Merge branch 'feature/disable-thinking-mode'

Adds `llm.thinking` config option (default: true) that when disabled:
- Injects /no_think into the last user message for Qwen 3.x compatibility
- Sends chat_template_kwargs in API payload for backends that support it
- Silently and immediately nudges on reasoning-only responses instead of
  showing warnings and wasting retry iterations

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

app/agent/loop.py

@@ -92,9 +92,29 @@ class AgentLoop:
         return prompt
 
     def _get_messages_with_system_prompt(self) -> list[Message]:
-        """Prepend the system prompt to conversation history."""
+        """Prepend the system prompt to conversation history.
+
+        When thinking is disabled, injects a /no_think tag into the last user
+        message so Qwen 3.x (and similar) chat templates see it regardless of
+        where tool-result messages fall in the history.
+        """
         system_msg = Message(role="system", content=self._system_prompt)
-        return [system_msg] + self._ctx.get_history()
+        history = self._ctx.get_history()
+
+        if not self._config.llm.thinking and history:
+            history = list(history)
+            for i in range(len(history) - 1, -1, -1):
+                if history[i].role == "user":
+                    original = history[i]
+                    history[i] = Message(
+                        role="user",
+                        content=(original.content or "") + " /no_think",
+                        tool_call_id=original.tool_call_id,
+                        name=original.name,
+                    )
+                    break
+
+        return [system_msg] + history
 
     async def run_turn(self, user_input: str) -> None:
         """Execute one full agent turn: add user message, loop until done.
@@ -161,6 +181,10 @@ class AgentLoop:
                 reasoning_only_streak += 1
                 self._ctx.pop_last_message()
 
+                # When thinking is disabled, reasoning-only is expected model noise.
+                # Nudge immediately and silently to avoid wasting iterations.
+                thinking_disabled = not self._config.llm.thinking
+
                 # If the last context messages are tool errors, nudge immediately
                 # rather than wasting retries — the model is likely confused by the error.
                 has_recent_tool_error = any(
@@ -168,9 +192,14 @@ class AgentLoop:
                     for m in self._ctx.get_history()[-3:]
                 )
 
-                if has_recent_tool_error or reasoning_only_streak >= _MAX_REASONING_RETRIES:
+                should_nudge = (
-                    # Nudge the model by injecting a user hint
+                    thinking_disabled
-                    if self._display:
+                    or has_recent_tool_error
+                    or reasoning_only_streak >= _MAX_REASONING_RETRIES
+                )
+
+                if should_nudge:
+                    if not thinking_disabled and self._display:
                         self._display.write_warning(
                             f"Model produced reasoning but no response {reasoning_only_streak} times. "
                             "Nudging model to respond..."

app/models/config.py

@@ -20,6 +20,10 @@ class LLMConfig(BaseModel):
     max_retries: int = Field(default=3, description="Max retry attempts on transient errors")
     retry_backoff_base: float = Field(default=1.0, description="Base seconds for exponential backoff")
     retry_backoff_max: float = Field(default=30.0, description="Maximum backoff seconds")
+    thinking: bool = Field(
+        default=True,
+        description="Enable model thinking/reasoning mode (disable to reduce reasoning-only loops)",
+    )
     extra_body: dict[str, Any] = Field(
         default_factory=dict,
         description="Extra parameters merged into the API request body (model-specific)",

app/services/llm.py

@@ -151,7 +151,11 @@ class LLMClient:
         if tools:
             payload["tools"] = tools
 
-        # Merge model-specific extra parameters (e.g., enable_thinking, reasoning_effort)
+        # When thinking is disabled, inject chat_template_kwargs for backends that support it
+        if not self._config.thinking:
+            payload.setdefault("chat_template_kwargs", {})["enable_thinking"] = False
+
+        # Merge model-specific extra parameters (e.g., reasoning_effort)
         if self._config.extra_body:
             payload.update(self._config.extra_body)
 

config/config.yaml

@@ -10,13 +10,11 @@ llm:
   max_retries: 3
   retry_backoff_base: 1.0
   retry_backoff_max: 30.0
+  thinking: false  # Disable model thinking/reasoning mode (reduces reasoning-only loops)
   # Extra parameters merged into the API request body (model-specific).
   # Examples:
-  #   Qwen 3.x:  enable_thinking: false
-  #   DeepSeek:   enable_thinking: false
   #   OpenAI: reasoning_effort: "low"
-  extra_body:
+  extra_body: {}
-    enable_thinking: false
 
 agent:
   max_iterations: 25