fix(wework): recover guidance turn races

docs/en/developer-guide/local-device-architecture.md

@@ -65,6 +65,8 @@ The runtime task `running` field represents only whether a model turn is current
 
 Goals have an independent lifecycle. An `active` goal means that its objective can continue in later turns; it does not mean that a model turn is currently executing. Keeping an active goal while a task is idle must not mark the task as running again. A user's next message creates a new turn directly instead of being sent as guidance to an in-progress turn.
 
+Codex guidance is sent to the active turn through the shared app-server. If that turn finishes or changes while guidance is being sent, the executor reports the race as `no_active_turn`; Wework then sends the same content as a normal follow-up message so user input is preserved without a misleading send failure.
+
 ### Backend Device Chat Task REST Entrypoint
 
 The web device chat page still sends messages through WebSocket. For external systems or curl-based callers that need to create the same kind of task, Backend exposes a REST entrypoint:

docs/zh/developer-guide/local-device-architecture.md

@@ -65,6 +65,8 @@ Backend 是可选能力，而不是本地 app 的必需依赖。需要登录、
 
 目标（goal）有独立的生命周期。目标为 `active` 表示其目标仍可在后续回合继续推进，不表示当前存在模型回合。因此，任务空闲时保留 active goal 不会将任务重新标记为运行中；用户发送下一条消息会直接创建新回合，而不是把消息作为对运行中回合的引导。
 
+Codex 引导通过共享 app-server 的活跃回合发送。若回合恰好在发送期间结束或切换，executor 会将该竞态报告为 `no_active_turn`；Wework 随后把同一内容作为普通后续消息发送，避免丢失用户输入或显示误导性的发送失败。
+
 ### 后端设备对话任务 REST 入口
 
 网页版设备对话页仍然通过 WebSocket 发送消息。对于需要从外部系统或 curl 创建同类任务的场景，Backend 提供 REST 入口：

executor/src/runtime_work/handler.rs

@@ -2176,14 +2176,17 @@ impl RuntimeWorkRpcHandler {
                 "turnId": turn_id,
                 "runtime": "codex",
             })),
-            Err(error) => Ok(json!({
-                "success": false,
-                "accepted": false,
-                "error": error,
-                "code": "guidance_failed",
-                "taskId": local_task_id,
-                "runtime": "codex",
-            })),
+            Err(error) => {
+                let code = codex_guidance_failure_code(&error);
+                Ok(json!({
+                    "success": false,
+                    "accepted": false,
+                    "error": error,
+                    "code": code,
+                    "taskId": local_task_id,
+                    "runtime": "codex",
+                }))
+            }
         }
     }
 
@@ -4972,6 +4975,17 @@ fn guidance_input_items(message: &str, attachments: Option<&Value>) -> Vec<Value
     inputs
 }
 
+fn codex_guidance_failure_code(error: &str) -> &'static str {
+    let normalized = error.to_ascii_lowercase();
+    if normalized.contains("no active turn to steer")
+        || (normalized.contains("expected active turn id") && normalized.contains("but found"))
+    {
+        "no_active_turn"
+    } else {
+        "guidance_failed"
+    }
+}
+
 fn copy_attachment_field(source: &Map<String, Value>, target: &mut Map<String, Value>, key: &str) {
     if let Some(value) = source.get(key).cloned() {
         target.insert(key.to_owned(), value);
@@ -6598,6 +6612,22 @@ mod tests {
         assert!(guidance_input_items("", None).is_empty());
     }
 
+    #[test]
+    fn codex_guidance_turn_races_are_reported_as_no_active_turn() {
+        assert_eq!(
+            codex_guidance_failure_code("no active turn to steer"),
+            "no_active_turn"
+        );
+        assert_eq!(
+            codex_guidance_failure_code("expected active turn id `turn-1` but found `turn-2`"),
+            "no_active_turn"
+        );
+        assert_eq!(
+            codex_guidance_failure_code("turn/steer response missing turnId"),
+            "guidance_failed"
+        );
+    }
+
     #[test]
     fn archived_cleanup_targets_do_not_delete_regular_project_root() {
         let link = RuntimeTaskLink::new_pending(

wework/src/features/workbench/WorkbenchProvider.test.tsx

@@ -6765,13 +6765,13 @@ describe('WorkbenchProvider runtime tasks', () => {
     expect(screen.getByTestId('queued-notices')).not.toHaveTextContent('正在引导当前对话')
   })
 
-  test('sends queued guidance as a follow-up when the pane is already idle', async () => {
+  test('sends queued guidance as a follow-up when the active turn is unavailable', async () => {
     let streamHandlers: ChatStreamHandlers = {}
     const subscribe = vi.fn((handlers: ChatStreamHandlers) => {
       if (hasRuntimeStreamHandler(handlers)) streamHandlers = handlers
       return vi.fn()
     })
-    const sendRuntimeMessage = vi.fn().mockResolvedValueOnce(false).mockResolvedValueOnce({
+    const sendRuntimeMessage = vi.fn().mockResolvedValueOnce(false).mockResolvedValue({
       accepted: true,
       taskId: 'runtime-a',
     })
@@ -6877,6 +6877,7 @@ describe('WorkbenchProvider runtime tasks', () => {
 
     await userEvent.click(screen.getByText('guide first queued'))
 
+    await waitFor(() => expect(guideRuntimeTask).toHaveBeenCalledTimes(1))
     await waitFor(() => expect(sendRuntimeMessage).toHaveBeenCalledTimes(2))
     expect(sendRuntimeMessage).toHaveBeenLastCalledWith({
       address: {