update ai dolboeba fix2

AI/talk_handler.py

@@ -28,6 +28,8 @@ SUMMARY_CHAR_BUDGET = 1_800
 MAX_INPUT_CHARS = 2_000
 SUMMARY_LINE_CHAR_LIMIT = 220
 HISTORY_LINE_CHAR_LIMIT = 450
+RETRY_MIN_MAX_TOKENS = 384
+RETRY_MAX_MAX_TOKENS = 768
 
 REPLY_RULES = {
     "mention": {"cooldown": 25, "min_user_messages": 1},
@@ -311,6 +313,7 @@ async def _call_llm(
     max_tokens: int,
     temperature: float,
     top_p: float,
+    disable_thinking: bool = False,
 ) -> str:
     url = f"{LLAMA_API_URL.rstrip('/')}/v1/chat/completions"
     payload = {
@@ -319,6 +322,17 @@ async def _call_llm(
         "temperature": temperature,
         "top_p": top_p,
     }
+    if disable_thinking:
+        payload.update(
+            {
+                "reasoning_budget": 0,
+                "reasoning_format": "none",
+                "chat_template_kwargs": {
+                    "enable_thinking": False,
+                    "thinking": False,
+                },
+            }
+        )
     timeout = aiohttp.ClientTimeout(total=120)
     async with aiohttp.ClientSession(timeout=timeout) as session:
         async with session.post(url, json=payload) as resp:
@@ -336,10 +350,33 @@ async def _call_llm(
     if not choices:
         logger.error("LLM API returned no choices: %s", _clip_text(str(data), 300))
         raise RuntimeError(f"LLM API returned no choices: {data}")
-    content = choices[0].get("message", {}).get("content", "")
+    choice = choices[0]
+    message = choice.get("message", {}) or {}
+    finish_reason = choice.get("finish_reason")
+    content = message.get("content", "")
+    reasoning_content = (message.get("reasoning_content") or "").strip()
     cleaned_content = content.strip()
+    if not cleaned_content and reasoning_content and not disable_thinking:
+        retry_max_tokens = min(max(max_tokens * 2, RETRY_MIN_MAX_TOKENS), RETRY_MAX_MAX_TOKENS)
+        logger.warning(
+            "LLM returned reasoning_content without final content. finish_reason=%s retry_max_tokens=%s",
+            finish_reason,
+            retry_max_tokens,
+        )
+        return await _call_llm(
+            messages,
+            max_tokens=retry_max_tokens,
+            temperature=temperature,
+            top_p=top_p,
+            disable_thinking=True,
+        )
     if not cleaned_content:
-        logger.warning("LLM returned empty content. Raw response: %s", _clip_text(raw_text, 300))
+        logger.warning(
+            "LLM returned empty content. finish_reason=%s disable_thinking=%s raw=%s",
+            finish_reason,
+            disable_thinking,
+            _clip_text(raw_text, 300),
+        )
     return cleaned_content