feat(agent): 任务执行加结构化日志(LLM推理/tool调用query与返回/产出)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
(cherry picked from commit 099f270cc2a092a4eae7e5d06d99c4d7e076465e)

agent/task_executor.py

@@ -342,6 +342,8 @@ Return ONLY the JSON array, no other text."""
         content = ""
         try:
             description = subtask["description"]
+            logger.info("════════ [任务开始] task=%s role=%s\n  描述: %s",
+                        task_id, context.get("specialist_role", "general"), description)
             workspace_files = self._summarize_workspace()
             workspace_context = self._collect_workspace_context()
             user_prompt = context.get("user_prompt") or context.get("run_goal") or context.get("root_task_description") or ""
@@ -437,6 +439,10 @@ Return your response as JSON with this structure:
 Return ONLY the JSON, no other text."""
             content = await self._complete(prompt, max_tokens=4000)
             result = self._parse_json_response(content)
+            logger.info("──────── [任务产出] task=%s status=%s 文件=%s\n  说明: %s",
+                        task_id, result.get("status"),
+                        [f.get("path") for f in (result.get("files") or [])],
+                        (result.get("changes") or result.get("summary") or "")[:600])
             if result.get("status") == "completed":
                 apply_result = await self._apply_file_changes(result.get("files", []))
                 result["files_modified"] = apply_result["files_modified"]
@@ -539,7 +545,7 @@ Return ONLY the JSON, no other text."""
         tools = await self._load_jina_tools()
         messages = [{"role": "user", "content": prompt}]
 
-        for _ in range(8):  # max 8 tool-call rounds
+        for round_i in range(8):  # max 8 tool-call rounds
             kwargs: dict = dict(
                 model=self.model,
                 messages=messages,
@@ -553,6 +559,10 @@ Return ONLY the JSON, no other text."""
             self._record_openai_usage(response)
             msg = response.choices[0].message
 
+            # 记录这一轮 LLM 的"思考/回复"(可观测 agent 怎么想的)
+            if msg.content:
+                logger.info("[LLM·第%d轮] 思考/回复:\n%s", round_i + 1, msg.content.strip()[:1200])
+
             if not msg.tool_calls:
                 return msg.content or ""
 
@@ -560,8 +570,13 @@ Return ONLY the JSON, no other text."""
             messages.append(msg.model_dump(exclude_unset=True))
             for tc in msg.tool_calls:
                 args = json.loads(tc.function.arguments or "{}")
+                # 记录调用了哪个 tool、query 是什么
+                logger.info("[TOOL·调用] %s 参数=%s", tc.function.name,
+                            json.dumps(args, ensure_ascii=False)[:400])
                 result = await self._call_jina_tool(tc.function.name, args)
-                logger.info("Jina tool %s → %d chars", tc.function.name, len(result))
+                # 记录 tool 返回了什么(搜索结果内容,截断)
+                logger.info("[TOOL·返回] %s (%d字):\n%s", tc.function.name, len(result),
+                            (result or "").strip()[:1000])
                 messages.append({
                     "role": "tool",
                     "tool_call_id": tc.id,