feat(agent): 运行时内部错误上报到日志(onReport)

- PiAgentRuntimeOptions.onReport:本次运行扩展的系统提示 section、能力工具、beforeTool / beforeRequest 钩子抛错
  先上报(runId + 位置 + 脱敏信息)再原样抛出;pi-durable 的 Harness onReport 也接进来(source=harness)
- server runner / 视觉子流程用 runtimeReporter 写 warn 日志
- 测试:钩子抛错时上报带 runId 与位置、信息已脱敏,调用照常被 block

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Cxwgnn2pbzQ7rXv3rq7fei

apps/server/src/modules/agent/challenge.ts

@@ -24,7 +24,7 @@ import { BrowserServiceError, type ChallengeVision, registerChallengeVision } fr
 import { resolveRunUpstream } from "../pi/index.ts";
 import type { ServerContext } from "../types.ts";
 import { parseFakeScripts } from "./fake-scripts.ts";
-import { type AgentEnv, agentSettings, eventSink, publishStatus, sessionDirOf } from "./runner.ts";
+import { type AgentEnv, agentSettings, eventSink, publishStatus, runtimeReporter, sessionDirOf } from "./runner.ts";
 
 // VIS-1 browser.solve_challenge:有界的视觉子流程(REQ §4.4、§10)。
 // - 服务端起一个受限的子 agent_run(父会话 = 调用它的 Agent 运行;白名单只有截图 / 坐标点击 / 输入 / 滚动 / 等待 / 读页面),
@@ -305,6 +305,7 @@ async function runChild(
   const onEvent = eventSink(env);
   const onEventError = (err: unknown) =>
     ctx.logger.warn({ runId: child.id, err: { message: (err as Error).message } }, "视觉子流程事件落库失败");
+  const onReport = runtimeReporter(ctx);
   let scripts: FakeScript[] = [FALLBACK_SCRIPT];
   if (fake) {
     try {
@@ -315,11 +316,12 @@ async function runChild(
     }
   }
   const runtime = fake
-    ? createFakeRuntime({ registry, scripts, onEvent, onEventError, fake: { vision: true } }).runtime
+    ? createFakeRuntime({ registry, scripts, onEvent, onEventError, onReport, fake: { vision: true } }).runtime
     : new PiAgentRuntime({
         registry,
         onEvent,
         onEventError,
+        onReport,
         upstream: { spec: upstream.spec, credentials: upstream.credentials },
       });
   const systemPrompt = buildSystemPrompt({

apps/server/src/modules/agent/runner.ts

@@ -10,6 +10,7 @@ import {
   type FakeScript,
   PiAgentRuntime,
   type RunOutcome,
+  type RuntimeReport,
   type SessionKind,
   type SkillDoc,
   selectSkills,
@@ -399,6 +400,15 @@ export async function buildRunPrompt(
   return { systemPrompt, kind, skills, available };
 }
 
+/**
+ * 运行时内部错误(本次运行扩展的系统提示 / 能力工具 / 白名单与预算钩子抛错,pi-durable 的 onReport)→ warn 日志。
+ * message 已由运行时脱敏;只记位置与信息,不记原始错误对象。
+ */
+export function runtimeReporter(ctx: ServerContext): (r: RuntimeReport) => void {
+  return (r) =>
+    ctx.logger.warn({ runId: r.runId, source: r.source, err: { message: r.message } }, "agent 运行时内部错误");
+}
+
 /** 读 agent_run_event 里最后一轮的结束事件(崩溃发生在"模型已结束、业务状态未写"之间时复用) */
 async function lastOutcomeEvent(env: AgentEnv, runId: string) {
   const r = await env.ctx.db
@@ -460,6 +470,7 @@ export async function drive(
   const onEvent = eventSink(env);
   const onEventError = (err: unknown, e: AgentRunEvent) =>
     ctx.logger.warn({ runId, seq: e.seq, err: { message: (err as Error).message } }, "agent 事件落库失败");
+  const onReport = runtimeReporter(ctx);
   let scripts: FakeScript[] = DEMO_SCRIPTS;
   if (fake) {
     try {
@@ -475,11 +486,12 @@ export async function drive(
     }
   }
   const runtime = fake
-    ? createFakeRuntime({ registry, scripts, onEvent, onEventError, fake: { vision: true } }).runtime
+    ? createFakeRuntime({ registry, scripts, onEvent, onEventError, onReport, fake: { vision: true } }).runtime
     : new PiAgentRuntime({
         registry,
         onEvent,
         onEventError,
+        onReport,
         ...(upstream ? { upstream: { spec: upstream.spec, credentials: upstream.credentials } } : {}),
       });
   const maxSeq = await ctx.db

packages/agent/src/index.ts

@@ -36,6 +36,7 @@ export {
   DEFAULT_COMPACTION_INSTRUCTIONS,
   PiAgentRuntime,
   type PiAgentRuntimeOptions,
+  type RuntimeReport,
   toolContent,
 } from "./pi/runtime.ts";
 export { LegacySessionError, sessionPathFor } from "./pi/session.ts";

packages/agent/src/pi/runtime.ts

@@ -1,3 +1,4 @@
+import type { Context } from "@earendil-works/chord";
 import { BACKGROUND_CONTEXT } from "@earendil-works/chord/context";
 import type { AssistantMessage, Message, ToolResultMessage } from "@earendil-works/pi-ai";
 import {
@@ -13,6 +14,7 @@ import {
   hook,
   type SettledSubmissionRecord,
   section,
+  type ToolExecutionApi,
   type ToolExecutionResult,
   type ToolRegistration,
   ToolTask,
@@ -66,9 +68,23 @@ export interface PiAgentRuntimeOptions {
    */
   onEvent?: (event: AgentRunEvent) => void | Promise<void>;
   onEventError?: (error: unknown, event: AgentRunEvent) => void;
+  /**
+   * 运行时内部错误(不影响本次调用结果、但应该看见的):本次运行扩展里的系统提示 section、能力工具、白名单 / 预算钩子抛错,
+   * 以及 pi-durable 的 onReport(提交 / 进度写入等失败)。message 已脱敏。不能抛错。
+   */
+  onReport?: (report: RuntimeReport) => void;
   now?: () => Date;
 }
 
+export interface RuntimeReport {
+  runId: string;
+  /** 出错的位置:section:umt、tool:<工具名>、hook:beforeTool、hook:beforeRequest、harness(pi-durable onReport) */
+  source: string;
+  /** 已脱敏的错误信息 */
+  message: string;
+  error: unknown;
+}
+
 export const DEFAULT_COMPACTION_INSTRUCTIONS =
   "保留:本次目标、约束与授权范围、证据引用(对象 ID/记录)、已完成的动作及其结果(已提交≠已入库)、剩余事项。不要保留任何凭证。";
 
@@ -113,7 +129,11 @@ export class PiAgentRuntime implements AgentRuntimePort {
     };
     let harness: Harness | undefined;
     try {
-      harness = await Harness.open(opened.storage, { models, registry, settings: harnessSettings }, ctx);
+      harness = await Harness.open(
+        opened.storage,
+        { models, registry, settings: harnessSettings, onReport: (error) => run.report("harness", error) },
+        ctx,
+      );
       const agent = { model: { provider: model.provider, modelId: model.id }, thinkingLevel };
       const root = await harness.root(ctx, { agent });
       // 已有会话:模型 / 思考档位跟随本次设置(换了上游或模型后继续)
@@ -192,27 +212,58 @@ class PiAgentRun implements AgentRunHandle {
     const systemPrompt = this.spec.systemPrompt;
     return defineExtension({
       name: "umt",
-      sections: [section("umt", () => systemPrompt, { tag: false })],
+      sections: [
+        section(
+          "umt",
+          this.reported("section:umt", () => systemPrompt),
+          { tag: false },
+        ),
+      ],
       tools: this.capabilities.map((c) => this.toTool(c)),
       hooks: [
-        hook(ToolTask, { beforeTool: (call) => this.beforeTool(call.id, call.name) }),
+        hook(ToolTask, {
+          beforeTool: this.reported("hook:beforeTool", (call: { id: string; name: string }) =>
+            this.beforeTool(call.id, call.name),
+          ),
+        }),
         hook(GenerationTask, {
-          beforeRequest: () => {
+          beforeRequest: this.reported("hook:beforeRequest", () => {
             this.checkDeadline();
             this.checkCost();
             if (this.budget) this.scheduleAbort();
             return undefined;
-          },
+          }),
         }),
       ],
     });
   }
 
+  /** 记一条运行时内部错误(脱敏后交给 onReport;onReport 自身出错忽略) */
+  report(source: string, error: unknown): void {
+    try {
+      const message = redactString(error instanceof Error ? error.message : String(error));
+      this.options.onReport?.({ runId: this.runId, source, message, error });
+    } catch {
+      // 上报失败不影响运行
+    }
+  }
+
+  /** 包一层:抛错时先上报(带位置)再原样抛出,pi-durable 的处理语义不变 */
+  private reported<A extends unknown[], R>(source: string, fn: (...args: A) => R): (...args: A) => R {
+    return (...args: A) => {
+      try {
+        return fn(...args);
+      } catch (error) {
+        this.report(source, error);
+        throw error;
+      }
+    };
+  }
+
   /** 能力 → Pi 工具。执行时统一走注册中心 invoke(schema/作用域/授权/幂等/审计) */
   toTool(entry: RegisteredCapability): ToolRegistration {
     const { descriptor } = entry;
     const registry = this.options.registry;
-    const actor = this.spec.actor;
     const gate = this.spec.replayGate;
     // 可挂起重放(见 CreateRunSpec.replayGate):对 harness 声明 safe,执行时用 memo 判断是否重放
     const gated = descriptor.replay === "never" && !!gate;
@@ -223,7 +274,25 @@ class PiAgentRun implements AgentRunHandle {
       replay: gated || descriptor.replay === "safe" ? "safe" : "unsafe",
       outputLimits: TOOL_OUTPUT_LIMITS,
       ...(descriptor.kind === "execute" ? { executionMode: "sequential" as const } : {}),
-      execute: async (params, api, context): Promise<ToolExecutionResult> => {
+      execute: (params, api, context) =>
+        this.executeTool(entry, gated, params, api, context).catch((error: unknown) => {
+          if (!this.closed && !context.abortSignal?.aborted) this.report(`tool:${entry.toolName}`, error);
+          throw error;
+        }),
+    };
+  }
+
+  private async executeTool(
+    entry: RegisteredCapability,
+    gated: boolean,
+    params: unknown,
+    api: ToolExecutionApi,
+    context: Context,
+  ): Promise<ToolExecutionResult> {
+    const { descriptor } = entry;
+    const registry = this.options.registry;
+    const actor = this.spec.actor;
+    const gate = this.spec.replayGate;
     const signals = [this.abortController.signal, ...(context.abortSignal ? [context.abortSignal] : [])];
     const signal = AbortSignal.any(signals);
     // 稳定的调用 ID:同一存储内唯一,挂起 / 崩溃后重放时不变(审批、子会话等待按它对应)
@@ -233,7 +302,7 @@ class PiAgentRun implements AgentRunHandle {
       if (gated) {
         const replayed = (await api.memo<boolean>(STARTED_MEMO, context)) === true;
         if (!replayed) await api.memo(STARTED_MEMO, true, context);
-            else if (!gate(descriptor.id, invocationId))
+        else if (!gate?.(descriptor.id, invocationId))
           return errorResult(
             JSON.stringify({
               error: {
@@ -275,8 +344,6 @@ class PiAgentRun implements AgentRunHandle {
     } finally {
       this.activeTools -= 1;
     }
-      },
-    };
   }
 
   async attach(harness: Harness, root: Conversation): Promise<void> {

packages/agent/test/runtime.test.ts

@@ -11,6 +11,7 @@ import {
   createFakeRuntime,
   type FakeScript,
   LegacySessionError,
+  type RuntimeReport,
 } from "../src/index.ts";
 import { crashScript, RUN_ID } from "./fixtures/crash-script.ts";
 import { type Fixture, makeFixture, settings, tempDir, types } from "./helpers.ts";
@@ -462,4 +463,43 @@ describe("PiAgentRuntime + FakeModelProvider", () => {
     expect(err).toBeInstanceOf(LegacySessionError);
     expect((err as Error).message).toContain("旧版本");
   });
+
+  it("扩展钩子抛错:上报 onReport(带 runId 与位置、已脱敏),pi-durable 照常按 block 处理", async () => {
+    const fx = makeFixture();
+    const reports: RuntimeReport[] = [];
+    const { runtime } = createFakeRuntime({
+      registry: fx.registry,
+      scripts: [
+        {
+          match: "Dune",
+          steps: [{ toolCalls: [{ capability: "inventory.query", input: { title: "Dune" } }] }],
+          finalText: (r) => `结果:${r.at(-1)?.isError}`,
+        },
+      ],
+      onReport: (r) => reports.push(r),
+    });
+    const run = await runtime.createRun({
+      runId: "run-report",
+      goal: "Dune",
+      systemPrompt: "x",
+      capabilities: ["inventory.query"],
+      settings: settings(),
+      sessionDir: await tempDir(),
+      actor: { actorId: "user:1", taskId: "t", scopeIds: [] },
+      externalUsage: () => {
+        throw new Error("子会话用量读取失败 passkey=abcdef123456");
+      },
+    });
+    opened.push(run);
+    const out = await run.prompt();
+    expect(out).toMatchObject({ status: "completed", finalText: "结果:true" });
+    expect(reports).toEqual([
+      expect.objectContaining({
+        runId: "run-report",
+        source: "hook:beforeTool",
+        message: expect.stringContaining("子会话用量读取失败"),
+      }),
+    ]);
+    expect(reports[0]?.message).not.toContain("abcdef123456");
+  });
 });