Repin modules/backend with the intent-test hardening

Why: the intent-session tests were load-sensitive in CI (409 races
against the prior run settling); the hardening rides the A2 pin.
Evidence: repair receipt (a1ff-run.json) and 3x race verification
(verify-a1ff.json exit=0) on the merged tip; full make test green
(verify-a2-combo3.json exit=0).
Impact: one component gitlink.
Refs: PR #704

.task/archive/feature-exec-recovery-diagnostics-a2.md

@@ -0,0 +1,167 @@
+# feature/exec-recovery-diagnostics-a2 任务卡
+
+任务主记录：PR #704 https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Dev/pulls/704；状态：已完成（合入前闭合）；owner：weichen
+
+## 基线与授权
+
+- 权威仓：ssh://git@git-ssh.u1680832.nyat.app:31411/weichen/Agent-System-Dev.git；目标分支：main；R0（initial_base_sha）= c92443280ef8e9ed29bf5bf5250df650c01b4478
+- 工作区：/home/weichen/projects/Agent-System-Dev-wt-feature-exec-recovery-diagnostics-a2（worktree）；分支：feature/exec-recovery-diagnostics-a2
+- 分支类型：standalone；root_target：main；immediate_target：main；accepting_requirements：false
+- 需求分流：verdict=NEW；selected_initiative=无；requirement_sha256=34e72676959d0e970f1a3cb0e90ac2067dfb31f3c2e7a6650b5de9f3f77923c7；triage_receipt=/tmp/triage-a2-1792672248.json；receipt_sha256=7924fdbaeba41d98994c2e11bc96026d4db3cdbd0b28a9d1e381400eae9a6de4
+
+- 授权与交付终点：用户已批准启动 A2 执行与恢复诊断（方案 Functional Follow-Ups 第二项）。范围内连续推进：设计、agent-per-unit 实施、验证、组合、Review、PR/CI/合入。未授权：控制台/前端界面、A3/A4、部署。
+- 设计决定（2026-10-22，host）：
+  1. 只读诊断 API：GET /api/runs/{id}/diagnostics（owner/project 授权，复用既有 run 鉴权），响应聚合：状态、当前责任方与等待原因（running / waiting_hitl / confirmation_pending / recovery_hold(原因+created_at) / reconciling / failed+reason / completed）、最后稳定 checkpoint（如可得）、已知副作用与关联 operation 记录（复用 cmdbus 审计信封查询）、下一步动作建议（确认卡片 / 解除 hold / 可安全重试 / 需人工核对 / 无）。
+  2. 纯读聚合，不改任何执行/恢复行为；复用既有 store 方法（GetRun/RecoveryHold/ListInterruptedSubagentRuns/审计查询）与 U7b 的 operation 查询。
+  3. 不建控制台、不加写接口；批量列表走既有会话 run 列表。
+- 交付模式：goal-first（目标、范围和交付终点一次确认；范围内连续推进）
+- 实施执行：agent-per-unit（配置：AGENT-EXECUTION.md）
+- 合入执行：agent（已授权交付范围内，Review 通过、冲突解决且当前 HEAD 的 CI 通过后合入）
+- 人工决策点：仅限目标/范围/验收变更、重大取舍、高后果操作或无法自动判定的业务验收
+- 当前阶段：需求拆解；当前增量：I1（待定义）
+- 流程依据：仓库根目录 policy/delivery.md（归档后仍按仓库根目录定位）
+
+## 目标与非目标
+
+**目标**：单只读端点聚合"卡在哪里/谁在等/为何等/最后 checkpoint/已知副作用/下一步"，覆盖 HITL 等待、确认等待、恢复 hold（批二 needs-reconciliation）、失败原因等全部可诊断状态。
+
+**非目标**：控制台/UI；任何写操作或状态变更端点；A3/A4； squad sqlite 协作层的跨库聚合（只读后端 store 与审计）。
+
+## Agent 实施收据
+
+路由与 fail-stop 规则见根目录 AGENT-EXECUTION.md 与 policy/delivery.md；每次尝试一行，原始日志留在源码外收据路径。
+
+| 单元 ID | 预期范围 | 配置摘要 | 角色 | 候选顺序 | attempt | CLI/模型 | 终态 | transport | fallback | 源码外收据路径 |
+|---|---|---|---|---|---|---|---|---|---|---|
+| 待定义 | — | NOT_RUN | NOT_RUN | NOT_RUN | NOT_RUN | NOT_RUN | NOT_RUN | NOT_RUN | NOT_RUN | — |
+
+## 关联方向与规划
+
+- 适用状态：待核对（无适用规划时说明原因）
+- 主记录路径/版本：待核对（路径按仓库根目录定位，归档后不变；只引用唯一正文）
+- 对应里程碑：待核对
+
+## 需求分析
+
+| ID | 描述 | 项目符合度及依据 | 优先级 | 验收条件 | 依赖 | 结论 |
+|---|---|---|---|---|---|---|
+| R1 | 待填写 | 待评估 | 待定 | 待定义 | 待确认 | 待评审 |
+
+## 方案与计划
+
+- 方案引用：待填写（含对象版本）
+- 实施计划引用：待填写（含增量与需求 ID）
+- 测试计划引用：待填写（含需求 ID 与验收条件）
+
+## 阶段记录
+
+| 阶段 | 对象版本 | 状态 | 决策/执行者 | 时间 | 依据 |
+|---|---|---|---|---|---|
+| 需求拆解 | 待定 | 待执行 | — | — | — |
+| 需求评估 | 待定 | 待执行 | — | — | — |
+| 需求表确认 | 待定 | 待执行 | — | — | — |
+| 方案设计 | 待定 | 待执行 | — | — | — |
+| 实施计划 | 待定 | 待执行 | — | — | — |
+| 实施与自检 | 待定 | 待执行 | — | — | — |
+| 正式测试与验收 | 待定 | 待执行 | — | — | — |
+| 代码 Review | 待定 | 待执行 | — | — | — |
+| 合入 | 待定 | 待执行 | — | — | — |
+| 合入核验与收口 | 待定 | 待执行 | — | — | — |
+
+## 探索试验与迭代
+
+| 增量/试验 | 假设 | 范围 | 判断和停止条件 | 授权/决策依据 | 结果 | 变更影响 | 需重审补测 |
+|---|---|---|---|---|---|---|---|
+| 待定义 | — | — | — | 待核对 | NOT_RUN | 待评估 | 待评估 |
+
+## 同步记录
+
+| 节点 | sync_base_sha | tested_head_sha | checked_at |
+|---|---|---|---|
+| 启动 | c92443280ef8e9ed29bf5bf5250df650c01b4478 | c92443280ef8e9ed29bf5bf5250df650c01b4478 | 2026-10-22T12:37:28Z |
+
+## 证据（Evidence）
+
+<!-- CMD="<命令>" make evidence 写入；手抄/转写输出视为无效证据（AGENTS.md 证据与验证门禁）。 -->
+| 时间(UTC) | 命令 | 被测HEAD | exit | 来源 | sha256(16) |
+|---|---|---|---|---|---|
+
+## 交付调度
+
+- 调度状态：未入队；调度记录：无
+- entry_id：无；attempt_id：无；窗口执行者：无
+- 固定候选：无（base/head、组件与相关配置引用实际尝试）
+- 调度依据：policy/delivery.md#合入调度
+- 固定后证据：源码外 RECEIPT；实际结果后按授权原样导入和闭合
+
+## 结果与验证
+
+- 结果：
+- 验证：<结论摘要 PASS/FAIL/SKIP/NOT_RUN 逐项；逐条证据由 make evidence 写入上方证据表>
+- 影响与剩余项：
+
+## 规划进展与下一步
+
+- 本次推进/剩余：待核对（依据实际成果更新适用方案/规划）
+- 首选下一任务：待核对（无适用规划、已完成或受阻时如实说明，不虚构任务）
+- 推荐依据：待填写
+- 前置条件/待确认：待核对
+- 预期产物/验收条件：待定义
+- 重复任务核对：待核对（已有/在途任务、PR 与规划记录）
+- 启动授权：未授权（建议不自动启动新任务，仍须明确授权）
+
+## 合入与清理
+
+- 实施收据：a2-diagnostics-api（pi attempt1，rundiag 聚合 + GET /api/runs/{id}/diagnostics，backend 06b0c9f，verify-a2b exit=0）；架构基线边登记（repair，a2arch-run.json）；backend pin 集成（a2git2-run.json，cursor 配额间歇故障重试一次后成功）。
+- 组合验证：全量 make test PASS（verify-a2-combo2.json exit=0，本轮 svc 既有 flake 未复现）。
+- Review 记录：独立 CLI 评审单元 a2-independent-review（第 2 次派发 success；首轮 pi-glm 终态记录不可解码，现场保留）；结论 APPROVE（0 blocker / 0 major）；报告 .task/reviews/2026-10-22-a2-diagnostics-independent-review.md；被审对象合并基线 vs main
+- 冲突记录：NOT_RUN（最新目标同步、冲突处置、组合兼容核验及复审/补测）
+- CI 记录：NOT_RUN（当前 HEAD 完整 SHA、全部必需检查、运行链接、执行身份及 GATE 结果）
+- 合入记录：NOT_RUN（交付授权范围、PR、被核验 HEAD、合入 SHA、目标分支及内容核验）
+- 清理记录：NOT_RUN（授权依据、写者退出确认、worktree/分支/任务卡处置）
+
+<!-- Local receipt: SELF. Digests detect corruption; they do not authenticate a reviewer or CI. -->
+```json
+{
+  "summary_schema": 1,
+  "source": "SELF",
+  "timestamp": "2026-10-22T12:30:58.698213+00:00",
+  "command": "make precheck",
+  "exit_code": 0,
+  "cwd": "/home/weichen/projects/Agent-System-Dev",
+  "before_head": "c92443280ef8e9ed29bf5bf5250df650c01b4478",
+  "after_head": "c92443280ef8e9ed29bf5bf5250df650c01b4478",
+  "branch": "fix/gitea-address-migration",
+  "target_branch": "main",
+  "target_head": "c92443280ef8e9ed29bf5bf5250df650c01b4478",
+  "root_target": "main",
+  "immediate_target": "main",
+  "before_tracked_diff_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+  "after_tracked_diff_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+  "output_tail": [
+    "make[1]: Entering directory '/home/weichen/projects/Agent-System-Dev'",
+    "[repo]   branch=fix/gitea-address-migration head=c92443280e",
+    "[remote] origin/main=c92443280e ahead=0 behind=0",
+    "[tree]   clean submodules_drift=0 uninit=0",
+    "[worktrees] linked=7",
+    "[recycle] merged un-recycled branches: fix/gitea-endpoint-facts",
+    "[initiatives] local cards (not authoritative): none",
+    "[hooks]   core.hooksPath=scripts/hooks（pre-commit/post-merge 实际接线）",
+    "[facts]   policy/facts.md 6 条事实均在有效期内",
+    "[required-reading] core/baseline.md, workflows/delivery.md §1, policy/delivery.md",
+    "[R0] c92443280e（完整：c92443280ef8e9ed29bf5bf5250df650c01b4478）",
+    "[worktree] before-sync: recursive component checks passed",
+    "[verdict] READY — 保留本次 precheck receipt 后建任务工作区：PRECHECK_RECEIPT=/tmp/task-precheck.json make task-start NAME=feature/your-branch",
+    "make[1]: Leaving directory '/home/weichen/projects/Agent-System-Dev'"
+  ],
+  "output_sha256": "d670ed210b367c72f877fb63bbbeb5b28411e95cb56fab782f27172fa481ed89",
+  "receipt_sha256": "2829e17065c329449fac8ef2690d50babf82015a1f546dcbfd3f27f67789dd0c",
+  "receipt_path": "/home/weichen/.cache/agent-system-dev/evidence/2829e17065c329449fac8ef2690d50babf82015a1f546dcbfd3f27f67789dd0c.json"
+}
+```
+
+## 闭合记录（合入前折叠）
+
+- 本卡在 PR #704（https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Dev/pulls/704）合入前闭合并归档，随该 PR 一同合入。
+- 该 PR 的合并提交即本任务的落地记录。
+- 合入后无需回填，也无需单独的闭合 PR。

.task/reviews/2026-10-22-a2-diagnostics-independent-review.md

@@ -0,0 +1,44 @@
+# Independent Code Review — feature/exec-recovery-diagnostics-a2 (A2 run diagnostics)
+
+- Reviewer: independent CLI review session (`a2-independent-review` unit; fresh session, not the implementer)
+- Date: 2026-10-22
+- Review object: this checkout @ `d0b2bde9` (branch `feature/exec-recovery-diagnostics-a2`) versus authoritative main. No `origin/*` refs exist in this sandbox; task card R0 = `c9244328` (= `d0b2bde9~5`, the merged main tip `c92443280ef8e9ed29bf5bf5250df650c01b4478`), so `git diff c9244328...d0b2bde9` ≡ `git diff origin/main...HEAD`.
+- Dev-level scope: task card `.task/feature-exec-recovery-diagnostics-a2.md` (new), `modules/backend` gitlink `9ee752c3` → `06b0c9f1`, `policy/architecture-baseline.json` +3 edges. Verified: dev diff is exactly these 3 files; the three edges (`backend/rundiag backend/cmdbus`, `backend/rundiag backend/store`, `backend/server backend/rundiag`) sit in correct sorted positions and match the actual import graph (see N5).
+- Backend scope: `git -C modules/backend diff 9ee752c..06b0c9f` — 7 files, +1418/−0: `internal/rundiag/rundiag.go` (488), `internal/rundiag/rundiag_test.go` (440), `internal/server/run_diagnostics.go` (32), `internal/server/run_diagnostics_test.go` (226), `internal/server/server.go` (+2), `internal/store/run_diagnostics.go` (110), `internal/store/run_diagnostics_test.go` (120).
+- Method: read-only review of the full diff plus the neighboring run-read/cancel routes, the pre-existing store methods the aggregator reuses, and the cmdbus pending-confirm write path. No code changed; the only write is this report.
+
+## Findings
+
+### Blockers
+
+None.
+
+### Major
+
+None.
+
+### Minor
+
+- **MI1 — Evidence-driven wait states outrank terminal run statuses; a completed/failed run can be reported as `waiting_hitl`.** In the responsibility switch the HITL arm `case ev.Run.Status == core.RunWaitingHITL || card != nil || decision != nil` (`modules/backend/internal/rundiag/rundiag.go:252`) is evaluated before `failed` (:281) and `completed` (:296). A run that ended while a decision card scoped only by owner+conversation (human_decisions has no run column; attributed by the time-window inference at :432-448) remained `pending`, or with a leftover unresolved run-scoped HITL card, yields `state=waiting_hitl`, `role=human`, `next_step.action=confirm_card` for a terminal run. This is exactly the fixed precedence the task card approved and the documented bounded attribution (`rundiag.go:15-19`), and `run.status` stays visible in the report — but a follow-up that suppresses evidence-driven wait mapping (or adds a note) when the run is terminal would remove the confusing "confirm card" advice on finished runs. Not blocking: matches the approved design; purely advisory output.
+- **MI2 — Silent truncation of side-effect evidence contradicts the package's own honesty contract.** `sideEffects` stops appending at `maxSideEffects = 50` with no note (`rundiag.go:342-366`, cap at :57), and `Gather` fetches only the 200 newest audit envelopes (`maxAuditScan`, `rundiag.go:175` → `ListAudits` orders `time DESC`, `internal/store/audit.go:147-152`), so a busy conversation can under-report a run's operations without any indication. Every other evidence gap in the package adds a note (even "absence of records is not proof of no effect", :313-315); a "side effects truncated" note would complete that contract. Diagnostic-only impact; recorded evidence is still correct as far as it goes.
+- **MI3 — Precedence is tested only at the top of the chain.** Of the ordered arms at `rundiag.go:224-310`, only "unreleased recovery hold wins over everything" (`rundiag_test.go:81-89`) exercises interleaved evidence; combinations such as released-hold+interrupted+open-HITL-card (→ `reconciling`) or decision+pending-confirm (→ `waiting_hitl`) are untested and rest on the switch order alone. Low risk because the order is literal and table-covered per state; suggesting coverage only because precedence is the stated spec center of this unit.
+
+### Notes
+
+- **N1 — Per-state derivation and precedence are complete and match the task-card contract.** All eight states are derivable: `unknown` fail-closed on unreadable hold (:226-233), `recovery_hold` on an open hold row (:234-240), `reconciling` for released-hold+`RunInterrupted` residue (:241-251, predicate at :402-403), `waiting_hitl` from status/run-scoped card/time-windowed decision (:252-273), `confirmation_pending` from an owner-scoped pending confirm parsed out of the dispatch payload (:274-280, :453-471), `failed` with retry-safety gated on unknown receipts (:281-295), `completed` (:296-298), `running` for running+pending (:299-303), `unknown` default for unmapped statuses (cancelled; interrupted without hold; :304-310). Precedence order is exactly `recovery_hold > reconciling > waiting_hitl > confirmation_pending > failed > completed > running > unknown`, with the fail-closed `HoldUnknown` degradation deliberately first.
+- **N2 — Security: scoping is byte-for-byte the neighboring run-read pattern; no cross-owner leakage path exists.** The handler resolves the run via `s.st.GetRun(s.owner(r), r.PathValue("id"))` (`internal/server/run_diagnostics.go:25`) — identical to `getRun` (`server.go:1592`) — where `owner` comes only from the auth context (`auth.go:220-225`), `GetRun` filters `WHERE id=? AND owner=?` with cross-tenant = `ErrNotFound` (`messages.go:307-316`), and `writeStoreErr` maps it to 404 indistinguishable from a missing run (`server.go:2617-2620`). Every follow-on query is keyed on the already-authorized run's own fields, never on request parameters: `GetRecoveryHold(run.ID)` (`recovery_holds.go:36`), `ActiveHITLCardsByRun(run.ID)` (`recovery.go:156`), `PendingHumanDecisionsByConversation(run.Owner, run.ConversationID)` with empty-owner defense (`store/run_diagnostics.go:51-53`), `ListOwnerPendingConfirms(run.Owner)` (:88-90), `LastRunCheckpoint(run.ConversationID, run.ID)` (:28-30), `ListAudits(run.Owner, run.ConversationID)` (`audit.go:134-140`). The `pending_confirms` payload is parsed only to extract command/owner/conversation identity; the raw payload is never echoed (`rundiag.go:453-471`).
+- **N3 — Read-only holds end to end.** All six store entry points the aggregator uses are SELECT-only (verified by reading each); `Build` is a pure function; the handler path contains no store write and the mux registers only the GET pattern (`server.go:396`). The author's `TestRunDiagnosticsIsReadOnly` additionally asserts the run row and hold are unchanged after a GET and that POST/PUT/DELETE/PATCH are not served (`run_diagnostics_test.go:153-193`).
+- **N4 — No route-registration conflicts.** `GET /api/runs/{id}/diagnostics` differs from `GET /api/runs/{id}` (:394) and `POST /api/runs/{id}/cancel` (:393) by a literal extra segment / method; ServeMux precedence is specificity-based, single registration, no duplicate pattern anywhere in the diff.
+- **N5 — Architecture-baseline edges are correct and validated.** `internal/rundiag` imports exactly `cmdbus` and `store` within the backend module (`rundiag.go:21-24`); the `agent-command` import is intentionally untracked — the baseline records only intra-repo package edges and has no deps-module edges anywhere (same precedent as `backend/cmdbus`, which also imports agent-command without an edge). `backend/server backend/rundiag` matches the handler import. Sorted positions are correct (`route < rundiag`; `ratelimit < rundiag < runmgr`), and `scripts/lint-architecture.go` compares this file against the collected import graph, so the edges are machine-checked in CI.
+- **N6 — Go tests NOT_RUN in this review workspace (environment, not code).** The backend's nested dependency checkouts (`modules/backend/deps/agent-command`, `deps/agent-system-core`, …) are empty gitlinks here and no `.git/modules` gitdirs exist; `git submodule init/update` is forbidden for this unit and network is off, so `go build ./internal/rundiag/ …` fails at module setup (`agent-command@v0.0.0 (replaced by ./deps/agent-command): reading deps/agent-command/go.mod: no such file or directory`), before any code compiles. Fallbacks applied: every changed file was fully line-read; all `core.Run*`/phase/event symbols referenced are used by pre-existing compiling backend code; `cmdbus.Request` carries `ID/Owner/ConversationID` (`cmdbus.go:163-177`) and the pending-confirm payload is the marshaled dispatch request (`cmdbus.go:522-523`), so `waitingConfirm`'s parsing is sound; the `events` table has `run_id`/`seq`/`ts`/`type` in both dialects (`migrate.go:47,71`) and the checkpoint query is served by `idx_events_conv(conversation_id, seq)`. Author-recorded receipts were reviewed and are internally consistent with the task card: `/tmp/agent-stack-b2/verify-a2b.json` (`go build ./... && go test -count=1 ./internal/server/... ./internal/rundiag/... ./internal/store/... && go vet …`, exit=0, `ok agent-system-backend/internal/{server,rundiag,store}`) and `/tmp/agent-stack-b2/verify-a2-combo2.json` (`make test`, exit=0) — author-claimed (SELF), not independently re-executed in this session.
+- **N7 — Attribution boundaries are documented and bounded; one sibling-run overlap remains.** Audit envelopes carry no runId today (host bus doesn't write it), so `belongsToRun` (`rundiag.go:371-383`) attributes by conversation + "not issued before run start". Concurrent sibling runs (subagents) in the same conversation have overlapping windows, so until the bus persists `runId`, an envelope from a sibling can appear in both runs' `side_effects`. The limitation is stated in the package doc (:17-19) and the report surfaces `operation_id`/`command_id` so a human can disambiguate; when the bus starts writing `env.RunID` the strict branch (:372-373) takes over automatically. Acceptable for a diagnostic surface.
+- **N8 — "No records ≠ no effect" honesty contract is consistently enforced** (unrecorded checkpoints → null + note at :311-312, tested at `rundiag_test.go:280-302`; empty side effects → note :313-315; unreadable hold → fail-closed unknown that can never yield `safe_retry` :226-233, tested :304-339). Only the truncation gap of MI2 falls short of it.
+
+## Lens summaries
+
+1. **Spec** — pass. State derivation and precedence are complete and per-state correct (N1); the JSON contract (snake_case, `last_checkpoint` explicit null, `side_effects` always an array, `next_step` never reason-less) is pinned by `TestReportJSONContract` (`rundiag_test.go:341-389`). Dev-level artifacts (task card, 3 baseline edges, gitlink `06b0c9f`) are exactly as claimed.
+2. **Security** — pass. Auth-context-only owner, tenant-filtered run fetch, 404 on cross-owner/missing (tested `run_diagnostics_test.go:62-97`), all downstream reads anchored on the authorized run, strictly read-only (N2/N3), no payload echo, empty-owner fail-closed store helpers.
+3. **Regressions** — pass. The diff is additive: 7 new files plus a 2-line route registration; no execution/recovery/runmgr path touched, no existing signature changed, no route conflicts (N4), no migration/schema/index change.
+4. **Tests** — pass with minor gaps. Per-state table of 12 cases covering all eight states plus degradation (`rundiag_test.go:78-225`), scoping at both server and store layers including cross-owner/cross-conversation/empty-owner (`run_diagnostics_test.go:62-97`, `store/run_diagnostics_test.go:13-120`), unknown-degradation at three levels (unreadable source, unmapped status, fail-closed hold; `run_diagnostics_test.go:195-226`), read-only and write-method rejection (`:153-193`), pre-run evidence exclusion (`rundiag_test.go:391-406`). Gaps: MI3 (chain precedence below the top arm), MI2 (truncation without note). Focused tests could not be re-run here (N6); author receipts for `go test ./internal/rundiag/... ./internal/server/... ./internal/store/...` and full `make test` are on record and uncontradicted.
+
+REVIEW VERDICT: APPROVE

policy/architecture-baseline.json

@@ -51,6 +51,8 @@
 "backend/rag backend/settings",
 "backend/rag backend/store",
 "backend/route backend/intentgate",
+"backend/rundiag backend/cmdbus",
+"backend/rundiag backend/store",
 "backend/runmgr backend/bus",
 "backend/runmgr backend/cardsvc",
 "backend/runmgr backend/docgen",
@@ -85,6 +87,7 @@
 "backend/server backend/prompts",
 "backend/server backend/ratelimit",
 "backend/server backend/route",
+"backend/server backend/rundiag",
 "backend/server backend/runmgr",
 "backend/server backend/settings",
 "backend/server backend/store",