Switch LLM backend from llama.cpp/Qwen to OpenAI

- Default models: gpt-4o-mini (triage), gpt-4o (escalation)
- Remove Qwen-specific /no_think hacks
- Reduce timeout from 600s to 120s, increase concurrency semaphore to 4
- Support empty LLM_BASE_URL to use OpenAI directly

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

utils/llm_client.py

@@ -132,12 +132,11 @@ class LLMClient:
         self.model = model
         self.host = base_url.rstrip("/")
         self._db = db
-        self._client = AsyncOpenAI(
-            base_url=f"{self.host}/v1",
-            api_key=api_key,
-            timeout=600.0,  # 10 min — first request loads model into VRAM
-        )
-        self._semaphore = asyncio.Semaphore(1)  # serialize requests to avoid overloading
+        client_kwargs = {"api_key": api_key, "timeout": 120.0}
+        if self.host:
+            client_kwargs["base_url"] = f"{self.host}/v1"
+        self._client = AsyncOpenAI(**client_kwargs)
+        self._semaphore = asyncio.Semaphore(4)
 
     def _log_llm(self, request_type: str, duration_ms: int, success: bool,
                  request: str, response: str | None = None, error: str | None = None,
@@ -169,7 +168,7 @@ class LLMClient:
             user_content += f"=== NOTES ABOUT THIS USER (from prior analysis) ===\n{user_notes}\n\n"
         if channel_context:
             user_content += f"=== CHANNEL INFO ===\n{channel_context}\n\n"
-        user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}\n/no_think"
+        user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}\n"
 
         req_json = json.dumps([
             {"role": "system", "content": SYSTEM_PROMPT[:500]},
@@ -300,16 +299,9 @@ class LLMClient:
         first content token arrives (useful for triggering the typing indicator
         only after the model starts generating).
         """
-        # Append /no_think to the last user message to disable thinking on Qwen3
-        patched = []
-        for m in messages:
-            patched.append(m)
-        if patched and patched[-1].get("role") == "user":
-            patched[-1] = {**patched[-1], "content": patched[-1]["content"] + "\n/no_think"}
-
         req_json = json.dumps([
             {"role": "system", "content": system_prompt[:500]},
-            *[{"role": m["role"], "content": str(m.get("content", ""))[:200]} for m in patched],
+            *[{"role": m["role"], "content": str(m.get("content", ""))[:200]} for m in messages],
         ], default=str)
         t0 = time.monotonic()
 
@@ -319,7 +311,7 @@ class LLMClient:
                     model=self.model,
                     messages=[
                         {"role": "system", "content": system_prompt},
-                        *patched,
+                        *messages,
                     ],
                     temperature=0.8,
                     max_tokens=2048,
@@ -363,7 +355,8 @@ class LLMClient:
         user_content: list[dict] = [
             {"type": "image_url", "image_url": {"url": data_url}},
         ]
-        user_content.append({"type": "text", "text": (user_text or "") + "\n/no_think"})
+        if user_text:
+            user_content.append({"type": "text", "text": user_text})
 
         req_json = json.dumps([
             {"role": "system", "content": system_prompt[:500]},
@@ -422,7 +415,7 @@ class LLMClient:
             user_content += f"=== NOTES ABOUT THIS USER (from prior analysis) ===\n{user_notes}\n\n"
         if channel_context:
             user_content += f"=== CHANNEL INFO ===\n{channel_context}\n\n"
-        user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}\n/no_think"
+        user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}\n"
 
         req_json = json.dumps([
             {"role": "system", "content": SYSTEM_PROMPT[:500]},