auto-sync: 2026-05-26 08:21:55

src/daemon/spawner.py

@@ -544,7 +544,12 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
 
     async def _handle_exit(self, session_id, agent_id, task_id, exit_code,
                            stdout_chunks, stderr_chunks, on_complete, db_path):
-        """情况 A：进程退出后的处理"""
+        """情况 A：进程退出后的处理
+
+        v2.7.2: 进程退出 = counter release（由 on_complete = wrapped_on_complete 保证）。
+        只有 A2/A3（gateway_timeout）触发续杯，其他都不 retry。
+        A9（api_error/429）额外推回 pending + 设冷却。
+        """
         stdout_text = b"".join(stdout_chunks).decode("utf-8", errors="replace")
         stderr_text = b"".join(stderr_chunks).decode("utf-8", errors="replace")
 
@@ -582,16 +587,45 @@ curl -X POST http://{api_host}:{api_port}/api/projects/{project_id}/tasks/{task_
         logger.info("Agent %s finished (session=%s, outcome=%s, exit=%d, task_status=%s)",
                     agent_id, session_id, outcome, exit_code, task_status)
 
-        if cls["release_counter"]:
-            await self._do_on_complete_async(on_complete, agent_id, outcome)
-        elif cls["should_retry"]:
-            # 续杯期间 counter 保持占用（设计文档规定）
-            # on_complete 传入续杯链：最终完成或超限时由 _do_retry 调用 release
+        if cls["should_retry"]:
+            # A2/A3: gateway_timeout → 续杯（on_complete 会 release counter）
             await self._do_retry(
                 session_id, agent_id, task_id, on_complete, db_path,
                 cls.get("retry_field", "retry_count")
             )
-        # else: 暂时性失败(A8/A9/A11)，不 release，不 retry，等 ticker
+        elif outcome == "api_error":
+            # A9: 429/API 错误 → release counter（on_complete）+ 推回 pending + 冷却
+            await self._do_on_complete_async(on_complete, agent_id, outcome)
+            if self.counter:
+                self.counter.set_cooldown(agent_id)
+            if db_path and task_id:
+                self._mark_task(db_path, task_id, "pending", {
+                    "reason": "api_error_retry",
+                })
+                logger.info("Task %s pushed back to pending (api_error)", task_id)
+        elif outcome == "fallback_timeout" and not cls["should_retry"]:
+            # A5/A6: fallback 不应出现，标 failed + escalate + context 日志
+            logger.error("UNEXPECTED FALLBACK: agent=%s session=%s task=%s "
+                         "transport=%s fallbackReason=%s counter_active=%s "
+                         "This indicates counter check failed to prevent concurrent spawn.",
+                         agent_id, session_id, task_id,
+                         meta.get("transport"), meta.get("fallbackReason"),
+                         self.counter.active_agents if self.counter else "N/A")
+            await self._do_on_complete_async(on_complete, agent_id, outcome)
+            if db_path and task_id:
+                self._mark_task(db_path, task_id, "failed", {
+                    "reason": "unexpected_fallback",
+                    "transport": meta.get("transport"),
+                    "fallback_reason": meta.get("fallbackReason"),
+                    "duration_ms": meta.get("durationMs"),
+                })
+        else:
+            # 其他：A1(completed), A4(agent_failed), A7(auth_failed),
+            #      A8(gateway_unreachable), A11(lock_conflict),
+            #      A10(compact_failed), A12(agent_error)
+            # 进程退出 → on_complete release counter
+            # 任务状态由各 outcome 自行处理（或等 ticker）
+            await self._do_on_complete_async(on_complete, agent_id, outcome)
 
     async def _handle_monitor_timeout(self, session_id, agent_id, task_id, proc,
                                       on_complete, db_path, stderr_chunks,