Merge pull request 'fix: 关 qwen 思考模式(消除 reasoning 回退) + 修 bounce 死锁 (BUG-A/BUG-B)' (#26) from fix/qwen-thinking-and-bounce-deadlock into main

Reviewed-on: #26

agent/task_executor.py

@@ -610,9 +610,19 @@ Return ONLY the JSON, no other text."""
             rc = extra.get("reasoning_content") if isinstance(extra, dict) else None
         return (rc or "").strip()
 
+    @staticmethod
+    def _thinking_extra_body() -> Optional[dict]:
+        """qwen3.7-max defaults to THINKING mode: on heavy impl subtasks it spends the token
+        budget on `reasoning_content` and never emits the final `content` JSON, so json.loads("")
+        fails ("Expecting value: line 1 column 1"). Disable thinking for deterministic structured
+        execution. Set AGENT_ENABLE_THINKING=1 to opt back in (e.g. for non-structured analysis)."""
+        enabled = os.getenv("AGENT_ENABLE_THINKING", "false").strip().lower() in ("1", "true", "yes", "on")
+        return None if enabled else {"enable_thinking": False}
+
     async def _complete(self, prompt: str, max_tokens: int) -> str:
         """Call the LLM with optional Jina MCP tools; handles the tool-call loop."""
         extra_headers = self._model_attribution_headers()
+        extra_body = self._thinking_extra_body()
         tools = await self._load_jina_tools()
         messages = [{"role": "user", "content": prompt}]
 
@@ -623,6 +633,8 @@ Return ONLY the JSON, no other text."""
                 max_tokens=max_tokens,
                 extra_headers=extra_headers or None,
             )
+            if extra_body:
+                kwargs["extra_body"] = extra_body
             if tools:
                 kwargs["tools"] = tools
                 kwargs["tool_choice"] = "auto"
@@ -656,10 +668,13 @@ Return ONLY the JSON, no other text."""
 
         # Fallback: ask for a final answer without tools
         messages.append({"role": "user", "content": "Please provide your final answer now."})
-        response = await self.client.chat.completions.create(
+        fallback_kwargs: dict = dict(
             model=self.model, messages=messages, max_tokens=max_tokens,
             extra_headers=extra_headers or None,
         )
+        if extra_body:
+            fallback_kwargs["extra_body"] = extra_body
+        response = await self.client.chat.completions.create(**fallback_kwargs)
         self._record_openai_usage(response)
         return self._message_text(response.choices[0].message)
 

orchestrator/main.py

@@ -776,13 +776,29 @@ async def refresh_swarm_run_status(run):
                 if collect_generated_files([t]).get("impl"):
                     if await task_queue.reopen_task(t.task_id):
                         reopened += 1
-            run.metadata["review_cycles"] = agg_cycles + 1
-            run.metadata["rework_reopens"] = int(run.metadata.get("rework_reopens", 0) or 0) + reopened
-            await swarm_runtime.save_run(run)
-            logger.info("aggregation: bounced run %s — reopened %d for rework (cycle %d, pass_rate=%s)",
-                        run.swarm_id, reopened, agg_cycles + 1,
-                        (agg.get("validation") or {}).get("pass_rate"))
-            return
+            # Deadlock guard: when the impl subtask(s) FAILED (0 impl files) and only test tasks
+            # produced output, the loop above finds nothing to reopen → reopens=0 → the run would
+            # spin 'running' forever. Reopen the FAILED tasks so the next cycle retries the impl.
+            if reopened == 0:
+                for t in tasks:
+                    if t.status == TaskStatus.FAILED and await task_queue.reopen_task(t.task_id):
+                        reopened += 1
+            if reopened > 0:
+                run.metadata["review_cycles"] = agg_cycles + 1
+                run.metadata["rework_reopens"] = int(run.metadata.get("rework_reopens", 0) or 0) + reopened
+                await swarm_runtime.save_run(run)
+                logger.info("aggregation: bounced run %s — reopened %d for rework (cycle %d, pass_rate=%s)",
+                            run.swarm_id, reopened, agg_cycles + 1,
+                            (agg.get("validation") or {}).get("pass_rate"))
+                return
+            # Nothing reopenable (no impl files and no failed tasks to retry): bouncing here would
+            # deadlock the run in 'running'. Stop bouncing and accept the existing artifacts so the
+            # run terminates via the normal completion/convergence path below.
+            agg["decision"] = "accept_no_rework"
+            run.metadata["aggregation"] = agg
+            logger.warning("aggregation: bounce on run %s but nothing reopenable (reopens=0) — "
+                           "accepting existing artifacts to avoid deadlock (cycle %d)",
+                           run.swarm_id, agg_cycles + 1)
         await swarm_runtime.save_run(run)
 
     if run.status == next_status: