Merge branch 'main' of github.com:NousResearch/hermes-agent into bb/gui

# Conflicts:
#	tui_gateway/server.py

agent/auxiliary_client.py

@@ -1680,26 +1680,48 @@ def _read_main_provider() -> str:
 # per turn — no lock needed. Cleared by ``clear_runtime_main()``.
 _RUNTIME_MAIN_PROVIDER: str = ""
 _RUNTIME_MAIN_MODEL: str = ""
+_RUNTIME_MAIN_BASE_URL: str = ""
+_RUNTIME_MAIN_API_KEY: str = ""
+_RUNTIME_MAIN_API_MODE: str = ""
 
 
-def set_runtime_main(provider: str, model: str) -> None:
-    """Record the live runtime provider/model for the current AIAgent.
+def set_runtime_main(
+    provider: str,
+    model: str,
+    *,
+    base_url: str = "",
+    api_key: str = "",
+    api_mode: str = "",
+) -> None:
+    """Record the live runtime provider/model/credentials for the current AIAgent.
 
     Called by ``run_agent.AIAgent._sync_runtime_main_for_aux_routing`` (or
     equivalent setter) at the top of each turn so that
     ``_read_main_provider`` / ``_read_main_model`` reflect CLI/gateway
     overrides instead of the stale config.yaml default.
+
+    For ``custom:`` providers, ``base_url`` and ``api_key`` must also be
+    recorded so that ``_resolve_auto`` can construct a valid client in
+    Step 1 instead of falling through to the aggregator chain.
     """
     global _RUNTIME_MAIN_PROVIDER, _RUNTIME_MAIN_MODEL
+    global _RUNTIME_MAIN_BASE_URL, _RUNTIME_MAIN_API_KEY, _RUNTIME_MAIN_API_MODE
     _RUNTIME_MAIN_PROVIDER = (provider or "").strip().lower()
     _RUNTIME_MAIN_MODEL = (model or "").strip()
+    _RUNTIME_MAIN_BASE_URL = (base_url or "").strip()
+    _RUNTIME_MAIN_API_KEY = api_key.strip() if isinstance(api_key, str) else ""
+    _RUNTIME_MAIN_API_MODE = (api_mode or "").strip()
 
 
 def clear_runtime_main() -> None:
     """Clear the runtime override (e.g. on session end)."""
     global _RUNTIME_MAIN_PROVIDER, _RUNTIME_MAIN_MODEL
+    global _RUNTIME_MAIN_BASE_URL, _RUNTIME_MAIN_API_KEY, _RUNTIME_MAIN_API_MODE
     _RUNTIME_MAIN_PROVIDER = ""
     _RUNTIME_MAIN_MODEL = ""
+    _RUNTIME_MAIN_BASE_URL = ""
+    _RUNTIME_MAIN_API_KEY = ""
+    _RUNTIME_MAIN_API_MODE = ""
 
 
 def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str], Optional[str]]:
@@ -2980,6 +3002,18 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
     runtime_api_key = runtime.get("api_key", "")
     runtime_api_mode = str(runtime.get("api_mode") or "")
 
+    # Fall back to process-local globals when main_runtime dict was not
+    # provided or was incomplete.  ``set_runtime_main()`` now records
+    # base_url/api_key/api_mode alongside provider/model, so custom:
+    # providers get the full credential surface in Step 1 of the
+    # auto-detect chain.
+    if not runtime_base_url and _RUNTIME_MAIN_BASE_URL:
+        runtime_base_url = _RUNTIME_MAIN_BASE_URL
+    if not runtime_api_key and _RUNTIME_MAIN_API_KEY:
+        runtime_api_key = _RUNTIME_MAIN_API_KEY
+    if not runtime_api_mode and _RUNTIME_MAIN_API_MODE:
+        runtime_api_mode = _RUNTIME_MAIN_API_MODE
+
     # ── Warn once if OPENAI_BASE_URL is set but config.yaml uses a named
     #    provider (not 'custom').  This catches the common "env poisoning"
     #    scenario where a user switches providers via `hermes model` but the

agent/chat_completion_helpers.py

@@ -1283,6 +1283,18 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
             agent._copy_reasoning_content_for_api(msg, api_msg)
             for internal_field in ("reasoning", "finish_reason", "_thinking_prefill"):
                 api_msg.pop(internal_field, None)
+            # Strict OpenAI-compatible gateways (Fireworks-backed OpenCode Go,
+            # Mistral, Moonshot/Kimi) reject any message key outside the Chat
+            # Completions schema. The main loop drops these via
+            # ChatCompletionsTransport.convert_messages(), but the summary path
+            # hand-builds messages and calls chat.completions.create() directly,
+            # bypassing the transport — so mirror that sanitization here:
+            # tool_name (SQLite FTS bookkeeping), the codex_* reasoning carriers,
+            # and every Hermes-internal underscore-prefixed scaffolding key.
+            for schema_foreign in ("tool_name", "codex_reasoning_items", "codex_message_items"):
+                api_msg.pop(schema_foreign, None)
+            for internal_key in [k for k in api_msg if isinstance(k, str) and k.startswith("_")]:
+                api_msg.pop(internal_key, None)
             if _needs_sanitize:
                 agent._sanitize_tool_calls_for_strict_api(api_msg)
             api_messages.append(api_msg)

agent/context_compressor.py

@@ -40,17 +40,47 @@ SUMMARY_PREFIX = (
     "window — treat it as background reference, NOT as active instructions. "
     "Do NOT answer questions or fulfill requests mentioned in this summary; "
     "they were already addressed. "
-    "Your current task is identified in the '## Active Task' section of the "
-    "summary — resume exactly from there. "
+    "Respond ONLY to the latest user message that appears AFTER this "
+    "summary — that message is the single source of truth for what to do "
+    "right now. "
+    "If the latest user message is consistent with the '## Active Task' "
+    "section, you may use the summary as background. If the latest user "
+    "message contradicts, supersedes, changes topic from, or in any way "
+    "diverges from '## Active Task' / '## In Progress' / '## Pending User "
+    "Asks' / '## Remaining Work', the latest message WINS — discard those "
+    "stale items entirely and do not 'wrap up the old task first'. "
+    "Reverse signals in the latest message (e.g. 'stop', 'undo', 'roll "
+    "back', 'just verify', 'don't do that anymore', 'never mind', a new "
+    "topic) must immediately end any in-flight work described in the "
+    "summary; do not re-surface it in later turns. "
     "IMPORTANT: Your persistent memory (MEMORY.md, USER.md) in the system "
     "prompt is ALWAYS authoritative and active — never ignore or deprioritize "
     "memory content due to this compaction note. "
-    "Respond ONLY to the latest user message "
-    "that appears AFTER this summary. The current session state (files, "
-    "config, etc.) may reflect work described here — avoid repeating it:"
+    "The current session state (files, config, etc.) may reflect work "
+    "described here — avoid repeating it:"
 )
 LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
 
+# Handoff prefixes that shipped in earlier releases. A summary persisted under
+# one of these can be inherited into a resumed lineage (#35344); when it is
+# re-normalized on re-compaction we must strip the OLD prefix too, otherwise the
+# stale directive it carried (e.g. "resume exactly from Active Task") survives
+# embedded in the body and keeps hijacking replies. Keep newest-first; entries
+# are matched literally. Add a frozen copy here whenever SUMMARY_PREFIX changes.
+_HISTORICAL_SUMMARY_PREFIXES = (
+    # Pre-#35344: contained the self-contradicting "resume exactly" directive.
+    "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
+    "into the summary below. This is a handoff from a previous context "
+    "window — treat it as background reference, NOT as active instructions. "
+    "Do NOT answer questions or fulfill requests mentioned in this summary; "
+    "they were already addressed. "
+    "Your current task is identified in the '## Active Task' section of the "
+    "summary — resume exactly from there. "
+    "Respond ONLY to the latest user message "
+    "that appears AFTER this summary. The current session state (files, "
+    "config, etc.) may reflect work described here — avoid repeating it:",
+)
+
 # Minimum tokens for the summary output
 _MIN_SUMMARY_TOKENS = 2000
 # Proportion of compressed content to allocate for summary
@@ -1236,11 +1266,27 @@ Summary generation was unavailable, so this is a best-effort deterministic fallb
 
         # Shared structured template (used by both paths).
         _template_sections = f"""## Active Task
-[THE SINGLE MOST IMPORTANT FIELD. Copy the user's most recent request or
-task assignment verbatim — the exact words they used. If multiple tasks
-were requested and only some are done, list only the ones NOT yet completed.
-Continuation should pick up exactly here. Example:
+[THE SINGLE MOST IMPORTANT FIELD. Capture the user's most recent unfulfilled
+input verbatim — the exact words they used. This includes:
+- Explicit task assignments ("refactor the auth module")
+- Questions awaiting an answer ("waarom staat X op Y?", "wat zijn de volgende stappen?")
+- Decisions awaiting input ("optie A of B?")
+- Ongoing discussions where the assistant owes the next substantive reply
+A conversation where the user just asked a question IS an active task — the
+task is "answer that question with full context". Do NOT write "None" merely
+because the user did not issue an imperative command; reserve "None" for the
+rare case where the last exchange was fully resolved and the user said
+something like "thanks, that's all".
+If multiple items are outstanding, list only the ones NOT yet completed.
+Continuation should pick up exactly here. Examples:
 "User asked: 'Now refactor the auth module to use JWT instead of sessions'"
+"User asked: 'Waarom stond provider ineens op openrouter?' — needs investigation + answer"
+"User chose option A; awaiting implementation of step 2"
+If the user's most recent message was a reverse signal (stop, undo, roll
+back, never mind, just verify, change of topic) that supersedes earlier
+work, write the reverse signal verbatim and DO NOT carry forward the
+cancelled task. Example: "User asked: 'Stop the i18n refactor and just
+verify the current diff' — earlier i18n in-flight work is cancelled."
 If no outstanding task exists, write "None."]
 
 ## Goal
@@ -1306,7 +1352,7 @@ PREVIOUS SUMMARY:
 NEW TURNS TO INCORPORATE:
 {content_to_summarize}
 
-Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new completed actions to the numbered list (continue numbering). Move items from "In Progress" to "Completed Actions" when done. Move answered questions to "Resolved Questions". Update "Active State" to reflect current state. Remove information only if it is clearly obsolete. CRITICAL: Update "## Active Task" to reflect the user's most recent unfulfilled request — this is the most important field for task continuity.
+Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new completed actions to the numbered list (continue numbering). Move items from "In Progress" to "Completed Actions" when done. Move answered questions to "Resolved Questions". Update "Active State" to reflect current state. Remove information only if it is clearly obsolete. CRITICAL: Update "## Active Task" to reflect the user's most recent unfulfilled input — this includes any question, decision request, or discussion turn that the assistant has not yet answered. Only write "None" if the last exchange was fully resolved.
 
 {_template_sections}"""
         else:
@@ -1470,9 +1516,16 @@ The user has requested that this compaction PRIORITISE preserving all informatio
 
     @staticmethod
     def _strip_summary_prefix(summary: str) -> str:
-        """Return summary body without the current or legacy handoff prefix."""
+        """Return summary body without the current, legacy, or any historical
+        handoff prefix.
+
+        Historical prefixes must be stripped too: a handoff persisted under an
+        older prefix can be inherited into a resumed lineage (#35344), and if we
+        only re-prepend the current prefix without removing the old one, the
+        stale directive it carried stays embedded in the body.
+        """
         text = (summary or "").strip()
-        for prefix in (SUMMARY_PREFIX, LEGACY_SUMMARY_PREFIX):
+        for prefix in (SUMMARY_PREFIX, LEGACY_SUMMARY_PREFIX, *_HISTORICAL_SUMMARY_PREFIXES):
             if text.startswith(prefix):
                 return text[len(prefix):].lstrip()
         return text
@@ -1486,7 +1539,9 @@ The user has requested that this compaction PRIORITISE preserving all informatio
     @staticmethod
     def _is_context_summary_content(content: Any) -> bool:
         text = _content_text_for_contains(content).lstrip()
-        return text.startswith(SUMMARY_PREFIX) or text.startswith(LEGACY_SUMMARY_PREFIX)
+        if text.startswith(SUMMARY_PREFIX) or text.startswith(LEGACY_SUMMARY_PREFIX):
+            return True
+        return any(text.startswith(p) for p in _HISTORICAL_SUMMARY_PREFIXES)
 
     @classmethod
     def _find_latest_context_summary(

agent/context_engine.py

@@ -115,6 +115,15 @@ class ContextEngine(ABC):
         """
         return False
 
+    def should_defer_preflight_to_real_usage(self, rough_tokens: int) -> bool:
+        """Return True when preflight should trust recent real usage instead.
+
+        Built-in compression uses this to avoid re-compacting from known-noisy
+        rough estimates after a compressed request has already fit. Third-party
+        engines can ignore it safely.
+        """
+        return False
+
     # -- Optional: manual /compress preflight ------------------------------
 
     def has_content_to_compress(self, messages: List[Dict[str, Any]]) -> bool:

agent/conversation_compression.py

@@ -575,19 +575,18 @@ def compress_context(
             force=True,
         )
 
-    # Update token estimate after compaction so pressure calculations
-    # use the post-compression count, not the stale pre-compression one.
-    # Use estimate_request_tokens_rough() so tool schemas are included —
-    # with 50+ tools enabled, schemas alone can add 20-30K tokens, and
-    # omitting them delays the next compression cycle far past the
-    # configured threshold (issue #14695).
+    # Keep the post-compression rough estimate for diagnostics, but do not
+    # treat it as provider-reported prompt usage. Schema-heavy rough estimates
+    # can remain above threshold even after the next real API request fits.
     _compressed_est = estimate_request_tokens_rough(
         compressed,
         system_prompt=new_system_prompt or "",
         tools=agent.tools or None,
     )
-    agent.context_compressor.last_prompt_tokens = _compressed_est
+    agent.context_compressor.last_compression_rough_tokens = _compressed_est
+    agent.context_compressor.last_prompt_tokens = -1
     agent.context_compressor.last_completion_tokens = 0
+    agent.context_compressor.awaiting_real_usage_after_compression = True
 
     # Clear the file-read dedup cache.  After compression the original
     # read content is summarised away — if the model re-reads the same
@@ -599,7 +598,7 @@ def compress_context(
         pass
 
     logger.info(
-        "context compression done: session=%s messages=%d->%d tokens=~%s",
+        "context compression done: session=%s messages=%d->%d rough_tokens=~%s awaiting_real_usage=true",
         agent.session_id or "none", _pre_msg_count, len(compressed),
         f"{_compressed_est:,}",
     )

agent/conversation_loop.py

@@ -392,6 +392,9 @@ def run_conversation(
         set_runtime_main(
             getattr(agent, "provider", "") or "",
             getattr(agent, "model", "") or "",
+            base_url=getattr(agent, "base_url", "") or "",
+            api_key=getattr(agent, "api_key", "") or "",
+            api_mode=getattr(agent, "api_mode", "") or "",
         )
     except Exception:
         pass
@@ -600,18 +603,50 @@ def run_conversation(
             system_prompt=active_system_prompt or "",
             tools=agent.tools or None,
         )
+        _compressor = agent.context_compressor
+        _defer_preflight = getattr(
+            _compressor,
+            "should_defer_preflight_to_real_usage",
+            lambda _tokens: False,
+        )
+        _preflight_deferred = _defer_preflight(_preflight_tokens)
 
-        if agent.context_compressor.should_compress(_preflight_tokens):
+        if not _preflight_deferred:
+            # Keep the CLI/ACP context display in sync with what preflight
+            # actually measured.  The status bar reads
+            # ``compressor.last_prompt_tokens``, which otherwise only updates
+            # from a *successful* API response.  When the conversation has grown
+            # since the last successful call — or when compression then fails
+            # (e.g. the auxiliary summary model times out) and no fresh usage
+            # arrives — the bar stays stuck at the old, smaller value while
+            # preflight reports a much larger number, looking out of sync.
+            # Seed it with the fresh estimate (only ever revising upward; a real
+            # ``update_from_response`` will correct it after the next API call).
+            # Skipped when deferring — a deferred estimate is known to over-count
+            # vs the last real provider prompt, so trusting it for the display
+            # would re-introduce the very desync we're avoiding.
+            if _preflight_tokens > (_compressor.last_prompt_tokens or 0):
+                _compressor.last_prompt_tokens = _preflight_tokens
+
+        if _preflight_deferred:
+            logger.info(
+                "Skipping preflight compression: rough estimate ~%s >= %s, "
+                "but last real provider prompt was %s after compression",
+                f"{_preflight_tokens:,}",
+                f"{_compressor.threshold_tokens:,}",
+                f"{_compressor.last_real_prompt_tokens:,}",
+            )
+        elif _compressor.should_compress(_preflight_tokens):
             logger.info(
                 "Preflight compression: ~%s tokens >= %s threshold (model %s, ctx %s)",
                 f"{_preflight_tokens:,}",
-                f"{agent.context_compressor.threshold_tokens:,}",
+                f"{_compressor.threshold_tokens:,}",
                 agent.model,
-                f"{agent.context_compressor.context_length:,}",
+                f"{_compressor.context_length:,}",
             )
             agent._emit_status(
                 f"📦 Preflight compression: ~{_preflight_tokens:,} tokens "
-                f">= {agent.context_compressor.threshold_tokens:,} threshold. "
+                f">= {_compressor.threshold_tokens:,} threshold. "
                 "This may take a moment."
             )
             # May need multiple passes for very large sessions with small
@@ -646,8 +681,8 @@ def run_conversation(
                     system_prompt=active_system_prompt or "",
                     tools=agent.tools or None,
                 )
-                if _preflight_tokens < agent.context_compressor.threshold_tokens:
-                    break  # Under threshold
+                if not _compressor.should_compress(_preflight_tokens):
+                    break  # Under threshold or anti-thrash guard stopped it
 
     # Plugin hook: pre_llm_call
     # Fired once per turn before the tool-calling loop.  Plugins can
@@ -1457,7 +1492,8 @@ def run_conversation(
                     
                     if retry_count >= max_retries:
                         # Try fallback before giving up
-                        agent._buffer_status(f"⚠️ Max retries ({max_retries}) for invalid responses — trying fallback...")
+                        if agent._has_pending_fallback():
+                            agent._buffer_status(f"⚠️ Max retries ({max_retries}) for invalid responses — trying fallback...")
                         if agent._try_activate_fallback():
                             retry_count = 0
                             compression_attempts = 0
@@ -3059,12 +3095,17 @@ def run_conversation(
                 ) and not is_context_length_error
 
                 if is_client_error:
-                    # Try fallback before aborting — a different provider
-                    # may not have the same issue (rate limit, auth, etc.)
-                    if classified.reason == FailoverReason.content_policy_blocked:
-                        agent._buffer_status("⚠️ Provider safety filter blocked this request — trying fallback...")
-                    else:
-                        agent._buffer_status(f"⚠️ Non-retryable error (HTTP {status_code}) — trying fallback...")
+                    # Try fallback before aborting — a different provider may
+                    # not have the same issue (rate limit, auth, etc.). Only
+                    # announce the attempt when a fallback chain actually
+                    # exists; otherwise "trying fallback..." is a lie and the
+                    # session looks like it's recovering when it's about to
+                    # abort silently (#35314, #17446).
+                    if agent._has_pending_fallback():
+                        if classified.reason == FailoverReason.content_policy_blocked:
+                            agent._buffer_status("⚠️ Provider safety filter blocked this request — trying fallback...")
+                        else:
+                            agent._buffer_status(f"⚠️ Non-retryable error (HTTP {status_code}) — trying fallback...")
                     if agent._try_activate_fallback():
                         retry_count = 0
                         compression_attempts = 0
@@ -3207,7 +3248,8 @@ def run_conversation(
                         retry_count = 0
                         continue
                     # Try fallback before giving up entirely
-                    agent._buffer_status(f"⚠️ Max retries ({max_retries}) exhausted — trying fallback...")
+                    if agent._has_pending_fallback():
+                        agent._buffer_status(f"⚠️ Max retries ({max_retries}) exhausted — trying fallback...")
                     if agent._try_activate_fallback():
                         retry_count = 0
                         compression_attempts = 0
@@ -3862,6 +3904,11 @@ def run_conversation(
                     # inflate completion_tokens with reasoning,
                     # causing premature compression.  (#12026)
                     _real_tokens = _compressor.last_prompt_tokens
+                elif _compressor.last_prompt_tokens == -1:
+                    # Compression just ran and no API-reported prompt count
+                    # has arrived yet. Avoid treating a schema-heavy rough
+                    # post-compression estimate as real context pressure.
+                    _real_tokens = 0
                 else:
                     # Include tool schemas — with 50+ tools enabled
                     # these add 20-30K tokens the messages-only
@@ -4443,6 +4490,55 @@ def run_conversation(
         except Exception as _ver_err:
             logger.debug("file-mutation verifier footer failed: %s", _ver_err)
 
+    # Turn-completion explainer.
+    # When a turn ends abnormally after substantive work — empty content
+    # after retries, a partial/truncated stream, a still-pending tool
+    # result, or an iteration/budget limit — the user otherwise gets a
+    # blank or fragmentary response box with no consolidated reason why
+    # the agent stopped (#34452).  Surface a single user-visible
+    # explanation derived from ``_turn_exit_reason``, mirroring the
+    # file-mutation verifier footer pattern above.
+    #
+    # Gate carefully so healthy turns stay quiet:
+    #   - ``text_response(...)`` exits never produce an explanation
+    #     (handled inside the formatter), so a terse ``Done.`` is silent.
+    #   - We only ACT when there is no genuinely usable reply this turn:
+    #     an empty response, the "(empty)" terminal sentinel, or a
+    #     suspiciously short partial fragment with no terminating
+    #     punctuation (e.g. "The").  A real short answer keeps its text.
+    if not interrupted:
+        try:
+            if agent._turn_completion_explainer_enabled():
+                _stripped = (final_response or "").strip()
+                _is_empty_terminal = _stripped == "" or _stripped == "(empty)"
+                # A short fragment that is not a normal text_response exit
+                # and lacks sentence-ending punctuation is treated as a
+                # truncated partial (the "The" case from #34452).
+                _is_partial_fragment = (
+                    not _is_empty_terminal
+                    and not str(_turn_exit_reason).startswith("text_response")
+                    and len(_stripped) <= 24
+                    and _stripped[-1:] not in {".", "!", "?", "。", "！", "？", "`", ")"}
+                )
+                if _is_empty_terminal or _is_partial_fragment:
+                    _explanation = agent._format_turn_completion_explanation(
+                        _turn_exit_reason
+                    )
+                    if _explanation:
+                        if _is_empty_terminal:
+                            # Replace the bare "(empty)"/blank sentinel with
+                            # the actionable explanation.
+                            final_response = _explanation
+                        else:
+                            # Keep the partial fragment, append the reason so
+                            # the user sees both what arrived and why it
+                            # stopped.
+                            final_response = (
+                                _stripped + "\n\n" + _explanation
+                            )
+        except Exception as _exp_err:
+            logger.debug("turn-completion explainer failed: %s", _exp_err)
+
     _response_transformed = False
 
     # Plugin hook: transform_llm_output

agent/lsp/cli.py

@@ -247,18 +247,13 @@ def _cmd_restart() -> int:
 
 
 def _cmd_which(server_id: str) -> int:
-    from agent.lsp.install import INSTALL_RECIPES, hermes_lsp_bin_dir
-    import shutil as _shutil
+    from agent.lsp.install import INSTALL_RECIPES, _existing_binary
 
     recipe = INSTALL_RECIPES.get(server_id)
     bin_name = (recipe or {}).get("bin", server_id)
-    staged = hermes_lsp_bin_dir() / bin_name
-    if staged.exists():
-        sys.stdout.write(str(staged) + "\n")
-        return 0
-    on_path = _shutil.which(bin_name)
-    if on_path:
-        sys.stdout.write(on_path + "\n")
+    resolved = _existing_binary(bin_name)
+    if resolved:
+        sys.stdout.write(resolved + "\n")
         return 0
     sys.stderr.write(f"{server_id}: not installed\n")
     return 1
@@ -292,11 +287,9 @@ def _backend_warnings() -> list:
     suggestion across common platforms.
     """
     import shutil as _shutil
-    from agent.lsp.install import hermes_lsp_bin_dir
+    from agent.lsp.install import _existing_binary
     notes: list = []
-    bash_installed = _shutil.which("bash-language-server") is not None or (
-        (hermes_lsp_bin_dir() / "bash-language-server").exists()
-    )
+    bash_installed = _existing_binary("bash-language-server") is not None
     if bash_installed and _shutil.which("shellcheck") is None:
         notes.append(
             "bash-language-server is installed but shellcheck is missing — "

agent/lsp/client.py

@@ -44,6 +44,7 @@ from __future__ import annotations
 import asyncio
 import logging
 import os
+import sys
 from pathlib import Path
 from typing import Any, Awaitable, Callable, Dict, List, Optional, Set
 from urllib.parse import quote, unquote
@@ -244,15 +245,27 @@ class LSPClient:
             await self._cleanup_process()
             raise
 
+    @staticmethod
+    def _win_wrap_cmd(cmd: List[str]) -> List[str]:
+        """On Windows, wrap .cmd/.bat shims so CreateProcess can run them."""
+        exe = cmd[0]
+        if exe.lower().endswith((".cmd", ".bat")):
+            return ["cmd.exe", "/c", *cmd]
+        return cmd
+
     async def _spawn(self) -> None:
         env = dict(os.environ)
         if self._env:
             env.update(self._env)
 
+        cmd = self._command
+        if sys.platform == "win32":
+            cmd = self._win_wrap_cmd(cmd)
+
         try:
             self._proc = await asyncio.create_subprocess_exec(
-                self._command[0],
-                *self._command[1:],
+                cmd[0],
+                *cmd[1:],
                 stdin=asyncio.subprocess.PIPE,
                 stdout=asyncio.subprocess.PIPE,
                 stderr=asyncio.subprocess.PIPE,
@@ -261,7 +274,7 @@ class LSPClient:
             )
         except FileNotFoundError as e:
             raise LSPProtocolError(
-                f"LSP server binary not found: {self._command[0]} ({e})"
+                f"LSP server binary not found: {cmd[0]} ({e})"
             ) from e
 
         # Drain stderr at debug level — if we don't, the pipe buffer

agent/lsp/install.py

@@ -108,6 +108,11 @@ INSTALL_RECIPES: Dict[str, Dict[str, Any]] = {
 _install_locks: Dict[str, threading.Lock] = {}
 _install_results: Dict[str, Optional[str]] = {}
 _install_lock_meta = threading.Lock()
+_WINDOWS_WRAPPER_SUFFIXES = (".cmd", ".exe", ".bat")
+
+
+def _is_windows() -> bool:
+    return os.name == "nt"
 
 
 def hermes_lsp_bin_dir() -> Path:
@@ -120,14 +125,33 @@ def hermes_lsp_bin_dir() -> Path:
     return p
 
 
+def _native_binary_candidates(base: Path) -> list[Path]:
+    """Return platform-native executable candidates for a staged binary."""
+    candidates = [base]
+    if _is_windows():
+        existing = {str(base).lower()}
+        for suffix in _WINDOWS_WRAPPER_SUFFIXES:
+            candidate = Path(str(base) + suffix)
+            key = str(candidate).lower()
+            if key not in existing:
+                candidates.append(candidate)
+                existing.add(key)
+    return candidates
+
+
 def _existing_binary(name: str) -> Optional[str]:
     """Probe the staging dir + PATH for a binary named ``name``."""
-    staged = hermes_lsp_bin_dir() / name
-    if staged.exists() and os.access(staged, os.X_OK):
-        return str(staged)
+    for staged in _native_binary_candidates(hermes_lsp_bin_dir() / name):
+        if staged.exists() and os.access(staged, os.X_OK):
+            return str(staged)
     on_path = shutil.which(name)
     if on_path:
         return on_path
+    if _is_windows():
+        for suffix in _WINDOWS_WRAPPER_SUFFIXES:
+            on_path = shutil.which(f"{name}{suffix}")
+            if on_path:
+                return on_path
     return None
 
 
@@ -250,12 +274,7 @@ def _install_npm(
 
     # Find the bin
     nm_bin = staging / "node_modules" / ".bin" / bin_name
-    if os.name == "nt":
-        # On Windows npm sometimes drops `.cmd` shims
-        candidates = [nm_bin, nm_bin.with_suffix(".cmd")]
-    else:
-        candidates = [nm_bin]
-    for c in candidates:
+    for c in _native_binary_candidates(nm_bin):
         if c.exists():
             # Symlink into our `lsp/bin/` for stable PATH access.
             link = hermes_lsp_bin_dir() / c.name
@@ -301,7 +320,7 @@ def _install_go(pkg: str, bin_name: str) -> Optional[str]:
         logger.warning("[install] go install errored for %s: %s", pkg, e)
         return None
     bin_path = staging / bin_name
-    if os.name == "nt":
+    if _is_windows():
         bin_path = bin_path.with_suffix(".exe")
     if bin_path.exists():
         return str(bin_path)
@@ -337,19 +356,24 @@ def _install_pip(pkg: str, bin_name: str) -> Optional[str]:
     except (subprocess.TimeoutExpired, OSError) as e:
         logger.warning("[install] pip install errored for %s: %s", pkg, e)
         return None
-    # Look for the script
-    bin_path = pip_target / "bin" / bin_name
-    if bin_path.exists():
-        link = hermes_lsp_bin_dir() / bin_name
-        if not link.exists():
-            try:
-                link.symlink_to(bin_path)
-            except (OSError, NotImplementedError):
-                try:
-                    shutil.copy2(bin_path, link)
-                except OSError:
-                    return str(bin_path)
-        return str(link if link.exists() else bin_path)
+    # Look for the console script.  POSIX wheels generally write to bin/,
+    # while native Windows installs use Scripts/.
+    script_dirs = [pip_target / "bin"]
+    if _is_windows():
+        script_dirs.append(pip_target / "Scripts")
+    for script_dir in script_dirs:
+        for bin_path in _native_binary_candidates(script_dir / bin_name):
+            if bin_path.exists():
+                link = hermes_lsp_bin_dir() / bin_path.name
+                if not link.exists():
+                    try:
+                        link.symlink_to(bin_path)
+                    except (OSError, NotImplementedError):
+                        try:
+                            shutil.copy2(bin_path, link)
+                        except OSError:
+                            return str(bin_path)
+                return str(link if link.exists() else bin_path)
     return None
 
 

agent/tool_executor.py

@@ -180,28 +180,9 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         except Exception:
             pass
 
-        # Checkpoint for file-mutating tools
-        if function_name in {"write_file", "patch"} and agent._checkpoint_mgr.enabled:
-            try:
-                file_path = function_args.get("path", "")
-                if file_path:
-                    work_dir = agent._checkpoint_mgr.get_working_dir_for_path(file_path)
-                    agent._checkpoint_mgr.ensure_checkpoint(work_dir, f"before {function_name}")
-            except Exception:
-                pass
-
-        # Checkpoint before destructive terminal commands
-        if function_name == "terminal" and agent._checkpoint_mgr.enabled:
-            try:
-                cmd = function_args.get("command", "")
-                if _is_destructive_command(cmd):
-                    cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd())
-                    agent._checkpoint_mgr.ensure_checkpoint(
-                        cwd, f"before terminal: {cmd[:60]}"
-                    )
-            except Exception:
-                pass
-
+        # ── Block evaluation (BEFORE checkpoint preflight) ───────────
+        # We must know whether the tool will execute before touching
+        # checkpoint state (dedup slot, real snapshots).
         block_result = None
         blocked_by_guardrail = False
         if _ts_scope_block is not None:
@@ -224,6 +205,30 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
                     block_result = agent._guardrail_block_result(guardrail_decision)
                     blocked_by_guardrail = True
 
+        # ── Checkpoint preflight (only for tools that will execute) ──
+        if block_result is None:
+            # Checkpoint for file-mutating tools
+            if function_name in {"write_file", "patch"} and agent._checkpoint_mgr.enabled:
+                try:
+                    file_path = function_args.get("path", "")
+                    if file_path:
+                        work_dir = agent._checkpoint_mgr.get_working_dir_for_path(file_path)
+                        agent._checkpoint_mgr.ensure_checkpoint(work_dir, f"before {function_name}")
+                except Exception:
+                    pass
+
+            # Checkpoint before destructive terminal commands
+            if function_name == "terminal" and agent._checkpoint_mgr.enabled:
+                try:
+                    cmd = function_args.get("command", "")
+                    if _is_destructive_command(cmd):
+                        cwd = function_args.get("workdir") or os.getenv("TERMINAL_CWD", os.getcwd())
+                        agent._checkpoint_mgr.ensure_checkpoint(
+                            cwd, f"before terminal: {cmd[:60]}"
+                        )
+                except Exception:
+                    pass
+
         parsed_calls.append((tool_call, function_name, function_args, block_result, blocked_by_guardrail))
 
     # ── Logging / callbacks ──────────────────────────────────────────
@@ -301,33 +306,38 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         # submit site below (GHSA-qg5c-hvr5-hjgr, #13617).
         start = time.time()
         try:
-            result = agent._invoke_tool(
-                function_name,
-                function_args,
-                effective_task_id,
-                tool_call.id,
-                messages=messages,
-                pre_tool_block_checked=True,
-            )
-        except Exception as tool_error:
-            result = f"Error executing tool '{function_name}': {tool_error}"
-            logger.error("_invoke_tool raised for %s: %s", function_name, tool_error, exc_info=True)
-        duration = time.time() - start
-        is_error, _ = _detect_tool_failure(function_name, result)
-        if is_error:
-            logger.info("tool %s failed (%.2fs): %s", function_name, duration, result[:200])
-        else:
-            logger.info("tool %s completed (%.2fs, %d chars)", function_name, duration, len(result))
-        results[index] = (function_name, function_args, result, duration, is_error, False)
-        # Tear down worker-tid tracking.  Clear any interrupt bit we may
-        # have set so the next task scheduled onto this recycled tid
-        # starts with a clean slate.
-        with agent._tool_worker_threads_lock:
-            agent._tool_worker_threads.discard(_worker_tid)
-        try:
-            _ra()._set_interrupt(False, _worker_tid)
-        except Exception:
-            pass
+            try:
+                result = agent._invoke_tool(
+                    function_name,
+                    function_args,
+                    effective_task_id,
+                    tool_call.id,
+                    messages=messages,
+                    pre_tool_block_checked=True,
+                )
+            except Exception as tool_error:
+                result = f"Error executing tool '{function_name}': {tool_error}"
+                logger.error("_invoke_tool raised for %s: %s", function_name, tool_error, exc_info=True)
+            duration = time.time() - start
+            is_error, _ = _detect_tool_failure(function_name, result)
+            if is_error:
+                logger.info("tool %s failed (%.2fs): %s", function_name, duration, result[:200])
+            else:
+                logger.info("tool %s completed (%.2fs, %d chars)", function_name, duration, len(result))
+            results[index] = (function_name, function_args, result, duration, is_error, False)
+        finally:
+            # Tear down worker-tid tracking.  Clear any interrupt bit we may
+            # have set so the next task scheduled onto this recycled tid
+            # starts with a clean slate.  This MUST be in a finally block
+            # because BaseException subclasses (CancelledError, KeyboardInterrupt)
+            # bypass ``except Exception`` and would otherwise leak the tid
+            # into _interrupted_threads, poisoning the recycled thread.
+            with agent._tool_worker_threads_lock:
+                agent._tool_worker_threads.discard(_worker_tid)
+            try:
+                _ra()._set_interrupt(False, _worker_tid)
+            except Exception:
+                pass
 
     # Start spinner for CLI mode (skip when TUI handles tool progress)
     spinner = None
@@ -753,10 +763,14 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         elif function_name == "delegate_task":
             tasks_arg = function_args.get("tasks")
             if tasks_arg and isinstance(tasks_arg, list):
-                spinner_label = f"🔀 delegating {len(tasks_arg)} tasks"
+                spinner_label = f"🔀 delegating {len(tasks_arg)} tasks · (/agents to monitor)"
             else:
                 goal_preview = (function_args.get("goal") or "")[:30]
-                spinner_label = f"🔀 {goal_preview}" if goal_preview else "🔀 delegating"
+                spinner_label = (
+                    f"🔀 {goal_preview} · (/agents to monitor)"
+                    if goal_preview
+                    else "🔀 delegating · (/agents to monitor)"
+                )
             spinner = None
             if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())