diff --git a/apps/ai-game-creator-shell/src-tauri/src/agent/claude_code_cli.rs b/apps/ai-game-creator-shell/src-tauri/src/agent/claude_code_cli.rs index 827b06682..9f5b91221 100644 --- a/apps/ai-game-creator-shell/src-tauri/src/agent/claude_code_cli.rs +++ b/apps/ai-game-creator-shell/src-tauri/src/agent/claude_code_cli.rs @@ -732,7 +732,9 @@ pub(crate) async fn direct_game_creator_claude_code_chat_at( client_turn_id: Option<&str>, observer: Option<&mut (dyn FnMut(DirectCodexTurnObservation) + Send)>, ) -> Result { + direct_turn_trace("claude-executor-enter"); let (mcp_url, mcp_token) = start_external_mcp_loopback(root, llm.web_search_enabled).await?; + direct_turn_trace("claude-mcp-loopback-ready"); let _mcp_guard = ExternalMcpStopGuard; let home = crate::project::resolve_local_project_path(root, ".agent/runtime/claude-code/home")?; std::fs::create_dir_all(&home) @@ -742,6 +744,7 @@ pub(crate) async fn direct_game_creator_claude_code_chat_at( .lock() .ok() .and_then(|sessions| sessions.get(&session_key).cloned()); + direct_turn_trace("claude-sidecar-start"); let payload = serde_json::json!({ "type": "turn", "id": client_turn_id.unwrap_or("claude-direct-turn"), @@ -769,6 +772,7 @@ pub(crate) async fn direct_game_creator_claude_code_chat_at( client_turn_id, ) .await?; + direct_turn_trace("claude-sidecar-done"); if let Some(session_id) = sidecar .result .get("session_id") @@ -785,6 +789,7 @@ pub(crate) async fn direct_game_creator_claude_code_chat_at( .filter_map(|event| serde_json::to_string(&event).ok()) .collect::>(); stream.push(String::new()); + direct_turn_trace("claude-parse-start"); parse_direct_stream_result(stream.join("\n").as_bytes(), observer) } diff --git a/apps/ai-game-creator-shell/src-tauri/src/agent/direct_runtime/mod.rs b/apps/ai-game-creator-shell/src-tauri/src/agent/direct_runtime/mod.rs index f5059ed82..92d673ebb 100644 --- a/apps/ai-game-creator-shell/src-tauri/src/agent/direct_runtime/mod.rs +++ b/apps/ai-game-creator-shell/src-tauri/src/agent/direct_runtime/mod.rs @@ -4294,6 +4294,24 @@ pub(crate) fn check_direct_turn_preconditions( .map_err(|detail| DirectTurnError::InputRejected { detail })?; Ok(()) } +/// Direct 回合在 `direct_execution::begin` 之后的分段 trace。 +/// +/// 起因:2026-10-02 连续四轮回合都停在 `begin` 之后——账本 `phase=working`、 +/// `nextSequence=0`、`active={}`、没有任何错误 sidecar、stderr 无 panic、连 +/// `agent.direct_codex.*` 都没有。窗口就夹在 `begin` 与执行器调用之间,需要知道 +/// 具体停在哪一步。测试态不写,避免污染本机 AppData。 +#[cfg(not(test))] +pub(crate) fn direct_turn_trace(stage: &str) { + let _ = crate::append_application_log( + "info".to_string(), + "direct-turn".to_string(), + format!("agent.direct_turn.trace stage={stage}"), + ); +} + +#[cfg(test)] +pub(crate) fn direct_turn_trace(_stage: &str) {} + pub(crate) async fn run_direct_game_creator_turn_at_with_creation_type_and_emitter( root: &Path, prompt: &str, @@ -4401,6 +4419,7 @@ async fn run_direct_game_creator_turn_inner( })?; let execution_session = execution_guard.session(); execution_session.set_analytics_capture(capture.clone()); + direct_turn_trace("session-ready"); let started = execution_session .newly_accepted .then(std::time::Instant::now); @@ -4411,6 +4430,7 @@ async fn run_direct_game_creator_turn_inner( } } } + direct_turn_trace("run-analytics-accepted"); // 在 guard 仍存活时冻结整体结果,避免 Drop 的中断收尾覆盖真实失败原因。 let result: Result = async { if let Some(report) = super::direct_delivery::terminal_report(&execution_session) { @@ -4444,6 +4464,7 @@ async fn run_direct_game_creator_turn_inner( ); } } + direct_turn_trace("goal-analytics-accepted"); emit_direct_game_creator_progress(root, "codex.turn", "陶泥儿正在处理这条消息"); if let Some(emitter) = turn_emitter { emitter.emit("running", Some("preparing")); @@ -4460,6 +4481,7 @@ async fn run_direct_game_creator_turn_inner( DirectTurnError::turn_failed(DirectCodexFailureStage::CodeGeneration, error) }, )?; + direct_turn_trace("system-prompt-ready"); // 三维请求:把"自选三维技术栈、解除 Phaser 固定约束"的合同放在系统提示最前, // 避免被长度上限截断,也不阻断任何工具。 let engine_contract = direct_engine_three_dimensional_contract(root, prompt); @@ -4471,6 +4493,7 @@ async fn run_direct_game_creator_turn_inner( None => base_system_prompt, }; let reply = if let Some(emitter) = turn_emitter { + direct_turn_trace("executor-dispatch"); let client_turn_id = emitter.turn_id().to_string(); let emitter = emitter.clone(); let mut observer = |observation: DirectCodexTurnObservation| { diff --git a/apps/ai-game-creator-shell/src-tauri/src/agent/thread_manager/dispatch.rs b/apps/ai-game-creator-shell/src-tauri/src/agent/thread_manager/dispatch.rs index 5eabb23b6..be5bc35cb 100644 --- a/apps/ai-game-creator-shell/src-tauri/src/agent/thread_manager/dispatch.rs +++ b/apps/ai-game-creator-shell/src-tauri/src/agent/thread_manager/dispatch.rs @@ -13,13 +13,16 @@ use std::path::{Path, PathBuf}; +use futures::FutureExt; + #[cfg(test)] use crate::agent::PendingTurn; use crate::agent::{ append_direct_project_user_message_at, claim_pending_turn, complete_turn_if_reserved, - direct_codex_user_item_to_prompt, direct_now_ms, redact_agent_runtime_error, - run_direct_game_creator_turn_at_with_creation_type_and_emitter, thread_id_for_project, - DirectGameCreatorTurnUpdateEmitter, DirectTurnError, DirectTurnTerminal, DispatchedTurn, + direct_codex_user_item_to_prompt, direct_now_ms, record_direct_codex_failure, + redact_agent_runtime_error, run_direct_game_creator_turn_at_with_creation_type_and_emitter, + thread_id_for_project, DirectCodexFailureStage, DirectGameCreatorTurnUpdateEmitter, + DirectTurnError, DirectTurnTerminal, DispatchedTurn, }; /// 一次放行的占用。持有它就代表这一轮还没收口。 @@ -84,13 +87,90 @@ impl Drop for TurnReservation { fn drop(&mut self) { // 兜底:任务 panic、future 被丢弃、或今后在终态之前新增的 `?` 早退。 // 这类失败说不出原因,只给分类;能说清原因的错误必须由调用方在更早的地方显式收口。 - let _ = self.finish_if_unfinished(DirectTurnTerminal::host_dropped()); + let finalized_by_guard = self.finish_if_unfinished(DirectTurnTerminal::host_dropped()); + // 兜底一旦真的收口,就说明这一轮**从未写下终态**:深层既没成功也没失败地退出了。 + // 这条必须留应用日志,否则离线只剩一个 `phase=working` 的账本,无从判断是哪一层 + // 提前退出(真实案例:2026-10-02 连续两轮只留下 working 账本,errors/ 与 + // application.log 全空)。测试态不写,避免用例污染本机 AppData。 + #[cfg(not(test))] + if finalized_by_guard { + // 项目侧也留一份脱敏诊断:用户在项目目录里就能看到这一轮的收场, + // 不必只依赖 AppData 的应用日志。 + let failure = DirectTurnError::TurnFailed { + stage: DirectCodexFailureStage::CodeGeneration, + detail: format!( + "DirectProject 宿主任务提前结束(panic、future 被丢弃或被取消),本轮未写下终态;threadId={} turnToken={}", + self.thread_id, self.token + ), + }; + let _ = record_direct_codex_failure(Path::new(&self.thread_id), &failure, None); + let _ = crate::append_application_log( + "error".to_string(), + "direct-turn".to_string(), + format!( + "agent.direct_turn.host_dropped threadId={} turnToken={} userItemId={} panicking={}", + self.thread_id, + self.token, + self.user_item_id.as_deref().unwrap_or("-"), + // true = 这一轮是被 panic 带走的;false = future 被丢弃或取消。 + std::thread::panicking() + ), + ); + } // 占用的释放就是"这个项目腾出了跑回合的位置":踢一脚,队里排着的消息不必等下一次用户动作。 // panic 也走这里,所以队列不会因为一个任务炸掉而永久停住。 kick_queue_dispatch(Path::new(&self.thread_id)); } } +// 运行段经 task-local 携带回合身份,panic hook 据此把 panic 位置与负载写进应用日志。 +// task-local 而非 thread-local:多线程 runtime 下 future 会跨 worker 迁移。 +tokio::task_local! { + static DIRECT_TURN_PANIC_CONTEXT: Option<(String, String)>; +} + +/// 兜底诊断:Direct 回合任务 panic 时,把位置与负载写进应用日志。 +/// +/// 没有它时,panic 只会让 `TurnReservation::drop` 把回合收成 `HostDropped`——那是"说不出原因" +/// 的分类,离线只留一个 `phase=working` 的账本,`.agent/runtime/errors/` 与 `application.log` +/// 全空,无法判断是哪一层退出的(真实案例:2026-10-02 连续两轮如此)。 +fn ensure_direct_turn_panic_hook() { + static ONCE: std::sync::Once = std::sync::Once::new(); + ONCE.call_once(|| { + let previous = std::panic::take_hook(); + std::panic::set_hook(Box::new(move |info| { + if let Ok(Some((thread_id, token))) = DIRECT_TURN_PANIC_CONTEXT.try_with(Clone::clone) { + let location = info + .location() + .map(|location| { + format!( + "{}:{}:{}", + location.file(), + location.line(), + location.column() + ) + }) + .unwrap_or_else(|| "未知位置".to_string()); + #[cfg(not(test))] + let _ = crate::append_application_log( + "error".to_string(), + "direct-turn".to_string(), + format!( + "agent.direct_turn.panic threadId={} turnToken={} location={} payload={}", + thread_id, + token, + location, + info.payload_as_str().unwrap_or("未知 panic 负载") + ), + ); + #[cfg(test)] + let _ = (thread_id, token, location); + } + previous(info); + })); + }); +} + /// 踢一脚:让队首那条待发消息有机会变成一轮真的回合。 /// /// 幂等,三个调用点:**入队之后**(队列空且没有回合在跑时,放行就是这一脚,不必再等一个调度周期)、 @@ -109,10 +189,23 @@ pub(crate) fn kick_queue_dispatch(root: &Path) { crate::platform_session::current_platform_session_write_state().identity_generation; let reservation = TurnReservation::resume(&thread_id, &dispatched); let root = root.to_path_buf(); - tauri::async_runtime::spawn(async move { + ensure_direct_turn_panic_hook(); + let panic_context = Some((thread_id.clone(), dispatched.token.clone())); + tauri::async_runtime::spawn(DIRECT_TURN_PANIC_CONTEXT.scope(panic_context, async move { run_dispatched_direct_turn(root, dispatched, reservation, release_identity_generation) .await; - }); + })); +} + +/// 把 panic 负载转成可读文本:`panic!("…")` 的负载是 `&str`,`panic!("{x}")` 是 `String`。 +fn direct_turn_panic_detail(payload: &(dyn std::any::Any + Send)) -> String { + if let Some(text) = payload.downcast_ref::<&str>() { + (*text).to_string() + } else if let Some(text) = payload.downcast_ref::() { + text.clone() + } else { + "未知 panic 负载".to_string() + } } /// 放行之后这一段:取这一轮的调用身份 → 落盘用户条目 → 下发 → 跑整轮。 @@ -149,16 +242,38 @@ async fn run_dispatched_direct_turn( crate::agent::codex_app_server::emit_direct_thread_user_item(&root, &canonical_user_item); let capture = crate::analytics::gui::capture_writer_context(); let emitter = DirectGameCreatorTurnUpdateEmitter::new(&root, turn_id); - let outcome = run_direct_game_creator_turn_at_with_creation_type_and_emitter( - &root, - &prompt, - dispatched.pending.creation_type.as_deref(), - Some(&emitter), - Some(canonical_user_item.clone()), - capture, - release_identity_generation, + // 回合体 panic 不能落到 `TurnReservation` 的 Drop 兜底:那会把这一轮收成"说不出原因"的 + // `HostDropped`,症状正是用户看到的「本轮执行已中断,请重试」,而项目侧诊断与应用日志 + // **一行都不会有**(真实案例:2026-10-02 连续两轮)。这里把 panic 转成分类失败, + // 走与正常失败同一条收口:先写脱敏诊断,再写终态。 + let outcome = match std::panic::AssertUnwindSafe( + run_direct_game_creator_turn_at_with_creation_type_and_emitter( + &root, + &prompt, + dispatched.pending.creation_type.as_deref(), + Some(&emitter), + Some(canonical_user_item.clone()), + capture, + release_identity_generation, + ), ) - .await; + .catch_unwind() + .await + { + Ok(result) => result, + Err(payload) => { + let failure = DirectTurnError::turn_failed( + DirectCodexFailureStage::CodeGeneration, + format!( + "DirectProject 宿主任务 panic:{}", + direct_turn_panic_detail(payload.as_ref()) + ), + ); + let stage = failure.turn_failure_stage(); + let detail = record_direct_codex_failure(&root, &failure, Some(turn_id.as_str())); + Err(DirectTurnError::TurnFailed { stage, detail }) + } + }; match outcome { Ok(_) => { // 深层的终态出口已经在 `run_turn` 里写出 `turn.completed`;这里只补最后一条回合更新。