feat(agent): W4-E2 in-grant auto recovery for decisions with verified outcome

- settings.agent.autoInvestigateMode investigate|recover, investigateMaxAttempts, recoverVerifySec
- new capability download.recheck (execute, replay never): qB force recheck / resume, no source swap/delete
- autoInvestigate moves to agent/investigate.ts: recover mode runs execute with a recovery allowlist
  only when an active grant (published policy if bound) covers the recovery actions; otherwise degrades
  to read-only investigation and leaves suggestion + recheck option on the decision
- runAgent verifies recovery against business facts (download_task/task/source_release) before finalize;
  verified -> decision resolved by system:auto-recover, run summary '已恢复(系统核实)'; otherwise stays
  awaiting_user; per-root-cause attempt cap
- e2e E6: missingFiles -> download_failed -> recheck -> verified; no grant degrades; swap rejected;
  dedupe; attempts exhausted; budget exceeded keeps attention

Co-Authored-By: Claude Opus 5.5 <[email protected]>

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

@@ -16,6 +16,7 @@ export const GRANT_ACTIONS: Record<string, readonly string[]> = {
   "identity.resolve": ["identity.resolve", "policy.identity"],
   "download.submit": ["download.submit", "policy.acquisition"],
   "organize.execute": ["organize.execute", "policy.acquisition", "policy.organize"],
+  "download.recheck": ["download.recheck", "policy.automation"],
   "task.retry": ["task.retry", "task.*"],
   "task.cancel": ["task.cancel", "task.*"],
   "policy.apply": ["policy.apply"],

apps/server/src/modules/agent/capabilities-download.ts

@@ -0,0 +1,100 @@
+import { createQbClient, type DownloaderClient } from "@umt/adapters";
+import { CapabilityError, type CapabilityRegistry } from "@umt/agent";
+import { capabilitySpec, type DownloadStatus } from "@umt/contracts";
+import type { DownloaderInstanceRow } from "@umt/db";
+import type { ServerContext } from "../types.ts";
+import type { AgentModuleOptions } from "./runner.ts";
+
+// download.recheck(REQ §8.4.2 / §10.7 授权内恢复):对已纳管的下载任务发起 qB 强制重新校验或恢复暂停。
+// 只动下载器里的这一个种子:不换源、不删除、不改保存路径 / 分类 / 标签;下载任务状态仍由后台同步按 qB 实际状态更新,
+// 是否"已恢复"由 resolveDecision 的核实步骤看 download_task 状态决定,本能力只返回"已发出请求"。
+
+/** recheck 允许的起点:文件缺失 / 出错 / 无进展 / 暂停(非阈值、非人工) */
+const RECHECK_FROM: DownloadStatus[] = ["missing", "failed", "stalled", "paused"];
+/** resume 允许的起点 */
+const RESUME_FROM: DownloadStatus[] = ["paused", "stalled", "failed"];
+
+async function defaultClient(ctx: ServerContext, inst: DownloaderInstanceRow): Promise<DownloaderClient> {
+  const settings = (inst.settings ?? {}) as Record<string, unknown>;
+  const secret = inst.secret_ref ? await ctx.secrets.resolve(inst.secret_ref) : undefined;
+  return createQbClient({
+    instanceId: inst.id,
+    baseUrl: inst.base_url,
+    ...(inst.username ? { username: inst.username } : {}),
+    ...(secret ? (settings.authKind === "api_key" ? { apiKey: secret } : { password: secret }) : {}),
+  });
+}
+
+/** 持续授权的对象范围:按下载任务 → 计划 → Slot → 作品核对(authorize 对非计划类能力拿不到目标作品) */
+async function checkGrantScope(ctx: ServerContext, grantId: string, planId: string | null) {
+  const g = await ctx.db.selectFrom("automation_grant").select("scope").where("id", "=", grantId).executeTakeFirst();
+  const scope = (g?.scope ?? {}) as { mediaEntityIds?: string[]; mediaKinds?: string[] };
+  if (!scope.mediaEntityIds?.length && !scope.mediaKinds?.length) return;
+  const e = planId
+    ? await ctx.db
+        .selectFrom("acquisition_plan as p")
+        .innerJoin("acquisition_slot as s", "s.id", "p.slot_id")
+        .innerJoin("media_entity as e", "e.id", "s.media_entity_id")
+        .select(["e.id", "e.kind"])
+        .where("p.id", "=", planId)
+        .executeTakeFirst()
+    : undefined;
+  if (!e) throw new CapabilityError("unauthorized", `授权 ${grantId} 限定了对象范围,但下载任务没有关联作品`);
+  if (scope.mediaEntityIds?.length && !scope.mediaEntityIds.includes(e.id))
+    throw new CapabilityError("unauthorized", `授权 ${grantId} 的对象范围不含目标作品`);
+  if (scope.mediaKinds?.length) {
+    const ok =
+      scope.mediaKinds.includes(e.kind) ||
+      (e.kind !== "movie" && e.kind !== "variety" && scope.mediaKinds.includes("series"));
+    if (!ok) throw new CapabilityError("unauthorized", `授权 ${grantId} 的媒体类型范围不含 ${e.kind}`);
+  }
+}
+
+export function registerDownloadRecoveryCapabilities(
+  registry: CapabilityRegistry,
+  ctx: ServerContext,
+  opts: AgentModuleOptions,
+): void {
+  registry.register(capabilitySpec("download.recheck"), async (input, c) => {
+    const dt = await ctx.db
+      .selectFrom("download_task")
+      .selectAll()
+      .where("id", "=", input.downloadTaskId)
+      .executeTakeFirst();
+    if (!dt) throw new CapabilityError("not_found", "下载任务不存在");
+    if (!dt.managed || dt.removed_at || dt.status === "removed")
+      throw new CapabilityError("precondition", "未纳管或已移除的下载任务不做自动恢复");
+    const status = dt.status as DownloadStatus;
+    if (input.action === "recheck" && !RECHECK_FROM.includes(status))
+      throw new CapabilityError("precondition", `当前状态 ${status} 不需要重新校验`);
+    if (input.action === "resume" && !RESUME_FROM.includes(status))
+      throw new CapabilityError("precondition", `当前状态 ${status} 不需要恢复`);
+    // 阈值暂停不被绕过;人工暂停只由人恢复
+    if (status === "paused" && (dt.pause_reason === "threshold" || dt.pause_reason === "user"))
+      throw new CapabilityError(
+        "precondition",
+        `${dt.pause_reason === "threshold" ? "缓存阈值" : "人工"}暂停不做自动恢复`,
+      );
+    if (c.authorization?.grantId) await checkGrantScope(ctx, c.authorization.grantId, dt.acquisition_plan_id);
+    const inst = await ctx.db
+      .selectFrom("downloader_instance")
+      .selectAll()
+      .where("id", "=", dt.instance_id)
+      .executeTakeFirst();
+    if (!inst) throw new CapabilityError("not_found", "下载器实例不存在");
+    const client = opts.downloader ? await opts.downloader(inst) : await defaultClient(ctx, inst);
+    try {
+      if (input.action === "recheck") await client.recheck([dt.torrent_hash]);
+      else await client.resume([dt.torrent_hash]);
+    } catch (err) {
+      throw new CapabilityError("internal", `下载器请求失败:${(err as Error).message}`);
+    }
+    return {
+      downloadTaskId: dt.id,
+      action: input.action,
+      statusBefore: status,
+      rawStateBefore: dt.raw_state,
+      requested: true,
+    };
+  });
+}

apps/server/src/modules/agent/capability-registry.ts

@@ -7,7 +7,7 @@ import {
   type IdempotencyKeyGenerator,
   stableStringify,
 } from "@umt/agent";
-import type { ActorType, CapabilityKind, CapabilitySpec } from "@umt/contracts";
+import { type ActorType, CAPABILITY_IDS, type CapabilityKind, type CapabilitySpec } from "@umt/contracts";
 import { type DB, type Kysely, sql } from "@umt/db";
 import type { Logger } from "pino";
 import type { z } from "zod";
@@ -20,7 +20,7 @@ import { createAuthorizer } from "./authorize.ts";
 const SIDE_EFFECT: readonly CapabilityKind[] = ["propose", "approve", "execute"];
 
 /** 首批能力总数:启动时断言注册数与此一致(漏挂处理器即启动失败) */
-export const EXPECTED_CAPABILITY_COUNT = 14;
+export const EXPECTED_CAPABILITY_COUNT = CAPABILITY_IDS.length;
 
 /**
  * 幂等键:能力 ID + 版本 + 所属运行(agent run / 任务)+ 规范化输入的 sha256。
@@ -187,7 +187,7 @@ export function createCapabilityRegistry(deps: CapabilityRegistryDeps): Capabili
   });
 }
 
-/** 启动断言:首批 14 个能力全部挂上处理器 */
+/** 启动断言:首批能力全部挂上处理器 */
 export function assertCapabilityCount(registry: CapabilityRegistry): void {
   const n = registry.list().length;
   if (n !== EXPECTED_CAPABILITY_COUNT) {

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

@@ -1,5 +1,6 @@
 import type { ServerContext, ServerModule } from "../types.ts";
 import { registerTaskCapabilities } from "./capabilities.ts";
+import { registerDownloadRecoveryCapabilities } from "./capabilities-download.ts";
 import { registerAgentRoutes } from "./routes.ts";
 import type { AgentEnv, AgentModuleOptions } from "./runner.ts";
 import { agentOutboxRoutes, registerAgentTasks } from "./tasks.ts";
@@ -34,6 +35,7 @@ export function createAgentModule(opts: AgentModuleOptions = {}): ServerModule &
     outboxRoutes: () => agentOutboxRoutes,
     registerCapabilities(registry, ctx) {
       registerTaskCapabilities(registry, ctx);
+      registerDownloadRecoveryCapabilities(registry, ctx, opts);
     },
   };
 }

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

@@ -0,0 +1,449 @@
+import type { DecisionOption, DownloadStatus } from "@umt/contracts";
+import { type DecisionRow, sql } from "@umt/db";
+import type { ServerContext } from "../types.ts";
+import { GRANT_ACTIONS } from "./authorize.ts";
+import { type DriveResult, INVESTIGATE_CAPABILITIES, RECOVER_CAPABILITIES, RECOVERY_ACTIONS } from "./runner.ts";
+import { startRun } from "./service.ts";
+
+// 后台自动调查与授权内恢复(REQ §8.4.2、§10.7、AT68 / AT71):
+//   decision.created → autoInvestigate:按 settings.agent.autoInvestigateMode 决定只读调查或授权内恢复,
+//     recover 需要有覆盖恢复动作的有效持续授权(策略版本须为 published),否则降级为只读调查并在事项上留建议;
+//     同一 decision 只一个 run(dedupe_key=investigate:<id>),同一根因(decision 去重键)的恢复次数有上限。
+//   runAgent 结束 → recoveryActions / checkRecovery / applyRecoveryOutcome:
+//     按业务事实(download_task / task / source_release 状态)核实恢复结果,核实成功才关闭事项;
+//     失败 / 超时 / 预算超限一律保持待处理(finalize 把事项转为 awaiting_user),不自动再试。
+
+const KINDS = ["identity_ambiguous", "download_failed", "download_stalled"] as const;
+export const AUTO_ACTOR = "system:auto-investigate";
+export const RECOVER_ACTOR = "system:auto-recover";
+
+/** 各类事项在 recover 模式下需要被授权覆盖的恢复动作 */
+const ACTIONS_FOR: Record<string, readonly string[]> = {
+  download_failed: ["download.recheck", "task.retry"],
+  download_stalled: ["download.recheck", "task.retry"],
+  identity_ambiguous: ["identity.resolve"],
+};
+
+/** 下载任务处于这些状态即视为已恢复(qB 重新开始下载或已经完成) */
+const DL_RECOVERED: DownloadStatus[] = ["downloading", "queued", "completed", "importing", "imported"];
+
+export interface AutoRecoverEvidence {
+  mode: "investigate" | "recover";
+  requestedMode: "investigate" | "recover";
+  /** 本根因第几次自动恢复(只在 recover 时计数) */
+  attempt: number;
+  maxAttempts: number;
+  recoverAttempted: boolean;
+  degraded?: { reason: string } | undefined;
+  skipped?: string | undefined;
+  runId?: string | null;
+  outcome?: RecoveryOutcomeKind;
+  actions?: RecoveryAction[];
+  denied?: { capabilityId: string; reason: string | null }[];
+  facts?: Record<string, unknown>;
+  suggestion?: string;
+  verifiedAt?: string;
+}
+
+type RecoveryOutcomeKind = "recovered" | "failed" | "unverified" | "not_attempted" | "budget_exceeded" | "cancelled";
+
+export interface RecoveryAction {
+  capabilityId: string;
+  input: Record<string, unknown>;
+  status: string;
+}
+
+export interface RecoveryCheck {
+  state: "recovered" | "failed" | "pending";
+  facts: Record<string, unknown>;
+}
+
+export interface RecoveryOutcome {
+  decisionId: string;
+  outcome: RecoveryOutcomeKind;
+  summary: string;
+}
+
+const decisionIdOf = (dedupeKey: string | null | undefined) =>
+  dedupeKey?.startsWith("investigate:") ? dedupeKey.slice("investigate:".length) : null;
+
+export const isInvestigationRun = (run: { dedupe_key: string | null }) => decisionIdOf(run.dedupe_key) !== null;
+
+async function mergeEvidence(ctx: ServerContext, decisionId: string, patch: Partial<AutoRecoverEvidence>) {
+  await ctx.db
+    .updateTable("decision")
+    .set({
+      evidence: sql`evidence || jsonb_build_object('autoRecover',
+        coalesce(evidence->'autoRecover', '{}'::jsonb) || ${JSON.stringify(patch)}::jsonb)`,
+    })
+    .where("id", "=", decisionId)
+    .execute();
+}
+
+/** 有没有覆盖这些动作的有效持续授权(未撤销、未过期;绑定策略时策略须仍是 published) */
+export async function recoveryGrants(ctx: ServerContext, capabilityIds: readonly string[]): Promise<string[]> {
+  const actions = [...new Set([...capabilityIds.flatMap((c) => GRANT_ACTIONS[c] ?? [c]), "*"])];
+  const rows = await ctx.db
+    .selectFrom("automation_grant as g")
+    .leftJoin("policy_revision as p", "p.id", "g.policy_revision_id")
+    .select(["g.id", "g.scope"])
+    .where("g.status", "=", "active")
+    .where("g.action", "in", actions)
+    .where((eb) => eb.or([eb("g.expires_at", "is", null), eb("g.expires_at", ">", new Date())]))
+    .where((eb) => eb.or([eb("g.policy_revision_id", "is", null), eb("p.status", "=", "published")]))
+    .execute();
+  return rows
+    .filter((g) => {
+      const caps = ((g.scope ?? {}) as { capabilities?: string[] }).capabilities;
+      return !caps?.length || caps.some((c) => capabilityIds.includes(c));
+    })
+    .map((g) => g.id);
+}
+
+/** 下载类事项关联的业务任务(提交下载 / 核验入库等),给模型 task.inspect / task.retry 用 */
+async function relatedTask(ctx: ServerContext, d: DecisionRow) {
+  if (d.task_id) return d.task_id;
+  if (d.subject_type !== "download_task" || !d.subject_id) return null;
+  const dt = await ctx.db
+    .selectFrom("download_task")
+    .select(["acquisition_plan_id"])
+    .where("id", "=", d.subject_id)
+    .executeTakeFirst();
+  const t = await ctx.db
+    .selectFrom("task")
+    .select("id")
+    .where((eb) =>
+      eb.or([
+        eb.and([eb("subject_type", "=", "download_task"), eb("subject_id", "=", d.subject_id!)]),
+        ...(dt?.acquisition_plan_id
+          ? [eb.and([eb("subject_type", "=", "acquisition_plan"), eb("subject_id", "=", dt.acquisition_plan_id)])]
+          : []),
+      ]),
+    )
+    .orderBy("created_at", "desc")
+    .executeTakeFirst();
+  return t?.id ?? null;
+}
+
+function suggestionFor(d: DecisionRow, reason: string): string {
+  if (d.kind === "identity_ambiguous")
+    return `${reason}。建议在待处理中心人工选择身份,或发布识别策略授权 identity.resolve。`;
+  return (
+    `${reason}。建议:在待处理中心选择"强制重新校验"(download.recheck),或为自动恢复发放持续授权` +
+    `(action=download.recheck / task.retry,可绑定已发布的 automation 策略);换源、删除仍需人工决定。`
+  );
+}
+
+/** 事项上补一个可直接执行的"强制重新校验"选项(人工点选仍经授权 / 审计 / 账本) */
+async function addRecheckOption(ctx: ServerContext, d: DecisionRow) {
+  if (d.subject_type !== "download_task" || !d.subject_id) return;
+  const opts = (d.options ?? []) as DecisionOption[];
+  if (opts.some((o) => o.key === "recheck")) return;
+  const next: DecisionOption[] = [
+    ...opts,
+    {
+      key: "recheck",
+      label: "强制重新校验",
+      description: "让下载器重新校验已有文件(不换源、不删除)",
+      capabilityId: "download.recheck",
+      input: { downloadTaskId: d.subject_id, action: "recheck" },
+    },
+  ];
+  await ctx.db
+    .updateTable("decision")
+    .set({ options: JSON.stringify(next) })
+    .where("id", "=", d.id)
+    .execute();
+}
+
+/**
+ * decision.created → 有界调查 / 授权内恢复。同一 decision 只一个 run;
+ * 同一根因(去重键)已自动恢复过 maxAttempts 次则不再唤醒模型,直接交还人工。
+ */
+export async function autoInvestigate(
+  ctx: ServerContext,
+  decisionId: string,
+): Promise<{ runId: string | null; created: boolean; mode?: "investigate" | "recover"; skipped?: string }> {
+  const d = await ctx.db.selectFrom("decision").selectAll().where("id", "=", decisionId).executeTakeFirst();
+  if (d?.status !== "open") return { runId: null, created: false };
+  if (!(KINDS as readonly string[]).includes(d.kind)) return { runId: null, created: false };
+  // Agent 自己引起的事项不再触发 Agent(防止递归自触发)
+  if (d.subject_type === "agent_run") return { runId: null, created: false };
+  const s = await ctx.settings.get("agent");
+  if (!s.autoInvestigate) return { runId: null, created: false };
+
+  const requested = s.autoInvestigateMode;
+  let mode: "investigate" | "recover" = requested;
+  let degraded: { reason: string } | undefined;
+  let attempt = 0;
+  if (requested === "recover") {
+    // 同根因的历史恢复次数(包括已关闭的事项)
+    const prior = await sql<{ n: number }>`
+      select count(*)::int as n from decision
+       where dedupe_key = ${d.dedupe_key} and id <> ${d.id}
+         and (evidence->'autoRecover'->>'recoverAttempted')::boolean is true`.execute(ctx.db);
+    attempt = (prior.rows[0]?.n ?? 0) + 1;
+    if (attempt > s.investigateMaxAttempts) {
+      const reason = `同一根因已自动恢复 ${attempt - 1} 次(上限 ${s.investigateMaxAttempts}),不再自动处理`;
+      await mergeEvidence(ctx, d.id, {
+        mode: "investigate",
+        requestedMode: requested,
+        attempt: attempt - 1,
+        maxAttempts: s.investigateMaxAttempts,
+        recoverAttempted: false,
+        skipped: "attempts_exhausted",
+        suggestion: suggestionFor(d, reason),
+      });
+      await addRecheckOption(ctx, d);
+      const r = await ctx.db
+        .updateTable("decision")
+        .set({ status: "awaiting_user" })
+        .where("id", "=", d.id)
+        .where("status", "=", "open")
+        .returning("id")
+        .executeTakeFirst();
+      if (r)
+        await ctx.outbox.emit(
+          ctx.db,
+          "decision.updated",
+          { decisionId: d.id, status: "awaiting_user" },
+          { idempotencyKey: `decision.updated:${d.id}:attempts_exhausted` },
+        );
+      return { runId: null, created: false, skipped: "attempts_exhausted" };
+    }
+    const grants = await recoveryGrants(ctx, ACTIONS_FOR[d.kind] ?? []);
+    if (grants.length === 0) {
+      mode = "investigate";
+      degraded = { reason: `没有覆盖 ${(ACTIONS_FOR[d.kind] ?? []).join(" / ")} 的有效持续授权` };
+    }
+  }
+
+  const dt =
+    d.subject_type === "download_task" && d.subject_id
+      ? await ctx.db
+          .selectFrom("download_task")
+          .select(["id", "name", "status", "raw_state", "torrent_hash"])
+          .where("id", "=", d.subject_id)
+          .executeTakeFirst()
+      : undefined;
+  const taskId = await relatedTask(ctx, d);
+  const goal = [
+    `调查待处理事项 ${d.id}(${d.kind}):${d.title}。`,
+    d.summary ? `摘要:${d.summary}。` : "",
+    d.subject_type && d.subject_id ? `对象:${d.subject_type} ${d.subject_id}。` : "",
+    dt ? `下载任务 ${dt.id}「${dt.name ?? ""}」当前 ${dt.status}(qB 状态 ${dt.raw_state ?? "未知"})。` : "",
+    taskId ? `关联任务 ${taskId}。` : "",
+    mode === "recover"
+      ? "在已授权范围内做有界恢复:先用 task.inspect 等只读能力查证,再按需调用 download.recheck(强制重新校验 / 恢复暂停)、" +
+        "task.retry(重试失败任务)或 identity.resolve,每种最多一次;不换源、不删除、不清理;恢复结果由系统核实。"
+      : "只做有界的只读调查:查证据、给出结论与处理建议;不要提交下载或变更文件。",
+  ].join("");
+  const { run, created } = await startRun(ctx, {
+    goal,
+    mode: mode === "recover" ? "execute" : "plan_only",
+    capabilityAllowlist: mode === "recover" ? [...RECOVER_CAPABILITIES] : [...INVESTIGATE_CAPABILITIES],
+    trigger: "event",
+    createdBy: mode === "recover" ? RECOVER_ACTOR : AUTO_ACTOR,
+    dedupeKey: `investigate:${d.id}`,
+    budget: { maxToolCalls: s.investigateMaxToolCalls, maxDurationSec: Math.min(600, s.maxDurationSec) },
+  });
+  if (created) {
+    const r = await ctx.db
+      .updateTable("decision")
+      .set({ status: "investigating", agent_run_id: run.id })
+      .where("id", "=", d.id)
+      .where("status", "=", "open")
+      .returning("id")
+      .executeTakeFirst();
+    if (r) {
+      await mergeEvidence(ctx, d.id, {
+        mode,
+        requestedMode: requested,
+        attempt,
+        maxAttempts: s.investigateMaxAttempts,
+        recoverAttempted: mode === "recover",
+        ...(degraded ? { degraded, suggestion: suggestionFor(d, degraded.reason) } : {}),
+        runId: run.id,
+      });
+      if (degraded) await addRecheckOption(ctx, d);
+    }
+  }
+  return { runId: run.id, created, mode };
+}
+
+/** 本 run 调用过的恢复动作(账本)与被拒绝的调用(审计) */
+export async function recoveryActions(
+  ctx: ServerContext,
+  runId: string,
+): Promise<{ actions: RecoveryAction[]; denied: { capabilityId: string; reason: string | null }[] }> {
+  const rows = await ctx.db
+    .selectFrom("capability_invocation")
+    .select(["capability_id", "input", "status"])
+    .where("agent_run_id", "=", runId)
+    .where("capability_id", "in", [...RECOVERY_ACTIONS])
+    .orderBy("created_at")
+    .execute();
+  const denied = await ctx.db
+    .selectFrom("audit_event")
+    .select(["target_id", sql<string | null>`payload->>'errorMessage'`.as("reason")])
+    .where("action", "=", "capability.denied")
+    .where("trace_id", "=", runId)
+    .where(sql<string>`payload->>'agentRunId'`, "=", runId)
+    .orderBy("at")
+    .execute();
+  return {
+    actions: rows.map((r) => ({
+      capabilityId: r.capability_id,
+      input: (r.input ?? {}) as Record<string, unknown>,
+      status: r.status,
+    })),
+    denied: denied.map((x) => ({ capabilityId: x.target_id ?? "", reason: x.reason })),
+  };
+}
+
+/** 按业务事实核实恢复动作:全部已恢复 → recovered;任一明确失败 → failed;否则 pending(继续等) */
+export async function checkRecovery(ctx: ServerContext, actions: RecoveryAction[]): Promise<RecoveryCheck> {
+  const facts: Record<string, unknown> = {};
+  const states: RecoveryCheck["state"][] = [];
+  for (const a of actions.filter((x) => x.status === "succeeded")) {
+    if (a.capabilityId === "download.recheck") {
+      const id = String(a.input.downloadTaskId ?? "");
+      const dt = await ctx.db
+        .selectFrom("download_task")
+        .select(["status", "raw_state", "progress"])
+        .where("id", "=", id)
+        .executeTakeFirst();
+      facts[`download_task:${id}`] = dt ? { status: dt.status, rawState: dt.raw_state } : null;
+      states.push(
+        !dt || dt.status === "removed"
+          ? "failed"
+          : DL_RECOVERED.includes(dt.status as DownloadStatus)
+            ? "recovered"
+            : "pending",
+      );
+    } else if (a.capabilityId === "task.retry") {
+      const id = String(a.input.taskId ?? "");
+      const t = await ctx.db.selectFrom("task").select(["status"]).where("id", "=", id).executeTakeFirst();
+      facts[`task:${id}`] = t ? { status: t.status } : null;
+      states.push(
+        t?.status === "succeeded"
+          ? "recovered"
+          : !t || ["failed", "dead", "cancelled"].includes(t.status)
+            ? "failed"
+            : "pending",
+      );
+    } else if (a.capabilityId === "identity.resolve") {
+      const id = String(a.input.sourceReleaseId ?? "");
+      const r = id
+        ? await ctx.db
+            .selectFrom("source_release")
+            .select(["resolution_status", "media_entity_id"])
+            .where("id", "=", id)
+            .executeTakeFirst()
+        : undefined;
+      facts[`source_release:${id}`] = r ? { resolutionStatus: r.resolution_status } : null;
+      states.push(r?.resolution_status === "resolved" && r.media_entity_id ? "recovered" : "pending");
+    }
+  }
+  const state = states.includes("failed") ? "failed" : states.every((x) => x === "recovered") ? "recovered" : "pending";
+  return { state, facts };
+}
+
+/**
+ * 核实结论写回:recovered → 事项 resolved(resolution.execution 记恢复动作,executeResolution 不会重复执行);
+ * 其余 → 只写证据,事项保持 investigating,由 finalize 转 awaiting_user(交还人工,不自动再试)。
+ */
+export async function applyRecoveryOutcome(
+  ctx: ServerContext,
+  runId: string,
+  res: DriveResult,
+  acts: { actions: RecoveryAction[]; denied: { capabilityId: string; reason: string | null }[] },
+  check: RecoveryCheck | null,
+): Promise<RecoveryOutcome | null> {
+  const run = await ctx.db
+    .selectFrom("agent_run")
+    .select(["dedupe_key", "cancel_requested_at"])
+    .where("id", "=", runId)
+    .executeTakeFirstOrThrow();
+  const decisionId = decisionIdOf(run.dedupe_key);
+  if (!decisionId) return null;
+  const d = await ctx.db.selectFrom("decision").selectAll().where("id", "=", decisionId).executeTakeFirst();
+  if (!d) return null;
+  const done = acts.actions.filter((a) => a.status === "succeeded");
+  let outcome: RecoveryOutcomeKind;
+  if (res.status === "cancelled" || run.cancel_requested_at) outcome = "cancelled";
+  else if (check?.state === "recovered") outcome = "recovered";
+  else if (check?.state === "failed") outcome = "failed";
+  else if (done.length > 0) outcome = "unverified";
+  else if (res.reason === "budget") outcome = "budget_exceeded";
+  else outcome = "not_attempted";
+  const what = done.map((a) => a.capabilityId).join(" + ");
+  const summary =
+    outcome === "recovered"
+      ? `已恢复(系统核实):执行 ${what} 后业务状态已恢复。`
+      : outcome === "failed"
+        ? `未恢复:执行 ${what} 后业务状态仍为失败,已交还待处理中心。`
+        : outcome === "unverified"
+          ? `未恢复:执行 ${what} 后在核实时限内未见业务状态恢复,已交还待处理中心。`
+          : outcome === "budget_exceeded"
+            ? `未恢复:调查超出预算(${res.message ?? ""}),未执行恢复动作,已交还待处理中心。`
+            : outcome === "cancelled"
+              ? "未恢复:运行已取消。"
+              : `未恢复:没有执行恢复动作${acts.denied.length ? `(被拒绝:${acts.denied.map((x) => x.capabilityId).join(", ")})` : ""},已交还待处理中心。`;
+  const evidence: Partial<AutoRecoverEvidence> = {
+    runId,
+    outcome,
+    actions: acts.actions,
+    denied: acts.denied,
+    facts: check?.facts ?? {},
+    verifiedAt: new Date().toISOString(),
+    ...(outcome === "recovered" ? {} : { suggestion: suggestionFor(d, summary.replace(/。$/, "")) }),
+  };
+  await mergeEvidence(ctx, decisionId, evidence);
+  if (outcome === "recovered") {
+    await ctx.db.transaction().execute(async (trx) => {
+      const r = await trx
+        .updateTable("decision")
+        .set({
+          status: "resolved",
+          resolved_by: RECOVER_ACTOR,
+          resolved_at: new Date(),
+          resolution: JSON.stringify({
+            optionKey: null,
+            by: RECOVER_ACTOR,
+            note: summary,
+            execution: { kind: "agent_recover", runId, actions: done, verified: true, facts: check?.facts ?? {} },
+          }),
+        })
+        .where("id", "=", decisionId)
+        .where("status", "in", ["investigating", "open", "awaiting_user"])
+        .returning("id")
+        .executeTakeFirst();
+      if (r)
+        await ctx.outbox.emit(
+          trx,
+          "decision.updated",
+          { decisionId, status: "resolved" },
+          { idempotencyKey: `decision.updated:${decisionId}:recovered:${runId}` },
+        );
+    });
+  } else {
+    const fresh = await ctx.db.selectFrom("decision").selectAll().where("id", "=", decisionId).executeTakeFirst();
+    if (fresh) await addRecheckOption(ctx, fresh);
+  }
+  return { decisionId, outcome, summary };
+}
+
+/** finalize 之后:把系统核实结论放在 run 汇总最前面(模型文本仍标注"未经核查") */
+export async function annotateRunSummary(ctx: ServerContext, runId: string, rec: RecoveryOutcome) {
+  const run = await ctx.db
+    .selectFrom("agent_run")
+    .select(["summary", "status"])
+    .where("id", "=", runId)
+    .executeTakeFirstOrThrow();
+  if (run.summary?.startsWith(rec.summary)) return { status: run.status, summary: run.summary };
+  const summary = `${rec.summary}\n${run.summary ?? ""}`.slice(0, 4000);
+  await ctx.db.updateTable("agent_run").set({ summary }).where("id", "=", runId).execute();
+  void ctx.events.publish("agent.run_status", { runId, status: run.status, summary: summary.slice(0, 2000) });
+  return { status: run.status, summary };
+}

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

@@ -1,4 +1,5 @@
 import { join, resolve } from "node:path";
+import type { DownloaderClient } from "@umt/adapters";
 import {
   type AgentRunHandle,
   type AgentSettings,
@@ -21,7 +22,7 @@ import {
   type GlobalSettings,
   KnownCapabilityId,
 } from "@umt/contracts";
-import { type AgentRunRow, sql } from "@umt/db";
+import { type AgentRunRow, type DownloaderInstanceRow, sql } from "@umt/db";
 import type { ServerContext } from "../types.ts";
 import { DEMO_SCRIPTS, parseFakeScripts } from "./fake-scripts.ts";
 import { buildGoalReport } from "./report.ts";
@@ -36,6 +37,8 @@ export interface AgentModuleOptions {
   fakeScripts?: FakeScript[];
   /** 覆盖 settings.agent.goalPollSec(测试用小值) */
   goalPollSec?: number;
+  /** download.recheck 用的下载器客户端;缺省按实例配置 + secret 现场创建 qB 客户端 */
+  downloader?: (inst: DownloaderInstanceRow) => Promise<DownloaderClient>;
 }
 
 export interface AgentEnv {
@@ -54,6 +57,14 @@ export const INVESTIGATE_CAPABILITIES = [
   "task.inspect",
   "identity.resolve",
 ] as const;
+/**
+ * 授权内恢复(settings.agent.autoInvestigateMode = recover)的能力白名单:只读调查 + 恢复类
+ * (重试失败任务、强制重新校验 / 恢复暂停)。换源(download.plan / download.submit)、删除 / 清理、
+ * 取消任务、整理执行、策略发布一律不在白名单内。实际能否执行仍由 authorize 按持续授权判定。
+ */
+export const RECOVER_CAPABILITIES = [...INVESTIGATE_CAPABILITIES, "task.retry", "download.recheck"] as const;
+/** 被视为"恢复动作"的能力(结果需系统核实) */
+export const RECOVERY_ACTIONS = ["task.retry", "download.recheck", "identity.resolve"] as const;
 
 export interface RunScope {
   capabilities: string[];
@@ -243,7 +254,12 @@ function systemPromptFor(run: AgentRunRow, scope: RunScope): string {
   ];
   if (run.mode === "plan_only")
     constraints.push("本次为“先给方案”:只调查与规划,不提交下载、不做文件变更、不发布策略。");
-  if (run.trigger === "event")
+  if (run.trigger === "event" && run.mode === "execute")
+    constraints.push(
+      "本任务由后台事件触发,是有界的授权内恢复:先查证据,再只在授权范围内执行恢复(task.retry / download.recheck / identity.resolve)," +
+        "每种恢复动作最多一次;不换源、不删除、不清理;恢复是否成功由系统按业务状态核实,不要自行宣称已恢复。",
+    );
+  else if (run.trigger === "event")
     constraints.push("本任务由后台事件触发,是有界的只读调查:收集证据、给出结论与建议,不要反复重试同一操作。");
   return buildSystemPrompt({
     role: run.trigger === "event" ? "媒体库异常调查员" : "媒体库自动化助手",

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

@@ -4,9 +4,19 @@ import { defineTask, type OutboxRouteFn, type TaskRegistry } from "@umt/tasks";
 import { z } from "zod";
 import type { ServerContext } from "../types.ts";
 import { retryTask } from "./capabilities.ts";
+import {
+  annotateRunSummary,
+  applyRecoveryOutcome,
+  autoInvestigate,
+  checkRecovery,
+  isInvestigationRun,
+  type RecoveryCheck,
+  type RecoveryOutcome,
+  recoveryActions,
+} from "./investigate.ts";
 import { buildGoalReport } from "./report.ts";
 import { type AgentEnv, type DriveResult, drive, goalSettled } from "./runner.ts";
-import { autoInvestigate, finalize } from "./service.ts";
+import { finalize } from "./service.ts";
 
 // 持久任务:runAgent(驱动一次有界 Agent 会话 → 等目标关联的下载/入库落定 → 以业务表汇报)
 // 与 resolveDecision(decision.created → 自动调查;decision.updated(resolved)→ 执行所选处理)。
@@ -37,10 +47,31 @@ export function registerAgentTasks(registry: TaskRegistry, env: () => AgentEnv)
         let settled: AgentGoalReport | null = null;
         const run = await e.ctx.db
           .selectFrom("agent_run")
-          .select(["mode"])
+          .select(["mode", "dedupe_key"])
           .where("id", "=", runId)
           .executeTakeFirstOrThrow();
-        if (res.status === "completed" && run.mode === "execute") {
+        // 后台授权内恢复:核实业务事实(下载 / 任务 / 身份状态)后再收尾;不等"目标入库"
+        const recovering = run.mode === "execute" && isInvestigationRun(run);
+        let recovery: RecoveryOutcome | null = null;
+        if (recovering) {
+          const acts = await t.step.run("recover-actions", () => recoveryActions(e.ctx, runId));
+          let check: RecoveryCheck | null = null;
+          if (res.status === "completed" && acts.actions.some((a) => a.status === "succeeded")) {
+            const s = await e.ctx.settings.get("agent");
+            const pollSec = Math.min(e.opts.goalPollSec ?? s.goalPollSec, 30);
+            check = await t.step.waitForFact<RecoveryCheck>("recover-verified", {
+              check: async () => {
+                const c = await checkRecovery(e.ctx, acts.actions);
+                return c.state === "pending" ? null : c;
+              },
+              timeout: `${pollSec}s`,
+              maxRounds: Math.max(1, Math.min(900, Math.ceil(s.recoverVerifySec / pollSec))),
+            });
+            // 超时:用最后一次观测作为证据(state 仍是 pending → unverified)
+            check ??= await t.step.run("recover-last-check", () => checkRecovery(e.ctx, acts.actions));
+          }
+          recovery = await t.step.run("recover-outcome", () => applyRecoveryOutcome(e.ctx, runId, res, acts, check));
+        } else if (res.status === "completed" && run.mode === "execute") {
           const s = await e.ctx.settings.get("agent");
           const pollSec = e.opts.goalPollSec ?? s.goalPollSec;
           const rounds = Math.max(1, Math.min(900, Math.ceil((s.goalWaitHours * 3600) / pollSec)));
@@ -59,7 +90,12 @@ export function registerAgentTasks(registry: TaskRegistry, env: () => AgentEnv)
             maxRounds: rounds,
           });
         }
-        return t.step.run("finalize", () => finalize(e, runId, res, settled));
+        const out = await t.step.run("finalize", () => finalize(e, runId, res, settled));
+        if (recovery) {
+          const rec = recovery;
+          return t.step.run("recover-summary", () => annotateRunSummary(e.ctx, runId, rec));
+        }
+        return out;
       },
     }),
     defineTask({

apps/server/test/e2e/e2-agent-delegate.test.ts

@@ -42,8 +42,8 @@ describe.skipIf(!hasPg)("E2 Agent 委托闭环(fake 模型 + 真实能力/授权
     return r.body.data.id as string;
   };
 
-  it("启动断言:14 个能力全部注册,并在能力目录里可见", async () => {
+  it("启动断言:15 个能力全部注册,并在能力目录里可见", async () => {
-    expect(e.booted.ctx.capabilities.list()).toHaveLength(14);
+    expect(e.booted.ctx.capabilities.list()).toHaveLength(15);
     const r = await e.api("GET", "/capabilities");
     expect(r.body.data.items.map((c: { id: string }) => c.id).sort()).toEqual(
       e.booted.ctx.capabilities

apps/server/test/e2e/e6-auto-recover.test.ts

@@ -0,0 +1,360 @@
+import { createHash } from "node:crypto";
+import type { FakeScript } from "@umt/agent";
+import { afterAll, beforeAll, describe, expect, it } from "vitest";
+import { RECOVER_CAPABILITIES } from "../../src/modules/agent/runner.ts";
+import { createE2E, type E2E, pgAvailable, runEvents, seedBase, waitFor } from "./harness.ts";
+
+// E6 授权内异常调查与恢复(REQ §8.4.2、§10.7、§15 M1、AT68 / AT71):
+// qB 任务 missingFiles → 同步写 download_task=missing + download.status_changed → decision.created(download_failed)
+// → resolveDecision → autoInvestigate(mode=recover):
+//   - 没有覆盖 download.recheck 的持续授权:降级为只读调查,不碰 qB(无 recheck 调用),事项留建议 + "强制重新校验"选项;
+//   - 有绑定已发布策略的授权:有界 run 调用 task.inspect → download.recheck(qB 收到 recheck 并恢复下载),
+//     换源类能力(download.submit)不在白名单 → tool_call_rejected;系统按 download_task 状态核实后关闭事项,
+//     run 汇报"已恢复(系统核实)";同一 decision 重复触发只一个 run;同根因再次出错超过尝试上限不再唤醒模型;
+//   - 预算超限:run budget_exceeded,事项保持待处理(awaiting_user),不重试。
+
+const hasPg = await pgAvailable();
+const ZERO = "00000000-0000-4000-8000-000000000000";
+
+/** 当前场景的对象(fake 脚本的输入从这里取;场景串行执行) */
+const cur: { taskId?: string; downloadTaskId?: string } = {};
+
+const recoverScript: FakeScript = {
+  name: "e6-recover",
+  match: /调查待处理事项/,
+  steps: [
+    {
+      text: "先看关联任务与下载时间线。",
+      toolCalls: [{ capability: "task.inspect", input: () => ({ taskId: cur.taskId ?? ZERO }) }],
+    },
+    {
+      text: "qB 报告文件缺失,强制重新校验。",
+      toolCalls: [
+        {
+          capability: "download.recheck",
+          input: () => ({ downloadTaskId: cur.downloadTaskId ?? ZERO, action: "recheck" }),
+        },
+      ],
+    },
+    // 换源类动作:不在恢复白名单内,必须被拒绝
+    { text: "顺手换个源重新下载。", toolCalls: [{ capability: "download.submit", input: { planId: ZERO } }] },
+  ],
+  finalText: "已恢复下载。",
+};
+
+describe.skipIf(!hasPg)("E6 授权内异常调查与恢复", () => {
+  let e: E2E;
+  let instanceId: string;
+  let grantId: string;
+
+  const decisionOf = (dtId: string, status?: string) =>
+    waitFor(
+      async () => {
+        let q = e.db
+          .selectFrom("decision")
+          .selectAll()
+          .where("kind", "=", "download_failed")
+          .where("subject_id", "=", dtId);
+        if (status) q = q.where("status", "=", status as never);
+        return q.orderBy("created_at", "desc").executeTakeFirst();
+      },
+      60_000,
+      `download_failed decision(${dtId})`,
+    );
+
+  const runOf = (decisionId: string) =>
+    waitFor(
+      async () =>
+        e.db
+          .selectFrom("agent_run")
+          .selectAll()
+          .where("dedupe_key", "=", `investigate:${decisionId}`)
+          .executeTakeFirst(),
+      60_000,
+      "自动调查 run",
+    );
+
+  const finished = (runId: string) =>
+    waitFor(
+      async () => {
+        const r = await e.db.selectFrom("agent_run").selectAll().where("id", "=", runId).executeTakeFirstOrThrow();
+        return ["succeeded", "failed", "budget_exceeded", "cancelled"].includes(r.status) && r.finished_at ? r : null;
+      },
+      90_000,
+      "run 结束",
+    ).catch(async (err) => {
+      const r = await e.db.selectFrom("agent_run").selectAll().where("id", "=", runId).executeTakeFirst();
+      const runs = await e.booted.ctx.engine.list({ taskName: "runAgent" });
+      console.log("DEBUG", JSON.stringify({ r, runs }, null, 1).slice(0, 6000));
+      throw err;
+    });
+
+  /** 纳管下载中的任务 + 关联业务任务;随后 qB 报 missingFiles */
+  async function failingDownload(n: number) {
+    const hash = createHash("sha1").update(`e6-${n}`).digest("hex");
+    const name = `E6.Sample.${n}.2024.1080p.WEB-DL`;
+    e.qb.seed({ hash, name, size: 10_000_000, progress: 0.3, tags: ["umt"], save_path: e.cache.root });
+    const dt = await e.db
+      .insertInto("download_task")
+      .values({
+        instance_id: instanceId,
+        torrent_hash: hash,
+        name,
+        save_path: e.cache.root,
+        tags: ["umt"],
+        managed: true,
+        status: "downloading",
+        raw_state: "downloading",
+        progress: 0.3,
+        downloaded_bytes: 3_000_000,
+        last_progress_at: new Date(),
+      })
+      .returning("id")
+      .executeTakeFirstOrThrow();
+    const task = await e.db
+      .insertInto("task")
+      .values({
+        kind: "submitDownload",
+        idempotency_key: `e6:submit:${dt.id}`,
+        input: JSON.stringify({ downloadTaskId: dt.id }),
+        subject_type: "download_task",
+        subject_id: dt.id,
+        status: "succeeded",
+        triggered_by: "system:e6",
+      })
+      .returning("id")
+      .executeTakeFirstOrThrow();
+    cur.taskId = task.id;
+    cur.downloadTaskId = dt.id;
+    e.qb.setMissingFiles(hash);
+    return { hash, dtId: dt.id, taskId: task.id };
+  }
+
+  const toolStarts = async (runId: string) =>
+    (await runEvents(e, runId))
+      .filter((x) => x.type === "tool_call_start")
+      .map((x) => (x.payload as { capabilityId: string }).capabilityId);
+  const toolRejected = async (runId: string) =>
+    (await runEvents(e, runId))
+      .filter((x) => x.type === "tool_call_rejected")
+      .map((x) => (x.payload as { toolName: string }).toolName);
+
+  beforeAll(async () => {
+    e = await createE2E("e6", { fakeScripts: [recoverScript] });
+    instanceId = (await seedBase(e)).instanceId;
+    // download.recheck 走生产路径:按实例用户名 + secret 现场建 qB 客户端
+    const ref = await e.booted.ctx.secrets.create({ purpose: "downloader_password", plaintext: "adminadmin" });
+    await e.db.updateTable("downloader_instance").set({ secret_ref: ref }).where("id", "=", instanceId).execute();
+    await e.settings({
+      agent: {
+        autoInvestigate: true,
+        autoInvestigateMode: "recover",
+        investigateMaxToolCalls: 6,
+        investigateMaxAttempts: 1,
+        recoverVerifySec: 60,
+      },
+    });
+  }, 120_000);
+
+  afterAll(async () => {
+    await e?.close();
+  }, 60_000);
+
+  it("没有持续授权:降级为只读调查,不调用 qB recheck,事项留建议与可执行选项", async () => {
+    const f = await failingDownload(1);
+    const d = await decisionOf(f.dtId);
+    const run = await runOf(d.id);
+    expect(run).toMatchObject({ trigger: "event", mode: "plan_only", created_by: "system:auto-investigate" });
+    const done = await finished(run.id);
+    expect(done.status).toBe("succeeded");
+    expect(await toolStarts(run.id)).toEqual(["task.inspect"]);
+    // 恢复与换源能力都不在只读白名单里
+    expect(await toolRejected(run.id)).toEqual(["download__recheck", "download__submit"]);
+    expect(e.qb.callsTo("torrents/recheck")).toHaveLength(0);
+
+    const d2 = await waitFor(
+      async () => {
+        const r = await e.db.selectFrom("decision").selectAll().where("id", "=", d.id).executeTakeFirstOrThrow();
+        return r.status === "awaiting_user" ? r : null;
+      },
+      30_000,
+      "事项交还人工",
+    );
+    const ar = (d2.evidence as { autoRecover: Record<string, unknown> }).autoRecover;
+    expect(ar).toMatchObject({ requestedMode: "recover", mode: "investigate", recoverAttempted: false });
+    expect((ar.degraded as { reason: string }).reason).toContain("download.recheck");
+    expect(ar.suggestion).toMatch(/强制重新校验/);
+    expect(d2.options).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          key: "recheck",
+          capabilityId: "download.recheck",
+          input: { downloadTaskId: f.dtId, action: "recheck" },
+        }),
+      ]),
+    );
+    const row = await e.db.selectFrom("download_task").select("status").where("id", "=", f.dtId).executeTakeFirst();
+    expect(row?.status).toBe("missing");
+  }, 180_000);
+
+  it("有绑定已发布策略的授权:task.inspect → download.recheck → 系统核实已恢复;换源被拒;重复触发只一个 run", async () => {
+    const policy = await e.db
+      .insertInto("policy_revision")
+      .values({
+        policy_key: "e6-auto-recover",
+        revision: 1,
+        kind: "automation",
+        status: "published",
+        title: "下载异常自动恢复",
+        published_at: new Date(),
+        published_by: "admin:e6",
+      })
+      .returning("id")
+      .executeTakeFirstOrThrow();
+    const g = await e.db
+      .insertInto("automation_grant")
+      .values({
+        action: "download.recheck",
+        policy_revision_id: policy.id,
+        granted_by: "admin:e6",
+        reason: "e6:自动恢复下载异常",
+      })
+      .returning("id")
+      .executeTakeFirstOrThrow();
+    grantId = g.id;
+
+    const f = await failingDownload(2);
+    const d = await decisionOf(f.dtId);
+    const run = await runOf(d.id);
+    expect(run).toMatchObject({ trigger: "event", mode: "execute", created_by: "system:auto-recover" });
+    expect(run.goal).toContain(f.dtId);
+    expect(run.goal).toContain(f.taskId);
+    expect(run.scope).toMatchObject({ capabilities: [...RECOVER_CAPABILITIES] });
+    expect(run.budget).toMatchObject({ maxToolCalls: 6 });
+
+    // qB 收到强制重新校验 → 校验完成后恢复下载
+    const calls = await waitFor(
+      async () => {
+        const c = e.qb.callsTo("torrents/recheck");
+        return c.length ? c : null;
+      },
+      60_000,
+      "qB recheck 调用",
+    );
+    expect(calls[0]!.form.hashes).toBe(f.hash);
+    e.qb.setState(f.hash, "downloading");
+
+    const done = await finished(run.id);
+    expect(done.status).toBe("succeeded");
+    expect(done.summary).toMatch(/^已恢复\(系统核实\)/);
+    expect(done.summary).toContain("模型汇报(未经核查)");
+    expect(await toolStarts(run.id)).toEqual(["task.inspect", "download.recheck"]);
+    expect(await toolRejected(run.id)).toEqual(["download__submit"]);
+    const inv = await e.db
+      .selectFrom("capability_invocation")
+      .selectAll()
+      .where("agent_run_id", "=", run.id)
+      .where("capability_id", "=", "download.recheck")
+      .executeTakeFirstOrThrow();
+    expect(inv).toMatchObject({ status: "succeeded", grant_id: grantId });
+    expect(e.qb.callsTo("torrents/recheck")).toHaveLength(1);
+
+    // 业务事实:下载恢复,事项由系统核实后关闭
+    const dt = await e.db.selectFrom("download_task").select("status").where("id", "=", f.dtId).executeTakeFirst();
+    expect(dt?.status).toBe("downloading");
+    const d2 = await e.db.selectFrom("decision").selectAll().where("id", "=", d.id).executeTakeFirstOrThrow();
+    expect(d2).toMatchObject({ status: "resolved", resolved_by: "system:auto-recover" });
+    expect((d2.resolution as { execution: Record<string, unknown> }).execution).toMatchObject({
+      kind: "agent_recover",
+      runId: run.id,
+      verified: true,
+    });
+    expect((d2.evidence as { autoRecover: Record<string, unknown> }).autoRecover).toMatchObject({
+      mode: "recover",
+      attempt: 1,
+      recoverAttempted: true,
+      outcome: "recovered",
+    });
+
+    // 同一 decision 的重复事件:仍只有一个 run
+    for (const k of ["dup-1", "dup-2"])
+      await e.booted.ctx.outbox.emit(
+        e.db,
+        "decision.created",
+        { decisionId: d.id, kind: "download_failed", severity: "warning", title: d.title },
+        { idempotencyKey: `e6:${k}` },
+      );
+    await waitFor(
+      async () => {
+        const rows = await e.db
+          .selectFrom("outbox_event")
+          .select("status")
+          .where("idempotency_key", "in", ["e6:dup-1", "e6:dup-2"])
+          .execute();
+        return rows.length === 2 && rows.every((r) => r.status === "dispatched");
+      },
+      30_000,
+      "重复事件已投递",
+    );
+    await waitFor(
+      async () => {
+        const runs = await e.booted.ctx.engine.list({ taskName: "resolveDecision" });
+        return runs.length > 0 && runs.every((r) => r.status === "completed") ? runs : null;
+      },
+      30_000,
+      "resolveDecision 完成",
+    );
+    expect(
+      await e.db.selectFrom("agent_run").select("id").where("dedupe_key", "=", `investigate:${d.id}`).execute(),
+    ).toHaveLength(1);
+
+    // 同根因再次出错:已达自动恢复上限(1 次)→ 不再唤醒模型,直接交还人工
+    const runsBefore = (await e.db.selectFrom("agent_run").select("id").execute()).length;
+    e.qb.setMissingFiles(f.hash);
+    const again = await waitFor(
+      async () =>
+        e.db
+          .selectFrom("decision")
+          .selectAll()
+          .where("dedupe_key", "=", `download_failed:${f.dtId}`)
+          .where("id", "<>", d.id)
+          .where("status", "=", "awaiting_user")
+          .executeTakeFirst(),
+      60_000,
+      "同根因新事项交还人工",
+    );
+    expect((again.evidence as { autoRecover: Record<string, unknown> }).autoRecover).toMatchObject({
+      skipped: "attempts_exhausted",
+      recoverAttempted: false,
+    });
+    expect((await e.db.selectFrom("agent_run").select("id").execute()).length).toBe(runsBefore);
+    expect(e.qb.callsTo("torrents/recheck")).toHaveLength(1);
+  }, 240_000);
+
+  it("预算超限:run budget_exceeded,未执行恢复,事项保持待处理", async () => {
+    await e.settings({ agent: { investigateMaxToolCalls: 1 } });
+    const f = await failingDownload(3);
+    const d = await decisionOf(f.dtId);
+    const run = await runOf(d.id);
+    expect(run.mode).toBe("execute");
+    const done = await finished(run.id);
+    expect(done.status).toBe("budget_exceeded");
+    expect(done.summary).toMatch(/^未恢复:调查超出预算/);
+    expect(await toolStarts(run.id)).toEqual(["task.inspect"]);
+    expect(e.qb.callsTo("torrents/recheck")).toHaveLength(1);
+    const d2 = await waitFor(
+      async () => {
+        const r = await e.db.selectFrom("decision").selectAll().where("id", "=", d.id).executeTakeFirstOrThrow();
+        return r.status === "awaiting_user" ? r : null;
+      },
+      30_000,
+      "事项保持待处理",
+    );
+    expect((d2.evidence as { autoRecover: Record<string, unknown> }).autoRecover).toMatchObject({
+      outcome: "budget_exceeded",
+      recoverAttempted: true,
+    });
+    const dt = await e.db.selectFrom("download_task").select("status").where("id", "=", f.dtId).executeTakeFirst();
+    expect(dt?.status).toBe("missing");
+  }, 180_000);
+});

docs/runbook/agent.md

@@ -6,7 +6,7 @@
 
 ## 能力
 
-启动时装配 14 个能力(`catalog.lookup`、`release.search`、`release.inspect`、`identity.resolve`、`download.plan`、`download.submit`、`organize.plan`、`organize.execute`、`task.inspect`、`task.retry`、`task.cancel`、`policy.preview`、`policy.apply`、`inventory.query`)。注册数不等于 14 时进程启动失败。Web(`POST /capabilities/{id}/invoke`)、Pi 工具和后台任务都走 `ctx.capabilities.invoke` 这一条路径。
+启动时装配 15 个能力(`catalog.lookup`、`release.search`、`release.inspect`、`identity.resolve`、`download.plan`、`download.submit`、`download.recheck`、`organize.plan`、`organize.execute`、`task.inspect`、`task.retry`、`task.cancel`、`policy.preview`、`policy.apply`、`inventory.query`)。注册数不等于 `CAPABILITY_IDS.length` 时进程启动失败。Web(`POST /capabilities/{id}/invoke`)、Pi 工具和后台任务都走 `ctx.capabilities.invoke` 这一条路径。
 
 - **授权**(`agent/authorize.ts`):propose / approve / execute 类必须满足以下之一:
   - 单次审批 `approval`:状态为 approved,绑定计划类型 + 计划 ID + payload hash + revision,且未过期。执行成功后审批标记为 consumed。
@@ -57,7 +57,7 @@ worker 崩溃后,新 worker 在引擎租约到期后(约 30s)接手,`drive` 步
 
 ## 后台触发与 resolveDecision
 
-- `decision.created` 路由到 `resolveDecision(trigger=created)`。满足以下全部条件时,会发起一次有界调查 run:`agent.autoInvestigate=true`、kind 为 `identity_ambiguous` / `download_failed` / `download_stalled`、事项仍为 open、事项不是由 Agent 自己引起的。这个 run 的参数:`trigger=event`、`mode=plan_only`,白名单为只读能力加 `identity.resolve`,工具调用上限为 `agent.investigateMaxToolCalls`,`dedupe_key=investigate:<decisionId>`。同一事项只会有一个 run,重复事件或重新打开都不会再建。调查结束后,事项转为 `awaiting_user`,`evidence.agentInvestigation` 里带上结论。
+- `decision.created` 路由到 `resolveDecision(trigger=created)`,由 `agent/investigate.ts` 的 `autoInvestigate` 处理。满足以下全部条件时,会发起一次有界 run:`agent.autoInvestigate=true`、kind 为 `identity_ambiguous` / `download_failed` / `download_stalled`、事项仍为 open、事项不是由 Agent 自己引起的。`trigger=event`,工具调用上限为 `agent.investigateMaxToolCalls`,时长上限 `min(600, agent.maxDurationSec)` 秒,`dedupe_key=investigate:<decisionId>`。同一事项只会有一个 run,重复事件或重新打开都不会再建。模式见下一节。调查结束后,没有被系统核实恢复的事项转为 `awaiting_user`,`evidence.agentInvestigation` 里带上结论。
 - `decision.updated(status=resolved)` 路由到 `resolveDecision(trigger=updated)`,执行用户选中的处理:
   - 选项带 `capabilityId` 时经注册中心调用(照样校验授权);
   - `retry` 重试关联任务;
@@ -66,6 +66,48 @@ worker 崩溃后,新 worker 在引擎租约到期后(约 30s)接手,`drive` 步
 
   执行结果写进 `decision.resolution.execution`。身份歧义类由 sources-identity 在处理端点里直接执行。
 
+## 自动调查与恢复
+
+依据 REQ §8.4.2(下载无进展默认不换源)、§10.7(有界任务、同根因去重、有限重试)、AT68 / AT71。
+
+**模式**(`agent.autoInvestigateMode`):
+
+| 模式 | run 参数 | 能做什么 |
+|---|---|---|
+| `investigate`(缺省) | `mode=plan_only`,白名单 = 只读能力 + `identity.resolve` | 查证据、给结论与建议,事项交还人工 |
+| `recover` | `mode=execute`,白名单 = 上面 + `task.retry` + `download.recheck`,`created_by=system:auto-recover` | 在持续授权范围内执行恢复动作,系统核实后自动关闭事项 |
+
+**恢复范围**:
+- `download.recheck`(execute,replay=never):对已纳管、未移除的下载任务调 qB 强制重新校验(`action=recheck`,起点 missing / failed / stalled / paused),或恢复暂停(`action=resume`)。缓存阈值暂停和人工暂停不自动恢复。只动这一个种子,不换源、不删除、不改路径 / 分类 / 标签。
+- `task.retry`:重试失败 / 死信的业务任务,副作用由各处理器的操作账本去重。
+- `identity.resolve`:按已有证据等级自动接受;证据不够时仍进待处理。
+
+**明确排除**:换源(`download.plan` / `download.submit`)、删除与缓存清理、`task.cancel`、`organize.execute`、`policy.apply`。这些能力不在恢复白名单里,模型调用会得到 `tool_call_rejected`。
+
+**授权**:每次恢复调用仍经 `authorize` 校验持续授权,模型声明无效。可用的 grant:
+- `action=download.recheck`,或发布 kind=automation 的策略时自动生成的 `policy.automation`;
+- `action=task.retry` / `task.*`;`identity.resolve` / `policy.identity`;或 `*`。
+
+grant 必须 active、未过期,绑定策略版本时该版本必须仍是 published。`scope.mediaKinds` / `scope.mediaEntityIds` 由 `download.recheck` 处理器按下载任务 → 计划 → Slot → 作品核对。注意:authorize 对非计划类能力拿不到目标作品,所以带 `mediaEntityIds` 的 grant 目前一律不覆盖 `download.recheck`,是保守拒绝。建 run 前先查有没有覆盖该类事项恢复动作的 grant(下载类看 `download.recheck` / `task.retry`,身份类看 `identity.resolve`)。没有就**降级为 investigate**,并在事项上写 `evidence.autoRecover.degraded` 和 `suggestion`,再追加一个可直接点选的"强制重新校验"选项(`capabilityId=download.recheck`,人工点选同样经授权、审计、账本)。
+
+授权示例(把恢复限定在某个已发布策略下):
+
+```sh
+curl -X POST localhost:4380/api/v1/automation-grants -H "Authorization: Bearer <admin>" -H 'content-type: application/json' \
+  -d '{"action":"download.recheck","policyRevisionId":"<published automation 策略版本>","reason":"下载异常自动重新校验"}'
+```
+
+**核实与收尾**(`runAgent` 里的 `recover-*` 步骤):
+1. 从账本(`capability_invocation`)取出本 run 成功的恢复动作,从审计取出被拒绝的调用。
+2. 按业务事实轮询核实,间隔 `min(goalPollSec, 30)` 秒,最长 `agent.recoverVerifySec`:
+   - `download_task` 进入 downloading / queued / completed / importing / imported 算恢复,removed 算失败;
+   - 重试的 `task` 为 succeeded 算恢复,failed / dead / cancelled 算失败;
+   - `source_release` 已 resolved 算恢复。
+3. 核实已恢复:事项转 `resolved`,`resolved_by=system:auto-recover`,`resolution.execution={kind:"agent_recover", verified:true, facts}`,发 `decision.updated`(executeResolution 看到 execution 已存在,不会重复执行)。run 汇总以"已恢复(系统核实)"开头,模型原话附在后面,标注"未经核查"。
+4. 失败、超时(unverified)、预算超限、没有执行恢复动作:只写 `evidence.autoRecover.outcome` 和建议,事项由 finalize 转 `awaiting_user`,不会自动再试。run 汇总以"未恢复:…"开头。
+
+**有界尝试**:`evidence.autoRecover.attempt` 记录该根因的第几次自动恢复。同一去重键(如 `download_failed:<downloadTaskId>`)下,历史上 `recoverAttempted=true` 的事项数达到 `agent.investigateMaxAttempts`(缺省 2)后,同根因的新事项不再唤醒模型,直接 `awaiting_user`,并写 `skipped=attempts_exhausted`。
+
 ## 配置上游
 
 在设置页或 `PATCH /settings` 里配置:
@@ -85,6 +127,8 @@ worker 崩溃后,新 worker 在引擎租约到期后(约 30s)接手,`drive` 步
     "maxToolCalls": 40, "maxDurationSec": 900,          // 单次运行预算
     "maxTokensPerRun": null, "maxCostUsdPerRun": null,
     "autoInvestigate": false, "investigateMaxToolCalls": 8,
+    "autoInvestigateMode": "investigate",               // 或 "recover"(授权内恢复,见上一节)
+    "investigateMaxAttempts": 2, "recoverVerifySec": 600,
     "goalPollSec": 60, "goalWaitHours": 72
   },
   "budget": { "agentMaxRunsPerDay": 50, "agentMaxCostUsdPerDay": 5 }
@@ -137,6 +181,7 @@ curl -X POST localhost:4380/api/v1/agent/runs -H "Authorization: Bearer umt-dev-
 | `e3-background-continues` | 关闭聊天后台照样推进 |
 | `e4-crash-recovery` | worker 子进程在 download.submit 时 kill -9,新 worker resume,qB 只收到一次 add |
 | `e5-auto-investigate` | 识别歧义 → 有界调查,去重 |
+| `e6-auto-recover` | qB missingFiles → download_failed → 授权内恢复(task.inspect → download.recheck → 系统核实关闭事项);无授权时降级只调查;换源被拒;去重;同根因尝试上限;预算超限保持待处理 |
 
 ## 已知限制
 

packages/contracts/src/capabilities-io.ts

@@ -3,7 +3,7 @@ import { MediaAsset } from "./assets.ts";
 import { CAPABILITY_CATALOG, type CapabilitySpec, type KnownCapabilityId } from "./capability.ts";
 import { Id, JsonObject, NonNegInt } from "./common.ts";
 import { MediaEntity, MediaKind } from "./content.ts";
-import { AcquisitionPlan, DownloadTask, Task, TaskStatus } from "./execution.ts";
+import { AcquisitionPlan, DownloadStatus, DownloadTask, Task, TaskStatus } from "./execution.ts";
 import { ExternalEntityType, ExternalIdentity } from "./identity.ts";
 import { ServerMapping } from "./ops.ts";
 import { OrganizationAction, OrganizationRecord } from "./organization.ts";
@@ -121,6 +121,22 @@ export const OrganizePlanOutput = z.object({
 export const OrganizeExecuteInput = z.object({ planId: Id, payloadHash: z.string().min(1), approvalId: Id.optional() });
 export const OrganizeExecuteOutput = Accepted.extend({ record: OrganizationRecord.nullable() });
 
+/** 授权内恢复:强制重新校验(missingFiles / error 之后)或恢复暂停;换源、删除、清理不在此能力内 */
+export const DownloadRecheckAction = z.enum(["recheck", "resume"]);
+export const DownloadRecheckInput = z.object({
+  downloadTaskId: Id,
+  action: DownloadRecheckAction.default("recheck"),
+  reason: z.string().max(500).optional(),
+});
+export const DownloadRecheckOutput = z.object({
+  downloadTaskId: Id,
+  action: DownloadRecheckAction,
+  statusBefore: DownloadStatus,
+  rawStateBefore: z.string().nullable(),
+  /** 已向下载器发出请求;是否恢复以后续同步到的 download_task 状态为准 */
+  requested: z.boolean(),
+});
+
 export const TaskInspectInput = z.object({ taskId: Id });
 export const TaskInspectOutput = z.object({
   task: Task,
@@ -178,6 +194,7 @@ export const CAPABILITY_IO = {
   "identity.resolve": { input: IdentityResolveInput, output: IdentityResolveOutput },
   "download.plan": { input: DownloadPlanInput, output: DownloadPlanOutput },
   "download.submit": { input: DownloadSubmitInput, output: DownloadSubmitOutput },
+  "download.recheck": { input: DownloadRecheckInput, output: DownloadRecheckOutput },
   "organize.plan": { input: OrganizePlanInput, output: OrganizePlanOutput },
   "organize.execute": { input: OrganizeExecuteInput, output: OrganizeExecuteOutput },
   "task.inspect": { input: TaskInspectInput, output: TaskInspectOutput },
@@ -223,6 +240,8 @@ export type OrganizeExecuteInput = z.infer<typeof OrganizeExecuteInput>;
 export type OrganizeExecuteOutput = z.infer<typeof OrganizeExecuteOutput>;
 export type TaskInspectInput = z.infer<typeof TaskInspectInput>;
 export type TaskInspectOutput = z.infer<typeof TaskInspectOutput>;
+export type DownloadRecheckInput = z.infer<typeof DownloadRecheckInput>;
+export type DownloadRecheckOutput = z.infer<typeof DownloadRecheckOutput>;
 export type TaskRetryInput = z.infer<typeof TaskRetryInput>;
 export type TaskRetryOutput = z.infer<typeof TaskRetryOutput>;
 export type TaskCancelInput = z.infer<typeof TaskCancelInput>;

packages/contracts/src/capability.ts

@@ -57,6 +57,7 @@ export const CAPABILITY_IDS = [
   "identity.resolve",
   "download.plan",
   "download.submit",
+  "download.recheck",
   "organize.plan",
   "organize.execute",
   "task.inspect",
@@ -171,6 +172,17 @@ export const CAPABILITY_CATALOG: Record<KnownCapabilityId, CapabilityDescriptor>
     true,
     false,
   ),
+  "download.recheck": d(
+    "download.recheck",
+    "重新校验/恢复下载",
+    "对已纳管的下载任务发起强制重新校验(action=recheck)或恢复非阈值、非人工的暂停(action=resume);不换源、不删除、不改保存路径",
+    "execute",
+    "DownloadRecheck",
+    "organize.execute",
+    "medium",
+    true,
+    false,
+  ),
   "organize.plan": d(
     "organize.plan",
     "生成整理计划",

packages/contracts/src/settings.ts

@@ -73,8 +73,22 @@ export type PiSettings = z.infer<typeof PiSettings>;
  * 上游/模型/思考档位在 pi 分组;pi.provider = "fake" 时使用脚本化假模型(不联网,演示与 e2e 用)。
  */
 export const AgentRuntimeSettings = z.object({
-  /** 识别歧义 / 下载失败诊断等待处理事项创建时,自动发起有界只读调查 */
+  /** 识别歧义 / 下载失败诊断等待处理事项创建时,自动发起有界调查 */
   autoInvestigate: z.boolean(),
+  /**
+   * 自动调查的模式:investigate = 只读调查,结论交还人工;
+   * recover = 在已发布策略 / 持续授权范围内可执行恢复类能力(task.retry、download.recheck、identity.resolve),
+   * 换源、删除、清理始终排除;没有覆盖的授权时自动降级为 investigate。
+   */
+  autoInvestigateMode: z.enum(["investigate", "recover"]),
+  /** 同一根因(decision 去重键)最多自动恢复几次;用尽后只交还人工,不再唤醒模型 */
+  investigateMaxAttempts: z.number().int().min(1).max(10),
+  /** 恢复动作后等待业务事实(下载状态 / 任务状态)核实的最长时间(秒) */
+  recoverVerifySec: z
+    .number()
+    .int()
+    .min(5)
+    .max(24 * 3600),
   /** 单次运行:工具调用次数上限 */
   maxToolCalls: z.number().int().min(1).max(500),
   /** 单次运行:时长上限(秒) */
@@ -165,6 +179,9 @@ export const DEFAULT_SETTINGS: GlobalSettings = {
   },
   agent: {
     autoInvestigate: false,
+    autoInvestigateMode: "investigate",
+    investigateMaxAttempts: 2,
+    recoverVerifySec: 600,
     maxToolCalls: 40,
     maxDurationSec: 900,
     maxTokensPerRun: null,