补齐 AGC Runtime 错误采集

Rust Runtime 终态失败、预算耗尽和启动确认失败统一进入错误队列

Direct Codex 与专业 Agent 前台裸 invoke 补充采集并统一上下文去重

归一化 Codex app-server 错误类别并新增回归测试与技术方案
This commit is contained in:
2026-09-01 21:43:03 +08:00
parent de671b2e0d
commit 35ec049138
7 changed files with 90 additions and 7 deletions
@@ -1410,6 +1410,7 @@ pub(crate) fn fail_game_creator_agent_runtime_turn_at(
state.next_step = "等待开发者处理失败".to_string();
let public_error = redact_agent_runtime_error(root, error, 500);
state.error = Some(public_error.clone());
let _ = crate::error_report::report_agent_runtime_error(&state.agent_id, error);
// The public terminal message is deliberately committed before the
// remaining Runtime projections. Even if a task/event/state write is the
// failing subsystem, the user still receives one stable failure outcome.
@@ -1469,6 +1470,7 @@ pub(crate) fn fail_game_creator_agent_runtime_budget_at(
state.next_step = "调整任务范围后重试".to_string();
let public_error = redact_agent_runtime_error(root, error, 500);
state.error = Some(public_error.clone());
let _ = crate::error_report::report_agent_runtime_error(&state.agent_id, error);
let public_status_result =
append_game_creator_agent_runtime_terminal_public_message_at(root, &state, &public_error);
write_non_terminal_isolated_child_cancel_tombstones_for_parent_at(
@@ -3779,6 +3781,7 @@ pub(crate) fn fail_game_creator_agent_runtime_public_start_status_at(
error: &str,
) -> Result<(), String> {
let error = redact_agent_runtime_project_paths(root, error, 500);
let _ = crate::error_report::report_agent_runtime_error(&record.agent_id, &error);
let failed_task = AgentRuntimeTaskRecord {
status: "failed".to_string(),
phase: "public-status-write-failed".to_string(),
@@ -5,4 +5,4 @@ mod sanitize;
pub use commands::{ack_error_reports, get_pending_error_reports, report_client_error};
pub use notifications::initialize_notifications;
pub use queue::report_diagnostic_error;
pub use queue::{report_agent_runtime_error, report_diagnostic_error};
@@ -48,6 +48,33 @@ fn now() -> String {
format!("{seconds}")
}
fn stable_codex_app_server_error_message(message: &str) -> Option<String> {
for marker in ["kind=codex-app-server-", "codex-app-server-error:"] {
let Some(start) = message.find(marker) else {
continue;
};
let kind_start = start + marker.len();
let kind = message[kind_start..]
.chars()
.take_while(|character| character.is_ascii_lowercase() || *character == '-')
.collect::<String>();
if !kind.is_empty() {
return Some(format!("codex-app-server-error:{kind}"));
}
}
None
}
pub fn report_agent_runtime_error(agent_id: &str, error: &str) -> Option<ErrorReportEvent> {
report_diagnostic_error(
"agent-runtime",
error,
None,
Some("agent-runtime"),
Some(agent_id),
)
}
pub fn report_diagnostic_error(
source: &str,
message: &str,
@@ -71,10 +98,12 @@ pub fn report_diagnostic_error(
})
});
let source = sanitize(source, MAX_SOURCE_CHARS);
let message = sanitize(message, MAX_MESSAGE_CHARS);
let stack = stack.map(|value| sanitize(value, MAX_STACK_CHARS));
let action = action.map(|value| sanitize(value, MAX_ACTION_CHARS));
let page = page.map(|value| sanitize(value, MAX_PAGE_CHARS));
let sanitized_message = sanitize(message, MAX_MESSAGE_CHARS);
let message =
stable_codex_app_server_error_message(&sanitized_message).unwrap_or(sanitized_message);
let stack = stack.map(|value| sanitize(value, MAX_STACK_CHARS));
if message.trim().is_empty() {
return None;
}
@@ -168,7 +197,7 @@ pub(crate) fn reset_for_tests() {
#[cfg(test)]
mod tests {
use super::{report_diagnostic_error, reset_for_tests, snapshot};
use super::{report_agent_runtime_error, report_diagnostic_error, reset_for_tests, snapshot};
#[test]
fn merges_events_by_call_site_without_storing_second_queue() {
@@ -199,4 +228,25 @@ mod tests {
assert!(report_diagnostic_error("test", "\n", None, None, None).is_none());
assert!(snapshot().is_empty());
}
#[test]
fn canonicalizes_codex_app_server_public_summaries_and_deduplicates_frontend_capture() {
reset_for_tests();
let runtime = report_agent_runtime_error(
"project-supervisor",
"runtime 调用 LLM 失败:kind=codex-app-server-other fingerprint=deadbeef chars=42",
)
.expect("runtime event");
let frontend = report_diagnostic_error(
"agent-runtime",
"codex-app-server-error:other",
None,
Some("agent-runtime"),
Some("project-supervisor"),
)
.expect("frontend event");
assert_eq!(runtime.event_id, frontend.event_id);
assert_eq!(snapshot()[0].count, 2);
assert_eq!(snapshot()[0].message, "codex-app-server-error:other");
}
}
+7 -1
View File
@@ -237,7 +237,10 @@ import { ProjectSupervisorView } from './features/project-workspace/ProjectSuper
import { ProjectWorkspaceChatPane } from './features/project-workspace/ProjectWorkspaceChatPane';
import { SupervisorChatOnlyView } from './features/project-workspace/SupervisorChatOnlyView';
import { RuntimeConfigDialog } from './features/runtime-config/RuntimeConfigDialog';
import { invokeDiagnostic } from './services/errorReporting';
import {
captureAgentRuntimeError,
invokeDiagnostic,
} from './services/errorReporting';
import type { HomeCreationType } from './view/home';
import {
type ProjectAgentResultSummary,
@@ -3388,6 +3391,7 @@ export function App({
);
setCommandLog((current) => [...current, 'conversation.write']);
} catch (error) {
void captureAgentRuntimeError(error, agent.id);
if (agentConversationLoadVersionRef.current !== saveVersion) {
return;
}
@@ -5541,6 +5545,7 @@ export function App({
}
return;
}
void captureAgentRuntimeError(error, PROJECT_SUPERVISOR_AGENT_ID);
const message =
error instanceof Error ? error.message : String(error);
const visibleMessage = projectRuntimeVisibleError(
@@ -5746,6 +5751,7 @@ export function App({
);
}
} catch (error) {
void captureAgentRuntimeError(error, PROJECT_SUPERVISOR_AGENT_ID);
if (localProjectPathRef.current !== nextProjectPath) {
return;
}
@@ -75,6 +75,14 @@ export async function captureClientError(
}).catch(() => undefined);
}
export function captureAgentRuntimeError(error: unknown, agentId: string) {
return captureClientError(error, {
source: 'agent-runtime',
action: 'agent-runtime',
page: agentId,
});
}
function formatConsoleArgument(value: unknown) {
if (value instanceof Error) return value.stack || value.message;
if (typeof value === 'string') return value;
@@ -81,6 +81,7 @@ import { invoke } from '@tauri-apps/api/core';
import { fetchClientHttp } from '../src/services/clientHttp';
import {
captureAgentRuntimeError,
captureClientError,
getPendingClientErrorEvents,
installWebviewLogBridge,
@@ -110,6 +111,21 @@ describe('客户端错误报告池', () => {
expect((await getPendingClientErrorEvents())[0]?.count).toBe(2);
});
it('使用统一上下文采集 Agent Runtime 错误', async () => {
await captureAgentRuntimeError(
new Error('kind=codex-app-server-other fingerprint=dynamic chars=12'),
'project-supervisor',
);
expect(invoke).toHaveBeenCalledWith('report_client_error', {
source: 'agent-runtime',
message: 'kind=codex-app-server-other fingerprint=dynamic chars=12',
stack: expect.any(String),
action: 'agent-runtime',
page: 'project-supervisor',
});
});
it('限制当前进程错误池最多保留 100 条', async () => {
for (let index = 0; index < 101; index += 1) {
await captureClientError(new Error(`错误 ${index}`), { source: 'test' });
@@ -6,13 +6,13 @@ AI Game Creator Shell 采用 IDEA 风格的当前进程错误报告:错误事
## 客户端
- 捕获 React render error、`window.onerror``unhandledrejection` 以及显式标记的 Tauri/API/Agent 错误。
- 捕获 React render error、`window.onerror``unhandledrejection` 以及显式标记的 Tauri/API/Agent 错误。Agent Runtime 的终态失败、预算耗尽和启动确认失败由 Rust 失败投影统一入池;Direct Codex 与专业 Agent 的前台裸 Tauri invoke catch 作为补充入口,重复事件由同一 fingerprint 合并,主动取消和“同一 turn 已在运行”不作为错误采集。
- 事件字段包括 eventId、fingerprint、source、message、stack、时间和次数;重复事件合并。不再携带 severity、errorCode、page、action、requestId 等无法稳定关联的字段。
- 指纹计算可使用调用方的 page/action 及脱敏后的首个调用点作为进程内区分输入,但这些上下文不会作为事件字段上传;消息与 stack 在入池前统一脱敏,WebCrypto 失败时降级为稳定可读指纹,采集本身不得产生新的未处理拒绝。
- 客户端 API 自动采集只覆盖网络错误、408 和 5xx;预期的 4xx 登录/鉴权失败不进入错误报告池。
- Rust 侧通过 `app_log!` 将普通文本日志同时输出到 stderr 和 AppData `diagnostics/application.log`,超出 256 KiB 滚动到 `application.previous.log`WebView 的 console 输出通过 `append_application_log` 镜像到同一 raw log,并在客户端桥接处再次脱敏;`read_diagnostic_logs` 只读取应用级日志。
- 报告面板只由自动诊断通知中的“查看并报告”打开,不提供聊天命令、崩溃页按钮或其他手动入口;默认选中最新事件,其他事件可勾选。允许填写最多 2,000 字中文描述并取消日志附件;本版本不支持截图或任意文件附件。
- Rust 是结构化错误队列的唯一真相源:`src-tauri/src/error_report/` 负责脱敏、调用点指纹、eventId、计数、100 条上限、5 秒聚合和 ack 生命周期;WebView 仅通过 Tauri bridge 上报、读取快照并保存短暂 React 展示状态,不维护第二份事件 Map。Rust emit 只作为无状态唤醒,携带单调递增的 generation,不携带错误正文或事件 ID。
- Rust 是结构化错误队列的唯一真相源:`src-tauri/src/error_report/` 负责脱敏、调用点指纹、eventId、计数、100 条上限、5 秒聚合和 ack 生命周期;WebView 仅通过 Tauri bridge 上报、读取快照并保存短暂 React 展示状态,不维护第二份事件 Map。Runtime 错误上报统一使用 `source=agent-runtime``action=agent-runtime` 和 Agent ID 作为 page;包含 `kind=codex-app-server-*``codex-app-server-error:*` 的消息在入池前归一为稳定类别(例如 `codex-app-server-error:other`),不把 public summary 中的动态 fingerprint/长度作为分桶输入。Rust emit 只作为无状态唤醒,携带单调递增的 generation,不携带错误正文或事件 ID。
- 错误事件先在当前进程内存池按 fingerprint 合并,经过 5 秒聚合后只发出一次非阻塞存在性唤醒;WebView 挂载、收到唤醒、重新获得焦点或恢复可见时都查询完整未 ack 快照,并在桥接暂时失败时做有限退避重试。通知支持“查看并报告”和“忽略”,同一 fingerprint 仅在新增时唤醒一次。通知不直接打开阻塞式报告面板。忽略只关闭当前 UI,不删除事件;提交成功后由 bridge ack/delete 选中事件。
- 上传失败只在当前进程显示失败并允许用户再次提交,不跨重启恢复事件池,不后台自动重试。