fix: cache thinking agent and add timeouts to prevent loop pane death (#535)

Co-authored-by: Kimi Agent <kimi@timmy.local>
Co-committed-by: Kimi Agent <kimi@timmy.local>

src/timmy/thinking.py

@@ -1124,21 +1124,37 @@ class ThinkingEngine:
             lines.append(f"- [{thought.seed_type}] {snippet}")
         return "\n".join(lines)
 
+    _thinking_agent = None  # cached agent — avoids per-call resource leaks (#525)
+
     async def _call_agent(self, prompt: str) -> str:
         """Call Timmy's agent to generate a thought.
 
-        Creates a lightweight agent with skip_mcp=True to avoid the cancel-scope
+        Reuses a cached agent with skip_mcp=True to avoid the cancel-scope
         errors that occur when MCP stdio transports are spawned inside asyncio
-        background tasks (#72).  The thinking engine doesn't need Gitea or
-        filesystem tools — it only needs the LLM.
+        background tasks (#72) and to prevent per-call resource leaks (httpx
+        clients, SQLite connections, model warmups) that caused the thinking
+        loop to die every ~10 min (#525).
+
+        Individual calls are capped at 120 s so a hung Ollama never blocks
+        the scheduler indefinitely.
 
         Strips ``<think>`` tags from reasoning models (qwen3, etc.) so that
         downstream parsers (fact distillation, issue filing) receive clean text.
         """
-        from timmy.agent import create_timmy
+        import asyncio
+
+        if self._thinking_agent is None:
+            from timmy.agent import create_timmy
+
+            self._thinking_agent = create_timmy(skip_mcp=True)
+
+        try:
+            async with asyncio.timeout(120):
+                run = await self._thinking_agent.arun(prompt, stream=False)
+        except TimeoutError:
+            logger.warning("Thinking LLM call timed out after 120 s")
+            return ""
 
-        agent = create_timmy(skip_mcp=True)
-        run = await agent.arun(prompt, stream=False)
         raw = run.content if hasattr(run, "content") else str(run)
         return _THINK_TAG_RE.sub("", raw) if raw else raw