feat(agent): add lmstudio integration

run_agent.py

@@ -1826,9 +1826,6 @@ class AIAgent:
                 )
                 _config_context_length = None
 
-        # Store for reuse in switch_model (so config override persists across model switches)
-        self._config_context_length = _config_context_length
-
         # Resolve custom_providers list once for reuse below (startup
         # context-length override and plugin context-engine init).
         try:
@@ -1887,7 +1884,14 @@ class AIAgent:
                                             file=sys.stderr,
                                         )
                         break
-        
+
+        # Persist for reuse on switch_model / fallback activation. Must come
+        # AFTER the custom_providers branch so per-model overrides aren't lost.
+        self._config_context_length = _config_context_length
+
+        self._ensure_lmstudio_runtime_loaded(_config_context_length)
+
+
         # Select context engine: config-driven (like memory providers).
         # 1. Check config.yaml context.engine setting
         # 2. Check plugins/context_engine/<name>/ directory (repo-shipped)
@@ -2129,6 +2133,24 @@ class AIAgent:
         if hasattr(self, "context_compressor") and self.context_compressor:
             self.context_compressor.on_session_reset()
     
+    def _ensure_lmstudio_runtime_loaded(self, config_context_length: Optional[int] = None) -> None:
+        """
+        Preload the LM Studio model with at least Hermes' minimum context.
+        """
+        if (self.provider or "").strip().lower() != "lmstudio":
+            return
+        try:
+            from agent.model_metadata import MINIMUM_CONTEXT_LENGTH
+            from hermes_cli.models import ensure_lmstudio_model_loaded
+            if config_context_length is None:
+                config_context_length = getattr(self, "_config_context_length", None)
+            target_ctx = max(config_context_length or 0, MINIMUM_CONTEXT_LENGTH)
+            ensure_lmstudio_model_loaded(
+                self.model, self.base_url, getattr(self, "api_key", ""), target_ctx,
+            )
+        except Exception as err:
+            logger.debug("LM Studio preload skipped: %s", err)
+
     def switch_model(self, new_model, new_provider, api_key='', base_url='', api_mode=''):
         """Switch the model/provider in-place for a live agent.
 
@@ -2224,6 +2246,9 @@ class AIAgent:
             )
         )
 
+        # ── LM Studio: preload before probing context length ──
+        self._ensure_lmstudio_runtime_loaded()
+
         # ── Update context compressor ──
         if hasattr(self, "context_compressor") and self.context_compressor:
             from agent.model_metadata import get_model_context_length
@@ -7327,6 +7352,9 @@ class AIAgent:
                 )
             )
 
+            # LM Studio: preload before probing the fallback's context length.
+            self._ensure_lmstudio_runtime_loaded()
+
             # Update context compressor limits for the fallback model.
             # Without this, compression decisions use the primary model's
             # context window (e.g. 200K) instead of the fallback's (e.g. 32K),
@@ -8047,6 +8075,7 @@ class AIAgent:
             or base_url_host_matches(self.base_url, "moonshot.cn")
         )
         _is_tokenhub = base_url_host_matches(self._base_url_lower, "tokenhub.tencentmaas.com")
+        _is_lmstudio = (self.provider or "").strip().lower() == "lmstudio"
 
         # Temperature: _fixed_temperature_for_model may return OMIT_TEMPERATURE
         # sentinel (temperature omitted entirely), a numeric override, or None.
@@ -8119,6 +8148,7 @@ class AIAgent:
             is_nvidia_nim=_is_nvidia,
             is_kimi=_is_kimi,
             is_tokenhub=_is_tokenhub,
+            is_lmstudio=_is_lmstudio,
             is_custom_provider=self.provider == "custom",
             ollama_num_ctx=self._ollama_num_ctx,
             provider_preferences=_prefs or None,
@@ -8129,6 +8159,7 @@ class AIAgent:
             omit_temperature=_omit_temp,
             supports_reasoning=self._supports_reasoning_extra_body(),
             github_reasoning_extra=self._github_models_reasoning_extra_body() if _is_gh else None,
+            lmstudio_reasoning_options=self._lmstudio_reasoning_options_cached() if _is_lmstudio else None,
             anthropic_max_output=_ant_max,
             provider_name=self.provider,
         )
@@ -8154,6 +8185,10 @@ class AIAgent:
                 return bool(github_model_reasoning_efforts(self.model))
             except Exception:
                 return False
+        if (self.provider or "").strip().lower() == "lmstudio":
+            opts = self._lmstudio_reasoning_options_cached()
+            # "off-only" (or absent) means no real reasoning capability.
+            return any(opt and opt != "off" for opt in opts)
         if "openrouter" not in self._base_url_lower:
             return False
         if "api.mistral.ai" in self._base_url_lower:
@@ -8171,6 +8206,48 @@ class AIAgent:
         )
         return any(model.startswith(prefix) for prefix in reasoning_model_prefixes)
 
+    def _lmstudio_reasoning_options_cached(self) -> list[str]:
+        """Probe LM Studio's published reasoning ``allowed_options`` once per
+        (model, base_url). The list (e.g. ``["off","on"]`` or
+        ``["off","minimal","low"]``) is needed both for the supports-reasoning
+        gate and for clamping the emitted ``reasoning_effort`` so toggle-style
+        models don't 400 on ``high``. Cache is keyed on (model, base_url) so
+        ``/model`` swaps and base-URL changes don't reuse a stale list, and an
+        empty result (transient probe failure) is *not* cached so the next call
+        retries instead of silently disabling reasoning for the rest of the
+        session.
+        """
+        cache = getattr(self, "_lm_reasoning_opts_cache", None)
+        if cache is None:
+            cache = self._lm_reasoning_opts_cache = {}
+        key = (self.model, self.base_url)
+        cached = cache.get(key)
+        if cached:
+            return cached
+        try:
+            from hermes_cli.models import lmstudio_model_reasoning_options
+            opts = lmstudio_model_reasoning_options(
+                self.model, self.base_url, getattr(self, "api_key", ""),
+            )
+        except Exception:
+            opts = []
+        if opts:
+            cache[key] = opts
+        return opts
+
+    def _resolve_lmstudio_summary_reasoning_effort(self) -> Optional[str]:
+        """Resolve a safe top-level ``reasoning_effort`` for LM Studio.
+
+        The iteration-limit summary path calls ``chat.completions.create()``
+        directly, bypassing the transport. Share the helper so the two paths
+        can't drift on effort resolution and clamping.
+        """
+        from agent.lmstudio_reasoning import resolve_lmstudio_effort
+        return resolve_lmstudio_effort(
+            self.reasoning_config,
+            self._lmstudio_reasoning_options_cached(),
+        )
+
     def _github_models_reasoning_extra_body(self) -> dict | None:
         """Format reasoning payload for GitHub Models/OpenAI-compatible routes."""
         try:
@@ -9692,7 +9769,19 @@ class AIAgent:
             _omit_summary_temperature = _raw_summary_temp is _OMIT_TEMP
             _summary_temperature = None if _omit_summary_temperature else _raw_summary_temp
             _is_nous = "nousresearch" in self._base_url_lower
-            if self._supports_reasoning_extra_body():
+            # LM Studio uses top-level `reasoning_effort` (not extra_body.reasoning).
+            # Mirror ChatCompletionsTransport.build_kwargs() so the summary path
+            # — which calls chat.completions.create() directly without going
+            # through the transport — sends the same shape the transport does.
+            _is_lmstudio_summary = (
+                (self.provider or "").strip().lower() == "lmstudio"
+                and self._supports_reasoning_extra_body()
+            )
+            _lm_reasoning_effort: str | None = (
+                self._resolve_lmstudio_summary_reasoning_effort()
+                if _is_lmstudio_summary else None
+            )
+            if not _is_lmstudio_summary and self._supports_reasoning_extra_body():
                 if self.reasoning_config is not None:
                     summary_extra_body["reasoning"] = self.reasoning_config
                 else:
@@ -9719,6 +9808,8 @@ class AIAgent:
                     summary_kwargs["temperature"] = _summary_temperature
                 if self.max_tokens is not None:
                     summary_kwargs.update(self._max_tokens_param(self.max_tokens))
+                if _lm_reasoning_effort is not None:
+                    summary_kwargs["reasoning_effort"] = _lm_reasoning_effort
 
                 # Include provider routing preferences
                 provider_preferences = {}
@@ -9784,6 +9875,8 @@ class AIAgent:
                         summary_kwargs["temperature"] = _summary_temperature
                     if self.max_tokens is not None:
                         summary_kwargs.update(self._max_tokens_param(self.max_tokens))
+                    if _lm_reasoning_effort is not None:
+                        summary_kwargs["reasoning_effort"] = _lm_reasoning_effort
                     if summary_extra_body:
                         summary_kwargs["extra_body"] = summary_extra_body