), "diff": raw}
        rendered = render_diff(raw, session.get("cols", 80))
        if rendered:
            payload["rendered"] = rendered
        return _ok(rid, payload)
    except Exception as e:
        return _err(rid, 5022, str(e))


# ── Methods: browser / plugins / cron / skills ───────────────────────


def _resolve_browser_cdp_url() -> str:
    """Return the configured browser CDP override without network I/O.

    ``/browser status`` must be fast — calling
    ``tools.browser_tool._get_cdp_override`` would invoke
    ``_resolve_cdp_override``, which performs an HTTP probe to
    ``.../json/version`` for discovery-style URLs.  That probe has
    a multi-second timeout and would block the TUI on a slow or
    unreachable host even though status only needs to report whether
    an override is set.

    Mirrors the env/config precedence of ``_get_cdp_override`` (env
    var first, then ``browser.cdp_url`` from config.yaml) without the
    websocket-resolution step, so the answer reflects user intent
    even when the configured host is not currently reachable.  The
    actual WS normalization happens in ``browser_navigate`` on the
    next tool call.
    """
    env_url = os.environ.get("BROWSER_CDP_URL", "").strip()
    if env_url:
        return env_url
    try:
        from hermes_cli.config import read_raw_config

        cfg = read_raw_config()
        browser_cfg = cfg.get("browser", {}) if isinstance(cfg, dict) else {}
        if isinstance(browser_cfg, dict):
            return str(browser_cfg.get("cdp_url", "") or "").strip()
    except Exception:
        pass
    return ""


def _is_default_local_cdp(parsed) -> bool:
    """Match the discovery-style local default; never the concrete WS form.

    A user-supplied ``ws://127.0.0.1:9222/devtools/browser/<id>`` is a
    real, connectable endpoint — collapsing it to bare ``http://...:9222``
    would strip the path and break the connect.
    """
    try:
        port = parsed.port or 80
    except ValueError:
        return False

    discovery_path = parsed.path in {"", "/", "/json", "/json/version"}
    return (
        parsed.scheme in {"http", "ws"}
        and parsed.hostname in {"127.0.0.1", "localhost"}
        and port == 9222
        and discovery_path
    )


def _http_ok(url: str, timeout: float) -> bool:
    import urllib.request

    try:
        with urllib.request.urlopen(url, timeout=timeout) as resp:
            return 200 <= getattr(resp, "status", 200) < 300
    except Exception:
        return False


def _probe_urls(parsed) -> list[str]:
    scheme = {"ws": "http", "wss": "https"}.get(parsed.scheme, parsed.scheme)
    root = f"{scheme}://{parsed.netloc}".rstrip("/")
    return [f"{root}/json/version", f"{root}/json"]


def _normalize_cdp_url(parsed) -> str:
    # Concrete ``/devtools/browser/<id>`` endpoints (Browserbase et al.)
    # are connectable as-is. Discovery-style inputs collapse to bare
    # ``scheme://host:port`` so ``_resolve_cdp_override`` can append
    # ``/json/version`` later without doubling the path.
    if parsed.path.startswith("/devtools/browser/"):
        return parsed.geturl()
    return parsed._replace(path="", params="", query="", fragment="").geturl()


def _failure_messages(url: str, port: int, system: str) -> list[str]:
    from hermes_cli.browser_connect import manual_chrome_debug_command

    command = manual_chrome_debug_command(port, system)
    hint = (
        ["Start a Chromium-family browser with remote debugging, then retry /browser connect:", command]
        if command
        else [
            "No supported Chromium-family browser executable was found in this environment.",
            f"Install one or start a Chromium-family browser with --remote-debugging-port={port}, then retry /browser connect.",
        ]
    )
    return [
        f"Browser CDP is not reachable at {url}.",
        *hint,
        "Browser not connected — start a Chromium-family browser with remote debugging and retry /browser connect",
    ]


@method("browser.manage")
def _(rid, params: dict) -> dict:
    action = params.get("action", "status")

    if action == "status":
        url = _resolve_browser_cdp_url()
        return _ok(rid, {"connected": bool(url), "url": url})

    if action == "disconnect":
        return _browser_disconnect(rid)

    if action != "connect":
        return _err(rid, 4015, f"unknown action: {action}")

    return _browser_connect(rid, params)


def _browser_connect(rid, params: dict) -> dict:
    import platform

    from hermes_cli.browser_connect import DEFAULT_BROWSER_CDP_URL
    from tools.browser_tool import cleanup_all_browsers
    from urllib.parse import urlparse

    raw_url = params.get("url")
    if raw_url is not None and not isinstance(raw_url, str):
        return _err(
            rid, 4015, f"browser url must be a string, got {type(raw_url).__name__}"
        )
    url = (raw_url or "").strip() or DEFAULT_BROWSER_CDP_URL

    sid = params.get("session_id") or ""
    system = platform.system()
    messages: list[str] = []

    def announce(message: str, *, level: str = "info") -> None:
        messages.append(message)
        # Without a session id the TUI prints `messages` from the
        # response; emitting an event would double-render. Only stream
        # progress when there's a real session to scope it to.
        if sid:
            _emit("browser.progress", sid, {"message": message, "level": level})

    parsed = urlparse(url if "://" in url else f"http://{url}")
    if parsed.scheme not in {"http", "https", "ws", "wss"}:
        return _err(rid, 4015, f"unsupported browser url: {url}")
    if not parsed.hostname:
        return _err(rid, 4015, f"missing host in browser url: {url}")
    try:
        port = parsed.port or (443 if parsed.scheme in {"https", "wss"} else 80)
    except ValueError:
        return _err(rid, 4015, f"invalid port in browser url: {url}")

    # Always normalize default-local to 127.0.0.1:9222 so downstream
    # comparisons + messaging match what we'll actually persist.
    if _is_default_local_cdp(parsed):
        url = DEFAULT_BROWSER_CDP_URL
        parsed = urlparse(url)
        port = parsed.port or 9222

    try:
        # ws[s]://.../devtools/browser/<id> endpoints (hosted CDP
        # providers) don't serve the HTTP discovery path; just check
        # TCP-level reachability and let browser_navigate handshake.
        if parsed.scheme in {"ws", "wss"} and parsed.path.startswith(
            "/devtools/browser/"
        ):
            import socket

            try:
                with socket.create_connection((parsed.hostname, port), timeout=2.0):
                    pass
            except OSError as e:
                return _err(rid, 5031, f"could not reach browser CDP at {url}: {e}")
        else:
            probes = _probe_urls(parsed)
            ok = any(_http_ok(p, timeout=2.0) for p in probes)

            if not ok and _is_default_local_cdp(parsed):
                from hermes_cli.browser_connect import try_launch_chrome_debug

                announce(
                    "Chromium-family browser isn't running with remote debugging — attempting to launch..."
                )

                if try_launch_chrome_debug(port, system):
                    for _ in range(20):
                        time.sleep(0.5)
                        if any(_http_ok(p, timeout=1.0) for p in probes):
                            ok = True
                            break

                if ok:
                    announce(f"Chromium-family browser launched and listening on port {port}")
                else:
                    for line in _failure_messages(url, port, system)[1:]:
                        announce(line, level="error")
                    return _ok(
                        rid, {"connected": False, "url": url, "messages": messages}
                    )
            elif not ok:
                return _err(rid, 5031, f"could not reach browser CDP at {url}")
            elif _is_default_local_cdp(parsed):
                announce(f"Chromium-family browser is already listening on port {port}")

        normalized = _normalize_cdp_url(parsed)

        # Order matters: reap sessions BEFORE publishing the new env
        # so an in-flight tool call sees the old supervisor closed,
        # then again AFTER so the default task's cached supervisor
        # is drained against the new URL.
        cleanup_all_browsers()
        os.environ["BROWSER_CDP_URL"] = normalized
        cleanup_all_browsers()
    except Exception as e:
        return _err(rid, 5031, str(e))

    payload: dict[str, object] = {"connected": True, "url": normalized}
    if messages:
        payload["messages"] = messages
    return _ok(rid, payload)


def _browser_disconnect(rid) -> dict:
    # Reap, drop the env override, reap again — closes the same swap
    # window covered by ``_browser_connect``.
    def reap() -> None:
        try:
            from tools.browser_tool import cleanup_all_browsers

            cleanup_all_browsers()
        except Exception:
            pass

    reap()
    os.environ.pop("BROWSER_CDP_URL", None)
    reap()
    return _ok(rid, {"connected": False})


@method("plugins.list")
def _(rid, params: dict) -> dict:
    try:
        from hermes_cli.plugins import get_plugin_manager

        return _ok(
            rid,
            {
                "plugins": [
                    {
                        "name": n,
                        "version": getattr(i, "version", "?"),
                        "enabled": getattr(i, "enabled", True),
                    }
                    for n, i in get_plugin_manager()._plugins.items()
                ]
            },
        )
    except Exception as e:
        return _err(rid, 5032, str(e))


@method("config.show")
def _(rid, params: dict) -> dict:
    try:
        cfg = _load_cfg()
        model = _resolve_model()
        api_key = os.environ.get("HERMES_API_KEY", "") or cfg.get("api_key", "")
        masked = f"****{api_key[-4:]}" if len(api_key) > 4 else "(not set)"
        base_url = os.environ.get("HERMES_BASE_URL", "") or cfg.get("base_url", "")

        sections = [
            {
                "title": "Model",
                "rows": [
                    ["Model", model],
                    ["Base URL", base_url or "(default)"],
                    ["API Key", masked],
                ],
            },
            {
                "title": "Agent",
                "rows": [
                    ["Max Turns", str(_cfg_max_turns(cfg, 90))],
                    ["Toolsets", ", ".join(cfg.get("enabled_toolsets", [])) or "all"],
                    ["Verbose", str(cfg.get("verbose", False))],
                ],
            },
            {
                "title": "Environment",
                "rows": [
                    ["Working Dir", os.getcwd()],
                    ["Config File", str(_hermes_home / "config.yaml")],
                ],
            },
        ]
        return _ok(rid, {"sections": sections})
    except Exception as e:
        return _err(rid, 5030, str(e))


@method("tools.list")
def _(rid, params: dict) -> dict:
    try:
        from toolsets import get_all_toolsets, get_toolset_info

        session = _sessions.get(params.get("session_id", ""))
        enabled = (
            set(getattr(session["agent"], "enabled_toolsets", []) or [])
            if session
            else set(_load_enabled_toolsets() or [])
        )

        items = []
        for name in sorted(get_all_toolsets().keys()):
            info = get_toolset_info(name)
            if not info:
                continue
            items.append(
                {
                    "name": name,
                    "description": info["description"],
                    "tool_count": info["tool_count"],
                    "enabled": name in enabled if enabled else True,
                    "tools": info["resolved_tools"],
                }
            )
        return _ok(rid, {"toolsets": items})
    except Exception as e:
        return _err(rid, 5031, str(e))


@method("tools.show")
def _(rid, params: dict) -> dict:
    try:
        from model_tools import get_toolset_for_tool, get_tool_definitions

        session = _sessions.get(params.get("session_id", ""))
        enabled = (
            getattr(session["agent"], "enabled_toolsets", None)
            if session
            else _load_enabled_toolsets()
        )
        tools = get_tool_definitions(enabled_toolsets=enabled, quiet_mode=True)
        sections = {}

        for tool in sorted(tools, key=lambda t: t["function"]["name"]):
            name = tool["function"]["name"]
            desc = str(tool["function"].get("description", "") or "").split("\n")[0]
            if ". " in desc:
                desc = desc[: desc.index(". ") + 1]
            sections.setdefault(get_toolset_for_tool(name) or "unknown", []).append(
                {
                    "name": name,
                    "description": desc,
                }
            )

        return _ok(
            rid,
            {
                "sections": [
                    {"name": name, "tools": rows}
                    for name, rows in sorted(sections.items())
                ],
                "total": len(tools),
            },
        )
    except Exception as e:
        return _err(rid, 5034, str(e))


@method("tools.configure")
def _(rid, params: dict) -> dict:
    action = str(params.get("action", "") or "").strip().lower()
    targets = [
        str(name).strip() for name in params.get("names", []) or [] if str(name).strip()
    ]
    if action not in {"disable", "enable"}:
        return _err(rid, 4017, f"unknown tools action: {action}")
    if not targets:
        return _err(rid, 4018, "names required")

    try:
        from hermes_cli.config import load_config, save_config
        from hermes_cli.tools_config import (
            CONFIGURABLE_TOOLSETS,
            _apply_mcp_change,
            _apply_toolset_change,
            _get_platform_tools,
            _get_plugin_toolset_keys,
        )

        cfg = load_config()
        valid_toolsets = {
            ts_key for ts_key, _, _ in CONFIGURABLE_TOOLSETS
        } | _get_plugin_toolset_keys()
        toolset_targets = [name for name in targets if ":" not in name]
        mcp_targets = [name for name in targets if ":" in name]
        unknown = [name for name in toolset_targets if name not in valid_toolsets]
        toolset_targets = [name for name in toolset_targets if name in valid_toolsets]

        if toolset_targets:
            _apply_toolset_change(cfg, "cli", toolset_targets, action)

        missing_servers = (
            _apply_mcp_change(cfg, mcp_targets, action) if mcp_targets else set()
        )
        save_config(cfg)

        session = _sessions.get(params.get("session_id", ""))
        info = (
            _reset_session_agent(params.get("session_id", ""), session)
            if session
            else None
        )
        enabled = sorted(
            _get_platform_tools(load_config(), "cli", include_default_mcp_servers=False)
        )
        changed = [
            name
            for name in targets
            if name not in unknown
            and (":" not in name or name.split(":", 1)[0] not in missing_servers)
        ]

        return _ok(
            rid,
            {
                "changed": changed,
                "enabled_toolsets": enabled,
                "info": info,
                "missing_servers": sorted(missing_servers),
                "reset": bool(session),
                "unknown": unknown,
            },
        )
    except Exception as e:
        return _err(rid, 5035, str(e))


@method("toolsets.list")
def _(rid, params: dict) -> dict:
    try:
        from toolsets import get_all_toolsets, get_toolset_info

        session = _sessions.get(params.get("session_id", ""))
        enabled = (
            set(getattr(session["agent"], "enabled_toolsets", []) or [])
            if session
            else set(_load_enabled_toolsets() or [])
        )

        items = []
        for name in sorted(get_all_toolsets().keys()):
            info = get_toolset_info(name)
            if not info:
                continue
            items.append(
                {
                    "name": name,
                    "description": info["description"],
                    "tool_count": info["tool_count"],
                    "enabled": name in enabled if enabled else True,
                }
            )
        return _ok(rid, {"toolsets": items})
    except Exception as e:
        return _err(rid, 5032, str(e))


@method("agents.list")
def _(rid, params: dict) -> dict:
    try:
        from tools.process_registry import process_registry

        procs = process_registry.list_sessions()
        return _ok(
            rid,
            {
                "processes": [
                    {
                        "session_id": p["session_id"],
                        "command": p["command"][:80],
                        "status": p["status"],
                        "uptime": p["uptime_seconds"],
                    }
                    for p in procs
                ]
            },
        )
    except Exception as e:
        return _err(rid, 5033, str(e))


@method("cron.manage")
def _(rid, params: dict) -> dict:
    action, jid = params.get("action", "list"), params.get("name", "")
    try:
        from tools.cronjob_tools import cronjob

        if action == "list":
            return _ok(rid, json.loads(cronjob(action="list")))
        if action == "add":
            return _ok(
                rid,
                json.loads(
                    cronjob(
                        action="create",
                        name=jid,
                        schedule=params.get("schedule", ""),
                        prompt=params.get("prompt", ""),
                    )
                ),
            )
        if action in {"remove", "pause", "resume"}:
            return _ok(rid, json.loads(cronjob(action=action, job_id=jid)))
        return _err(rid, 4016, f"unknown cron action: {action}")
    except Exception as e:
        return _err(rid, 5023, str(e))


@method("skills.manage")
def _(rid, params: dict) -> dict:
    action, query = params.get("action", "list"), params.get("query", "")
    try:
        if action == "list":
            from hermes_cli.banner import get_available_skills

            return _ok(rid, {"skills": get_available_skills()})
        if action == "search":
            from tools.skills_hub import (
                GitHubAuth,
                create_source_router,
                unified_search,
            )

            raw = (
                unified_search(
                    query,
                    create_source_router(GitHubAuth()),
                    source_filter="all",
                    limit=20,
                )
                or []
            )
            return _ok(
                rid,
                {
                    "results": [
                        {"name": r.name, "description": r.description} for r in raw
                    ]
                },
            )
        if action == "install":
            from hermes_cli.skills_hub import do_install

            class _Q:
                def print(self, *a, **k):
                    pass

            do_install(query, skip_confirm=True, console=_Q())
            return _ok(rid, {"installed": True, "name": query})
        if action == "browse":
            from hermes_cli.skills_hub import browse_skills

            pg = int(params.get("page", 0) or 0) or (
                int(query) if query.isdigit() else 1
            )
            return _ok(
                rid, browse_skills(page=pg, page_size=int(params.get("page_size", 20)))
            )
        if action == "inspect":
            from hermes_cli.skills_hub import inspect_skill

            return _ok(rid, {"info": inspect_skill(query) or {}})
        return _err(rid, 4017, f"unknown skills action: {action}")
    except Exception as e:
        return _err(rid, 5024, str(e))


@method("skills.reload")
def _(rid, params: dict) -> dict:
    try:
        from agent.skill_commands import reload_skills

        result = reload_skills()
        added = result.get("added") or []
        removed = result.get("removed") or []
        total = int(result.get("total") or 0)

        lines = ["Reloading skills..."]
        if not added and not removed:
            lines.append("No new skills detected.")
        if added:
            lines.append("Added skills:")
            lines.extend(f"  - {item.get('name', '')}" for item in added)
        if removed:
            lines.append("Removed skills:")
            lines.extend(f"  - {item.get('name', '')}" for item in removed)
        lines.append(f"{total} skill(s) available")
        return _ok(rid, {"output": "\n".join(lines), "result": result})
    except Exception as e:
        return _err(rid, 5025, str(e))


# ── Methods: shell ───────────────────────────────────────────────────


@method("shell.exec")
def _(rid, params: dict) -> dict:
    cmd = params.get("command", "")
    if not cmd:
        return _err(rid, 4004, "empty command")
    try:
        from tools.approval import detect_dangerous_command

        is_dangerous, _, desc = detect_dangerous_command(cmd)
        if is_dangerous:
            return _err(
                rid, 4005, f"blocked: {desc}. Use the agent for dangerous commands."
            )
    except ImportError:
        pass
    try:
        r = subprocess.run(
            cmd, shell=True, capture_output=True, text=True, timeout=30, cwd=os.getcwd()
        )
        return _ok(
            rid,
            {
                "stdout": r.stdout[-4000:],
                "stderr": r.stderr[-2000:],
                "code": r.returncode,
            },
        )
    except subprocess.TimeoutExpired:
        return _err(rid, 5002, "command timed out (30s)")
    except Exception as e:
        return _err(rid, 5003, str(e))