Pin the nested agent-native to the M5 fix

Why: composition consistency - Dev pins agent-native at the M5 preflight tip; the squad container's nested copy must match (check-pins). Evidence: pin only; adapters/native tests green on the new pin. Impact: nested gitlink pointer change. Refs: feature/audit-followups-pin-gate-m5
Preserve frame arrival order in the native event probe
2026-10-22 06:41:21 +00:00 · 2026-10-22 03:59:55 +00:00 · 2026-10-22 03:25:28 +00:00 · 2026-10-22 02:57:17 +00:00 · 2026-10-22 02:57:17 +00:00 · 2026-10-21 18:05:51 +00:00
59 changed files with 7545 additions and 168 deletions
@@ -0,0 +1,150 @@
+name: governance
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  governance:
+    runs-on: ubuntu-latest
+    env:
+      GOPROXY: https://goproxy.cn,direct
+    steps:
+      - name: Checkout
+        env:
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          SUBMODULE_TOKEN: ${{ secrets.SUBMODULE_TOKEN }}
+        run: |
+          set -euo pipefail
+          # 只物化两个治理 gitlink（Guidelines 与 deps/development-manager），非递归；
+          # 其它 gitlink 不物化、不取回。凭据只以 URL 限定的 extraheader 存在于本进程树内。
+          python3 - <<'PY_CHECKOUT'
+          import os
+          from pathlib import Path
+          import re
+          import subprocess
+          import sys
+          from urllib.parse import urlsplit
+
+          # 服务端基址取运行器注入的 GITHUB_SERVER_URL（平台控制、候选 PR 不可影响）。
+          SERVER = os.environ.get("GITHUB_SERVER_URL", "").rstrip("/")
+          parsed_server = urlsplit(SERVER)
+          if parsed_server.scheme not in {"http", "https"} or not parsed_server.netloc:
+              sys.exit("Checkout server must be the platform-injected http(s) URL")
+          SSH = "ssh://git@git-ssh.u1680832.nyat.app:31411/weichen/"
+          # 审查过的白名单，独立于候选 .gitmodules：仅列出两个治理 gitlink。
+          REPOSITORIES = {
+              "deps/development-guidelines": "Development-Guidelines",
+              "deps/development-manager": "Development-Manager",
+          }
+          TARGET = "weichen/Agent-Squad"
+          root_token = os.environ.pop("TOKEN", "")
+          component_token = os.environ.pop("SUBMODULE_TOKEN", "")
+          if not root_token or not component_token:
+              sys.exit("Checkout requires TOKEN and SUBMODULE_TOKEN")
+          if os.environ.get("GITHUB_REPOSITORY") != TARGET:
+              sys.exit("Checkout repository is outside the reviewed allowlist")
+          sha = os.environ.get("GITHUB_SHA", "")
+          if not re.fullmatch(r"[0-9a-fA-F]{40}", sha):
+              sys.exit("Checkout requires a fixed commit SHA")
+
+          # 丢弃继承的 Git 追踪/配置/SSH 钩子；凭据只存在于单个 Git 进程树的精确 URL 头中。
+          clean_env = {k: v for k, v in os.environ.items() if not k.startswith("GIT_")}
+          clean_env.update(GIT_CONFIG_NOSYSTEM="1", GIT_CONFIG_SYSTEM="/dev/null",
+                           GIT_CONFIG_GLOBAL="/dev/null", GIT_TERMINAL_PROMPT="0")
+          config = {
+              "core.hooksPath": "/dev/null", "core.fsmonitor": "false",
+              "init.templateDir": "", "credential.helper": "",
+              "fetch.recurseSubmodules": "false", "submodule.recurse": "false",
+              "http.followRedirects": "false", "protocol.allow": "never",
+              "protocol." + urlsplit(SERVER).scheme + ".allow": "always",
+          }
+
+
+          def git(repo, *args, extra=None):
+              values = config | (extra or {})
+              env = clean_env | {"GIT_CONFIG_COUNT": str(len(values))}
+              for index, (key, value) in enumerate(values.items()):
+                  env[f"GIT_CONFIG_KEY_{index}"] = key
+                  env[f"GIT_CONFIG_VALUE_{index}"] = value
+              return subprocess.run(["git", "-C", str(repo), *args], env=env,
+                                    check=True, stdout=subprocess.PIPE).stdout
+
+
+          def materialize_governance(root):
+              # -r 列出任意深度的 gitlink，但不进入 submodule 内部；其它 gitlink 只记录、从不取回。
+              links = {}
+              for entry in git(root, "ls-tree", "-rz", "HEAD").split(b"\0"):
+                  if not entry:
+                      continue
+                  metadata, path = entry.split(b"\t", 1)
+                  mode, kind, commit = metadata.decode().split()
+                  if mode == "160000":
+                      links[path.decode()] = commit
+              for name in REPOSITORIES:
+                  if name not in links:
+                      raise ValueError("Governance gitlink is missing from the fixed commit")
+              raw = git(root, "config", "--no-includes", "--blob", "HEAD:.gitmodules", "--null", "--list")
+              entries = {}
+              for item in raw.split(b"\0"):
+                  if not item:
+                      continue
+                  key, value = item.decode().split("\n", 1)
+                  match = re.fullmatch(r"submodule\.(.+)\.(path|url)", key)
+                  if not match:
+                      continue
+                  name, field = match.groups()
+                  if name not in REPOSITORIES:
+                      continue
+                  if field in entries.setdefault(name, {}):
+                      raise ValueError("Duplicate candidate .gitmodules setting")
+                  entries[name][field] = value
+              # 先整体校验两个治理条目，再进行任何带凭据的取回。
+              for name, repository in REPOSITORIES.items():
+                  fields = entries.get(name, {})
+                  if fields.get("path") != name:
+                      raise ValueError("Governance submodule path is outside the reviewed allowlist")
+                  trusted_url = f"{SERVER}/weichen/{repository}.git"
+                  if fields.get("url") not in {SSH + repository + ".git", trusted_url}:
+                      raise ValueError("Governance submodule URL is outside the reviewed allowlist")
+              for name, repository in REPOSITORIES.items():
+                  url = f"{SERVER}/weichen/{repository}.git"
+                  # 仅持久化审查过的公开 URL，使 Git 在进程级配置消失后仍识别该 submodule。
+                  git(root, "config", f"submodule.{name}.url", url)
+                  git(root, "submodule", "update", "--init", "--checkout", "--", name,
+                      extra={f"submodule.{name}.url": url,
+                             f"submodule.{name}.update": "checkout",
+                             f"http.{url}.extraheader": "AUTHORIZATION: token " + component_token})
+                  child = root / name
+                  if git(child, "rev-parse", "HEAD").decode().strip() != links[name]:
+                      raise ValueError("Materialized governance gitlink does not match the fixed SHA")
+
+
+          try:
+              root = Path.cwd()
+              # 必须是干净的临时工作区；不复用既有 Git 状态。
+              if (root / ".git").exists():
+                  raise ValueError("Checkout requires a fresh workspace")
+              git(root, "init", "-q")
+              url = f"{SERVER}/{TARGET}.git"
+              git(root, "remote", "add", "origin", url)
+              git(root, "fetch", "--depth=1", "origin", sha,
+                  extra={f"http.{url}.extraheader": "AUTHORIZATION: token " + root_token})
+              git(root, "checkout", "-q", "--detach", "FETCH_HEAD")
+              materialize_governance(root)
+          except (ValueError, OSError, subprocess.CalledProcessError) as error:
+              sys.exit(f"Checkout refused: {error}")
+          print("Checkout complete; credentials were not persisted in Git configuration")
+          PY_CHECKOUT
+          unset TOKEN SUBMODULE_TOKEN
+      - name: Install Go 1.27.1
+        run: |
+          set -euo pipefail
+          curl -fsSL -o "$RUNNER_TEMP/go.tgz" https://mirrors.aliyun.com/golang/go1.27.1.linux-amd64.tar.gz
+          tar -C /usr/local -xzf "$RUNNER_TEMP/go.tgz"
+          ln -sfn /usr/local/go/bin/go /usr/local/bin/go
+          ln -sfn /usr/local/go/bin/gofmt /usr/local/bin/gofmt
+      - name: Validate agent execution configuration
+        run: scripts/agent-exec validate --repo .
+      - name: Governance check
+        run: scripts/dm governance check --repo .
@@ -13,3 +13,6 @@
 [submodule "deps/agent-intent"]
 	path = deps/agent-intent
 	url = ssh://git@git-ssh.u1680832.nyat.app:31411/weichen/Agent-Intent.git
+[submodule "deps/development-manager"]
+	path = deps/development-manager
+	url = ssh://git@git-ssh.u1680832.nyat.app:31411/weichen/Development-Manager.git
@@ -0,0 +1,58 @@
+# Agent-Squad 采纳 development-governance-v2 治理套件 任务卡
+
+任务主记录：[Agent-Squad PR #54](https://git-home.u1680832.nyat.app:30319/weichen/Agent-Squad/pulls/54)（本文件随 PR 描述维护）。
+状态：已完成（闭合随本 PR 合入；合入以平台记录为准）；owner：weichen。
+
+## 基线与授权
+
+- 权威仓：weichen/Agent-Squad；root_target：main；immediate_target：main
+- branch_kind：initiative；initiative_id：governance-adoption-Agent-Squad
+- R0（initial_main_sha）：490b0b2dfa1567cb16ff7a7230fd78e52bc2938e；工作区：/home/weichen/projects/Agent-Squad-wt-feature-governance-adoption（worktree）；分支：feature/governance-adoption
+- 父任务信息（仅子任务填写）：无
+- accepting_requirements：false（单仓采纳 initiative，范围固定为下列十步，不接收新需求）
+- 需求分流：verdict=NEW；selected_initiative=无；候选=无；理由=development-governance-v2 Phase 6 每个正式仓库独立 initiative；triage 收据=无（本仓采纳前尚无 triage 入口）
+- 授权与交付终点：owner 已授权 development-governance-v2 全部实施（Agent-System-Dev 主记录 docs/superpowers/plans/2026-10-16-development-governance-v2-master.md）；终点为本 PR 合入 main 且采纳状态 active。分支保护与 tag 由用户执行。
+- 实施执行：agent-per-unit（配置：AGENT-EXECUTION.md；外部 Agent 一单元一新进程/会话/工作区）
+
+## 目标与非目标
+
+目标（Phase 6 十步）：固定 Guidelines SHA（deps/development-guidelines @ 2688fd2be6a52c0bf110e7e6366b6f291ab5e4eb）；增加 AGENT-EXECUTION.md；增加共享执行器薄入口（deps/development-manager @ bd3a7de57a3a9e1a4495b94c5fa5c08d71e7be8c + scripts/dm、scripts/agent-exec、scripts/task）；四条路由（implementation/repair/test/review）；任务卡模板与分支元数据；CI 中启用 schema 校验（.gitea/workflows/governance.yml：agent-exec validate + governance check）；一次正例运行；负例（缺配置、坏收据、越界路径、未知副作用）；独立 Review + CI；采纳状态 active。
+
+非目标：不改业务代码与既有测试；不改分支保护（由用户执行）；不发布 tag。
+
+范围说明（Review 阻塞项 2）：`scripts/submodule-consistency.sh` 要求 deps/agent-native、deps/agent-intent 内嵌的 Guidelines/错误码/日志 gitlink 与本仓一致；本 PR 把 Guidelines 固定到 2688fd2 后，按最小修复把这两个 gitlink 升到各自已合入治理采纳的 main（Agent-Native 14d4869、Agent-Intent 651a2e1；两者相对原固定版本只增治理文件、不改 Go 代码），其余 gitlink 不动。核对见证据 `scripts/submodule-consistency.sh`。
+
+引导说明：AGENT-EXECUTION.md 由 host 以固定版本治理套件（dm-server governance init @ bd3a7de57a3a9e1a4495b94c5fa5c08d71e7be8c）生成并作为引导配置提交（执行器运行前必须存在有效配置；沿用 DM 采纳先例）；其余文件由实施单元生成。
+
+## Agent 实施收据
+
+| 单元 ID | 预期范围 | 角色 | 候选顺序 | attempt | CLI/模型 | 终态 | transport | fallback | 源码外收据路径 |
+|---|---|---|---|---|---|---|---|---|---|
+| governance-adoption-Agent-Squad-r1 | 13 个允许路径（治理文件 + 两个 gitlink） |  | implementation | cursor→antigravity→pi | run 71696608（dm-server pinned，attempts=1） | cursor / not-observable | failed-unknown | 无（apply 快进） | 否 | /tmp/gov-receipt-Agent-Squad-r1.json |
+| 对账（r1） | 规范 worktree 仍为 2a2a1fd 且干净；无 transport ref | — | — | — | — | — | 已对账；全新运行 r2（单元包追加禁止固定共享 /tmp 路径约束） | — | — | — |
+| governance-adoption-Agent-Squad-r2 | 13 个允许路径（治理文件 + 两个 gitlink） | b83905a8 | implementation | cursor→antigravity→pi | run 145e30ce（dm-server 42e675f，attempts=1） | cursor / not-observable | success | 468878f62ede（apply 快进） | 否 | /tmp/gov-receipt-Agent-Squad-r2.json |
+| （无效启动 repair-r0） | — | repair | — | 无 attempt | — | invalid-config（单元包尚未写出即启动，执行器读取失败；无 attempt、无副作用，worktree 仍为 e29284e 且干净） | 无 | 否 | /tmp/gov-receipt-Agent-Squad-repair-r0-invalid-unit-path.json |
+| governance-adoption-Agent-Squad-repair-r1 | policy/guidelines-adoption.md、PROJECT.md、deps/agent-native、deps/agent-intent | repair | cursor→antigravity→pi | run fadb9c92（attempts=1） | cursor / not-observable | success | d0f9081f47eb（apply 快进） | 否 | /tmp/gov-receipt-Agent-Squad-repair-r1.json |
+
+## 同步记录
+
+| 节点 | sync_base_sha | tested_head_sha | checked_at | 结果 |
+|---|---|---|---|---|
+| 启动 | 490b0b2dfa1567cb16ff7a7230fd78e52bc2938e | 490b0b2dfa1567cb16ff7a7230fd78e52bc2938e | 2026-10-17T07:17:41Z | 一致 |
+
+## 结果与验证（完成后填写）
+
+- 结果：治理套件已采纳——Guidelines 与 Development-Manager gitlink 固定、AGENT-EXECUTION.md 四条路由、共享执行器薄入口、pre-commit admit、governance CI；采纳状态 active（随合入生效）。
+- 验证：见下表（命令输出由 /tmp/gov-neg/ev.sh 写入 .task/evidence/governance-adoption/，SELF 除非另注）。
+- 影响与剩余项：不改业务代码；本地启用钩子需由人执行 `git config core.hooksPath scripts/hooks`；将 governance 检查设为分支保护必需项属用户双钥匙事项（命令见 Agent-System-Dev 治理主记录）。
+
+| 时间(UTC) | 命令 | 被测 HEAD | exit | 来源 | 摘要/位置 |
+|---|---|---|---|---|---|
+| 20261017T080150Z | `scripts/dm governance check --repo .` | 468878f62ede | 0 | SELF | ok fixtures；[输出](../evidence/governance-adoption/00-20261017T080150Z.log) sha256:b07d207012a2fd6d |
+| 20261017T080150Z | `scripts/agent-exec validate --repo .` | 468878f62ede | 0 | SELF | validate: ok；[输出](../evidence/governance-adoption/01-20261017T080150Z.log) sha256:b3a91246221bb5f8 |
+| 20261017T080150Z | `python3 /tmp/gov-neg/adopt_cases.py .` | 468878f62ede | 0 | SELF | [verdict] ALL PASS cases=5；[输出](../evidence/governance-adoption/02-20261017T080150Z.log) sha256:109906d5e57fe039 |
+| 20261017T094306Z | `scripts/dm governance check --repo .` | d0f9081f47eb | 0 | SELF | ok fixtures；[输出](../evidence/governance-adoption/03-20261017T094306Z.log) sha256:b904686133da5f80 |
+| 20261017T094306Z | `scripts/agent-exec validate --repo .` | d0f9081f47eb | 0 | SELF | validate: ok；[输出](../evidence/governance-adoption/04-20261017T094306Z.log) sha256:19dc223efda318d6 |
+| 20261017T094308Z | `scripts/submodule-consistency.sh` | 05b509827274 | 0 | SELF | ok deps/agent-intent 651a2e1e03c4326888ef9f742ce592140ae30ab8；[输出](../evidence/governance-adoption/05-20261017T094308Z.log) sha256:bc528fe029ed41d2 |
+| — | 候选 HEAD CI | b3e7da4a11b9 | 0 | GATE | success ci / docker-checkout-hostgateway (pull_request)=success（/weichen/Agent-Squad/actions/runs/3672/jobs/5787）；governance / governance (pull_request)=success（/weichen/Agent-Squad/actions/runs/3673/jobs/5788）；ci / vet-and-test (pull_request)=success（/weichen/Agent-Squad/actions/runs/3672/jobs/5786） |
+| — | 独立 Review | b3e7da4a11b9 | 0 | REPLAY | cursor-agent-stable --mode ask 全新只读会话（fresh_session=true，session id not-observable）闭合重审 VERDICT: APPROVE（对象 b3e7da4） |
@@ -0,0 +1,19 @@
+command: scripts/dm governance check --repo . 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: 468878f62eded8347d18a1d6c1b7c5f7e24a9c4c
+exit: 0
+utc: 20261017T080150Z
+---
+ok gitlink deps/development-guidelines
+ok gitlink deps/development-manager
+ok agent-execution
+ok file AGENTS.md
+ok file PROJECT.md
+ok file policy/guidelines-adoption.md
+ok file .task/templates/task-card.md
+ok executable scripts/agent-exec
+ok executable scripts/hooks/pre-commit
+ok executable scripts/dm
+ok executable scripts/task
+ok adoption status
+ok fixtures
@@ -0,0 +1,7 @@
+command: scripts/agent-exec validate --repo . 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: 468878f62eded8347d18a1d6c1b7c5f7e24a9c4c
+exit: 0
+utc: 20261017T080150Z
+---
+validate: ok
@@ -0,0 +1,13 @@
+command: python3 /tmp/gov-neg/adopt_cases.py . 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: 468878f62eded8347d18a1d6c1b7c5f7e24a9c4c
+exit: 0
+utc: 20261017T080150Z
+---
+[object] Agent-Squad-wt-feature-governance-adoption HEAD=468878f62eded8347d18a1d6c1b7c5f7e24a9c4c
+PASS positive-run-apply: run exit=0 terminal=success apply exit=0 head=312d80d409a1
+PASS missing-config: run exit=2 terminal=invalid-config validate exit=2
+PASS bad-receipt: apply exit=4 head_unchanged=True msg=apply: no local integrity record: changed paths do not match the transport diff
+PASS out-of-scope-path: run exit=5 terminal=scope-expansion transport=False
+PASS unknown-side-effect: run exit=4 terminal=failed-unknown transport=False
+[verdict] ALL PASS cases=5
@@ -0,0 +1,19 @@
+command: scripts/dm governance check --repo . 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: d0f9081f47eb59565323f731cb8085e5891dd7ba
+exit: 0
+utc: 20261017T094306Z
+---
+ok gitlink deps/development-guidelines
+ok gitlink deps/development-manager
+ok agent-execution
+ok file AGENTS.md
+ok file PROJECT.md
+ok file policy/guidelines-adoption.md
+ok file .task/templates/task-card.md
+ok executable scripts/agent-exec
+ok executable scripts/hooks/pre-commit
+ok executable scripts/dm
+ok executable scripts/task
+ok adoption status
+ok fixtures
@@ -0,0 +1,7 @@
+command: scripts/agent-exec validate --repo . 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: d0f9081f47eb59565323f731cb8085e5891dd7ba
+exit: 0
+utc: 20261017T094306Z
+---
+validate: ok
@@ -0,0 +1,14 @@
+command: scripts/submodule-consistency.sh 
+cwd: /home/weichen/projects/Agent-Squad-wt-feature-governance-adoption
+head: 05b5098272741987767efa2e78b8fe01aa84b6e5
+exit: 0
+utc: 20261017T094308Z
+---
+ok development-guidelines@deps/agent-native 2688fd2be6a52c0bf110e7e6366b6f291ab5e4eb
+ok error-code-system@deps/agent-native a3ee67bba269e1276c14e22c026171ecf2cf6190
+ok logging-system@deps/agent-native 49b5d8be604d05e4246dffa0d10fa1c2b7ad45d7
+ok development-guidelines@deps/agent-intent 2688fd2be6a52c0bf110e7e6366b6f291ab5e4eb
+ok error-code-system@deps/agent-intent a3ee67bba269e1276c14e22c026171ecf2cf6190
+ok logging-system@deps/agent-intent 49b5d8be604d05e4246dffa0d10fa1c2b7ad45d7
+ok deps/agent-native 14d48697a69ddba9ebc67ba7daa588f1b082d096
+ok deps/agent-intent 651a2e1e03c4326888ef9f742ce592140ae30ab8
@@ -0,0 +1,66 @@
+# <任务标题> 任务卡
+
+任务主记录：<Issue/PR 链接，或注明"本文件随 PR 描述维护">。
+状态：<进行中 / 待审核 / 已合入 / 已放弃>；owner：<填写>。
+
+## 基线与授权
+
+- 权威仓：<仓 URL 或逻辑 ID>；root_target：main；immediate_target：<main 或 feature/initiative>
+- branch_kind：<initiative | child | standalone>；initiative_id：<稳定 ID，普通任务为空>
+- R0（initial_main_sha）：<完整 SHA>；工作区：<路径>（worktree / clone）；分支：<feature|fix|hotfix/...>
+- 父任务信息（仅子任务填写）：parent_branch=<父分支>；parent_head_sha=<SHA>；parent_pr=<父 PR>
+- accepting_requirements：<true | false>（仅 initiative 可为 true；冻结后为 false）
+- 需求分流：verdict=<JOIN | NEW>；selected_initiative=<分支或无>；候选=<分支@HEAD 列表或无>；理由=<不适配/适配依据>；triage 收据=<源码外路径与摘要>
+- 授权与交付终点：<本次允许的动作与终点；未授权事项写明>
+- 实施执行：agent-per-unit（配置：AGENT-EXECUTION.md；外部 Agent 一单元一新进程/会话/工作区）
+
+## 目标与非目标
+
+<预期结果、受影响入口/模块/契约；明确不做什么。>
+
+## Agent 实施收据
+
+按根目录 AGENT-EXECUTION.md 路由；外部 Agent 一单元一新进程/会话/隔离工作区。实施、修复与测试必须通过登记的候选执行；未配置或候选安全耗尽停在 unavailable/blocked，禁止 host 隐式代写。
+
+| 单元 ID | 预期范围 | 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 | — |
+
+## 复用材料（task-start 自动检索结果摘录）
+
+- 相关近期提交：<git log -- <paths> 命中；无则写无>。
+- 相关经验条目：<检索命中；无则写无>。
+
+## 同步记录
+
+| 节点 | sync_base_sha | tested_head_sha | checked_at | 结果 |
+|---|---|---|---|---|
+| 启动 | <R0> | <HEAD> | <日期时间> | <一致/已同步> |
+| 验收前 | <填写> | <填写> | <填写> | <填写> |
+
+字段没有发生的写"无"，不省略。本卡与 PR 描述同源：开 PR 时将"目标与非目标、结果与验证、影响与剩余项"复制进 PR 描述，不另写第二份。
+
+## 交付调度（有多个就绪任务时）
+
+- 调度记录 / entry_id / 当前 attempt_id：<平台或调度服务链接及标识；未入队写无>。
+- 入队源 HEAD / owner / 依赖：<完整 SHA、身份、依赖或无>。
+- 固定候选引用：<调度尝试中的 base/head、组件与配置；未固定写无>。
+- 窗口、最终 CI、Review 与真实合入结果：<引用调度记录及原始证据，不复制全队列>。
+
+定义见 [合入调度与 CI 分配](../standards/merge-queue.md)。本段在候选固定前整理；固定后先使用源码之外的记录保存运行结果，再按项目闭合流程回填。
+
+## 结果与验证（完成后填写）
+
+- 结果：<交付了什么；未发生写无>。
+- 验证：<结论性摘要：PASS/FAIL/SKIP/NOT_RUN 逐项；逐条证据见下表，来源标注与写入规则见 [按影响选择验证与证据](../standards/verification.md)>。
+- 影响与剩余项：<影响面、遗留风险、后续任务>。
+
+| 时间(UTC) | 命令 | 被测 HEAD | exit | 来源 | 摘要/位置 |
+|---|---|---|---|---|---|
+| <时间> | <命令；由项目 evidence 工具写入，无工具时附命令+对象 SHA+输出 digest 并标 SELF> | <HEAD> | <exit> | GATE/REPLAY/SELF | <摘要/位置> |
+
+## 回收（task-finish）
+
+- 合入核对：<分支已包含于目标分支的核对输出>。
+- 未跟踪/ignored 文件处置：<归档/可再生/存疑 逐项；无则写无>。
+- 回收动作：<worktree remove、分支删除、prune；放弃路径注明原因>。
@@ -0,0 +1,73 @@
+# Project Agent Execution Configuration
+# 项目工程 Agent 执行路由配置
+
+适用范围：定义本项目中由 AI / Agent 主导的代码实施、修复、测试与审阅所调用的 Agent CLI 路由。
+解析器识别正文中唯一一个 ```agent-execution 代码块。正文外的说明文字可根据项目需要调整。
+
+通用硬规则见 Development-Guidelines 的 [Agent 执行与协作契约](deps/development-guidelines/standards/agent-execution.md)。
+本配置为项目本地路由权威，声明具体的候选 profile、尝试次数上限、超时、集成入口与豁免路径；
+代码块须通过 `deps/development-guidelines/templates/schemas/agent-execution.schema.json` 校验后才能登记为 active。
+`fallback_to_host` 只能为 `never`；旧值 `all-unavailable` 仅供迁移工具读取，不能通过准入。
+下例的 CLI 名称、命令与路径均为占位，按项目实际执行器替换。
+
+```agent-execution
+schema: agent-execution/v1
+
+execution:
+  fresh_process_per_unit: true
+  permission: run-everything
+  fallback_to_host: never
+  max_parallel: 3
+  timeout_seconds: 7200
+  idle_timeout_seconds: 600
+  summary_bytes: 16384
+  raw_output_bytes: 10485760
+  max_attempts_per_profile: 3
+  emergency_profile: codex
+
+roles:
+  implementation: [cursor, antigravity, pi]
+  repair: [cursor, antigravity, pi]
+  test: [pi, antigravity, cursor]
+  review: [pi, antigravity, cursor, codex]
+
+profiles:
+  cursor:
+    adapter: cursor
+    executable: "cursor-agent-stable"
+    model: auto
+    # 2026-10-21 临时禁用：U3B 连续两次 fail-stop（首次 TLS 会话前故障 553B 日志、
+    # 二次 "Cannot use this model: auto"，与 Backend 配置已记录的月配额耗尽一致），
+    # 两次均未建立会话、零改动（现场保留）。恢复时改回 enabled: true。
+    enabled: false
+
+  antigravity:
+    adapter: antigravity
+    executable: agy
+    model: auto
+    # 2026-10-21 临时禁用：U3B 尝试被 400 FAILED_PRECONDITION "User location is
+    # not supported for the API use" 拒绝，0 token、未建立会话（现场保留）；
+    # 与 Backend 配置已记录的 Gemini 端点不稳一致。恢复时改回 enabled: true。
+    enabled: false
+
+  pi:
+    adapter: pi
+    executable: pi
+    model: auto
+    enabled: true
+
+  codex:
+    adapter: codex
+    executable: codex
+    model: auto
+    enabled: true
+
+integration:
+  exempt_paths: [".task/**", "docs/superpowers/specs/**", "docs/superpowers/plans/**"]
+  transport:
+    command: "scripts/agent-exec apply"
+  admission:
+    command: "scripts/agent-exec admit"
+  cleanup:
+    command: "scripts/agent-exec cleanup"
+```
@@ -34,3 +34,21 @@
 - **SSH key**：`~/.ssh/home_gitea`（ED25519，公钥指纹
  `SHA256:GYXaOXIMeloQ1fKSfdJ/ndgPQzs/Ix7Hk62ekfkfQ20`，Gitea 侧密钥名 GMKtec）；
  `~/.ssh/config` 已有 `git-ssh.u1680832.nyat.app` 主机条目，克隆/推送直接可用。
+
+## Agent 执行与新项目准入（强制）
+
+1. 缺 Guidelines 固定版本（`deps/development-guidelines` gitlink 未就绪或 HEAD 不符）时，Agent 只能只读分析，不得实施。
+2. 缺有效 `AGENT-EXECUTION.md`（`scripts/dm governance check` 未通过）时禁止 AI 实施。
+3. 缺共享执行器（`scripts/agent-exec` 或 `deps/development-manager` 不可用）时，状态为 `blocked/unavailable`，不得改由 host 替代。
+4. 不允许 host 代写实施单元；实施、修复与测试必须经 `AGENT-EXECUTION.md` 登记的候选执行。
+5. 人类直接开发按其自身授权与项目规则进行，不受本条款约束。
+6. 分支保护与规范 tag 由用户负责设置与维护，Agent 不得代为设置。
+
+入口：
+
+- 需求分流：`scripts/task triage discover|decide|verify`
+- 任务启动：`scripts/task start …`
+- PR 检查：`scripts/task pr-check …`
+- Agent 执行：`scripts/agent-exec run|apply|admit|validate`
+
+AI 实施单元只经 `scripts/agent-exec run` 按 AGENT-EXECUTION.md 运行，并经 `scripts/agent-exec apply` 集成。
@@ -12,8 +12,8 @@
 - 提交/审核/合并/发布身份与权限：仓库所有者 weichen。
 - 任务唯一来源：Gitea `weichen/Agent-Squad` 的 Issue 与 Pull Request。
 - 规范性文档目录：`policy/`。
- 过程文档仓/目录：`docs/records/`（项目内豁免，见采纳记录）。
- 交接记录位置：随任务主记录；需要落盘的交接放入 `docs/records/`。
+- 过程文档仓/目录：任务卡与交付证据在 `.task/`（模板 `.task/templates/task-card.md`）；其他过程记录在 `docs/records/`（项目内豁免，见采纳记录）。
+- 交接记录位置：随任务主记录；需要落盘的任务交接与交付证据放入 `.task/`，其他落盘交接放入 `docs/records/`。
 - 凭证查找入口及所需字段：LLM 使用环境变量 `AGENT_SQUAD_LLM_BASE_URL`、`AGENT_SQUAD_LLM_API_KEY`、`AGENT_SQUAD_LLM_MODEL`。日志采集 token 使用日志库自己的配置。文档与仓库不记录这些值。

 ## 依赖组合
@@ -45,6 +45,7 @@ git clone --recurse-submodules ssh://git@git-ssh.u1680832.nyat.app:31411/weichen
 | 模块/接口 | 仓库根：`GOWORK=off go vet ./... && GOWORK=off go test -count=1 ./...` | 需要已初始化的 Agent-Native 子模块；默认 stub，不调用真实模型 |
 | 数据/权限/恢复 | conformance 与 store 测试覆盖运行落盘、HITL 恢复 | 无共享数据库 |
 | UI/执行后端 | `GOWORK=off go test -count=1 ./server/... ./conformance/...` | 不覆盖浏览器；SSE 由测试客户端验证 |
+| 治理与 Agent 执行配置 | 仓库根：`scripts/agent-exec validate --repo .` 与 `scripts/dm governance check --repo .` | 需 `deps/development-manager` 已按 gitlink 检出（首次 `dm` 调用会构建 dm-server）；只核验 AGENT-EXECUTION.md 与治理文件，不证明实施单元正确 |

 ## 环境与数据

@@ -0,0 +1,179 @@
+// Package a2a 把小队接入 A2A（Agent2Agent）1.0 互通面（Q3 WP4）：
+//   - 发布侧：BuildCard 从角色注册表与蓝图注册表渲染 Agent Card（描述本
+//     团队能力），CardHandler 以 JSON 服务（挂载在
+//     /.well-known/agent-card.json）；
+//   - 消费侧：FetchCard 发现远端 Agent Card（fail-closed 校验必备字段），
+//     Executor 实现 squad.Executor——message/send 发起任务，tasks/get 轮询
+//     A2A 任务状态机（submitted/working/input-required → completed/failed/
+//     canceled），终态映射 ExecResult：completed 的 artifacts（TextPart/
+//     FilePart）落成 patches/目标文件；failed/canceled/input-required 一律
+//     Success=false（跨进程挂起不在本适配面内，fail-closed 交引擎重试/降级
+//     语义处理）。
+//
+// 传输为 JSON-RPC 2.0 over HTTP（A2A 1.0 形态）；不引入协议之外的依赖。
+package a2a
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"agent-squad/squad"
+)
+
+// ProtocolVersion 对齐的 A2A 协议版本。
+const ProtocolVersion = "1.0"
+
+// WellKnownPath 是 Agent Card 的发现路径（A2A 1.0）。
+const WellKnownPath = "/.well-known/agent-card.json"
+
+// AgentCard 是 A2A 1.0 Agent Card（发布与消费共用形态）。
+type AgentCard struct {
+	ProtocolVersion string `json:"protocolVersion"`
+	Name            string `json:"name"`
+	Description     string `json:"description,omitempty"`
+	// URL 是 JSON-RPC 服务端点（消费侧据此发 message/send）。
+	URL         string   `json:"url"`
+	Version     string   `json:"version"`
+	Capabilities CardCapabilities `json:"capabilities"`
+	DefaultInputModes  []string `json:"defaultInputModes,omitempty"`
+	DefaultOutputModes []string `json:"defaultOutputModes,omitempty"`
+	Skills      []AgentSkill `json:"skills,omitempty"`
+}
+
+// CardCapabilities 能力声明（本适配面：非流式、无推送、无状态历史）。
+type CardCapabilities struct {
+	Streaming              bool `json:"streaming"`
+	PushNotifications      bool `json:"pushNotifications"`
+	StateTransitionHistory bool `json:"stateTransitionHistory"`
+}
+
+// AgentSkill 是一项可被远端调用的能力。
+type AgentSkill struct {
+	ID          string   `json:"id"`
+	Name        string   `json:"name"`
+	Description string   `json:"description,omitempty"`
+	Tags        []string `json:"tags,omitempty"`
+	Examples    []string `json:"examples,omitempty"`
+}
+
+// CardOptions 覆盖 BuildCard 的身份字段（空则用缺省）。
+type CardOptions struct {
+	Name        string
+	Description string
+	// URL JSON-RPC 端点（必填：远端没有端点就无法调用本团队）。
+	URL string
+	// Version 卡片版本（空 = "1.0.0"）。
+	Version string
+	// Skills 显式能力清单；空 = 从注册表派生。
+	Skills []AgentSkill
+}
+
+// BuildCard 从角色注册表与蓝图注册表渲染 Agent Card（描述本团队能力）。
+// 角色 → 技能（关键词作 tags）；蓝图 → 技能（match 词作 examples）。
+func BuildCard(reg *squad.RoleRegistry, bps *squad.BlueprintRegistry, opts CardOptions) (AgentCard, error) {
+	if strings.TrimSpace(opts.URL) == "" {
+		return AgentCard{}, fmt.Errorf("a2a: Agent Card 需要 JSON-RPC 端点 URL")
+	}
+	card := AgentCard{
+		ProtocolVersion: ProtocolVersion,
+		Name:            orDefault(opts.Name, "agent-squad"),
+		Description:     opts.Description,
+		URL:             opts.URL,
+		Version:         orDefault(opts.Version, "1.0.0"),
+		DefaultInputModes:  []string{"text/plain"},
+		DefaultOutputModes: []string{"text/plain"},
+		Skills:          append([]AgentSkill(nil), opts.Skills...),
+	}
+	if len(card.Skills) == 0 {
+		if reg != nil {
+			for _, role := range reg.List() {
+				desc := firstLine(role.Instructions)
+				if desc == "" {
+					desc = role.Name
+				}
+				card.Skills = append(card.Skills, AgentSkill{
+					ID: role.Role, Name: orDefault(role.Name, role.Role),
+					Description: desc, Tags: append([]string(nil), role.Keywords...),
+				})
+			}
+		}
+		if bps != nil {
+			for _, bp := range bps.List() {
+				card.Skills = append(card.Skills, AgentSkill{
+					ID: bp.Key, Name: orDefault(bp.Title, bp.Key),
+					Description: bp.Description, Examples: append([]string(nil), bp.Match...),
+				})
+			}
+		}
+	}
+	return card, nil
+}
+
+// CardHandler 以 JSON 服务 Agent Card（调用方挂载到 WellKnownPath）。
+func CardHandler(card AgentCard) http.Handler {
+	body, _ := json.Marshal(card)
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write(body)
+	})
+}
+
+// FetchCard 发现并校验远端 Agent Card（fail-closed：必备字段缺失即拒）。
+func FetchCard(client *http.Client, wellKnownURL string) (AgentCard, error) {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	resp, err := client.Get(wellKnownURL)
+	if err != nil {
+		return AgentCard{}, fmt.Errorf("a2a: 获取 Agent Card: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return AgentCard{}, fmt.Errorf("a2a: Agent Card 端点返回 %d", resp.StatusCode)
+	}
+	var card AgentCard
+	if err := json.NewDecoder(resp.Body).Decode(&card); err != nil {
+		return AgentCard{}, fmt.Errorf("a2a: Agent Card 解析失败: %w", err)
+	}
+	if err := validateCard(card); err != nil {
+		return AgentCard{}, err
+	}
+	return card, nil
+}
+
+// validateCard 消费侧最小校验（A2A 1.0 必备字段）。
+func validateCard(card AgentCard) error {
+	if card.ProtocolVersion == "" {
+		return fmt.Errorf("a2a: Agent Card 缺 protocolVersion")
+	}
+	if strings.TrimSpace(card.Name) == "" {
+		return fmt.Errorf("a2a: Agent Card 缺 name")
+	}
+	if strings.TrimSpace(card.URL) == "" {
+		return fmt.Errorf("a2a: Agent Card 缺 url")
+	}
+	if strings.TrimSpace(card.Version) == "" {
+		return fmt.Errorf("a2a: Agent Card 缺 version")
+	}
+	return nil
+}
+
+func orDefault(v, def string) string {
+	if strings.TrimSpace(v) != "" {
+		return v
+	}
+	return def
+}
+
+func firstLine(s string) string {
+	s = strings.TrimSpace(s)
+	if i := strings.IndexByte(s, '\n'); i >= 0 {
+		s = s[:i]
+	}
+	if len(s) > 120 {
+		s = s[:120] + "…"
+	}
+	return s
+}
@@ -0,0 +1,345 @@
+package a2a
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"agent-squad/squad"
+)
+
+// conformance_test.go — Q3 WP4 验收：双向 conformance（本地假 A2A 服务端）。
+// 发布侧：BuildCard + CardHandler 在 well-known 路径服务合法 Agent Card；
+// 消费侧：FetchCard 发现校验 + Executor 走完任务状态机
+// （submitted→working→completed / failed / input-required），终态对齐
+// squad.ExecResult。
+
+// fakeA2AServer 最小 A2A 1.0 服务端：well-known 卡片 + JSON-RPC
+// （message/send 建任务；tasks/get 按脚本推进状态）。
+type fakeA2AServer struct {
+	mu     sync.Mutex
+	card   AgentCard
+	tasks  map[string]*Task
+	plan   []string // 每次 tasks/get 推进到的状态序列（走完停在末态）
+	calls  map[string]int
+	artifacts []Artifact
+	statusMsg string
+}
+
+func newFakeServer(t *testing.T, card AgentCard, plan []string, artifacts []Artifact) *fakeA2AServer {
+	t.Helper()
+	s := &fakeA2AServer{card: card, tasks: map[string]*Task{}, plan: plan,
+		calls: map[string]int{}, artifacts: artifacts}
+	return s
+}
+
+func (s *fakeA2AServer) handler() http.Handler {
+	mux := http.NewServeMux()
+	mux.HandleFunc(WellKnownPath, func(w http.ResponseWriter, r *http.Request) {
+		s.mu.Lock()
+		body, _ := json.Marshal(s.card)
+		s.mu.Unlock()
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write(body)
+	})
+	mux.HandleFunc("/rpc", func(w http.ResponseWriter, r *http.Request) {
+		var req rpcRequest
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			writeRPCError(w, 0, -32700, "parse error")
+			return
+		}
+		s.mu.Lock()
+		defer s.mu.Unlock()
+		s.calls[req.Method]++
+		switch req.Method {
+		case "message/send":
+			id := fmt.Sprintf("task_%d", len(s.tasks)+1)
+			state := StateSubmitted
+			if len(s.plan) > 0 {
+				state = s.plan[0]
+			}
+			task := &Task{Kind: "task", ID: id, ContextID: "ctx_1",
+				Status: TaskStatus{State: state, Message: s.statusMessage()}}
+			s.tasks[id] = task
+			writeRPCResult(w, req.ID, task)
+		case "tasks/get":
+			var params struct {
+				ID string `json:"id"`
+			}
+			raw, _ := json.Marshal(req.Params)
+			_ = json.Unmarshal(raw, &params)
+			task, ok := s.tasks[params.ID]
+			if !ok {
+				writeRPCError(w, req.ID, -32001, "task not found")
+				return
+			}
+			// 按脚本推进状态机（停留在终态）。
+			if !isTerminal(task.Status.State) && len(s.plan) > 0 {
+				n := s.calls["tasks/get#"+params.ID]
+				s.calls["tasks/get#"+params.ID] = n + 1
+				idx := n + 1
+				if idx >= len(s.plan) {
+					idx = len(s.plan) - 1
+				}
+				task.Status.State = s.plan[idx]
+				task.Status.Message = s.statusMessage()
+				if task.Status.State == StateCompleted {
+					task.Artifacts = s.artifacts
+				}
+			}
+			writeRPCResult(w, req.ID, task)
+		default:
+			writeRPCError(w, req.ID, -32601, "method not found")
+		}
+	})
+	return mux
+}
+
+func (s *fakeA2AServer) statusMessage() *Message {
+	if s.statusMsg == "" {
+		return nil
+	}
+	return &Message{Kind: "message", Role: "agent",
+		Parts: []Part{{Kind: "text", Text: s.statusMsg}}}
+}
+
+func writeRPCResult(w http.ResponseWriter, id int64, result any) {
+	body, _ := json.Marshal(result)
+	env, _ := json.Marshal(rpcResponse{JSONRPC: "2.0", ID: id, Result: body})
+	w.Header().Set("Content-Type", "application/json")
+	_, _ = w.Write(env)
+}
+
+func writeRPCError(w http.ResponseWriter, id int64, code int, msg string) {
+	env, _ := json.Marshal(rpcResponse{JSONRPC: "2.0", ID: id, Error: &rpcError{Code: code, Message: msg}})
+	w.Header().Set("Content-Type", "application/json")
+	_, _ = w.Write(env)
+}
+
+// ---- 发布侧 ----
+
+func TestBuildCardFromRegistries(t *testing.T) {
+	reg := squad.NewRoleRegistry()
+	for _, r := range []squad.RoleSpec{
+		{Role: "researcher", Name: "调研员", Keywords: []string{"调研", "检索"},
+			Instructions: "围绕目标调研关键问题。\n第二行不进摘要。"},
+		{Role: "writer", Name: "撰稿员", Keywords: []string{"撰写"}},
+	} {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	bps := squad.NewBlueprintRegistry()
+	if err := bps.Register(squad.Blueprint{Key: "brief", Title: "简报",
+		Description: "生成简报", Match: []string{"简报", "brief"},
+		Steps: []squad.StepSpec{{ID: "s1", Role: "writer", Title: "成文"}}}); err != nil {
+		t.Fatal(err)
+	}
+	card, err := BuildCard(reg, bps, CardOptions{
+		Name: "示例小队", Description: "演示团队能力", URL: "https://team.example/rpc",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if card.ProtocolVersion != ProtocolVersion || card.Version != "1.0.0" {
+		t.Fatalf("card = %+v", card)
+	}
+	byID := map[string]AgentSkill{}
+	for _, s := range card.Skills {
+		byID[s.ID] = s
+	}
+	if byID["researcher"].Name != "调研员" || byID["researcher"].Description != "围绕目标调研关键问题。" {
+		t.Fatalf("role skill = %+v", byID["researcher"])
+	}
+	if len(byID["researcher"].Tags) != 2 {
+		t.Fatalf("tags = %v", byID["researcher"].Tags)
+	}
+	if byID["brief"].Name != "简报" || strings.Join(byID["brief"].Examples, ",") != "简报,brief" {
+		t.Fatalf("blueprint skill = %+v", byID["brief"])
+	}
+	if _, err := BuildCard(reg, bps, CardOptions{}); err == nil {
+		t.Fatal("缺 URL 必须 fail-closed")
+	}
+}
+
+// CardHandler 以 JSON 服务卡片（发布端点形态）。
+func TestCardHandlerServesJSON(t *testing.T) {
+	card := AgentCard{ProtocolVersion: ProtocolVersion, Name: "小队", URL: "https://x/rpc", Version: "1.0.0"}
+	ts := httptest.NewServer(CardHandler(card))
+	defer ts.Close()
+	resp, err := http.Get(ts.URL + WellKnownPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	if ct := resp.Header.Get("Content-Type"); ct != "application/json" {
+		t.Fatalf("content-type = %s", ct)
+	}
+	var got AgentCard
+	if err := json.NewDecoder(resp.Body).Decode(&got); err != nil {
+		t.Fatal(err)
+	}
+	if got.Name != "小队" || got.ProtocolVersion != ProtocolVersion {
+		t.Fatalf("card = %+v", got)
+	}
+}
+
+// ---- 消费侧：发现 + 状态机 ----
+
+func TestFetchCardAndExecuteCompleted(t *testing.T) {
+	fileBytes := base64.StdEncoding.EncodeToString([]byte("notes 内容"))
+	artifacts := []Artifact{{
+		ArtifactID: "art1", Name: "调研结果",
+		Parts: []Part{
+			{Kind: "text", Text: "# 调研结论\n\n结论正文，长度足够作为交付物内容。"},
+			{Kind: "file", File: &FilePart{Name: "notes.md", Bytes: fileBytes, MimeType: "text/markdown"}},
+		},
+	}}
+	card := AgentCard{ProtocolVersion: ProtocolVersion, Name: "远端小队",
+		URL: "PLACEHOLDER", Version: "1.0.0"}
+	srv := newFakeServer(t, card, []string{StateWorking, StateWorking, StateCompleted}, artifacts)
+	ts := httptest.NewServer(srv.handler())
+	defer ts.Close()
+	srv.mu.Lock()
+	srv.card.URL = ts.URL + "/rpc"
+	srv.mu.Unlock()
+
+	// 发现：well-known 路径取卡并校验。
+	got, err := FetchCard(ts.Client(), ts.URL+WellKnownPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Name != "远端小队" || got.URL != ts.URL+"/rpc" {
+		t.Fatalf("card = %+v", got)
+	}
+
+	// 执行：submitted/working → completed，工件映射 patches。
+	var states []string
+	ex := NewExecutor(got)
+	ex.Client = ts.Client()
+	ex.PollInterval = 5 * time.Millisecond
+	ex.OnState = func(s string) { states = append(states, s) }
+	res, err := ex.Execute(context.Background(), squad.ExecRequest{
+		Prompt: "调研主题 X", Targets: []string{"output.md"},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !res.Success {
+		t.Fatalf("result = %+v", res)
+	}
+	if string(res.Patches["notes.md"]) != "notes 内容" {
+		t.Fatalf("file part 未落成 patch：%q", res.Patches["notes.md"])
+	}
+	if !strings.Contains(string(res.Patches["output.md"]), "调研结论") {
+		t.Fatalf("文本部件未写入目标：%q", res.Patches["output.md"])
+	}
+	joined := strings.Join(states, ",")
+	if !strings.Contains(joined, StateWorking) || !strings.HasSuffix(joined, StateCompleted) {
+		t.Fatalf("状态机迁移 = %v", states)
+	}
+	srv.mu.Lock()
+	sends := srv.calls["message/send"]
+	srv.mu.Unlock()
+	if sends != 1 {
+		t.Fatalf("message/send = %d", sends)
+	}
+}
+
+func TestExecuteMapsFailedAndInputRequired(t *testing.T) {
+	card := AgentCard{ProtocolVersion: ProtocolVersion, Name: "远端", URL: "PLACEHOLDER", Version: "1"}
+	srv := newFakeServer(t, card, []string{StateWorking, StateFailed}, nil)
+	srv.statusMsg = "远端执行崩溃"
+	ts := httptest.NewServer(srv.handler())
+	defer ts.Close()
+	srv.mu.Lock()
+	srv.card.URL = ts.URL + "/rpc"
+	srv.mu.Unlock()
+
+	ex := NewExecutor(card)
+	ex.BaseURL = ts.URL + "/rpc"
+	ex.Client = ts.Client()
+	ex.PollInterval = 5 * time.Millisecond
+	res, err := ex.Execute(context.Background(), squad.ExecRequest{Prompt: "x", Targets: []string{"output.md"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Success || !strings.Contains(res.ErrorReason, "failed") || !strings.Contains(res.ErrorReason, "远端执行崩溃") {
+		t.Fatalf("failed 映射 = %+v", res)
+	}
+
+	// input-required：fail-closed（跨进程人工输入不在适配面内）。
+	srv2 := newFakeServer(t, card, []string{StateInputRequired}, nil)
+	srv2.statusMsg = "请补充授权"
+	ts2 := httptest.NewServer(srv2.handler())
+	defer ts2.Close()
+	ex2 := &Executor{BaseURL: ts2.URL + "/rpc", Client: ts2.Client(), PollInterval: 5 * time.Millisecond}
+	res2, err := ex2.Execute(context.Background(), squad.ExecRequest{Prompt: "x", Targets: []string{"output.md"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Success || !strings.Contains(res2.ErrorReason, "input-required") {
+		t.Fatalf("input-required 映射 = %+v", res2)
+	}
+}
+
+// 直接回 Message（无任务态）→ 文本落目标。
+func TestExecuteDirectMessageResult(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var req rpcRequest
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		writeRPCResult(w, req.ID, Message{Kind: "message", Role: "agent",
+			Parts: []Part{{Kind: "text", Text: "直接回答正文"}}})
+	}))
+	defer ts.Close()
+	ex := &Executor{BaseURL: ts.URL, Client: ts.Client()}
+	res, err := ex.Execute(context.Background(), squad.ExecRequest{Prompt: "x", Targets: []string{"output.md"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !res.Success || !strings.Contains(string(res.Patches["output.md"]), "直接回答正文") {
+		t.Fatalf("res = %+v", res)
+	}
+}
+
+// 传输/协议错误 → error（引擎按瞬态评估），不吞成成功。
+func TestExecuteProtocolErrors(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		writeRPCError(w, 1, -32601, "method not found")
+	}))
+	defer ts.Close()
+	ex := &Executor{BaseURL: ts.URL, Client: ts.Client()}
+	if _, err := ex.Execute(context.Background(), squad.ExecRequest{Prompt: "x"}); err == nil ||
+		!strings.Contains(err.Error(), "协议错误") {
+		t.Fatalf("err = %v", err)
+	}
+	// ctx 取消传播。
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	slow := newFakeServer(t, AgentCard{ProtocolVersion: "1.0", Name: "x", URL: ts.URL, Version: "1"},
+		[]string{StateWorking}, nil)
+	ts2 := httptest.NewServer(slow.handler())
+	defer ts2.Close()
+	ex2 := &Executor{BaseURL: ts2.URL + "/rpc", Client: ts2.Client(), PollInterval: time.Second}
+	if _, err := ex2.Execute(ctx, squad.ExecRequest{Prompt: "x"}); err == nil {
+		t.Fatal("取消必须传播")
+	}
+}
+
+// FetchCard fail-closed：缺必备字段拒绝。
+func TestFetchCardValidates(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"protocolVersion":"1.0","name":"","url":"","version":""}`))
+	}))
+	defer ts.Close()
+	if _, err := FetchCard(ts.Client(), ts.URL); err == nil || !strings.Contains(err.Error(), "name") {
+		t.Fatalf("err = %v", err)
+	}
+}
@@ -0,0 +1,331 @@
+package a2a
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"sync/atomic"
+	"time"
+
+	"agent-squad/squad"
+)
+
+// executor.go — 远端智能体 Executor 客户端（A2A 任务状态机对齐 ExecResult）。
+
+// 任务状态（A2A 1.0 TaskState）。
+const (
+	StateSubmitted    = "submitted"
+	StateWorking      = "working"
+	StateInputRequired = "input-required"
+	StateCompleted    = "completed"
+	StateCanceled     = "canceled"
+	StateFailed       = "failed"
+	StateRejected     = "rejected"
+)
+
+// DefaultPollInterval 是 tasks/get 轮询缺省间隔。
+const DefaultPollInterval = 500 * time.Millisecond
+
+// Executor 把远端 A2A 智能体接成 squad.Executor：一次步骤执行 = 一个
+// A2A 任务（message/send 发起，tasks/get 轮询到终态）。
+type Executor struct {
+	// BaseURL 远端 JSON-RPC 端点（Agent Card 的 url）。
+	BaseURL string
+	// Client HTTP 客户端（nil = http.DefaultClient）。
+	Client *http.Client
+	// PollInterval 轮询间隔（0 = DefaultPollInterval）。
+	PollInterval time.Duration
+	// OnState 可选状态回调（观测任务状态机迁移）。
+	OnState func(state string)
+
+	callID atomic.Int64
+}
+
+// NewExecutor 按 Agent Card 构造远端执行器。
+func NewExecutor(card AgentCard) *Executor {
+	return &Executor{BaseURL: card.URL}
+}
+
+// Name 实现 squad.Executor。
+func (e *Executor) Name() string { return "a2a" }
+
+// ---- A2A 1.0 报文形态（JSON-RPC 2.0） ----
+
+type rpcRequest struct {
+	JSONRPC string `json:"jsonrpc"`
+	ID      int64  `json:"id"`
+	Method  string `json:"method"`
+	Params  any    `json:"params"`
+}
+
+type rpcError struct {
+	Code    int    `json:"code"`
+	Message string `json:"message"`
+}
+
+type rpcResponse struct {
+	JSONRPC string          `json:"jsonrpc"`
+	ID      int64           `json:"id"`
+	Result  json.RawMessage `json:"result,omitempty"`
+	Error   *rpcError       `json:"error,omitempty"`
+}
+
+// Part 是消息/工件的部件（text | file | data）。
+type Part struct {
+	Kind string     `json:"kind"`
+	Text string     `json:"text,omitempty"`
+	File *FilePart  `json:"file,omitempty"`
+	Data *json.RawMessage `json:"data,omitempty"`
+}
+
+// FilePart 文件部件：bytes（base64）或 uri 二选一。
+type FilePart struct {
+	Name     string `json:"name"`
+	Bytes    string `json:"bytes,omitempty"`
+	URI      string `json:"uri,omitempty"`
+	MimeType string `json:"mimeType,omitempty"`
+}
+
+// Message 是一条 A2A 消息。
+type Message struct {
+	Kind      string `json:"kind"`
+	Role      string `json:"role"`
+	Parts     []Part `json:"parts"`
+	MessageID string `json:"messageId"`
+}
+
+// TaskStatus 任务状态。
+type TaskStatus struct {
+	State    string   `json:"state"`
+	Message  *Message `json:"message,omitempty"`
+	Timestamp string  `json:"timestamp,omitempty"`
+}
+
+// Artifact 任务工件。
+type Artifact struct {
+	ArtifactID string `json:"artifactId,omitempty"`
+	Name       string `json:"name,omitempty"`
+	Parts      []Part `json:"parts"`
+}
+
+// Task 是 A2A 任务（状态机载体）。
+type Task struct {
+	Kind      string     `json:"kind"`
+	ID        string     `json:"id"`
+	ContextID string     `json:"contextId,omitempty"`
+	Status    TaskStatus `json:"status"`
+	Artifacts []Artifact `json:"artifacts,omitempty"`
+}
+
+// Execute 实现 squad.Executor：发起任务并轮询到终态，映射 ExecResult。
+// 传输/协议错误返回 error（引擎按瞬态失败评估）；任务级失败
+// （failed/canceled/rejected/input-required）返回 Success=false 的结果
+// （fail-closed：跨进程人工输入不在本适配面内）。
+func (e *Executor) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
+	if strings.TrimSpace(e.BaseURL) == "" {
+		return nil, fmt.Errorf("a2a: executor 缺 BaseURL")
+	}
+	prompt := strings.TrimSpace(req.Prompt)
+	if prompt == "" {
+		prompt = strings.TrimSpace(req.Instructions)
+	}
+	msg := Message{
+		Kind: "message", Role: "user",
+		Parts:     []Part{{Kind: "text", Text: prompt}},
+		MessageID: fmt.Sprintf("msg-%d-%d", time.Now().UnixNano(), e.callID.Add(1)),
+	}
+	raw, err := e.call(ctx, "message/send", map[string]any{"message": msg})
+	if err != nil {
+		return nil, err
+	}
+	// 服务端可直接回 Message（无任务态）或 Task。
+	var probe struct {
+		Kind string `json:"kind"`
+	}
+	_ = json.Unmarshal(raw, &probe)
+	if probe.Kind == "message" {
+		var m Message
+		if err := json.Unmarshal(raw, &m); err != nil {
+			return nil, fmt.Errorf("a2a: message 结果解析失败: %w", err)
+		}
+		return e.resultFromParts(m.Parts, req.Targets), nil
+	}
+	var task Task
+	if err := json.Unmarshal(raw, &task); err != nil {
+		return nil, fmt.Errorf("a2a: task 结果解析失败: %w", err)
+	}
+	e.observe(task.Status.State)
+	if shouldStop(task.Status.State) {
+		return e.resultFromTask(task, req.Targets), nil
+	}
+	// 轮询任务状态机直至终态（ctx 取消传播）。
+	interval := e.PollInterval
+	if interval <= 0 {
+		interval = DefaultPollInterval
+	}
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		case <-ticker.C:
+		}
+		raw, err := e.call(ctx, "tasks/get", map[string]any{"id": task.ID})
+		if err != nil {
+			return nil, err
+		}
+		var next Task
+		if err := json.Unmarshal(raw, &next); err != nil {
+			return nil, fmt.Errorf("a2a: tasks/get 解析失败: %w", err)
+		}
+		if next.Status.State != task.Status.State {
+			task = next
+			e.observe(task.Status.State)
+		} else {
+			task = next
+		}
+		if shouldStop(task.Status.State) {
+			return e.resultFromTask(task, req.Targets), nil
+		}
+	}
+}
+
+// call 发一次 JSON-RPC 调用并取 result。
+func (e *Executor) call(ctx context.Context, method string, params any) (json.RawMessage, error) {
+	client := e.Client
+	if client == nil {
+		client = http.DefaultClient
+	}
+	body, _ := json.Marshal(rpcRequest{
+		JSONRPC: "2.0", ID: e.callID.Add(1), Method: method, Params: params,
+	})
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, e.BaseURL, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("a2a: %s 传输失败: %w", method, err)
+	}
+	defer resp.Body.Close()
+	data, err := io.ReadAll(io.LimitReader(resp.Body, 8<<20))
+	if err != nil {
+		return nil, fmt.Errorf("a2a: %s 读取响应失败: %w", method, err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("a2a: %s 返回 HTTP %d", method, resp.StatusCode)
+	}
+	var env rpcResponse
+	if err := json.Unmarshal(data, &env); err != nil {
+		return nil, fmt.Errorf("a2a: %s 响应不是合法 JSON-RPC: %w", method, err)
+	}
+	if env.Error != nil {
+		return nil, fmt.Errorf("a2a: %s 协议错误 %d: %s", method, env.Error.Code, env.Error.Message)
+	}
+	return env.Result, nil
+}
+
+func (e *Executor) observe(state string) {
+	if e.OnState != nil {
+		e.OnState(state)
+	}
+}
+
+func isTerminal(state string) bool {
+	switch state {
+	case StateCompleted, StateFailed, StateCanceled, StateRejected:
+		return true
+	}
+	return false
+}
+
+// shouldStop 轮询停止条件：真终态，或 input-required（跨进程人工输入不在
+// 本适配面内，fail-closed 交引擎语义处理，不死等）。
+func shouldStop(state string) bool {
+	return isTerminal(state) || state == StateInputRequired
+}
+
+// resultFromTask 终态任务 → ExecResult。
+func (e *Executor) resultFromTask(task Task, targets []string) *squad.ExecResult {
+	switch task.Status.State {
+	case StateCompleted:
+		var parts []Part
+		for _, art := range task.Artifacts {
+			parts = append(parts, art.Parts...)
+		}
+		return e.resultFromParts(parts, targets)
+	case StateInputRequired:
+		return &squad.ExecResult{Success: false,
+			ErrorReason: "a2a: 远端要求人工输入（input-required）：" + statusText(task.Status)}
+	case StateCanceled:
+		return &squad.ExecResult{Success: false,
+			ErrorReason: "a2a: 远端任务被取消（canceled）：" + statusText(task.Status)}
+	case StateRejected:
+		return &squad.ExecResult{Success: false,
+			ErrorReason: "a2a: 远端拒绝任务（rejected）：" + statusText(task.Status)}
+	default: // failed 及未知终态一律 fail-closed
+		return &squad.ExecResult{Success: false,
+			ErrorReason: "a2a: 远端任务失败（" + task.Status.State + "）：" + statusText(task.Status)}
+	}
+}
+
+// resultFromParts 部件流 → patches：FilePart(bytes) 按名落 patch；
+// 文本部件拼接后写入全部声明目标（无文件工件时的交付物兜底，
+// 与其他后端「正文写目标文件」契约一致）。
+func (e *Executor) resultFromParts(parts []Part, targets []string) *squad.ExecResult {
+	patches := map[string][]byte{}
+	var text strings.Builder
+	for _, p := range parts {
+		switch p.Kind {
+		case "text":
+			if text.Len() > 0 {
+				text.WriteString("\n")
+			}
+			text.WriteString(p.Text)
+		case "file":
+			if p.File == nil || p.File.Name == "" {
+				continue
+			}
+			data, err := base64.StdEncoding.DecodeString(p.File.Bytes)
+			if err != nil {
+				continue // uri 形态与坏 base64 不落 patch（消费侧最小面）
+			}
+			patches[p.File.Name] = data
+		}
+	}
+	body := strings.TrimSpace(text.String())
+	if len(targets) == 0 {
+		targets = []string{"output.md"}
+	}
+	if body != "" {
+		for _, t := range targets {
+			if _, ok := patches[t]; !ok {
+				patches[t] = []byte(body + "\n")
+			}
+		}
+	}
+	if len(patches) == 0 {
+		return &squad.ExecResult{Success: false, ErrorReason: "a2a: 完成任务无任何工件"}
+	}
+	return &squad.ExecResult{Success: true, Patches: patches, Logs: body}
+}
+
+func statusText(st TaskStatus) string {
+	if st.Message == nil {
+		return ""
+	}
+	var b strings.Builder
+	for _, p := range st.Message.Parts {
+		if p.Kind == "text" {
+			b.WriteString(p.Text)
+		}
+	}
+	return b.String()
+}
@@ -0,0 +1,11 @@
+module agent-squad/adapters/a2a
+
+go 1.27
+
+require agent-squad v0.0.0
+
+require error-code-system v0.0.0 // indirect
+
+replace agent-squad => ../..
+
+replace error-code-system => ../../deps/error-code-system
@@ -3,8 +3,13 @@ package native

 import (
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
 	"fmt"
+	"reflect"
 	"strings"
+	"sync"
 	"time"

 	nativexec "agent-native/adapters/native"
@@ -52,12 +57,58 @@ type Config struct {
 	PermissionAsk func(ctx context.Context, call llm.ToolCall) runtime.AskVerdict
 }

+// Identity 返回塑造 Executor 的配置摘要（模型/客户端/工具/Exec 等）。
+// RegisterFull 经 BackendResolver.Register 替换工厂时会失效缓存；角色侧
+// BashPermission 等由 squad.ConfigIdentity 纳入 Resolve 缓存键，故 allow→deny
+// 在下一次 Resolve 即生效，无需重启进程。
+func (c Config) Identity() string {
+	var b strings.Builder
+	b.WriteString("model=")
+	b.WriteString(c.DefaultModel)
+	b.WriteString("\x00turns=")
+	b.WriteString(fmt.Sprintf("%d", c.MaxTurns))
+	b.WriteString("\x00timeout=")
+	b.WriteString(c.Timeout.String())
+	if c.Client != nil {
+		b.WriteString("\x00base=")
+		b.WriteString(c.Client.BaseURL)
+		b.WriteString("\x00key=")
+		b.WriteString(c.Client.APIKey)
+	}
+	if c.Exec != nil {
+		b.WriteString("\x00exec.bash=")
+		b.WriteString(c.Exec.Bash)
+		b.WriteString("\x00exec.python=")
+		b.WriteString(c.Exec.Python)
+	}
+	if c.Search != nil {
+		b.WriteString("\x00search=1")
+	}
+	if c.Fetch != nil {
+		b.WriteString("\x00fetch=1")
+	}
+	if c.MCP != nil {
+		b.WriteString("\x00mcp=1")
+	}
+	if len(c.SkillsRoots) > 0 {
+		b.WriteString("\x00skills=")
+		b.WriteString(strings.Join(c.SkillsRoots, ","))
+	}
+	if c.CodeOrchestration {
+		b.WriteString("\x00code=1")
+	}
+	sum := sha256.Sum256([]byte(b.String()))
+	return hex.EncodeToString(sum[:16])
+}
+
 // Register 把 "agent-native" 工厂注册进解析器。
 func Register(r *squad.BackendResolver, client *llm.Client, defaultModel string, maxTurns int, timeout time.Duration) {
 	RegisterFull(r, Config{Client: client, DefaultModel: defaultModel, MaxTurns: maxTurns, Timeout: timeout})
 }

 // RegisterFull 注册带工具装配的 "agent-native" 工厂。
+// 再次调用会替换工厂并失效该后端缓存（见 BackendResolver.Register）；
+// 配置摘要见 Config.Identity，角色侧见 squad.ConfigIdentity。
 func RegisterFull(r *squad.BackendResolver, cfg Config) {
 	r.Register("agent-native", func(role squad.RoleSpec) (squad.Executor, error) {
 		opts, err := buildOptions(cfg, role)
@@ -134,12 +185,19 @@ func toNative(req squad.ExecRequest) exec.Request {
 	}
 	if req.OnEvent != nil {
 		prev := req.OnEvent
+		// 一帧一回调：同步转发，不缓存、不批量重发、不按调用分组——下游看到的
+		// 就是 native 的到达序。富化在原帧上就地完成（逐帧独立取值），
+		// 并行工具调用互不串字段，也就不会把交错帧重排成按调用分组。
 		out.OnEvent = func(ev exec.ProcessEvent) {
+			// CallID/Status/ErrorReason 是 native 侧的可选字段（U11a 落地）：
+			// 缺席 = legacy，取零值透传，绝不伪造。
+			extras := probeNativeEventExtras(ev)
 			prev(squad.ProcessEvent{
 				Type: ev.Type, Text: ev.Text, Tool: ev.Tool, Detail: ev.Detail,
 				PromptTokens: ev.PromptTokens, CompletionTokens: ev.CompletionTokens,
 				LastPromptTokens: ev.LastPromptTokens, LastCompletionTokens: ev.LastCompletionTokens,
 				CacheReadTokens: ev.CacheReadTokens, CacheWriteTokens: ev.CacheWriteTokens,
+				CallID: extras.CallID, Status: extras.Status, ErrorReason: extras.ErrorReason,
 			})
 		}
 	}
@@ -147,28 +205,148 @@ func toNative(req squad.ExecRequest) exec.Request {
 }

 func fromNative(res *exec.Result) *squad.ExecResult {
-	return &squad.ExecResult{
-		Success: res.Success, Patches: res.Patches, Logs: res.Logs, ErrorReason: res.ErrorReason,
-		Retryable: res.Retryable, PromptTokens: res.PromptTokens, CompletionTokens: res.CompletionTokens,
+	if res == nil {
+		return nil
+	}
+	out := &squad.ExecResult{
+		Success:          res.Success,
+		Patches:          res.Patches,
+		Logs:             res.Logs,
+		ErrorReason:      res.ErrorReason,
+		Retryable:        res.Retryable,
+		PromptTokens:     res.PromptTokens,
+		CompletionTokens: res.CompletionTokens,
+	}
+	// F05/U3b（Squad 侧 contract）：
+	//   - native exec.Result 在后续 release 会新增 SideEffects、Canceled、Aborted 字段；
+	//     本地 deps gitlink 若尚未跟进，从 JSON 探测拿不到这些字段时回落到
+	//     「legacy reading: side effects = possible」，与既有调用方旧行为一致。
+	//   - 在不影响 ExecResult 默认值的前提下，用 JSON 探测而非硬性修改 deps。
+	probe, err := probeNativeResultExtras(res)
+	if err == nil {
+		out.SideEffects = probe.SideEffects
+		out.Canceled = probe.Canceled
+		out.Aborted = probe.Aborted
+	}
+	return out
+}
+
+// nativeResultExtras 探测 native exec.Result 上的可选字段（结构化 JSON）。
+// 缺省（字段不存在或为零值时）按 legacy reading：SideEffects 空串视为 "possible"；
+// 调用方侧 ExecResult.IsRetryableSideEffectReports() 已经处理该回退。
+type nativeResultExtras struct {
+	SideEffects squad.ExecSideEffects `json:"side_effects,omitempty"`
+	Canceled    bool                  `json:"canceled,omitempty"`
+	Aborted     bool                  `json:"aborted,omitempty"`
+}
+
+// probeNativeResultExtras 通过 JSON 探测 native Result 上的可选字段。
+// 探测出 cancel/abort/side_effects 任何一个就使用它们；探测失败或全部为零值视为
+// legacy reading（SideEffects 空 = possible 兜底）。
+func probeNativeResultExtras(res *exec.Result) (nativeResultExtras, error) {
+	if res == nil {
+		return nativeResultExtras{}, fmt.Errorf("nil result")
+	}
+	v := reflect.ValueOf(res).Elem()
+	if !v.IsValid() {
+		return nativeResultExtras{}, fmt.Errorf("invalid result")
+	}
+	// 用 JSON 做结构化探测：这一路兼容 deps 升级（增字段）、不增字段两种情形。
+	data, err := json.Marshal(v.Interface())
+	if err != nil {
+		return nativeResultExtras{}, err
+	}
+	var out nativeResultExtras
+	if err := json.Unmarshal(data, &out); err != nil {
+		return nativeResultExtras{}, err
+	}
+	return out, nil
+}
+
+// nativeEventExtras 是 native exec.ProcessEvent 上的可选字段：工具调用身份与
+// 真实结局（U11b，语义与 native 侧一致：空串 = legacy/未知）。
+type nativeEventExtras struct {
+	CallID      string
+	Status      string
+	ErrorReason string
+}
+
+// probeNativeEventExtras 读取 native 过程帧上的 CallID/Status/ErrorReason。
+//
+// 与 probeNativeResultExtras 同一取法（可选字段缺席 = legacy 零值），但改用
+// 反射按字段名读取而不是 JSON：本地 deps gitlink 上的 exec.ProcessEvent 尚未
+// 跟进这三个字段（U11a 落地），硬写 ev.CallID 会编译不过；反射让同一份代码
+// 两侧都成立——字段缺席取零值（既有行为逐字节不变），字段落地后无需改本文件
+// 即生效，也不受 native 是否给字段加 JSON tag 影响。每次过程帧只做三次字段
+// 定位，不构造中间对象；native 可并行发帧时读到的仍是该帧自己的值。
+func probeNativeEventExtras(v any) nativeEventExtras {
+	rv := reflect.ValueOf(v)
+	for rv.Kind() == reflect.Pointer || rv.Kind() == reflect.Interface {
+		if rv.IsNil() {
+			return nativeEventExtras{}
+		}
+		rv = rv.Elem()
+	}
+	if rv.Kind() != reflect.Struct {
+		return nativeEventExtras{}
+	}
+	return nativeEventExtras{
+		CallID:      optionalStringField(rv, "CallID"),
+		Status:      optionalStringField(rv, "Status"),
+		ErrorReason: optionalStringField(rv, "ErrorReason"),
 	}
 }

+// optionalStringField 按字段名取字符串字段；缺席、嵌套或类型不符都返回 ""
+// （legacy 零值）。只取直接字段，不做嵌套指针下钻。
+func optionalStringField(rv reflect.Value, name string) string {
+	field, ok := rv.Type().FieldByName(name)
+	if !ok || len(field.Index) != 1 {
+		return ""
+	}
+	value := rv.Field(field.Index[0])
+	if value.Kind() != reflect.String {
+		return ""
+	}
+	return value.String()
+}
+
+// eventTap 在组合根侧再挂一层过程事件观察。帧按值传递（squad.ProcessEvent
+// 只含值类型字段），两个下游各自拿到整帧副本，并行工具调用不会互相串字段；
+// 回调保持同步，不排队、不缓存。
+//
+// 扇出以 mu 串行：「一帧 → 全部下游」是一个整体。这不是可有可无的细节：
+// native 对并行工具调用是每调用一个 goroutine 发帧，两次帧到达可以真正并发。
+// 不串行时，帧 F 交给组合根侧 sink 之后、在交给请求侧 sink 之前，可能被后到
+// 的帧 G 抢先递完，于是同一串到达帧在两个下游呈现出不同的顺序——实测就是
+// 组合根侧看到 [start(err), start(ok), end(ok), end(err)]（到达序），而请求侧
+// 看到 [start(ok), end(ok), start(err), end(err)]（被按调用重排/分组）。
+// 串行扇出后每个下游看到的都是同一串到达序：一个上游帧 → 每个下游恰好一次、
+// 顺序不变，且事件富化（CallID/Status/ErrorReason）在桥里就地完成、不在这里
+// 重新分帧。mu 只在本次扇出期间持有，不跨帧持有。
 type eventTap struct {
 	next squad.Executor
 	on   func(squad.ProcessEvent)
+	mu   sync.Mutex
 }

 func (b *eventTap) Name() string { return b.next.Name() }

 func (b *eventTap) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
-	prev := req.OnEvent
-	req.OnEvent = func(ev squad.ProcessEvent) {
+	req.OnEvent = b.hook(req.OnEvent)
+	return b.next.Execute(ctx, req)
+}
+
+// hook 把一帧扇出到组合根侧 sink 与请求侧 sink，顺序不变。
+func (b *eventTap) hook(prev func(squad.ProcessEvent)) func(squad.ProcessEvent) {
+	return func(ev squad.ProcessEvent) {
+		b.mu.Lock()
+		defer b.mu.Unlock()
 		b.on(ev)
 		if prev != nil {
 			prev(ev)
 		}
 	}
-	return b.next.Execute(ctx, req)
 }

 // LLM 把 llm.Client 适配成 ModelClient。
@@ -0,0 +1,427 @@
+package native
+
+import (
+	"context"
+	"reflect"
+	"sync"
+	"testing"
+	"time"
+
+	"agent-native/exec"
+	"agent-squad/squad"
+)
+
+// nativeEventWithIDs 模拟 U11a 落地后的 native exec.ProcessEvent 形态：
+// 多出的 CallID / Status / ErrorReason，且带 snake_case JSON tag
+// （证明 squad 侧读取不受 native 的 tag 写法影响）。
+type nativeEventWithIDs struct {
+	Type        string `json:"type"`
+	Text        string `json:"text"`
+	Tool        string `json:"tool"`
+	Detail      string `json:"detail"`
+	CallID      string `json:"call_id,omitempty"`
+	Status      string `json:"status,omitempty"`
+	ErrorReason string `json:"error_reason,omitempty"`
+}
+
+func TestBridgeMapsLegacyNativeEventWithoutFabricatingFields(t *testing.T) {
+	var got []squad.ProcessEvent
+	out := toNative(squad.ExecRequest{OnEvent: func(ev squad.ProcessEvent) { got = append(got, ev) }})
+	if out.OnEvent == nil {
+		t.Fatal("bridge must install its own event hook when the request has one")
+	}
+	out.OnEvent(exec.ProcessEvent{
+		Type: "tool_execution_end", Text: "found", Tool: "grep", Detail: `{"q":"x"}`,
+		PromptTokens: 11, CompletionTokens: 22,
+		LastPromptTokens: 3, LastCompletionTokens: 4, CacheReadTokens: 5, CacheWriteTokens: 6,
+	})
+	if len(got) != 1 {
+		t.Fatalf("want 1 squad event, got %d", len(got))
+	}
+	want := squad.ProcessEvent{
+		Type: "tool_execution_end", Text: "found", Tool: "grep", Detail: `{"q":"x"}`,
+		PromptTokens: 11, CompletionTokens: 22,
+		LastPromptTokens: 3, LastCompletionTokens: 4, CacheReadTokens: 5, CacheWriteTokens: 6,
+	}
+	if !reflect.DeepEqual(got[0], want) {
+		t.Fatalf("legacy native event must map byte-identically:\n got %+v\nwant %+v", got[0], want)
+	}
+	// 无请求侧回调时不装钩子（既有行为不变）。
+	if toNative(squad.ExecRequest{}).OnEvent != nil {
+		t.Fatal("bridge must not install an event hook when the request has none")
+	}
+}
+
+func TestProbeNativeEventExtrasReadsNewFieldsWhenNativeHasThem(t *testing.T) {
+	// legacy 帧：字段缺席（老 pin）或在场但未填（当前 pin 的零值）→ 一律零值，
+	// 不臆造身份与结局。
+	if extras := probeNativeEventExtras(exec.ProcessEvent{Type: "tool_execution_end", Tool: "grep"}); extras != (nativeEventExtras{}) {
+		t.Fatalf("legacy native event must yield zero extras, got %+v", extras)
+	}
+	// U11a 落地后：按字段名读到真实值（与 JSON tag 写法无关）。
+	future := nativeEventWithIDs{Type: "tool_execution_end", Tool: "grep", CallID: "call_err",
+		Status: "error", ErrorReason: "grep: permission denied"}
+	want := nativeEventExtras{CallID: "call_err", Status: "error", ErrorReason: "grep: permission denied"}
+	if extras := probeNativeEventExtras(future); extras != want {
+		t.Fatalf("identity/outcome not carried through: %+v", extras)
+	}
+	if extras := probeNativeEventExtras(&future); extras != want {
+		t.Fatalf("pointer and value frames must probe identically: %+v", extras)
+	}
+	// 缺席/非字符串字段都回落到零值。
+	odd := struct {
+		CallID      int
+		Status      string
+		ErrorReason bool
+	}{CallID: 7, Status: "ok", ErrorReason: true}
+	if extras := probeNativeEventExtras(odd); extras != (nativeEventExtras{Status: "ok"}) {
+		t.Fatalf("non-string/absent fields must stay zero-valued, got %+v", extras)
+	}
+	var nilFrame *nativeEventWithIDs
+	if extras := probeNativeEventExtras(nilFrame); extras != (nativeEventExtras{}) {
+		t.Fatalf("nil frame must stay zero-valued, got %+v", extras)
+	}
+	if extras := probeNativeEventExtras("not a frame"); extras != (nativeEventExtras{}) {
+		t.Fatalf("non-struct input must stay zero-valued, got %+v", extras)
+	}
+	// 嵌套（嵌入）字段不探测：不沿 nil 嵌入指针下钻，也不当作直接字段读取。
+	nested := struct {
+		*nativeEventWithIDs
+		Status string
+	}{Status: "ok"}
+	if extras := probeNativeEventExtras(nested); extras != (nativeEventExtras{Status: "ok"}) {
+		t.Fatalf("nested optional fields must stay zero-valued, got %+v", extras)
+	}
+}
+
+// interleavedToolFrames 是并行工具调用的固定交错帧序（squad 侧形态）：
+//
+//	start(call_err) → start(call_ok) → end(call_ok, ok) → end(call_err, error)
+//
+// 按到达序转发时逐帧等值；任何「按调用分组重排」「按调用缓存后再发」都会把它
+// 变成 [start, end, start, end]，因此它能把重排/重发类缺陷钉成硬失败。
+func interleavedToolFrames() []squad.ProcessEvent {
+	return []squad.ProcessEvent{
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_err"},
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_ok"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_ok",
+			Status: squad.ToolStatusOK, Text: "found"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_err",
+			Status: squad.ToolStatusError, ErrorReason: "grep: permission denied"},
+	}
+}
+
+// interleavedNativeFrames 是 interleavedToolFrames 的 native 侧形态：U11a 落地后
+// （当前 pin 即此形态）的帧自带 CallID/Status/ErrorReason，桥读到的就是它们。
+func interleavedNativeFrames() []exec.ProcessEvent {
+	return []exec.ProcessEvent{
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_err"},
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_ok"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_ok",
+			Status: squad.ToolStatusOK, Text: "found"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_err",
+			Status: squad.ToolStatusError, ErrorReason: "grep: permission denied"},
+	}
+}
+
+// legacyToolFrames 是同一个交错序列的 legacy squad 形态：身份与结局字段为空串
+// （上游未填充），失败真相只在 Text 里——空串 = 未知，不是「无名调用」。
+func legacyToolFrames() []squad.ProcessEvent {
+	return []squad.ProcessEvent{
+		{Type: "tool_execution_start", Tool: "grep"},
+		{Type: "tool_execution_start", Tool: "grep"},
+		{Type: "tool_execution_end", Tool: "grep", Text: "found"},
+		{Type: "tool_execution_end", Tool: "grep", Text: "grep: permission denied"},
+	}
+}
+
+// legacyNativeFrames 是 legacy native 事件流：帧上没有 CallID/Status/ErrorReason
+// （字段不存在，或存在但未填——对探测而言都是零值）。桥必须原样透传。
+func legacyNativeFrames() []exec.ProcessEvent {
+	return []exec.ProcessEvent{
+		{Type: "tool_execution_start", Tool: "grep"},
+		{Type: "tool_execution_start", Tool: "grep"},
+		{Type: "tool_execution_end", Tool: "grep", Text: "found"},
+		{Type: "tool_execution_end", Tool: "grep", Text: "grep: permission denied"},
+	}
+}
+
+// assertArrivalOrder 断言帧序逐帧等于到达序：一个上游帧对应下游一帧，顺序、
+// 身份与结局全部逐帧等值（squad.ProcessEvent 只含值类型字段，可直接比较）。
+func assertArrivalOrder(t *testing.T, tag string, got, want []squad.ProcessEvent) {
+	t.Helper()
+	if len(got) != len(want) {
+		t.Fatalf("%s: frames = %d, want %d (每帧一次转发，不多不少不重发):\n%+v", tag, len(got), len(want), got)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("%s: frame %d = %+v\nwant %+v\n（帧序必须与到达序逐帧一致）", tag, i, got[i], want[i])
+		}
+	}
+}
+
+// concurrentToolExecutor 并发发两次同名工具调用的过程帧（CallID 不同、一成一败），
+// 模拟 native 侧并行工具执行时的帧序交错。
+type concurrentToolExecutor struct{}
+
+func (concurrentToolExecutor) Name() string { return "concurrent-tool" }
+
+func (concurrentToolExecutor) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
+	calls := []struct {
+		callID, status, errorReason, text string
+	}{
+		{callID: "call_ok", status: squad.ToolStatusOK, text: "found"},
+		{callID: "call_err", status: squad.ToolStatusError, errorReason: "grep: permission denied"},
+	}
+	var wg sync.WaitGroup
+	for _, call := range calls {
+		wg.Add(1)
+		go func(callID, status, errorReason, text string) {
+			defer wg.Done()
+			if req.OnEvent == nil {
+				return
+			}
+			req.OnEvent(squad.ProcessEvent{Type: "tool_execution_start", Tool: "grep", CallID: callID})
+			req.OnEvent(squad.ProcessEvent{Type: "tool_execution_end", Tool: "grep", CallID: callID,
+				Status: status, ErrorReason: errorReason, Text: text})
+		}(call.callID, call.status, call.errorReason, call.text)
+	}
+	wg.Wait()
+	return &squad.ExecResult{Success: true}, nil
+}
+
+// assertPaired 断言按 CallID 配对：每个 call_id 一帧 start、一帧 end，
+// 结局属于该调用自身（并行工具调用不串字段）。
+func assertPaired(t *testing.T, tag string, frames []squad.ProcessEvent) {
+	t.Helper()
+	starts := map[string]int{}
+	ends := map[string]squad.ProcessEvent{}
+	for _, ev := range frames {
+		if ev.CallID == "" {
+			t.Fatalf("%s: frame without CallID: %+v", tag, ev)
+		}
+		switch ev.Type {
+		case "tool_execution_start":
+			starts[ev.CallID]++
+		case "tool_execution_end":
+			if _, dup := ends[ev.CallID]; dup {
+				t.Fatalf("%s: duplicate end for %s: %+v", tag, ev.CallID, ev)
+			}
+			ends[ev.CallID] = ev
+		}
+	}
+	if len(ends) != 2 {
+		t.Fatalf("%s: distinct CallIDs = %d, want 2 (%+v)", tag, len(ends), frames)
+	}
+	for _, want := range []struct {
+		callID, status, errorReason string
+	}{
+		{callID: "call_ok", status: squad.ToolStatusOK},
+		{callID: "call_err", status: squad.ToolStatusError, errorReason: "grep: permission denied"},
+	} {
+		if starts[want.callID] != 1 {
+			t.Fatalf("%s: call %s starts = %d, want 1", tag, want.callID, starts[want.callID])
+		}
+		end := ends[want.callID]
+		if end.Status != want.status || end.ErrorReason != want.errorReason {
+			t.Fatalf("%s: call %s outcome = %s/%q, want %s/%q", tag, want.callID,
+				end.Status, end.ErrorReason, want.status, want.errorReason)
+		}
+	}
+}
+
+// TestEventTapKeepsConcurrentToolCallIdentityAndOutcome 在真实并发帧源下验证三件事：
+// 身份与结局不串（严格按 CallID 配对）、帧不多发不少发（每个调用一 start 一 end）、
+// 所有下游观察到同一串到达序（并发帧源下到达序本身不确定，但必须唯一且一致）。
+func TestEventTapKeepsConcurrentToolCallIdentityAndOutcome(t *testing.T) {
+	var tapped, requested []squad.ProcessEvent
+	var mu sync.Mutex
+	tap := &eventTap{next: concurrentToolExecutor{}, on: func(ev squad.ProcessEvent) {
+		mu.Lock()
+		defer mu.Unlock()
+		tapped = append(tapped, ev)
+	}}
+	_, err := tap.Execute(context.Background(), squad.ExecRequest{
+		OnEvent: func(ev squad.ProcessEvent) {
+			mu.Lock()
+			defer mu.Unlock()
+			requested = append(requested, ev)
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	assertPaired(t, "tap", tapped)
+	assertPaired(t, "request", requested)
+	if !reflect.DeepEqual(tapped, requested) {
+		t.Fatalf("every sink must observe the same arrival order (one frame per upstream frame, unbuffered):\ntap     %+v\nrequest %+v", tapped, requested)
+	}
+}
+
+// TestNativeFramesReachBothTapSinksInArrivalOrder 走完整一跳：native 帧 → 桥的
+// OnEvent（就地富化 CallID/Status/ErrorReason）→ 事件 tap → 组合根侧 sink 与
+// 请求侧 sink。两种 native（当前 pin 的带字段帧 / legacy 无字段帧）都要逐帧
+// 等值、顺序不变：既不能被按调用分组重排，也不能缓存后再批量发。
+func TestNativeFramesReachBothTapSinksInArrivalOrder(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		native []exec.ProcessEvent
+		want   []squad.ProcessEvent
+	}{
+		{name: "current-native", native: interleavedNativeFrames(), want: interleavedToolFrames()},
+		{name: "legacy-native", native: legacyNativeFrames(), want: legacyToolFrames()},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			var mu sync.Mutex
+			var tapped, requested []squad.ProcessEvent
+			tap := &eventTap{on: func(ev squad.ProcessEvent) {
+				mu.Lock()
+				defer mu.Unlock()
+				tapped = append(tapped, ev)
+			}}
+			nativeHook := toNative(squad.ExecRequest{OnEvent: tap.hook(func(ev squad.ProcessEvent) {
+				mu.Lock()
+				defer mu.Unlock()
+				requested = append(requested, ev)
+			})}).OnEvent
+			if nativeHook == nil {
+				t.Fatal("bridge must install its own event hook when the request has one")
+			}
+			for _, ev := range tc.native {
+				nativeHook(ev)
+			}
+			assertArrivalOrder(t, tc.name+"/tap", tapped, tc.want)
+			assertArrivalOrder(t, tc.name+"/request", requested, tc.want)
+			if tc.name == "current-native" {
+				assertPaired(t, tc.name, tapped)
+				assertPaired(t, tc.name, requested)
+			}
+		})
+	}
+}
+
+// TestBridgeForwardsNativeFramesInArrivalOrder 直接钉住桥这一跳的帧序与富化：
+// 当前 pin（帧带 CallID/Status/ErrorReason）就地把值带上来，legacy 帧（无这三个
+// key）就以空串原样透传——两种情况都必须逐帧等值、顺序不变。
+func TestBridgeForwardsNativeFramesInArrivalOrder(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		native []exec.ProcessEvent
+		want   []squad.ProcessEvent
+	}{
+		{name: "current-native", native: interleavedNativeFrames(), want: interleavedToolFrames()},
+		{name: "legacy-native", native: legacyNativeFrames(), want: legacyToolFrames()},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			var got []squad.ProcessEvent
+			hook := toNative(squad.ExecRequest{OnEvent: func(ev squad.ProcessEvent) {
+				got = append(got, ev)
+			}}).OnEvent
+			if hook == nil {
+				t.Fatal("bridge must install its own event hook when the request has one")
+			}
+			for _, ev := range tc.native {
+				hook(ev)
+			}
+			assertArrivalOrder(t, tc.name, got, tc.want)
+			if tc.name == "current-native" {
+				assertPaired(t, tc.name, got)
+			}
+		})
+	}
+}
+
+// fanoutWindow 制造扇出竞态窗口：支线 A 逐帧发 call_ok 的两帧；当 A 的第二帧进入
+// 组合根侧 sink 时（entered 关闭）支线 B 才发 call_err 的两帧，并在自己的扇出
+// 全部返回后关闭 finished。A 在 sink 里等 finished，于是「A 已收下第二帧、还没
+// 交给请求侧 sink」这个窗口被固定复现——扇出不串行时 B 会抢先穿过这个窗口。
+type fanoutWindow struct {
+	entered  chan struct{}
+	finished chan struct{}
+}
+
+func (fanoutWindow) Name() string { return "fanout-window" }
+
+func (w fanoutWindow) Execute(ctx context.Context, req squad.ExecRequest) (*squad.ExecResult, error) {
+	if req.OnEvent == nil {
+		return &squad.ExecResult{Success: true}, nil
+	}
+	var wg sync.WaitGroup
+	wg.Add(2)
+	go func() {
+		defer wg.Done()
+		req.OnEvent(squad.ProcessEvent{Type: "tool_execution_start", Tool: "grep", CallID: "call_ok"})
+		req.OnEvent(squad.ProcessEvent{Type: "tool_execution_end", Tool: "grep", CallID: "call_ok",
+			Status: squad.ToolStatusOK, Text: "found"})
+	}()
+	go func() {
+		defer wg.Done()
+		<-w.entered
+		req.OnEvent(squad.ProcessEvent{Type: "tool_execution_start", Tool: "grep", CallID: "call_err"})
+		req.OnEvent(squad.ProcessEvent{Type: "tool_execution_end", Tool: "grep", CallID: "call_err",
+			Status: squad.ToolStatusError, ErrorReason: "grep: permission denied"})
+		close(w.finished)
+	}()
+	wg.Wait()
+	return &squad.ExecResult{Success: true}, nil
+}
+
+// TestEventTapFansOutOneFramePerSinkInOneOrder 把「扇出是一个整体」钉成确定性失败：
+// 组合根侧 sink 在第二帧上主动等另一条支线走完自己的扇出。没有扇出串行化时，
+// 请求侧 sink 会先收到 call_err 的两帧，组合根侧 sink 却先收到 call_ok 的两帧——
+// 同一串到达帧在两个下游呈现为两种顺序（这正是并发下把帧重排的入口）。
+func TestEventTapFansOutOneFramePerSinkInOneOrder(t *testing.T) {
+	window := fanoutWindow{entered: make(chan struct{}), finished: make(chan struct{})}
+	var mu sync.Mutex
+	var tapped, requested []squad.ProcessEvent
+	tap := &eventTap{next: window, on: func(ev squad.ProcessEvent) {
+		mu.Lock()
+		tapped = append(tapped, ev)
+		seen := len(tapped)
+		mu.Unlock()
+		if seen == 2 {
+			// 不持有采集锁去等：让另一条支线真的能穿过扇出窗口（否则本测试
+			// 只会测到自己）。
+			close(window.entered)
+			select {
+			case <-window.finished:
+			case <-time.After(250 * time.Millisecond):
+			}
+		}
+	}}
+	_, err := tap.Execute(context.Background(), squad.ExecRequest{
+		OnEvent: func(ev squad.ProcessEvent) {
+			mu.Lock()
+			defer mu.Unlock()
+			requested = append(requested, ev)
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := []squad.ProcessEvent{
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_ok"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_ok", Status: squad.ToolStatusOK, Text: "found"},
+		{Type: "tool_execution_start", Tool: "grep", CallID: "call_err"},
+		{Type: "tool_execution_end", Tool: "grep", CallID: "call_err",
+			Status: squad.ToolStatusError, ErrorReason: "grep: permission denied"},
+	}
+	assertArrivalOrder(t, "tap", tapped, want)
+	assertArrivalOrder(t, "request", requested, want)
+	assertPaired(t, "tap", tapped)
+	assertPaired(t, "request", requested)
+}
+
+func TestEventTapWithoutRequestHookStillTaps(t *testing.T) {
+	var mu sync.Mutex
+	var got []squad.ProcessEvent
+	tap := &eventTap{next: concurrentToolExecutor{}, on: func(ev squad.ProcessEvent) {
+		mu.Lock()
+		defer mu.Unlock()
+		got = append(got, ev)
+	}}
+	if _, err := tap.Execute(context.Background(), squad.ExecRequest{}); err != nil {
+		t.Fatal(err)
+	}
+	assertPaired(t, "tap-only", got)
+}
@@ -47,3 +47,85 @@ func TestBuildOptionsBashPermissionGating(t *testing.T) {
 		t.Fatal("missing model must fail-closed")
 	}
 }
+
+func TestConfigIdentityReflectsToolAndClient(t *testing.T) {
+	base := Config{
+		Client:       &llm.Client{BaseURL: "http://x", APIKey: "k"},
+		DefaultModel: "m1", MaxTurns: 5, Timeout: time.Minute,
+		Exec: &tools.ExecConfig{Bash: "/bin/bash"},
+	}
+	otherExec := base
+	otherExec.Exec = &tools.ExecConfig{Bash: "/bin/zsh"}
+	if base.Identity() == otherExec.Identity() {
+		t.Fatal("Exec tool-config change must change Config.Identity")
+	}
+	otherClient := base
+	otherClient.Client = &llm.Client{BaseURL: "http://y", APIKey: "k"}
+	if base.Identity() == otherClient.Identity() {
+		t.Fatal("client change must change Config.Identity")
+	}
+	same := base
+	if base.Identity() != same.Identity() {
+		t.Fatal("identical config must keep Identity stable")
+	}
+}
+
+func TestRegisterFullAllowDenyTakesEffectOnNextResolve(t *testing.T) {
+	r := squad.NewBackendResolver()
+	cfg := Config{
+		Client:       &llm.Client{BaseURL: "http://x", APIKey: "k"},
+		DefaultModel: "m1", MaxTurns: 1, Timeout: time.Second,
+		Exec: &tools.ExecConfig{Bash: "/bin/bash"},
+	}
+	RegisterFull(r, cfg)
+
+	allowRole := squad.RoleSpec{Role: "eng", Backend: "agent-native", Model: "m1", BashPermission: squad.BashPermissionAllow}
+	denyRole := squad.RoleSpec{Role: "eng", Backend: "agent-native", Model: "m1", BashPermission: squad.BashPermissionDeny}
+
+	eAllow, _, err := r.Resolve(allowRole, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	eDeny, _, err := r.Resolve(denyRole, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if eAllow == eDeny {
+		t.Fatal("allow→deny must not reuse cached executor on next Resolve")
+	}
+	// Identity of role configs must diverge so cache keys differ.
+	if squad.ConfigIdentity(allowRole) == squad.ConfigIdentity(denyRole) {
+		t.Fatal("role ConfigIdentity must include BashPermission")
+	}
+}
+
+func TestRegisterFullToolConfigReplacementTakesEffect(t *testing.T) {
+	r := squad.NewBackendResolver()
+	cfg1 := Config{
+		Client:       &llm.Client{BaseURL: "http://x", APIKey: "k"},
+		DefaultModel: "m1", MaxTurns: 1, Timeout: time.Second,
+		Exec: &tools.ExecConfig{Bash: "/bin/bash"},
+	}
+	cfg2 := Config{
+		Client:       &llm.Client{BaseURL: "http://x", APIKey: "k"},
+		DefaultModel: "m1", MaxTurns: 1, Timeout: time.Second,
+		Exec: &tools.ExecConfig{Bash: "/bin/zsh"},
+	}
+	if cfg1.Identity() == cfg2.Identity() {
+		t.Fatal("precondition: tool configs must differ in Identity")
+	}
+	RegisterFull(r, cfg1)
+	role := squad.RoleSpec{Role: "eng", Backend: "agent-native", Model: "m1", BashPermission: squad.BashPermissionAllow}
+	e1, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	RegisterFull(r, cfg2)
+	e2, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if e1 == e2 {
+		t.Fatal("RegisterFull tool-config replacement must invalidate cache")
+	}
+}
@@ -189,8 +189,33 @@ func (s *Store) PromoteDeferred(at time.Time) (int, error) {
 	return int(n), nil
 }

+// claimSQL is the whole claim: one atomic conditional UPDATE whose two
+// guarded counts bound the agent by its own max_concurrency and the runtime by
+// max_concurrency over every in-flight task of that runtime (an agent bound to
+// it, or a task this runtime already leased). A count that cannot be read
+// fails the statement, so the claim is refused instead of allowed.
+const claimSQL = `UPDATE task SET status='dispatched', runtime_id=?, lease_until=?, updated_at=?
+	WHERE id=? AND status='queued'
+	AND (SELECT COUNT(*) FROM task WHERE agent_id=? AND status IN ('dispatched','running')) < ?
+	AND (SELECT COUNT(*) FROM task t WHERE t.status IN ('dispatched','running')
+		AND (t.runtime_id=? OR EXISTS (SELECT 1 FROM agent a WHERE a.id=t.agent_id AND a.runtime_id=?))) < ?`
+
+// claimCapacitySQL counts the agent's own and the runtime's in-flight tasks in
+// one read, so the pre-check below matches the guards of claimSQL.
+const claimCapacitySQL = `SELECT
+	(SELECT COUNT(*) FROM task WHERE agent_id=? AND status IN ('dispatched','running')),
+	(SELECT COUNT(*) FROM task t WHERE t.status IN ('dispatched','running')
+		AND (t.runtime_id=? OR EXISTS (SELECT 1 FROM agent a WHERE a.id=t.agent_id AND a.runtime_id=?)))`
+
 // Claim takes one queued task whose agent and runtime are under their caps.
 // The task becomes dispatched with a five-minute lease.
+//
+// Both caps live inside the single conditional UPDATE, so concurrent claims
+// can overshoot neither the agent cap nor the runtime total: the agent count
+// is bounded by agent.max_concurrency and the runtime count (every in-flight
+// task whose agent belongs to this runtime, plus every task this runtime
+// already leased) by runtime.max_concurrency. A read error is a refusal, never
+// a silent allow: the candidate is skipped, and no task is handed out.
 func (s *Store) Claim(runtimeID string, at time.Time) (Task, bool, error) {
 	rt, err := s.getRuntime(runtimeID)
 	if err != nil {
@@ -223,20 +248,20 @@ func (s *Store) Claim(runtimeID string, at time.Time) (Task, bool, error) {
 		if agent.RuntimeID != "" && agent.RuntimeID != runtimeID {
 			continue
 		}
-		cap := agent.MaxConcurrency
-		if rt.MaxConcurrency < cap {
-			cap = rt.MaxConcurrency
+		// Pre-check only: the UPDATE below re-counts both caps under the write
+		// lock. A count that cannot be read means unknown capacity, so refuse.
+		var agentBusy, runtimeBusy int
+		err = s.db.QueryRow(claimCapacitySQL, cand.agentID, runtimeID, runtimeID).Scan(&agentBusy, &runtimeBusy)
+		if err != nil {
+			return Task{}, false, fmt.Errorf("claim capacity count: %w", err)
 		}
-		var running int
-		_ = s.db.QueryRow(`SELECT COUNT(*) FROM task WHERE agent_id=? AND status IN ('dispatched','running')`, cand.agentID).Scan(&running)
-		if running >= cap {
+		if agentBusy >= agent.MaxConcurrency || runtimeBusy >= rt.MaxConcurrency {
 			continue
 		}
 		lease := at.UTC().Add(5 * time.Minute).Format(time.RFC3339Nano)
-		res, err := s.db.Exec(`UPDATE task SET status='dispatched', runtime_id=?, lease_until=?, updated_at=?
-			WHERE id=? AND status='queued'
-			AND (SELECT COUNT(*) FROM task WHERE agent_id=? AND status IN ('dispatched','running')) < ?`,
-			runtimeID, lease, now(), cand.id, cand.agentID, cap)
+		res, err := s.db.Exec(claimSQL,
+			runtimeID, lease, now(), cand.id, cand.agentID, agent.MaxConcurrency,
+			runtimeID, runtimeID, rt.MaxConcurrency)
 		if err != nil {
 			return Task{}, false, err
 		}
@@ -0,0 +1,325 @@
+package platform
+
+import (
+	"path/filepath"
+	"sync"
+	"testing"
+	"time"
+)
+
+// openQueue opens a hermetic store in a temp dir and returns it with its path.
+func openQueue(t *testing.T) (*Store, string) {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "queue.db")
+	s, err := Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = s.Close() })
+	return s, path
+}
+
+// newQueueRuntime registers a runtime with the given cap.
+func newQueueRuntime(t *testing.T, s *Store, workspaceID, name string, cap int) Runtime {
+	t.Helper()
+	rt, err := s.RegisterRuntime(Runtime{
+		WorkspaceID: workspaceID, Name: name, Command: "in-process",
+		Protocol: "agent-native", MaxConcurrency: cap,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return rt
+}
+
+// newQueueAgent creates an agent with the given cap, pinned to runtimeID
+// (an empty runtimeID leaves the agent unbound, claimable by any runtime).
+func newQueueAgent(t *testing.T, s *Store, workspaceID, name, runtimeID string, cap int) Agent {
+	t.Helper()
+	agent, err := s.CreateAgent(Agent{WorkspaceID: workspaceID, Name: name, RuntimeID: runtimeID, MaxConcurrency: cap})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return agent
+}
+
+// queueTask creates an issue and queues one task for agentID.
+func queueTask(t *testing.T, s *Store, workspaceID, agentID, title string) Task {
+	t.Helper()
+	issue, err := s.CreateIssue(Issue{WorkspaceID: workspaceID, Title: title, Body: title, Status: "todo"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	task, err := s.enqueue(issue, agentID, false, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return task
+}
+
+// taskStatus reads the durable status of one task.
+func taskStatus(t *testing.T, s *Store, id string) string {
+	t.Helper()
+	var status string
+	if err := s.db.QueryRow(`SELECT status FROM task WHERE id=?`, id).Scan(&status); err != nil {
+		t.Fatal(err)
+	}
+	return status
+}
+
+// R6: the runtime cap bounds every agent on that runtime, not one agent only.
+// A per-agent cap of 4 each still yields a single in-flight task at cap 1.
+func TestClaimEnforcesRuntimeCapAcrossAgents(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		bindToRT bool
+	}{
+		{name: "agents bound to the runtime", bindToRT: true},
+		{name: "agents not bound to the runtime", bindToRT: false},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			s, _ := openQueue(t)
+			ws, err := s.CreateWorkspace("lab")
+			if err != nil {
+				t.Fatal(err)
+			}
+			rt := newQueueRuntime(t, s, ws.ID, "rt", 1)
+			runtimeID := ""
+			if tc.bindToRT {
+				runtimeID = rt.ID
+			}
+			first := newQueueAgent(t, s, ws.ID, "a1", runtimeID, 4)
+			second := newQueueAgent(t, s, ws.ID, "a2", runtimeID, 4)
+			t1 := queueTask(t, s, ws.ID, first.ID, "one")
+			t2 := queueTask(t, s, ws.ID, second.ID, "two")
+
+			claimed, ok, err := s.Claim(rt.ID, time.Now())
+			if err != nil || !ok || claimed.ID != t1.ID || claimed.RuntimeID != rt.ID {
+				t.Fatalf("first claim %+v ok=%v err=%v want task %s on %s", claimed, ok, err, t1.ID, rt.ID)
+			}
+			if _, ok, err := s.Claim(rt.ID, time.Now()); err != nil || ok {
+				t.Fatalf("runtime cap 1 must refuse the other agent: ok=%v err=%v", ok, err)
+			}
+			if got := taskStatus(t, s, t2.ID); got != "queued" {
+				t.Fatalf("second task %q must stay queued while the runtime is full", got)
+			}
+
+			if err := s.FinishTask(claimed.ID, "completed", ""); err != nil {
+				t.Fatal(err)
+			}
+			next, ok, err := s.Claim(rt.ID, time.Now())
+			if err != nil || !ok || next.ID != t2.ID {
+				t.Fatalf("claim after release %+v ok=%v err=%v want task %s", next, ok, err, t2.ID)
+			}
+		})
+	}
+}
+
+// The agent cap holds when two runtimes compete for the same agent.
+func TestClaimEnforcesAgentCapAcrossRuntimes(t *testing.T) {
+	s, _ := openQueue(t)
+	ws, err := s.CreateWorkspace("lab")
+	if err != nil {
+		t.Fatal(err)
+	}
+	rtA := newQueueRuntime(t, s, ws.ID, "rt-a", 2)
+	rtB := newQueueRuntime(t, s, ws.ID, "rt-b", 2)
+	agent := newQueueAgent(t, s, ws.ID, "solo", "", 1)
+	t1 := queueTask(t, s, ws.ID, agent.ID, "one")
+	t2 := queueTask(t, s, ws.ID, agent.ID, "two")
+
+	claimed, ok, err := s.Claim(rtA.ID, time.Now())
+	if err != nil || !ok || claimed.ID != t1.ID {
+		t.Fatalf("rt-a claim %+v ok=%v err=%v want task %s", claimed, ok, err, t1.ID)
+	}
+	if _, ok, err := s.Claim(rtB.ID, time.Now()); err != nil || ok {
+		t.Fatalf("agent cap 1 must bound the second runtime: ok=%v err=%v", ok, err)
+	}
+	if got := taskStatus(t, s, t2.ID); got != "queued" {
+		t.Fatalf("second task %q must stay queued while the agent is busy", got)
+	}
+}
+
+// Concurrent claims never overshoot the runtime cap or an agent cap.
+func TestClaimConcurrentNeverExceedsCaps(t *testing.T) {
+	s, _ := openQueue(t)
+	ws, err := s.CreateWorkspace("lab")
+	if err != nil {
+		t.Fatal(err)
+	}
+	rt := newQueueRuntime(t, s, ws.ID, "rt", 2)
+	tight := newQueueAgent(t, s, ws.ID, "tight", rt.ID, 1)
+	loose := newQueueAgent(t, s, ws.ID, "loose", rt.ID, 8)
+	other := newQueueAgent(t, s, ws.ID, "other", rt.ID, 8)
+	for _, agent := range []Agent{tight, loose, other} {
+		queueTask(t, s, ws.ID, agent.ID, "task for "+agent.Name)
+	}
+
+	const workers = 12
+	var wg sync.WaitGroup
+	var mu sync.Mutex
+	claimed := map[string]string{} // task id -> agent id
+	failures := []string{}
+	for i := 0; i < workers; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			task, ok, err := s.Claim(rt.ID, time.Now())
+			mu.Lock()
+			defer mu.Unlock()
+			switch {
+			case err != nil:
+				failures = append(failures, err.Error())
+			case ok:
+				claimed[task.ID] = task.AgentID
+			}
+		}()
+	}
+	wg.Wait()
+
+	if len(failures) > 0 {
+		t.Fatalf("concurrent claims failed: %v", failures)
+	}
+	if len(claimed) == 0 {
+		t.Fatal("no claim landed")
+	}
+	if len(claimed) > rt.MaxConcurrency {
+		t.Fatalf("runtime cap %d exceeded: %v", rt.MaxConcurrency, claimed)
+	}
+	perAgent := map[string]int{}
+	for _, agentID := range claimed {
+		perAgent[agentID]++
+	}
+	caps := map[string]int{tight.ID: tight.MaxConcurrency, loose.ID: loose.MaxConcurrency, other.ID: other.MaxConcurrency}
+	for agentID, n := range perAgent {
+		if cap, known := caps[agentID]; !known {
+			t.Fatalf("unknown agent %s claimed", agentID)
+		} else if n > cap {
+			t.Fatalf("agent cap %d exceeded by %d for %s: %v", cap, n, agentID, claimed)
+		}
+	}
+	if got := perAgent[tight.ID]; got > 1 {
+		t.Fatalf("agent cap 1 exceeded by %d: %v", got, claimed)
+	}
+}
+
+// A completed or lease-expired task gives its slot back.
+func TestClaimCapacityFreedByReleaseAndReclaim(t *testing.T) {
+	s, _ := openQueue(t)
+	ws, err := s.CreateWorkspace("lab")
+	if err != nil {
+		t.Fatal(err)
+	}
+	rt := newQueueRuntime(t, s, ws.ID, "rt", 1)
+	agent := newQueueAgent(t, s, ws.ID, "solo", rt.ID, 1)
+	t1 := queueTask(t, s, ws.ID, agent.ID, "one")
+	t2 := queueTask(t, s, ws.ID, agent.ID, "two")
+
+	claimed, ok, err := s.Claim(rt.ID, time.Now())
+	if err != nil || !ok || claimed.ID != t1.ID {
+		t.Fatalf("first claim %+v ok=%v err=%v want task %s", claimed, ok, err, t1.ID)
+	}
+	if _, ok, err := s.Claim(rt.ID, time.Now()); err != nil || ok {
+		t.Fatalf("the agent is at its cap: ok=%v err=%v", ok, err)
+	}
+
+	if err := s.FinishTask(claimed.ID, "completed", ""); err != nil {
+		t.Fatal(err)
+	}
+	second, ok, err := s.Claim(rt.ID, time.Now())
+	if err != nil || !ok || second.ID != t2.ID {
+		t.Fatalf("claim after release %+v ok=%v err=%v want task %s", second, ok, err, t2.ID)
+	}
+
+	n, err := s.RecoverOrphans(time.Now().Add(10 * time.Minute))
+	if err != nil || n != 1 {
+		t.Fatalf("reclaim %d %v", n, err)
+	}
+	if got := taskStatus(t, s, t2.ID); got != "queued" {
+		t.Fatalf("reclaimed task %q must be queued again", got)
+	}
+	third, ok, err := s.Claim(rt.ID, time.Now())
+	if err != nil || !ok || third.ID != t2.ID || third.Attempts != 1 {
+		t.Fatalf("claim after reclaim %+v ok=%v err=%v want task %s with one attempt", third, ok, err, t2.ID)
+	}
+}
+
+// A capacity read that fails refuses the claim; it never allows it.
+func TestClaimRefusesWhenReadFails(t *testing.T) {
+	t.Run("unreadable store", func(t *testing.T) {
+		s, path := openQueue(t)
+		ws, err := s.CreateWorkspace("lab")
+		if err != nil {
+			t.Fatal(err)
+		}
+		rt := newQueueRuntime(t, s, ws.ID, "rt", 2)
+		agent := newQueueAgent(t, s, ws.ID, "a", rt.ID, 2)
+		task := queueTask(t, s, ws.ID, agent.ID, "one")
+		if err := s.Close(); err != nil {
+			t.Fatal(err)
+		}
+		if claimed, ok, err := s.Claim(rt.ID, time.Now()); ok {
+			t.Fatalf("an unreadable store handed out %+v (err=%v)", claimed, err)
+		}
+		again, err := Open(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+		defer again.Close()
+		if got := taskStatus(t, again, task.ID); got != "queued" {
+			t.Fatalf("task %q after a failed capacity read", got)
+		}
+	})
+
+	t.Run("agent lookup fails", func(t *testing.T) {
+		s, _ := openQueue(t)
+		ws, err := s.CreateWorkspace("lab")
+		if err != nil {
+			t.Fatal(err)
+		}
+		rt := newQueueRuntime(t, s, ws.ID, "rt", 2)
+		agent := newQueueAgent(t, s, ws.ID, "a", rt.ID, 2)
+		task := queueTask(t, s, ws.ID, agent.ID, "one")
+		if _, err := s.db.Exec(`DROP TABLE agent`); err != nil {
+			t.Fatal(err)
+		}
+		claimed, ok, err := s.Claim(rt.ID, time.Now())
+		if ok || err != nil {
+			t.Fatalf("a failed agent read must skip the candidate: %+v ok=%v err=%v", claimed, ok, err)
+		}
+		if _, err := s.db.Exec(`CREATE TABLE agent (id TEXT PRIMARY KEY, runtime_id TEXT)`); err != nil {
+			t.Fatal(err)
+		}
+		if got := taskStatus(t, s, task.ID); got != "queued" {
+			t.Fatalf("task %q after a skipped candidate", got)
+		}
+	})
+
+	// The guarded count runs inside the atomic claim: when it cannot be read
+	// the statement must fail rather than hand the task out.
+	t.Run("guard count fails", func(t *testing.T) {
+		s, _ := openQueue(t)
+		ws, err := s.CreateWorkspace("lab")
+		if err != nil {
+			t.Fatal(err)
+		}
+		rt := newQueueRuntime(t, s, ws.ID, "rt", 2)
+		agent := newQueueAgent(t, s, ws.ID, "a", rt.ID, 2)
+		task := queueTask(t, s, ws.ID, agent.ID, "one")
+		if _, err := s.db.Exec(`DROP TABLE agent`); err != nil {
+			t.Fatal(err)
+		}
+		lease := time.Now().UTC().Add(5 * time.Minute).Format(time.RFC3339Nano)
+		res, err := s.db.Exec(claimSQL, rt.ID, lease, now(), task.ID, agent.ID, 2, rt.ID, rt.ID, rt.MaxConcurrency)
+		if err == nil {
+			n, _ := res.RowsAffected()
+			t.Fatalf("a failed guard count claimed %d task(s)", n)
+		}
+		if _, err := s.db.Exec(`CREATE TABLE agent (id TEXT PRIMARY KEY, runtime_id TEXT)`); err != nil {
+			t.Fatal(err)
+		}
+		if got := taskStatus(t, s, task.ID); got != "queued" {
+			t.Fatalf("task %q after a failed guard count", got)
+		}
+	})
+}
@@ -22,13 +22,19 @@ func (s *Store) RunStore() *SQLiteStore {
 	return &SQLiteStore{db: s.db}
 }

+// runSchemaVersion 是 squad run 存储的 schema 版本（squad_run_schema_version 记档）。
+// v1 起步，v2 补齐协作态列（mailbox/handoffs/board）。迁移只做增量：
+// 新库直接建到最新形态，旧库经 ALTER TABLE ADD COLUMN 补列（既有行读出为空默认，不丢数据）。
+const runSchemaVersion = 2
+
 const runSchema = `
 CREATE TABLE IF NOT EXISTS squad_run_state (
  ws_dir TEXT PRIMARY KEY,
  meta TEXT NOT NULL, workflow TEXT NOT NULL, squad TEXT NOT NULL, bb TEXT NOT NULL,
  template TEXT, approved TEXT, repair_budget_remaining INTEGER NOT NULL DEFAULT 0,
  repair_budget_granted INTEGER NOT NULL DEFAULT 0,
-  clarifications TEXT, steering TEXT, usage TEXT, updated_at TEXT NOT NULL
+  clarifications TEXT, steering TEXT, usage TEXT, updated_at TEXT NOT NULL,
+  mailbox TEXT, handoffs TEXT, board TEXT
 );
 CREATE TABLE IF NOT EXISTS squad_run_gate (
  ws_dir TEXT PRIMARY KEY, gate TEXT NOT NULL
@@ -41,13 +47,90 @@ CREATE TABLE IF NOT EXISTS squad_run_checkpoint (
 CREATE TABLE IF NOT EXISTS squad_run_leader (
  ws_dir TEXT PRIMARY KEY, data TEXT NOT NULL
 );
+CREATE TABLE IF NOT EXISTS squad_run_schema_version (
+  id INTEGER PRIMARY KEY CHECK (id = 1), version INTEGER NOT NULL
+);
 `

+// runStateV2Columns 是 v2 在 squad_run_state 上新增的协作态列（列名为代码内常量，非外部输入）。
+var runStateV2Columns = []string{"mailbox", "handoffs", "board"}
+
+// migrate 建表并按版本步进：老库补列后再记版本号，逐次读写都是幂等的
+// （每次 Save/Load 都调；补列只在版本落后时发生，且既有行不被改写）。
 func (s *SQLiteStore) migrate() error {
-	_, err := s.db.Exec(runSchema)
+	if _, err := s.db.Exec(runSchema); err != nil {
+		return err
+	}
+	cur, err := s.runSchemaVersion()
+	if err != nil {
+		return err
+	}
+	if cur >= runSchemaVersion {
+		return nil
+	}
+	if err := s.addRunStateColumns(runStateV2Columns); err != nil {
+		return err
+	}
+	_, err = s.db.Exec(`INSERT INTO squad_run_schema_version (id, version) VALUES (1, ?)
+	  ON CONFLICT(id) DO UPDATE SET version = excluded.version`, runSchemaVersion)
 	return err
 }

+// runSchemaVersion 读取当前 schema 版本；无记录 = 0（v1 及更早的库）。
+func (s *SQLiteStore) runSchemaVersion() (int, error) {
+	var v int
+	err := s.db.QueryRow(`SELECT version FROM squad_run_schema_version WHERE id = 1`).Scan(&v)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return 0, nil
+		}
+		return 0, err
+	}
+	return v, nil
+}
+
+// runStateColumns 读出 squad_run_state 现存列名（迁移判存在性用）。
+func (s *SQLiteStore) runStateColumns() (map[string]bool, error) {
+	rows, err := s.db.Query(`PRAGMA table_info(squad_run_state)`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[string]bool{}
+	for rows.Next() {
+		var cid, notnull, pk int
+		var name, ctype string
+		var dflt sql.NullString
+		if err := rows.Scan(&cid, &name, &ctype, &notnull, &dflt, &pk); err != nil {
+			return nil, err
+		}
+		out[name] = true
+	}
+	return out, rows.Err()
+}
+
+// addRunStateColumns 为缺失的列补 ALTER TABLE ADD COLUMN（既有行保持原值，
+// 新列读出为 NULL → 空默认）。
+func (s *SQLiteStore) addRunStateColumns(cols []string) error {
+	have, err := s.runStateColumns()
+	if err != nil {
+		return err
+	}
+	for _, col := range cols {
+		if have[col] {
+			continue
+		}
+		if _, err := s.db.Exec(`ALTER TABLE squad_run_state ADD COLUMN ` + col + ` TEXT`); err != nil {
+			// 并发迁移：另一进程可能刚补上同一列，复核后不算失败。
+			if again, err2 := s.runStateColumns(); err2 == nil && again[col] {
+				continue
+			}
+			return fmt.Errorf("migrate squad_run_state: add column %s: %w", col, err)
+		}
+	}
+	return nil
+}
+
 // OpenSQLiteStore 独立打开一个 squad.Store（不经 platform.Store 的场景，如测试）。
 func OpenSQLiteStore(path string) (*SQLiteStore, error) {
 	db, err := sql.Open("sqlite", path+"?_pragma=busy_timeout(5000)&_pragma=journal_mode(WAL)")
@@ -70,54 +153,114 @@ func j(v any) string {
 	return string(b)
 }

+// runStateColumn 是 squad_run_state 的一列：列名 + 待编码值（编码失败即报错，
+// 不用空串顶替——空串读回来是「空默认」，等于静默丢数据）。
+type runStateColumn struct {
+	name string
+	val  any
+}
+
+// runStateColumnsOf 列出 RunState 的全部持久化字段（含协作态 mailbox/handoffs/board）。
+func runStateColumnsOf(st *squad.RunState) []runStateColumn {
+	return []runStateColumn{
+		{"meta", st.Meta},
+		{"workflow", st.Plan.Workflow},
+		{"squad", st.Plan.Squad},
+		{"bb", st.BB},
+		{"approved", st.Approved},
+		{"clarifications", st.Clarifications},
+		{"steering", st.Steering},
+		{"usage", st.Usage},
+		{"mailbox", st.Mailbox},
+		{"handoffs", st.Handoffs},
+		{"board", st.Board},
+	}
+}
+
 func (s *SQLiteStore) SaveRun(st *squad.RunState) error {
 	if err := s.migrate(); err != nil {
 		return err
 	}
+	encoded := map[string]string{}
+	for _, col := range runStateColumnsOf(st) {
+		data, err := json.Marshal(col.val)
+		if err != nil {
+			return fmt.Errorf("save run %s: encode column %s: %w", st.Meta.WorkDir, col.name, err)
+		}
+		encoded[col.name] = string(data)
+	}
 	_, err := s.db.Exec(`INSERT INTO squad_run_state
-	  (ws_dir, meta, workflow, squad, bb, template, approved, repair_budget_remaining, repair_budget_granted, clarifications, steering, usage, updated_at)
-	  VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?)
+	  (ws_dir, meta, workflow, squad, bb, template, approved, repair_budget_remaining, repair_budget_granted, clarifications, steering, usage, mailbox, handoffs, board, updated_at)
+	  VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)
 	  ON CONFLICT(ws_dir) DO UPDATE SET
 	    meta=excluded.meta, workflow=excluded.workflow, squad=excluded.squad, bb=excluded.bb,
 	    template=excluded.template, approved=excluded.approved,
 	    repair_budget_remaining=excluded.repair_budget_remaining,
 	    repair_budget_granted=excluded.repair_budget_granted,
 	    clarifications=excluded.clarifications, steering=excluded.steering,
-	    usage=excluded.usage, updated_at=excluded.updated_at`,
-		st.Meta.WorkDir, j(st.Meta), j(st.Plan.Workflow), j(st.Plan.Squad), j(st.BB),
-		st.Template, j(st.Approved), st.RepairBudgetRemaining, st.RepairBudgetGranted,
-		j(st.Clarifications), j(st.Steering), j(st.Usage), now())
+	    usage=excluded.usage, mailbox=excluded.mailbox, handoffs=excluded.handoffs,
+	    board=excluded.board, updated_at=excluded.updated_at`,
+		st.Meta.WorkDir, encoded["meta"], encoded["workflow"], encoded["squad"], encoded["bb"],
+		st.Template, encoded["approved"], st.RepairBudgetRemaining, st.RepairBudgetGranted,
+		encoded["clarifications"], encoded["steering"], encoded["usage"],
+		encoded["mailbox"], encoded["handoffs"], encoded["board"], now())
 	return err
 }

+// LoadRun 按列位置读回 RunState。解码目标逐列具名指定——绝不用 JSON 文本当键：
+// 文本相同的两列（例如两条一模一样的澄清/补充指令）会互相覆盖，把一列丢掉。
 func (s *SQLiteStore) LoadRun(wsDir string) (*squad.RunState, error) {
 	if err := s.migrate(); err != nil {
 		return nil, err
 	}
 	st := &squad.RunState{}
-	var meta, wf, sq, bb, template, approved, clar, steering, usage string
-	row := s.db.QueryRow(`SELECT meta, workflow, squad, bb, template, approved,
-	  repair_budget_remaining, repair_budget_granted, clarifications, steering, usage
-	  FROM squad_run_state WHERE ws_dir = ?`, wsDir)
+	// 列 → 扫描目标（顺序与 SQL 一一对应）。
+	var tpl sql.NullString
 	var granted int
-	err := row.Scan(&meta, &wf, &sq, &bb, &template, &approved,
-		&st.RepairBudgetRemaining, &granted, &clar, &steering, &usage)
-	if err != nil {
+	var meta, wf, sq, bb, approved, clar, steering, usage, mailbox, handoffs, board sql.NullString
+	row := s.db.QueryRow(`SELECT meta, workflow, squad, bb, template, approved,
+	  repair_budget_remaining, repair_budget_granted, clarifications, steering, usage,
+	  mailbox, handoffs, board
+	  FROM squad_run_state WHERE ws_dir = ?`, wsDir)
+	if err := row.Scan(&meta, &wf, &sq, &bb, &tpl, &approved,
+		&st.RepairBudgetRemaining, &granted, &clar, &steering, &usage,
+		&mailbox, &handoffs, &board); err != nil {
 		if err == sql.ErrNoRows {
 			return nil, fmt.Errorf("load run %s: %w", wsDir, os.ErrNotExist)
 		}
 		return nil, err
 	}
 	st.RepairBudgetGranted = granted
-	for pair, dst := range map[string]any{
-		meta: &st.Meta, wf: &st.Plan.Workflow, sq: &st.Plan.Squad, bb: &st.BB,
-		approved: &st.Approved, clar: &st.Clarifications, steering: &st.Steering, usage: &st.Usage,
+	for _, dec := range []struct {
+		col string
+		raw sql.NullString
+		dst any
+	}{
+		{"meta", meta, &st.Meta},
+		{"workflow", wf, &st.Plan.Workflow},
+		{"squad", sq, &st.Plan.Squad},
+		{"bb", bb, &st.BB},
+		{"approved", approved, &st.Approved},
+		{"clarifications", clar, &st.Clarifications},
+		{"steering", steering, &st.Steering},
+		{"usage", usage, &st.Usage},
+		{"mailbox", mailbox, &st.Mailbox},
+		{"handoffs", handoffs, &st.Handoffs},
+		{"board", board, &st.Board},
 	} {
-		if err := json.Unmarshal([]byte(pair), dst); err != nil {
-			return nil, fmt.Errorf("load run %s: parse: %w", wsDir, err)
+		// 旧行缺列（NULL）或空串 = 该字段的空默认，不报错、不影响其他列。
+		if !dec.raw.Valid || dec.raw.String == "" {
+			continue
+		}
+		if err := json.Unmarshal([]byte(dec.raw.String), dec.dst); err != nil {
+			return nil, fmt.Errorf("load run %s: parse column %s: %w", wsDir, dec.col, err)
 		}
 	}
-	st.Template = template
+	st.Template = tpl.String
+	if st.Meta.WorkDir == "" {
+		// 行键（ws_dir）即权威工作区路径：meta 旧数据缺该字段时由行键兜底。
+		st.Meta.WorkDir = wsDir
+	}
 	if st.Approved == nil {
 		st.Approved = map[string]bool{}
 	}
@@ -2,8 +2,14 @@ package platform

 import (
 	"context"
+	"database/sql"
+	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
+	"reflect"
+	"strings"
+	"sync"
 	"testing"
 	"time"

@@ -11,10 +17,12 @@ import (
 )

 func newTestRunState(wsDir string) *squad.RunState {
+	// 固定时刻：多次构造出的状态除工作区路径外完全一致，可直接比对两个后端的读回结果。
+	now := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
 	return &squad.RunState{
 		Meta: squad.RunMetadata{
 			ID: "run_test", Status: "executing", WorkDir: wsDir,
-			CreatedAt: time.Now(), UpdatedAt: time.Now(),
+			CreatedAt: now, UpdatedAt: now,
 			Metadata: map[string]string{"mode": ""},
 		},
 		Plan: squad.SquadPlan{
@@ -27,14 +35,28 @@ func newTestRunState(wsDir string) *squad.RunState {
 			},
 			Squad: squad.SquadConfig{SquadName: "demo"},
 		},
-		BB:                    squad.BlackboardState{Facts: []squad.FactItem{{ID: "f1", Claim: "x"}}},
+		BB:                    squad.BlackboardState{Facts: []squad.FactItem{{ID: "f1", Claim: "x"}}, Fields: map[string]string{"topic": "demo"}},
 		Template:              "tpl",
 		Approved:              map[string]bool{"step:s1": true},
 		RepairBudgetRemaining: 1,
 		RepairBudgetGranted:   0,
-		Clarifications:        []string{"c1"},
-		Steering:              []string{"st1"},
-		Usage:                 map[string]*squad.Usage{"writer": {PromptTokens: 10, CompletionTokens: 5, Steps: 1}},
+		// 内容完全相同的重复条目（且 clarifications 与 steering 的 JSON 文本一样）：
+		// 内容键寻址的解码会把它们互相覆盖掉——这里逐条都要活下来。
+		Clarifications: []string{"同一句澄清", "同一句澄清"},
+		Steering:       []string{"同一句澄清", "同一句澄清"},
+		Mailbox: []squad.MailMessage{
+			{ID: "m1", From: "researcher", To: "writer", Text: "同一句澄清", At: now},
+			{ID: "m2", From: "researcher", To: "writer", Text: "同一句澄清", At: now},
+		},
+		Handoffs: []squad.HandoffRecord{
+			{FromRole: "researcher", FromStep: "s1", ToRole: "writer", Output: "同一句澄清", At: now},
+			{FromRole: "writer", FromStep: "s2", ToRole: "reviewer", Output: "同一句澄清", At: now},
+		},
+		Board: []squad.BoardTask{
+			{ID: "s1", Status: "done", Attempts: 1},
+			{ID: "s2", Status: "claimed", ClaimedBy: "writer", LeaseExpires: now.Add(time.Minute), Attempts: 2},
+		},
+		Usage: map[string]*squad.Usage{"writer": {PromptTokens: 10, CompletionTokens: 5, Steps: 1}},
 	}
 }

@@ -57,9 +79,20 @@ func TestSQLiteStoreRoundTrip(t *testing.T) {
 	if got.Meta.ID != "run_test" || got.Plan.Workflow.WorkflowID != "wf1" || got.Template != "tpl" {
 		t.Fatalf("loaded = %+v", got.Meta)
 	}
-	if !got.Approved["step:s1"] || got.RepairBudgetRemaining != 1 || len(got.Clarifications) != 1 || len(got.Steering) != 1 {
+	if !got.Approved["step:s1"] || got.RepairBudgetRemaining != 1 ||
+		len(got.Clarifications) != 2 || len(got.Steering) != 2 {
 		t.Fatalf("scalars = %+v", got)
 	}
+	// 协作态必须真的落库（F11a）：重启后邮箱/移交/任务板不能静默丢。
+	if len(got.Mailbox) != 2 || got.Mailbox[0].ID != "m1" || got.Mailbox[1].ID != "m2" {
+		t.Fatalf("mailbox = %+v", got.Mailbox)
+	}
+	if len(got.Handoffs) != 2 || got.Handoffs[1].ToRole != "reviewer" {
+		t.Fatalf("handoffs = %+v", got.Handoffs)
+	}
+	if len(got.Board) != 2 || got.Board[1].Status != "claimed" || got.Board[1].ClaimedBy != "writer" {
+		t.Fatalf("board = %+v", got.Board)
+	}
 	if got.Usage["writer"].PromptTokens != 10 {
 		t.Fatalf("usage = %+v", got.Usage)
 	}
@@ -127,7 +160,7 @@ func TestEngineRunsOnSQLiteStore(t *testing.T) {
 		}
 	}
 	backends := squad.NewBackendResolver()
-	backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{Success: false, Logs: "ok", Patches: map[string][]byte{"output.md": []byte(goodReportLite), "notes.md": []byte(goodReportLite), "fields.json": []byte("{}")}}})
+	backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{Success: true, Logs: "ok", Patches: map[string][]byte{"output.md": []byte(goodReportLite), "notes.md": []byte(goodReportLite), "fields.json": []byte("{}")}}})
 	blueprints := squad.NewBlueprintRegistry()
 	if err := blueprints.Register(squad.Blueprint{
 		Key: "demo", Title: "入库演示", Match: []string{"入库演示"},
@@ -185,4 +218,542 @@ func TestEngineRunsOnSQLiteStore(t *testing.T) {
 	}
 }

+// TestEngineFailedResultWithCompletePatchesStaysFailedOnSQLiteStore：F05/U3B 的
+// 不变式在 SQLite-store 集成层同样成立——executor 报失败就是失败，patches
+// （交付物）齐全也不得把运行翻成 completed：文件只是诊断信息，不是权威结果。
+func TestEngineFailedResultWithCompletePatchesStaysFailedOnSQLiteStore(t *testing.T) {
+	runStore, err := OpenSQLiteStore(filepath.Join(t.TempDir(), "runs.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer runStore.Close()
+	dir := t.TempDir()
+	reg := squad.NewRoleRegistry()
+	if err := reg.Register(squad.RoleSpec{Role: "researcher", Backend: "script"}); err != nil {
+		t.Fatal(err)
+	}
+	backends := squad.NewBackendResolver()
+	// 失败结果但携带全套交付物：旧语义会把它提升为成功，U3B 起必须保持失败。
+	backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{
+		Success: false, Logs: "stub-refused",
+		Patches: map[string][]byte{
+			"output.md":   []byte(goodReportLite),
+			"notes.md":    []byte(goodReportLite),
+			"fields.json": []byte("{}"),
+		},
+	}})
+	blueprints := squad.NewBlueprintRegistry()
+	if err := blueprints.Register(squad.Blueprint{
+		Key: "faildemo", Title: "失败演示", Match: []string{"失败演示"},
+		Steps: []squad.StepSpec{{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"output.md"}}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	eng, err := squad.NewEngine(squad.Options{
+		DataDir: dir, Registry: reg, Blueprints: blueprints,
+		Planner:  squad.NewRulePlanner(),
+		Backends: backends, Sink: squad.NopSink{},
+		Store: runStore,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, err := eng.Run(context.Background(), squad.Goal{Text: "做一个失败演示", HitlMode: squad.HitlAuto}, squad.RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status == squad.RunCompleted {
+		t.Fatalf("failed executor result must not complete (patches are diagnostic only)")
+	}
+	if res.Status != squad.RunFailed {
+		t.Fatalf("status = %s, want failed", res.Status)
+	}
+	if res.Plan == nil || res.Plan.Steps[0].Status != "failed" {
+		t.Fatalf("s1 should be failed, plan = %+v", res.Plan)
+	}
+	if !strings.Contains(res.Error, "stub-refused") {
+		t.Fatalf("failure reason not preserved: %q", res.Error)
+	}
+	// 同一断言穿过 Store 端口：SQLite 里落盘的也是失败，而非 completed。
+	st, err := runStore.LoadRun(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if st.Plan.Workflow.Status != "failed" {
+		t.Fatalf("persisted workflow status = %s, want failed", st.Plan.Workflow.Status)
+	}
+	if got := st.Plan.Workflow.Steps[0]; got.Status != "failed" || !strings.Contains(got.Error, "stub-refused") {
+		t.Fatalf("persisted step = %+v", got)
+	}
+	// patches 确已物化（文件齐全），依然不翻转状态：权威是执行结果，不是文件。
+	sub := filepath.FromSlash(st.Plan.Workflow.Steps[0].SubWorkspace)
+	if _, err := os.Stat(filepath.Join(res.WorkDir, sub, "output.md")); err != nil {
+		t.Fatalf("diagnostic patch should still be materialized: %v", err)
+	}
+}
+
 const goodReportLite = "# 测试交付物\n\n## 概述\n\n这是由脚本后端生成的合格交付物，长度超过五十个字符以满足最小质检要求。\n"
+
+// ---- F11a/F11b：SQLite 与 FileStore 同一契约、重复条目不丢、旧库升级 ----
+
+// assertSameRunState 逐字段比对两个后端读回的运行状态（同一契约）。
+// 字段按规范 JSON 比对：时间等值字段（创建/更新时间、消息时刻、租约到期）在两侧
+// 都经历过 JSON 往返，按墙钟等价判定，不受内部表示（时区指针/单调时钟）影响。
+func assertSameRunState(t *testing.T, got, want *squad.RunState) {
+	t.Helper()
+	sameJSON := func(label string, a, b any) {
+		t.Helper()
+		ab, err := json.Marshal(a)
+		if err != nil {
+			t.Fatalf("%s: marshal got: %v", label, err)
+		}
+		bb, err := json.Marshal(b)
+		if err != nil {
+			t.Fatalf("%s: marshal want: %v", label, err)
+		}
+		if string(ab) != string(bb) {
+			t.Fatalf("%s differs:\n got %s\nwant %s", label, ab, bb)
+		}
+	}
+	sameJSON("meta", got.Meta, want.Meta)
+	sameJSON("plan", got.Plan, want.Plan)
+	sameJSON("blackboard", got.BB, want.BB)
+	sameJSON("template", got.Template, want.Template)
+	sameJSON("approved", got.Approved, want.Approved)
+	sameJSON("repair_budget_remaining", got.RepairBudgetRemaining, want.RepairBudgetRemaining)
+	sameJSON("repair_budget_granted", got.RepairBudgetGranted, want.RepairBudgetGranted)
+	sameJSON("clarifications", got.Clarifications, want.Clarifications)
+	sameJSON("steering", got.Steering, want.Steering)
+	sameJSON("mailbox", got.Mailbox, want.Mailbox)
+	sameJSON("handoffs", got.Handoffs, want.Handoffs)
+	sameJSON("board", got.Board, want.Board)
+	sameJSON("usage", got.Usage, want.Usage)
+}
+
+// TestRunStoreContractParity：FileStore 与 SQLiteStore 满足同一份 Store 契约——
+// 同一份运行状态（含 mailbox/handoffs/board）写进去，读回来必须一致。
+func TestRunStoreContractParity(t *testing.T) {
+	sq, err := OpenSQLiteStore(filepath.Join(t.TempDir(), "runs.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sq.Close()
+
+	fileWS := filepath.Join(t.TempDir(), "runs", "run_file")
+	sqlWS := filepath.Join(t.TempDir(), "runs", "run_sql")
+	// FileStore 侧的模板文件由工作区装配期写出（PrepareWorkspace），先补齐，
+	// 使两侧的 SaveRun/LoadRun 覆盖同一批字段。
+	if err := os.MkdirAll(fileWS, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(fileWS, "template.md"), []byte("tpl"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	want := newTestRunState(fileWS)
+	if err := (squad.FileStore{}).SaveRun(want); err != nil {
+		t.Fatal(err)
+	}
+	gotFile, err := (squad.FileStore{}).LoadRun(fileWS)
+	if err != nil {
+		t.Fatal(err)
+	}
+	sqlState := newTestRunState(sqlWS)
+	if err := sq.SaveRun(sqlState); err != nil {
+		t.Fatal(err)
+	}
+	gotSQL, err := sq.LoadRun(sqlWS)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// SQLite 读回的内容与写进去的一致。
+	assertSameRunState(t, gotSQL, sqlState)
+	// 两个后端的读回结果逐字段一致（除工作区路径这一运行自身的键）。
+	assertSameRunState(t, gotFile, want)
+	fileMeta, sqlMeta := gotFile.Meta, gotSQL.Meta
+	fileMeta.WorkDir, sqlMeta.WorkDir = "", ""
+	fileJSON, _ := json.Marshal(fileMeta)
+	sqlJSON, _ := json.Marshal(sqlMeta)
+	if string(fileJSON) != string(sqlJSON) {
+		t.Fatalf("meta parity broken:\n file %s\n  sql %s", fileJSON, sqlJSON)
+	}
+}
+
+// TestSQLiteRunStoreKeepsIdenticalEntries：内容相同的两条澄清/补充指令必须都活着
+// （旧实现拿 JSON 文本当键，两条同文条目只剩一条）。
+func TestSQLiteRunStoreKeepsIdenticalEntries(t *testing.T) {
+	sq, err := OpenSQLiteStore(filepath.Join(t.TempDir(), "runs.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sq.Close()
+	ws := filepath.Join(t.TempDir(), "runs", "run_dup")
+	st := newTestRunState(ws)
+	if err := sq.SaveRun(st); err != nil {
+		t.Fatal(err)
+	}
+	got, err := sq.LoadRun(ws)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !reflect.DeepEqual(got.Clarifications, []string{"同一句澄清", "同一句澄清"}) {
+		t.Fatalf("clarifications = %#v, want both identical entries", got.Clarifications)
+	}
+	if !reflect.DeepEqual(got.Steering, []string{"同一句澄清", "同一句澄清"}) {
+		t.Fatalf("steering = %#v, want both identical entries", got.Steering)
+	}
+	// 邮箱正文相同、ID 不同：两条都要在（不靠正文寻址）。
+	if len(got.Mailbox) != 2 || got.Mailbox[0].ID != "m1" || got.Mailbox[1].ID != "m2" {
+		t.Fatalf("mailbox = %+v", got.Mailbox)
+	}
+}
+
+// legacyRunStateDDL 是 F11 修复前的 squad_run_state（无协作态列），用于验证升级路径。
+const legacyRunStateDDL = `
+CREATE TABLE squad_run_state (
+  ws_dir TEXT PRIMARY KEY,
+  meta TEXT NOT NULL, workflow TEXT NOT NULL, squad TEXT NOT NULL, bb TEXT NOT NULL,
+  template TEXT, approved TEXT, repair_budget_remaining INTEGER NOT NULL DEFAULT 0,
+  repair_budget_granted INTEGER NOT NULL DEFAULT 0,
+  clarifications TEXT, steering TEXT, usage TEXT, updated_at TEXT NOT NULL
+);
+`
+
+// TestSQLiteRunStoreUpgradesLegacySchema：旧库（缺 mailbox/handoffs/board 列）打开后
+// 补列升级，既有一行原样读出，协作态为空默认不报错、不丢其他字段。
+func TestSQLiteRunStoreUpgradesLegacySchema(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "runs.db")
+	legacy, err := sql.Open("sqlite", path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := legacy.Exec(legacyRunStateDDL); err != nil {
+		t.Fatal(err)
+	}
+	ws := filepath.Join(t.TempDir(), "runs", "run_legacy")
+	meta, _ := json.Marshal(squad.RunMetadata{ID: "run_legacy", WorkDir: ws, Status: "executing"})
+	wf, _ := json.Marshal(squad.WorkflowPlan{WorkflowID: "wf_old", Status: "in_progress"})
+	sqCfg, _ := json.Marshal(squad.SquadConfig{SquadName: "legacy"})
+	bb, _ := json.Marshal(squad.BlackboardState{Fields: map[string]string{"k": "v"}})
+	approved, _ := json.Marshal(map[string]bool{"step:s1": true})
+	clar, _ := json.Marshal([]string{"旧澄清", "旧澄清"})
+	usage, _ := json.Marshal(map[string]*squad.Usage{"writer": {PromptTokens: 7, CompletionTokens: 3, Steps: 2}})
+	if _, err := legacy.Exec(`INSERT INTO squad_run_state
+	  (ws_dir, meta, workflow, squad, bb, template, approved, repair_budget_remaining,
+	   repair_budget_granted, clarifications, steering, usage, updated_at)
+	  VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?)`,
+		ws, meta, wf, sqCfg, bb, "旧模板", approved, 2, 1, clar, `["旧补充"]`, usage, "2026-01-01T00:00:00Z"); err != nil {
+		t.Fatal(err)
+	}
+	if err := legacy.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	sq, err := OpenSQLiteStore(path)
+	if err != nil {
+		t.Fatalf("open legacy db: %v", err)
+	}
+	defer sq.Close()
+
+	got, err := sq.LoadRun(ws)
+	if err != nil {
+		t.Fatalf("load upgraded row: %v", err)
+	}
+	// 旧数据无损。
+	if got.Meta.ID != "run_legacy" || got.Plan.Workflow.WorkflowID != "wf_old" ||
+		got.Plan.Squad.SquadName != "legacy" || got.Template != "旧模板" {
+		t.Fatalf("legacy row corrupted: %+v", got)
+	}
+	if !got.Approved["step:s1"] || got.RepairBudgetRemaining != 2 || got.RepairBudgetGranted != 1 {
+		t.Fatalf("legacy scalars lost: %+v", got)
+	}
+	if !reflect.DeepEqual(got.Clarifications, []string{"旧澄清", "旧澄清"}) {
+		t.Fatalf("legacy clarifications lost: %#v", got.Clarifications)
+	}
+	if !reflect.DeepEqual(got.Steering, []string{"旧补充"}) {
+		t.Fatalf("legacy steering lost: %#v", got.Steering)
+	}
+	if got.Usage["writer"].PromptTokens != 7 || got.BB.Fields["k"] != "v" {
+		t.Fatalf("legacy usage/blackboard lost: %+v %+v", got.Usage, got.BB)
+	}
+	// 缺列读出为空默认。
+	if len(got.Mailbox) != 0 || len(got.Handoffs) != 0 || len(got.Board) != 0 {
+		t.Fatalf("new columns should default to empty: %+v %+v %+v", got.Mailbox, got.Handoffs, got.Board)
+	}
+	// 版本已记档，补列只做一次。
+	var version int
+	if err := sq.db.QueryRow(`SELECT version FROM squad_run_schema_version WHERE id = 1`).Scan(&version); err != nil {
+		t.Fatalf("schema version not recorded: %v", err)
+	}
+	if version != runSchemaVersion {
+		t.Fatalf("schema version = %d, want %d", version, runSchemaVersion)
+	}
+	// 升级后写入协作态并再次读回。
+	got.Mailbox = []squad.MailMessage{{ID: "m1", From: "a", To: "b", Text: "升级后写入"}}
+	got.Board = []squad.BoardTask{{ID: "s1", Status: "done"}}
+	got.Meta.Status = "completed"
+	if err := sq.SaveRun(got); err != nil {
+		t.Fatalf("save after upgrade: %v", err)
+	}
+	again, err := sq.LoadRun(ws)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(again.Mailbox) != 1 || again.Mailbox[0].Text != "升级后写入" ||
+		len(again.Board) != 1 || again.Board[0].ID != "s1" {
+		t.Fatalf("post-upgrade collaboration state lost: %+v %+v", again.Mailbox, again.Board)
+	}
+	if !reflect.DeepEqual(again.Clarifications, []string{"旧澄清", "旧澄清"}) {
+		t.Fatalf("upgrade write clobbered legacy clarifications: %#v", again.Clarifications)
+	}
+}
+
+// TestSQLiteStoreKeepsCollaborationStateAcrossRestart：进程重启（新连接）后，
+// 邮箱消息、所有权移交记录与任务板状态都还在——恢复不必从头重建协作上下文。
+func TestSQLiteStoreKeepsCollaborationStateAcrossRestart(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "runs.db")
+	runStore, err := OpenSQLiteStore(dbPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer runStore.Close()
+
+	dir := t.TempDir()
+	reg := squad.NewRoleRegistry()
+	for _, role := range []string{"researcher", "writer", "reviewer"} {
+		if err := reg.Register(squad.RoleSpec{Role: role, Backend: "script"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := squad.NewBackendResolver()
+	backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{
+		Success: true, Logs: "ok",
+		Patches: map[string][]byte{"output.md": []byte(goodReportLite), "notes.md": []byte(goodReportLite), "fields.json": []byte("{}")},
+	}})
+	blueprints := squad.NewBlueprintRegistry()
+	if err := blueprints.Register(squad.Blueprint{
+		Key: "handoff", Title: "移交演示", Match: []string{"移交演示"},
+		Steps: []squad.StepSpec{
+			{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"notes.md", "fields.json"}},
+			{ID: "s2", Role: "writer", Title: "成文", Targets: []string{"output.md"}, Handoff: "reviewer"},
+			{ID: "s3", Role: "reviewer", Title: "终审", Targets: []string{"output.md"}},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := blueprints.Register(squad.Blueprint{
+		Key: "board", Title: "任务板演示", Match: []string{"任务板演示"},
+		Steps: []squad.StepSpec{
+			{ID: "a", Role: "researcher", Title: "任务a", Targets: []string{"notes.md"}},
+			{ID: "b", Role: "writer", Title: "任务b", Needs: []string{"a"}, Targets: []string{"output.md"}},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	eng, err := squad.NewEngine(squad.Options{
+		DataDir: dir, Registry: reg, Blueprints: blueprints,
+		Planner:  squad.NewRulePlanner(),
+		Backends: backends, Sink: squad.NopSink{},
+		Store: runStore,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	ctx := context.Background()
+	handed, err := eng.Run(ctx, squad.Goal{Text: "做一个移交演示", HitlMode: squad.HitlAuto}, squad.RunOptions{})
+	if err != nil || handed.Status != squad.RunCompleted {
+		t.Fatalf("handoff run: %v %+v", err, handed)
+	}
+	if err := eng.PostMessage(ctx, handed.RunID, squad.MailMessage{From: "human", To: "writer", Text: "补充口径"}); err != nil {
+		t.Fatalf("post message: %v", err)
+	}
+	boarded, err := eng.Run(ctx, squad.Goal{Text: "做一个任务板演示", HitlMode: squad.HitlAuto},
+		squad.RunOptions{Topology: squad.TopologyTaskBoard})
+	if err != nil || boarded.Status != squad.RunCompleted {
+		t.Fatalf("board run: %v %+v", err, boarded)
+	}
+
+	// 重启 = 新连接（新进程会重新 migrate）。
+	restarted, err := OpenSQLiteStore(dbPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer restarted.Close()
+
+	gotHanded, err := restarted.LoadRun(handed.WorkDir)
+	if err != nil {
+		t.Fatalf("reload handoff run: %v", err)
+	}
+	if len(gotHanded.Handoffs) != 1 || gotHanded.Handoffs[0].FromRole != "writer" || gotHanded.Handoffs[0].ToRole != "reviewer" {
+		t.Fatalf("handoffs lost across restart: %+v", gotHanded.Handoffs)
+	}
+	if len(gotHanded.Mailbox) != 1 || gotHanded.Mailbox[0].Text != "补充口径" || gotHanded.Mailbox[0].To != "writer" {
+		t.Fatalf("mailbox lost across restart: %+v", gotHanded.Mailbox)
+	}
+	gotBoarded, err := restarted.LoadRun(boarded.WorkDir)
+	if err != nil {
+		t.Fatalf("reload board run: %v", err)
+	}
+	if len(gotBoarded.Board) != 2 {
+		t.Fatalf("board lost across restart: %+v", gotBoarded.Board)
+	}
+	for _, task := range gotBoarded.Board {
+		if task.Status != "done" {
+			t.Fatalf("board task %s = %s, want done", task.ID, task.Status)
+		}
+	}
+}
+
+// faultingStore 在真实 SQLiteStore 之上注入落盘故障（其余方法透传）。
+type faultingStore struct {
+	inner          squad.Store
+	failCheckpoint bool
+	failTerminal   bool
+}
+
+func (f *faultingStore) SaveRun(st *squad.RunState) error {
+	if f.failTerminal && squad.RunStatus(st.Meta.Status).Terminal() {
+		return fmt.Errorf("injected sqlite run state save failure (%s)", st.Meta.Status)
+	}
+	return f.inner.SaveRun(st)
+}
+
+func (f *faultingStore) LoadRun(wsDir string) (*squad.RunState, error) { return f.inner.LoadRun(wsDir) }
+
+func (f *faultingStore) SaveCheckpoint(wsDir string, cp squad.Checkpoint) error {
+	if f.failCheckpoint {
+		return fmt.Errorf("injected sqlite checkpoint save failure (superstep %d)", cp.Superstep)
+	}
+	return f.inner.SaveCheckpoint(wsDir, cp)
+}
+
+func (f *faultingStore) LoadLatestCheckpoint(wsDir string) (*squad.Checkpoint, error) {
+	return f.inner.LoadLatestCheckpoint(wsDir)
+}
+
+func (f *faultingStore) SaveGate(wsDir string, gate *squad.HitlGate) error {
+	return f.inner.SaveGate(wsDir, gate)
+}
+
+func (f *faultingStore) LoadGate(wsDir string) (*squad.HitlGate, error) {
+	return f.inner.LoadGate(wsDir)
+}
+
+func (f *faultingStore) SaveLeader(wsDir string, data []byte) error {
+	return f.inner.SaveLeader(wsDir, data)
+}
+
+func (f *faultingStore) LoadLeader(wsDir string) ([]byte, error) { return f.inner.LoadLeader(wsDir) }
+
+// eventTally 收集事件类型计数（断言没有成功事件）。
+type eventTally struct {
+	mu   sync.Mutex
+	seen []string
+	sink squad.ProgressSink
+}
+
+func newEventTally() *eventTally {
+	t := &eventTally{}
+	t.sink = squad.ProgressFunc(func(ev squad.Event) {
+		t.mu.Lock()
+		defer t.mu.Unlock()
+		t.seen = append(t.seen, ev.Type)
+	})
+	return t
+}
+
+func (t *eventTally) has(typ string) bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	for _, got := range t.seen {
+		if got == typ {
+			return true
+		}
+	}
+	return false
+}
+
+func (t *eventTally) types() []string {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return append([]string(nil), t.seen...)
+}
+
+// TestSQLiteStoreSurfacesSaveFailures：SQLite 后端上的落盘故障必须可见——
+// 三档 checkpoint 模式下都不发 run_completed，运行要么报错、要么以失败收场。
+func TestSQLiteStoreSurfacesSaveFailures(t *testing.T) {
+	for _, mode := range []string{"sync", "async", "on_exit"} {
+		for _, fault := range []string{"checkpoint", "terminal"} {
+			t.Run(mode+"/"+fault, func(t *testing.T) {
+				runStore, err := OpenSQLiteStore(filepath.Join(t.TempDir(), "runs.db"))
+				if err != nil {
+					t.Fatal(err)
+				}
+				defer runStore.Close()
+				dir := t.TempDir()
+				reg := squad.NewRoleRegistry()
+				for _, role := range []string{"researcher", "writer", "reviewer"} {
+					if err := reg.Register(squad.RoleSpec{Role: role, Backend: "script"}); err != nil {
+						t.Fatal(err)
+					}
+				}
+				backends := squad.NewBackendResolver()
+				backends.RegisterInstance("script", stubExecutor{result: &squad.ExecResult{
+					Success: true, Logs: "ok",
+					Patches: map[string][]byte{"output.md": []byte(goodReportLite), "notes.md": []byte(goodReportLite), "fields.json": []byte("{}")},
+				}})
+				blueprints := squad.NewBlueprintRegistry()
+				if err := blueprints.Register(squad.Blueprint{
+					Key: "demo", Title: "故障演示", Match: []string{"故障演示"},
+					Steps: []squad.StepSpec{
+						{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"notes.md", "fields.json"}},
+						{ID: "s2", Role: "writer", Title: "成文", Targets: []string{"output.md"}},
+						{ID: "s3", Role: "reviewer", Title: "终审", Targets: []string{"output.md"}},
+					},
+				}); err != nil {
+					t.Fatal(err)
+				}
+				store := &faultingStore{inner: runStore,
+					failCheckpoint: fault == "checkpoint", failTerminal: fault == "terminal"}
+				tally := newEventTally()
+				eng, err := squad.NewEngine(squad.Options{
+					DataDir: dir, Registry: reg, Blueprints: blueprints,
+					Planner: squad.NewRulePlanner(), Backends: backends,
+					Sink: tally.sink, Store: store,
+				})
+				if err != nil {
+					t.Fatal(err)
+				}
+				res, err := eng.Run(context.Background(),
+					squad.Goal{Text: "做一个故障演示", HitlMode: squad.HitlAuto},
+					squad.RunOptions{CheckpointMode: mode})
+				if tally.has(squad.EvRunCompleted) {
+					t.Fatalf("run_completed emitted despite %s save failure: %v", fault, tally.types())
+				}
+				if fault == "checkpoint" && mode == "sync" {
+					if err == nil {
+						t.Fatalf("sync boundary failure must surface as error: %+v", res)
+					}
+					if !strings.Contains(err.Error(), "persistence failure") {
+						t.Fatalf("error not marked as persistence failure: %v", err)
+					}
+					return
+				}
+				if err != nil {
+					t.Fatalf("%s/%s: unexpected infra error: %v", mode, fault, err)
+				}
+				if res.Status != squad.RunFailed {
+					t.Fatalf("%s/%s: want failed, got %s", mode, fault, res.Status)
+				}
+				if !strings.Contains(res.Error, "persistence failure") {
+					t.Fatalf("%s/%s: failure not explained: %q", mode, fault, res.Error)
+				}
+				if !tally.has(squad.EvRunFailed) {
+					t.Fatalf("%s/%s: no run_failed event: %v", mode, fault, tally.types())
+				}
+			})
+		}
+	}
+}
@@ -1,24 +1,26 @@
 # 通用规范采纳记录

- 项目/负责人：Agent-Squad / 仓库所有者 weichen
+- 项目/负责人：Agent-Squad / weichen
 - 状态：active
 - 规范位置：`deps/development-guidelines`
 - 规范性文档目录：`policy/`
- 过程文档位置：`docs/records/`（项目内固定目录豁免。个人维护的编排库，任务主记录仍在 Gitea Issue/PR。过程记录不进规范库，也不写成与 Issue 平行的第二状态。）
- 版本权威：父仓 gitlink；可读标签：无（内容版本 `0.1.0-draft.6`，提交 `26b562d2e34fbc560711296a309c91c74ffc55b1`）
- 批准/生效依据：本仓接入提交
- 替代范围和迁移说明：本仓此前没有根 `AGENTS.md`。Agent-Native 子模块已从仓库根迁到 `deps/agent-native`，迁移说明见 `docs/records/2026-09-29-guidelines-and-observability.md`。
+- 过程文档位置：任务卡在 `.task/`（`.task/feature-governance-adoption.md`）；任务主记录在 Gitea PR 中维护。既有过程记录仍按项目内固定目录豁免留在 `docs/records/`。过程记录不进规范库，也不写成与 Issue/PR 平行的第二状态。
+- 版本权威：本仓 gitlink deps/development-guidelines @ 2688fd2be6a5（完整 `2688fd2be6a52c0bf110e7e6366b6f291ab5e4eb`，内容版本 `0.1.0-draft.12`，无已发布标签）
+- 批准/生效依据：`feature/governance-adoption PR（Agent-Squad）`
+- 替代范围和迁移说明：本次替代 2026-09-29 采纳的固定版本 `26b562d2e34f`（内容版本 `0.1.0-draft.6`），规则条目按 `0.1.0-draft.12` 重新核对。新增 `AGENT-EXECUTION.md` 执行路由、`deps/development-manager` gitlink（`bd3a7de57a3a`）、`scripts/dm`、`scripts/agent-exec`、`scripts/task`、`scripts/hooks/pre-commit`、`.gitea/workflows/governance.yml` 与任务卡模板。不改业务代码，不改既有 `.gitea/workflows/ci.yml`。此前迁移说明见 `docs/records/2026-09-29-guidelines-and-observability.md`。

 | 规范/工作流 | 适用条件与范围 | 采纳或不适用理由 | 执行方式/证据 |
 |---|---|---|---|
 | core/baseline.md | 日常工作 | 编排库对外有失败关闭语义 | manual |
 | workflows/delivery.md、debugging.md、handoff.md | 交付、排障与交接 | 跨仓组合需要可交接记录 | manual |
 | standards/collaboration.md | 并行共同起点与组合升级 | 父仓 gitlink 锁定 Agent-Native 与基础库 | manual |
-| standards/verification.md | 验证与结论 | `GOWORK=off go test` | manual；命令见 PROJECT.md |
+| standards/merge-queue.md | 多个就绪任务的最终同步、CI、Review 与合入调度 | 多个 agent 单元并行开发，合入须串行同步；协调者为仓库所有者 | manual |
+| standards/verification.md | 验证与结论 | `GOWORK=off go test` | manual；命令见 PROJECT.md（含 `scripts/agent-exec validate --repo .`、`scripts/dm governance check --repo .`） |
 | workflows/design-verification.md、standards/testability.md、standards/failure-containment.md | 规划、执行、HITL | conformance 覆盖公开 API | manual |
 | standards/module-contracts.md | squad 与 agent-native 边界 | 依赖方向只允许 squad → agent-native | manual |
 | standards/observability.md | HTTP 信封、trace、事件 | 接入错误码库与日志库 | manual |
-| standards/agent-execution.md | 角色绑定的执行后端 | 本库不实现 Agent | manual |
+| standards/agent-execution.md | 工程 Agent CLI 执行路由、会话隔离与受控集成 | 凡允许 AI 协助工程变更必须采纳；路由配置见 AGENT-EXECUTION.md；本库不实现 Agent | 执行方式：AGENT-EXECUTION.md + scripts/agent-exec（validate / admit）；证据：`governance check + 正负例（.task/feature-governance-adoption.md）` |
+| standards/tool-adaptation.md | 为具体工具创建规则与钩子 | 本次引入 `scripts/hooks/pre-commit`（启用由 host/人执行，`core.hooksPath`） | manual；启用后以本仓 pre-commit 检查为证据 |
 | standards/multi-repo.md | 多级 submodule | 本仓 replace 是编译时的唯一依赖解析 | manual |
 | standards/data-reliability.md | 运行目录、HITL 恢复、幂等回应 | 磁盘为权威 | manual |
 | standards/security.md | bash 权限挂起、脱敏 | 事件落盘沿用现有脱敏；凭证走环境变量 | manual |
@@ -31,14 +33,17 @@

 | 原规则/章节 | 项目采用方式 | 原因与风险 | 负责人/复查条件 |
 |---|---|---|---|
-| documentation.md 过程仓默认独立仓库 | 过程记录放 `docs/records/` | 同 Agent-Native：避免个人项目再维护一个过程仓。风险是证据与代码同库 | 多人协作或对外发布时复查 |
+| documentation.md 过程仓默认独立仓库 | 其他过程记录放 `docs/records/`（既有过程记录的项目内固定目录豁免） | 同 Agent-Native：避免个人项目再维护一个过程仓。风险是过程记录与代码同库 | 多人协作或对外发布时复查 |
+| documentation.md 过程仓默认独立仓库 | 任务卡与交付证据放 `.task/`（模板 `.task/templates/task-card.md`） | 任务状态仍只认 Gitea Issue/PR，`.task/` 不写成平行状态；风险是任务卡与代码同库 | 多人协作或对外发布时复查 |
 | multi-repo.md 组件说明不手写第二套 SHA | PROJECT.md 只写路径与角色，SHA 以 `git submodule status` 为准 | 手抄 SHA 会与 gitlink 漂移 | 每次依赖升级核对脚本 |

 ## 本次升级或首次接入验证

- `deps/development-guidelines` 固定 `26b562d2e34fbc560711296a309c91c74ffc55b1`。
- 执行方式为 manual。
- 回退该规范挂载时，同时确认 Agent-Native 子模块路径是否已迁移；只还原规范 gitlink 不会把 `deps/agent-native` 移回根目录。
+- `deps/development-guidelines` 固定 `2688fd2be6a52c0bf110e7e6366b6f291ab5e4eb`（内容版本 `0.1.0-draft.12`）；`deps/development-manager` 固定 `bd3a7de57a3a9e1a4495b94c5fa5c08d71e7be8c`。
+- 工程 Agent 执行：`AGENT-EXECUTION.md` 为唯一路由权威；AI 实施单元只经 `scripts/agent-exec run` 执行，并经 `scripts/agent-exec apply` 集成。`scripts/agent-exec` 与 `scripts/task` 只是转调 `scripts/dm` 的薄入口。
+- 登记为 active 的前置证据：一次成功执行正例（有效 `completed` 收据经 transport/admission 集成）；缺配置、无效终态收据、越权路径、未知副作用四类负例均被拒绝或 fail-stop。证据按 `.task/feature-governance-adoption.md` 记录。
+- 入口核验：`scripts/dm governance check --repo .` 与 `scripts/agent-exec validate --repo .`（PROJECT.md「验证」节）。CI 由 `.gitea/workflows/governance.yml` 执行；既有 `ci.yml` 未改。
+- 回退该规范挂载时，恢复上一已批准 gitlink `26b562d2e34f`，并同时还原本次新增的治理脚本与 `governance.yml`；如已启用 `core.hooksPath`，需一并关闭。只还原规范 gitlink 不会把 `deps/agent-native` 移回根目录。

 ### 2026-09-29 追加：agent-intent 可选意图闸门

@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+# 共享 Agent 执行器薄入口：转调 scripts/dm 的 agent-exec 子命令（validate / run / apply / admit / cleanup）。
+set -euo pipefail
+exec "$(dirname "$0")/dm" agent-exec "$@"
@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+# 共享 dm-server 薄入口：读取 deps/development-manager 的 gitlink 钉住提交，
+# 按该提交构建并缓存 dm-server，再 exec 其子命令（agent-exec、task、governance）。执行语义与退出码以该二进制为准。
+set -euo pipefail
+
+root=$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd -P)
+dm="$root/deps/development-manager"
+
+die() { echo "dm: $*" >&2; exit 2; }
+
+# git 钩子会导出仓库本地变量（GIT_DIR、GIT_INDEX_FILE 等），它们会让 `git -C "$dm"` 解析到外层仓库。
+# 所有针对 DM 检出的 git 与 go 调用都在清除这些变量后执行；读取外层 gitlink 与最终 exec 仍保留继承环境。
+dm_local_env="GIT_DIR GIT_WORK_TREE GIT_INDEX_FILE GIT_PREFIX GIT_OBJECT_DIRECTORY GIT_ALTERNATE_OBJECT_DIRECTORIES GIT_COMMON_DIR GIT_GRAFT_FILE GIT_NAMESPACE GIT_CONFIG GIT_CONFIG_PARAMETERS GIT_CONFIG_COUNT GIT_SHALLOW_FILE GIT_IMPLICIT_WORK_TREE $(git rev-parse --local-env-vars 2>/dev/null || true)"
+dm_clean_env() { for v in $dm_local_env; do unset "$v"; done; }
+dmgit() ( dm_clean_env; git -C "$dm" "$@" )
+
+# 1. 钉住版本只来自索引中的 gitlink（mode 160000），不读取工作区文件。
+entry=$(git -C "$root" ls-files -s -- deps/development-manager 2>/dev/null || true)
+[ -n "$entry" ] || die "索引中缺少 deps/development-manager 的 gitlink，无法确定钉住的 DM 版本"
+read -r mode pin _stage _path <<<"$entry"
+[ "$mode" = 160000 ] || die "deps/development-manager 不是 gitlink（mode=$mode）"
+[[ "$pin" =~ ^[0-9a-f]{40}$ ]] || die "gitlink 提交格式无效：$pin"
+
+# 2. 检出必须存在且确为 DM 独立仓库（避免 git -C 误解析到外层仓库），HEAD 必须等于钉住提交。
+[ -e "$dm/.git" ] || die "deps/development-manager 未检出；请按钉住提交 $pin 检出该 gitlink"
+top=$(dmgit rev-parse --show-toplevel 2>/dev/null || true)
+[ -n "$top" ] && [ "$top" = "$(cd "$dm" && pwd -P)" ] \
+  || die "deps/development-manager 不是独立检出；请按钉住提交 $pin 检出该 gitlink"
+head=$(dmgit rev-parse HEAD 2>/dev/null || true)
+[ "$head" = "$pin" ] \
+  || die "deps/development-manager HEAD（${head:-未知}）与钉住提交 $pin 不一致；请按钉住提交检出"
+
+# 3. 检出工作树必须干净（忽略嵌套 submodule 状态）。
+dirty=$(dmgit status --porcelain --ignore-submodules=all) \
+  || die "无法读取 deps/development-manager 的工作树状态"
+[ -z "$dirty" ] || die "deps/development-manager 存在未提交改动，拒绝运行（请先清理该检出）"
+
+# 4. 二进制按钉住提交缓存；缺失时在缓存目录内临时构建，完成后原子改名，避免留下半成品。
+cache="${DM_CACHE:-${XDG_CACHE_HOME:-$HOME/.cache}/development-manager/dm-server}"
+bin="$cache/dm-server-$pin"
+if [ ! -x "$bin" ]; then
+  command -v go >/dev/null 2>&1 || die "缺少 go，无法构建 dm-server（$pin）"
+  mkdir -p "$cache"
+  cache=$(cd "$cache" && pwd -P)
+  bin="$cache/dm-server-$pin"
+  work=$(mktemp -d "$cache/build-$pin.XXXXXX")
+  if ! (dm_clean_env; cd "$dm" && GOWORK=off go build -trimpath -o "$work/dm-server" ./cmd/dm-server); then
+    rm -rf "$work"
+    die "构建 dm-server 失败（$pin）"
+  fi
+  mv -f "$work/dm-server" "$bin"
+  rm -rf "$work"
+fi
+
+exec "$bin" "$@"
@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# 本地检查（pre-commit）：集成分支直提拒绝 + 分支命名 + 本地完整性核对。权威准入在 PR（baseline.md:26）。
+branch=$(git rev-parse --abbrev-ref HEAD)
+case "$branch" in
+  main|master)
+    echo "pre-commit: 拒绝——集成分支 ${branch} 不得直接提交，走 feature 分支 + PR（collaboration.md）" >&2
+    exit 1 ;;
+esac
+echo "$branch" | grep -qE '^(feature|fix|hotfix|release)/[A-Za-z0-9._-]+$' || {
+  echo "pre-commit: 拒绝——分支名须为 feature|fix|hotfix|release/<slug>，当前：${branch}" >&2
+  exit 1
+}
+root=$(git rev-parse --show-toplevel)
+# admit 只是本地完整性/篡改证据（SELF）：同一 OS 用户可读取本地 HMAC 密钥并绕过本钩子，
+# 它不是 GATE，也不是可信身份。缺失或无效的本地记录仍拒绝；交付信任来自精确 HEAD 的 CI 或独立 REPLAY。
+"$root/scripts/agent-exec" admit --repo "$root"
@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+# 通用任务入口：转调 scripts/dm 的 task 子命令（triage / start / pr-check）。
+set -euo pipefail
+exec "$(dirname "$0")/dm" task "$@"
@@ -2,12 +2,18 @@ package squad

 import (
 	"context"
+	"encoding/json"
 	"os"
 	"path/filepath"
 	"sync"
 	"testing"
+	"time"
 )

+// jsonForTest 包装 json.Marshal 让测试里构型参数不拉锁。是仅为占用 import 走
+// 公共路径，不污染生产代码。
+func jsonForTest(v any) ([]byte, error) { return json.Marshal(v) }
+
 // perStepRecordingBackend 按子工作区归因计数（steps 的 role→tasks/NN_role 布局）。
 type perStepRecordingBackend struct {
 	mu    sync.Mutex
@@ -59,6 +65,9 @@ func newAttemptTestEngine(t *testing.T, backend *perStepRecordingBackend, dir st
 	return eng
 }

+// flipToInterrupted 把该步骤手动置为 running，同时清除 trusted terminal
+// record（模拟「崩溃于执行中」状态）。清除让该路径能够测试 F05 走的
+// 「无 record」分支。
 func flipToInterrupted(t *testing.T, workDir string, stepIdx int) {
 	t.Helper()
 	st, err := loadRunState(workDir)
@@ -75,6 +84,16 @@ func flipToInterrupted(t *testing.T, workDir string, stepIdx int) {
 	if err := st.save(); err != nil {
 		t.Fatal(err)
 	}
+	// 同时清除该步骤的 trusted terminal record（模拟未达记录写入点崩溃）。
+	terminals := loadStepTerminalsForWorkDir(workDir)
+	if terminals != nil {
+		stepID := st.Plan.Workflow.Steps[stepIdx].ID
+		if _, ok := terminals[stepID]; ok {
+			delete(terminals, stepID)
+			data, _ := jsonForTest(terminals)
+			_ = os.WriteFile(filepath.Join(workDir, "step_terminals.json"), data, 0o644)
+		}
+	}
 }

 // TestAttemptIDFormat：attempt id 稳定落盘且形如 runID/stepID/N（Q1 WP3）。
@@ -98,8 +117,15 @@ func TestAttemptIDFormat(t *testing.T) {
 }

 // TestKilledAtSuperstepBoundarySideEffectsOnce：杀进程于 superstep 边界后恢复，
-// 已完成步骤不重新执行（执行次数=1）；running 且产物齐全的步骤按 attempt 判定
-// 完成、不 rerun——Q1 验收第一条（WP3）。
+// 已完成步骤不重新执行（执行次数=1）。
+//
+// F05/U3b：running 崩溃、文件齐全、却无 trusted terminal record（U3b 前被视为
+// “副作用真发生”自动完成；U3b 后仅记录 step.Status=needs_reconciliation，
+// 后台不再自动 rerun/完成。本测试验证该迁变后的语义不变量：
+//
+//   - running 且无 record：需要 reconciliation，不隐默 rerun；
+//   - file presence 不替 trusted record 完成步骤。
+//   - 若事先写入 trusted record（手工合成），即便文件缺失也恢复为 completed。
 func TestKilledAtSuperstepBoundarySideEffectsOnce(t *testing.T) {
 	backend := &perStepRecordingBackend{}
 	dir := t.TempDir()
@@ -125,22 +151,78 @@ func TestKilledAtSuperstepBoundarySideEffectsOnce(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if res.Status != RunCompleted {
-		t.Fatalf("resume: %+v", res)
+	// F05：需 reconciliation（auto 模式），不等于 completed。
+	if res.Status != RunWaitingReconciliation {
+		t.Fatalf("want waiting_reconciliation (auto mode), got %s", res.Status)
 	}
+	if res.Plan == nil || len(res.Plan.Steps) == 0 {
+		t.Fatal("missing plan")
+	}
+	var recon int
 	for _, s := range res.Plan.Steps {
-		if s.Status != "completed" && s.Status != "skipped" {
-			t.Fatalf("step %s not finished: %s", s.ID, s.Status)
+		if s.ID == "s2" && s.Status == StepStatusNeedsReconciliation {
+			recon++
 		}
+	}
+	if recon != 1 {
+		t.Fatalf("step s2 not in needs_reconciliation (got plan=%+v)", res.Plan.Steps)
+	}
+	// 重要：重复执行的次数不应发生隐默 rerun（本测试用 auto 模式运行走到 reconciliation）。
+	for _, s := range res.Plan.Steps {
 		got := backend.countOf(filepath.Base(filepath.FromSlash(s.SubWorkspace)))
 		if got != before[s.ID] {
-			t.Fatalf("step %s executed %d times after kill (before=%d)——已完成/running 步骤被重复执行", s.ID, got, before[s.ID])
+			t.Fatalf("step %s executed %d times after kill (before=%d)——running 步骤被隐默 rerun", s.ID, got, before[s.ID])
 		}
 	}
 }

-// TestInterruptedMissingProductsRetries：running 崩溃且产物缺失时按 auto 一次重试（WP3）。
-func TestInterruptedMissingProductsRetries(t *testing.T) {
+// TestTrustedRecordCompletesMissingFiles：trusted terminal record 存在 + 文件缺失
+// → 恢复为 completed（files 仍作诊断，state 是权威）。
+func TestTrustedRecordCompletesMissingFiles(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "trusted record 完成", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	// 制造「第二步走 running 但 trusted record 已存在」的场景：先写入 record，再
+	// flip 到 running + 删 file。
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	sub2 := filepath.FromSlash(st.Plan.Workflow.Steps[1].SubWorkspace)
+	flipToInterrupted(t, res.WorkDir, 1)
+	if err := os.Remove(filepath.Join(res.WorkDir, sub2, "output.md")); err != nil {
+		t.Fatal(err)
+	}
+	// 手工写入 trusted record（这能覆盖「后端状态已落盘但文件被外部删」该返回
+	// 已完成的合法语义）。
+	terminals := map[string]stepTerminalRecord{
+		"s2": {StepID: "s2", AttemptID: st.Plan.Workflow.Steps[1].AttemptID, Terminal: "completed", SideEffects: string(ExecSideEffectsNone), At: time.Now()},
+	}
+	data, _ := jsonForTest(terminals)
+	if err := os.WriteFile(filepath.Join(res.WorkDir, "step_terminals.json"), data, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	before := backend.countOf(filepath.Base(sub2))
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res, err = eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want completed via trusted record, got %s", res.Status)
+	}
+	if got := backend.countOf(filepath.Base(sub2)); got != before {
+		t.Fatalf("trusted-record step reran (before=%d after=%d)——record 需被信任为 completed", before, got)
+	}
+}
+
+// TestInterruptedMissingProductsEntersReconciliation：F05/U3b：running 崩溃 + 无
+// trusted record + 文件缺失 → auto 模式不再隐默 rerun，进入 waiting_reconciliation。
+func TestInterruptedMissingProductsEntersReconciliation(t *testing.T) {
 	backend := &perStepRecordingBackend{}
 	dir := t.TempDir()
 	eng := newAttemptTestEngine(t, backend, dir)
@@ -148,7 +230,6 @@ func TestInterruptedMissingProductsRetries(t *testing.T) {
 	if err != nil || res.Status != RunCompleted {
 		t.Fatalf("setup run: %v %+v", err, res)
 	}
-	// 第二步停 running 并删掉其交付物（副作用未发生的形态）。
 	flipToInterrupted(t, res.WorkDir, 1)
 	st, err := loadRunState(res.WorkDir)
 	if err != nil {
@@ -158,16 +239,24 @@ func TestInterruptedMissingProductsRetries(t *testing.T) {
 	if err := os.Remove(filepath.Join(res.WorkDir, sub, "output.md")); err != nil {
 		t.Fatal(err)
 	}
+	before := backend.countOf(filepath.Base(sub))
 	eng2 := newAttemptTestEngine(t, backend, dir)
 	res, err = eng2.Resume(context.Background(), res.RunID)
 	if err != nil {
 		t.Fatal(err)
 	}
-	if res.Status != RunCompleted {
-		t.Fatalf("resume: %+v", res)
+	if res.Status != RunWaitingReconciliation {
+		t.Fatalf("want waiting_reconciliation, got %s", res.Status)
 	}
-	got := backend.countOf(filepath.Base(sub))
-	if got != 2 {
-		t.Fatalf("missing-product step executed %d times, want 2（auto 一次重试）", got)
+	// 后续 Resume 仍旧是 reconciliation，不再 rerun。
+	res2, err := eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunWaitingReconciliation {
+		t.Fatalf("second resume shouldn't rerun, got %s", res2.Status)
+	}
+	if got := backend.countOf(filepath.Base(sub)); got != before {
+		t.Fatalf("missing-files step reran (before=%d after=%d)——record 缺失不应 rerun", before, got)
 	}
 }
@@ -1,7 +1,9 @@
 package squad

 import (
+	"context"
 	"fmt"
+	"strings"
 	"sync"

 	"error-code-system/errcode"
@@ -16,11 +18,20 @@ type BackendFactory func(role RoleSpec) (Executor, error)
 // agent-native 自研内核，产品也可注册 "pi-agent" 等外部智能体工厂。
 // 未知名字 fail-closed，绝不静默落 stub。角色声明 MaxConcurrentTasks 时
 // 以并发限流包装实现对象级串行（每个角色一份缓存实例，限流跨步骤共享）。
+//
+// 缓存键含 ConfigIdentity 与工厂世代，故 BashPermission/工具配置收紧或工厂
+// 替换在下一次 Resolve 即生效。并发 gate 按「后端名+角色」独立于执行器版本
+// 持有：配置换代不会另起限流器而绕过角色上限。
+//
+// 边界：已在飞的任务不会被回滚或取消；收紧权限/降低并发只影响后续 Resolve
+// 与后续 Acquire。需要即时撤权的宿主应在每次工具调用前再校验。
 type BackendResolver struct {
 	mu         sync.RWMutex
 	factories  map[string]BackendFactory
 	underlying map[string]Executor // 无工厂直挂的后端实例
-	wrapped    map[string]Executor // 角色（含模型绑定）→ 已包装实例缓存
+	wrapped    map[string]Executor // 执行器配置身份 → 已包装实例缓存
+	factoryGen map[string]uint64   // Register/RegisterInstance 世代，替换即递增
+	gates      map[string]*roleConcurrencyGate
 }

 // NewBackendResolver 创建解析器并注册内置后端。
@@ -29,23 +40,40 @@ func NewBackendResolver() *BackendResolver {
 		factories:  make(map[string]BackendFactory),
 		underlying: make(map[string]Executor),
 		wrapped:    make(map[string]Executor),
+		factoryGen: make(map[string]uint64),
+		gates:      make(map[string]*roleConcurrencyGate),
 	}
 	r.RegisterInstance("stub", &StubExecutor{})
 	return r
 }

 // Register 注册后端工厂（名字如 "agent-native"、"pi-agent"）。
+// 替换同名工厂会使该后端下已缓存的包装执行器失效。
 func (r *BackendResolver) Register(name string, f BackendFactory) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	r.factories[name] = f
+	r.invalidateLocked(name)
 }

 // RegisterInstance 直接注册后端实例（忽略角色级绑定）。
+// 替换同名实例会使该后端下已缓存的包装执行器失效；"stub" 语义保持不变。
 func (r *BackendResolver) RegisterInstance(name string, b Executor) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	r.underlying[name] = b
+	r.invalidateLocked(name)
+}
+
+func (r *BackendResolver) invalidateLocked(name string) {
+	r.factoryGen[name]++
+	prefix := name + "\x00"
+	for k := range r.wrapped {
+		if strings.HasPrefix(k, prefix) {
+			delete(r.wrapped, k)
+		}
+	}
+	// gates 按角色身份保留：换工厂/配置不得丢掉进行中的并发计数。
 }

 // ListBackends 返回可用后端名（排序稳定）。
@@ -73,8 +101,8 @@ func (r *BackendResolver) ListBackends() []string {
 }

 // Resolve 解析角色的执行后端：Role.Backend → Options.DefaultBackend → "stub"。
-// MaxConcurrentTasks > 1 时以 agent-native 并发限流包装（per-role 缓存，跨步骤共享；
-// 创建段持写锁防并行首解析竞态——否则并发步骤会各拿一份限流器）。
+// MaxConcurrentTasks >= 1 时以共享角色 gate 包装（跨执行器版本共用；
+// 创建段持写锁防并行首解析竞态——否则并发步骤会各拿一份执行器）。
 func (r *BackendResolver) Resolve(role RoleSpec, defaultBackend string) (Executor, string, error) {
 	name := role.Backend
 	if name == "" {
@@ -83,30 +111,44 @@ func (r *BackendResolver) Resolve(role RoleSpec, defaultBackend string) (Executo
 	if name == "" {
 		name = "stub"
 	}
-	cacheKey := name + "\x00" + role.Role + "\x00" + role.Model

 	// 后端名快照在锁外获取（Resolve 持锁期间再调 ListBackends 会自死锁）。
 	names := r.ListBackends()
+
 	r.mu.RLock()
+	gen := r.factoryGen[name]
+	cacheKey := resolveCacheKey(name, role, gen)
 	cached, ok := r.wrapped[cacheKey]
-	factory, hasFactory := r.factories[name]
-	under, hasUnder := r.underlying[name]
 	r.mu.RUnlock()
 	if ok {
+		if max := role.MaxConcurrentTasks; max >= 1 {
+			r.mu.Lock()
+			if g := r.gates[limiterKey(name, role.Role)]; g != nil {
+				g.setMax(max)
+			}
+			r.mu.Unlock()
+		}
 		return cached, name, nil
 	}

 	r.mu.Lock()
 	defer r.mu.Unlock()
+	gen = r.factoryGen[name]
+	cacheKey = resolveCacheKey(name, role, gen)
 	// 双检：等写锁期间可能已被并发解析填充。
 	if b, ok := r.wrapped[cacheKey]; ok {
+		if max := role.MaxConcurrentTasks; max >= 1 {
+			if g := r.gates[limiterKey(name, role.Role)]; g != nil {
+				g.setMax(max)
+			}
+		}
 		return b, name, nil
 	}

 	var b Executor
-	if hasUnder {
+	if under, hasUnder := r.underlying[name]; hasUnder {
 		b = under
-	} else if hasFactory {
+	} else if factory, hasFactory := r.factories[name]; hasFactory {
 		var err error
 		b, err = factory(role)
 		if err != nil {
@@ -117,9 +159,114 @@ func (r *BackendResolver) Resolve(role RoleSpec, defaultBackend string) (Executo
 	}

 	// MaxConcurrentTasks>=1 即包装限流（1 = 对象级串行；0 = 不限）。
+	// gate 按角色身份共享，与 ConfigIdentity / 工厂世代无关。
 	if max := role.MaxConcurrentTasks; max >= 1 {
-		b = WrapWithConcurrencyLimit(b, max)
+		g := r.gateLocked(name, role.Role, max)
+		b = &gatedExecutor{inner: b, gate: g}
 	}
 	r.wrapped[cacheKey] = b
 	return b, name, nil
 }
+
+func resolveCacheKey(name string, role RoleSpec, gen uint64) string {
+	return name + "\x00" + role.Role + "\x00" + ConfigIdentity(role) + "\x00" +
+		fmt.Sprintf("%d", gen) + "\x00" + limiterStamp(role.MaxConcurrentTasks)
+}
+
+func limiterKey(backend, role string) string {
+	return backend + "\x00" + role
+}
+
+func (r *BackendResolver) gateLocked(backend, role string, max int) *roleConcurrencyGate {
+	key := limiterKey(backend, role)
+	g, ok := r.gates[key]
+	if !ok {
+		g = newRoleConcurrencyGate(max)
+		r.gates[key] = g
+		return g
+	}
+	g.setMax(max)
+	return g
+}
+
+// roleConcurrencyGate 是按角色共享的并发上限。
+// 执行器配置换代仍复用同一 gate，故旧版本在飞任务计入同一上限。
+type roleConcurrencyGate struct {
+	mu     sync.Mutex
+	max    int
+	active int
+	notify chan struct{} // 关闭并替换以唤醒等待者
+}
+
+func newRoleConcurrencyGate(max int) *roleConcurrencyGate {
+	if max <= 0 {
+		max = 1
+	}
+	return &roleConcurrencyGate{max: max, notify: make(chan struct{})}
+}
+
+func (g *roleConcurrencyGate) setMax(max int) {
+	if max <= 0 {
+		max = 1
+	}
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.max = max
+	g.signalLocked()
+}
+
+func (g *roleConcurrencyGate) signalLocked() {
+	close(g.notify)
+	g.notify = make(chan struct{})
+}
+
+func (g *roleConcurrencyGate) acquire(ctx context.Context) error {
+	for {
+		g.mu.Lock()
+		if g.active < g.max {
+			g.active++
+			g.mu.Unlock()
+			return nil
+		}
+		notify := g.notify
+		g.mu.Unlock()
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-notify:
+		}
+	}
+}
+
+func (g *roleConcurrencyGate) release() {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.active > 0 {
+		g.active--
+	}
+	g.signalLocked()
+}
+
+// gatedExecutor 把共享角色 gate 套在具体执行器上。
+type gatedExecutor struct {
+	inner Executor
+	gate  *roleConcurrencyGate
+}
+
+func (e *gatedExecutor) Name() string {
+	if e.inner == nil {
+		return "gated"
+	}
+	return e.inner.Name()
+}
+
+func (e *gatedExecutor) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	if err := e.gate.acquire(ctx); err != nil {
+		return &ExecResult{
+			Success:     false,
+			ErrorReason: fmt.Sprintf("agent concurrency wait canceled: %v", err),
+		}, err
+	}
+	defer e.gate.release()
+	return e.inner.Execute(ctx, req)
+}
@@ -0,0 +1,405 @@
+package squad
+
+import (
+	"context"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+type labeledExec struct {
+	label string
+	perm  string
+	tools string
+	delay time.Duration
+	enter func()
+	leave func()
+}
+
+func (e *labeledExec) Name() string { return e.label }
+
+func (e *labeledExec) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	if e.enter != nil {
+		e.enter()
+	}
+	if e.leave != nil {
+		defer e.leave()
+	}
+	if e.delay > 0 {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		case <-time.After(e.delay):
+		}
+	}
+	return &ExecResult{Success: true, Logs: e.label + "|" + e.perm + "|" + e.tools}, nil
+}
+
+func TestConfigIdentityChangesWithBashAndTools(t *testing.T) {
+	a := RoleSpec{Role: "r", Model: "m", BashPermission: BashPermissionAllow}
+	d := RoleSpec{Role: "r", Model: "m", BashPermission: BashPermissionDeny}
+	if ConfigIdentity(a) == ConfigIdentity(d) {
+		t.Fatal("allow and deny must differ")
+	}
+	t1 := RoleSpec{Role: "r", Model: "m", Tools: []string{"a", "b"}}
+	t2 := RoleSpec{Role: "r", Model: "m", Tools: []string{"b", "a"}}
+	if ConfigIdentity(t1) != ConfigIdentity(t2) {
+		t.Fatal("tool order must not change identity")
+	}
+	t3 := RoleSpec{Role: "r", Model: "m", Tools: []string{"a", "c"}}
+	if ConfigIdentity(t1) == ConfigIdentity(t3) {
+		t.Fatal("tool set change must change identity")
+	}
+	same := RoleSpec{Role: "r", Model: "m", BashPermission: BashPermissionAllow, MaxConcurrentTasks: 1}
+	bump := RoleSpec{Role: "r", Model: "m", BashPermission: BashPermissionAllow, MaxConcurrentTasks: 8}
+	if ConfigIdentity(same) != ConfigIdentity(bump) {
+		t.Fatal("MaxConcurrentTasks must not affect ConfigIdentity")
+	}
+}
+
+func TestResolveAllowDenyTighteningAndLoosening(t *testing.T) {
+	var builds atomic.Int32
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		builds.Add(1)
+		return &labeledExec{label: "cap", perm: role.BashPermission}, nil
+	})
+
+	allow := RoleSpec{Role: "writer", Backend: "cap", Model: "m1", BashPermission: BashPermissionAllow}
+	e1, _, err := r.Resolve(allow, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ := e1.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "cap|allow|" {
+		t.Fatalf("want allow executor, got %q", res.Logs)
+	}
+
+	deny := RoleSpec{Role: "writer", Backend: "cap", Model: "m1", BashPermission: BashPermissionDeny}
+	e2, _, err := r.Resolve(deny, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ = e2.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "cap|deny|" {
+		t.Fatalf("allow→deny must take effect on next Resolve, got %q", res.Logs)
+	}
+	if e1 == e2 {
+		t.Fatal("tightened config must not reuse cached allow executor")
+	}
+
+	e3, _, err := r.Resolve(allow, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ = e3.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "cap|allow|" {
+		t.Fatalf("deny→allow must take effect, got %q", res.Logs)
+	}
+	if builds.Load() < 2 {
+		t.Fatalf("factory must rebuild on permission change, builds=%d", builds.Load())
+	}
+}
+
+func TestResolveFactoryReplacementTakesEffect(t *testing.T) {
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		return &labeledExec{label: "v1", perm: role.BashPermission}, nil
+	})
+	role := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionAllow}
+	e1, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ := e1.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "v1|allow|" {
+		t.Fatalf("got %q", res.Logs)
+	}
+
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		return &labeledExec{label: "v2", perm: role.BashPermission}, nil
+	})
+	e2, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ = e2.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "v2|allow|" {
+		t.Fatalf("factory replacement must take effect, got %q", res.Logs)
+	}
+}
+
+func TestResolveToolConfigChangeTakesEffect(t *testing.T) {
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		tools := ""
+		if len(role.Tools) > 0 {
+			tools = role.Tools[0]
+		}
+		return &labeledExec{label: "cap", tools: tools}, nil
+	})
+	base := RoleSpec{Role: "writer", Backend: "cap", Model: "m", Tools: []string{"read"}}
+	e1, _, err := r.Resolve(base, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ := e1.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "cap||read" {
+		t.Fatalf("got %q", res.Logs)
+	}
+	next := RoleSpec{Role: "writer", Backend: "cap", Model: "m", Tools: []string{"write"}}
+	e2, _, err := r.Resolve(next, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, _ = e2.Execute(context.Background(), ExecRequest{})
+	if res.Logs != "cap||write" {
+		t.Fatalf("tool-config change must take effect, got %q", res.Logs)
+	}
+}
+
+func TestResolveUnchangedConfigReusesCache(t *testing.T) {
+	var builds atomic.Int32
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		builds.Add(1)
+		return &labeledExec{label: "cap", perm: role.BashPermission}, nil
+	})
+	role := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionAsk}
+	e1, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	e2, _, err := r.Resolve(role, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if e1 != e2 {
+		t.Fatal("unchanged config must reuse cached executor")
+	}
+	if builds.Load() != 1 {
+		t.Fatalf("want 1 factory call, got %d", builds.Load())
+	}
+}
+
+func TestResolveParallelFirstDoesNotDuplicate(t *testing.T) {
+	var builds atomic.Int32
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		builds.Add(1)
+		time.Sleep(20 * time.Millisecond) // widen race window
+		return &labeledExec{label: "cap"}, nil
+	})
+	role := RoleSpec{Role: "writer", Backend: "cap", Model: "m", MaxConcurrentTasks: 2}
+
+	const n = 32
+	var wg sync.WaitGroup
+	got := make([]Executor, n)
+	errs := make([]error, n)
+	wg.Add(n)
+	for i := 0; i < n; i++ {
+		i := i
+		go func() {
+			defer wg.Done()
+			got[i], _, errs[i] = r.Resolve(role, "")
+		}()
+	}
+	wg.Wait()
+	for i, err := range errs {
+		if err != nil {
+			t.Fatalf("resolve %d: %v", i, err)
+		}
+	}
+	for i := 1; i < n; i++ {
+		if got[i] != got[0] {
+			t.Fatalf("parallel first Resolve must share one executor")
+		}
+	}
+	if builds.Load() != 1 {
+		t.Fatalf("want single factory construction, got %d", builds.Load())
+	}
+}
+
+func TestSharedGateAcrossVersionsAndLimiterChange(t *testing.T) {
+	var (
+		mu      sync.Mutex
+		current int
+		peak    int
+	)
+	hold := make(chan struct{})
+	entered := make(chan struct{}, 8)
+	trackEnter := func() {
+		mu.Lock()
+		current++
+		if current > peak {
+			peak = current
+		}
+		mu.Unlock()
+		entered <- struct{}{}
+		<-hold
+	}
+	trackLeave := func() {
+		mu.Lock()
+		current--
+		mu.Unlock()
+	}
+
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		return &labeledExec{
+			label: "cap",
+			perm:  role.BashPermission,
+			enter: trackEnter,
+			leave: trackLeave,
+		}, nil
+	})
+
+	allow := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionAllow, MaxConcurrentTasks: 1}
+	eOld, _, err := r.Resolve(allow, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	errCh := make(chan error, 4)
+	go func() { _, err := eOld.Execute(context.Background(), ExecRequest{}); errCh <- err }()
+	<-entered // old holds the sole slot
+
+	deny := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionDeny, MaxConcurrentTasks: 1}
+	eNew, _, err := r.Resolve(deny, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if eOld == eNew {
+		t.Fatal("version bump must produce a new executor")
+	}
+
+	secondDone := make(chan struct{})
+	go func() {
+		_, err := eNew.Execute(context.Background(), ExecRequest{})
+		errCh <- err
+		close(secondDone)
+	}()
+	select {
+	case <-secondDone:
+		t.Fatal("new-version execute must wait on shared role cap while old is in-flight")
+	case <-time.After(40 * time.Millisecond):
+	}
+
+	// Raising MaxConcurrentTasks must update the shared gate (not spawn a bypassing limiter).
+	deny2 := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionDeny, MaxConcurrentTasks: 2}
+	eLim, _, err := r.Resolve(deny2, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	go func() { _, err := eLim.Execute(context.Background(), ExecRequest{}); errCh <- err }()
+
+	// After setMax(2): waiting eNew can enter (active=2). eLim must still wait.
+	select {
+	case <-entered: // eNew
+	case <-time.After(time.Second):
+		t.Fatal("limiter raise must admit the waiting new-version task")
+	}
+	select {
+	case <-entered:
+		t.Fatal("third task must not bypass shared cap after limiter change")
+	case <-time.After(40 * time.Millisecond):
+	}
+
+	close(hold)
+	for i := 0; i < 3; i++ {
+		if err := <-errCh; err != nil {
+			t.Fatal(err)
+		}
+	}
+	mu.Lock()
+	p := peak
+	mu.Unlock()
+	if p > 2 {
+		t.Fatalf("shared gate must cap total in-flight across versions; peak=%d", p)
+	}
+}
+
+func TestInFlightOldVersionCannotExceedCap(t *testing.T) {
+	var (
+		mu      sync.Mutex
+		current int
+		peak    int
+	)
+	enter := func() {
+		mu.Lock()
+		current++
+		if current > peak {
+			peak = current
+		}
+		mu.Unlock()
+	}
+	leave := func() {
+		mu.Lock()
+		current--
+		mu.Unlock()
+	}
+
+	r := NewBackendResolver()
+	r.Register("cap", func(role RoleSpec) (Executor, error) {
+		return &labeledExec{
+			label: "cap",
+			perm:  role.BashPermission,
+			delay: 80 * time.Millisecond,
+			enter: enter,
+			leave: leave,
+		}, nil
+	})
+
+	v1 := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionAllow, MaxConcurrentTasks: 2}
+	e1, _, err := r.Resolve(v1, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	v2 := RoleSpec{Role: "writer", Backend: "cap", Model: "m", BashPermission: BashPermissionDeny, MaxConcurrentTasks: 2}
+	e2, _, err := r.Resolve(v2, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var wg sync.WaitGroup
+	// 2 from old version + 2 from new version; shared cap=2 ⇒ peak must stay ≤2.
+	for _, e := range []Executor{e1, e1, e2, e2} {
+		wg.Add(1)
+		e := e
+		go func() {
+			defer wg.Done()
+			_, _ = e.Execute(context.Background(), ExecRequest{})
+		}()
+	}
+	wg.Wait()
+
+	mu.Lock()
+	p := peak
+	mu.Unlock()
+	if p > 2 {
+		t.Fatalf("in-flight old+new must not exceed role cap; peak=%d", p)
+	}
+}
+
+func TestRegisterInstanceStubSemantics(t *testing.T) {
+	r := NewBackendResolver()
+	e, name, err := r.Resolve(RoleSpec{Role: "x"}, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if name != "stub" {
+		t.Fatalf("want stub, got %s", name)
+	}
+	if e.Name() != "stub" {
+		t.Fatalf("want stub executor, got %s", e.Name())
+	}
+	custom := &labeledExec{label: "stub-custom"}
+	r.RegisterInstance("stub", custom)
+	e2, _, err := r.Resolve(RoleSpec{Role: "x"}, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if e2.Name() != "stub-custom" {
+		t.Fatalf("RegisterInstance stub replace must take effect, got %s", e2.Name())
+	}
+}
@@ -0,0 +1,41 @@
+package squad
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+// brief_test.go — Q3 WP1a：StepSpec.Brief 渲染进 TASK.md「任务指令」节。
+
+func TestStepBriefRendersIntoTaskFile(t *testing.T) {
+	rec := &promptRecordingBackend{prompts: map[string][]string{}}
+	engine := newRecordingEngine(t, rec, []StepSpec{
+		{ID: "s1", Role: "researcher", Title: "起草", Targets: []string{"output.md"},
+			Brief: "写一份关于「太空电梯」的简报，篇幅：短。\n必须包含单独一行「核对标记：P4」。"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "带任务指令"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	var prompt string
+	for _, p := range rec.prompts["01_researcher"] {
+		prompt += p
+	}
+	if !strings.Contains(prompt, "## 任务指令") {
+		t.Fatalf("TASK.md 缺任务指令节：%q", prompt)
+	}
+	if !strings.Contains(prompt, "核对标记：P4") {
+		t.Fatalf("任务指令正文缺失：%q", prompt)
+	}
+	// 指令节在期望产出之后、上下文表之前。
+	iOut := strings.Index(prompt, "期望产出")
+	iBrief := strings.Index(prompt, "## 任务指令")
+	iCtx := strings.Index(prompt, "## 任务上下文资源")
+	if !(iOut < iBrief && iBrief < iCtx) {
+		t.Fatalf("任务指令节位置错误：%d/%d/%d", iOut, iBrief, iCtx)
+	}
+}
@@ -2,9 +2,11 @@ package squad

 import (
 	"context"
+	"fmt"
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"testing"
 )

@@ -111,3 +113,225 @@ func TestGraphContinuityRejectsStructuralChange(t *testing.T) {
 		t.Fatalf("expected continuity rejection, got %v", err)
 	}
 }
+
+// ---- F11c：落盘失败的诚实性 ----
+
+// faultStore 包装底层 Store 注入落盘故障，其余路径照常透传。
+// failCheckpoint：任何 SaveCheckpoint 失败（async 档在出口 flush 时暴露）。
+// failTerminalSave：Meta 进入终态时的 SaveRun 失败（finalize/manualPass 的收口写）。
+type faultStore struct {
+	Store
+	mu               sync.Mutex
+	failCheckpoint   bool
+	failTerminalSave bool
+	checkpointWrites int
+	terminalAttempts int
+}
+
+func (f *faultStore) SaveCheckpoint(wsDir string, cp Checkpoint) error {
+	f.mu.Lock()
+	f.checkpointWrites++
+	shouldFail := f.failCheckpoint
+	f.mu.Unlock()
+	if shouldFail {
+		return fmt.Errorf("injected checkpoint write failure (superstep %d)", cp.Superstep)
+	}
+	return f.Store.SaveCheckpoint(wsDir, cp)
+}
+
+func (f *faultStore) SaveRun(st *RunState) error {
+	f.mu.Lock()
+	f.terminalAttempts++
+	shouldFail := f.failTerminalSave && RunStatus(st.Meta.Status).Terminal()
+	f.mu.Unlock()
+	if shouldFail {
+		return fmt.Errorf("injected terminal run state failure (%s)", st.Meta.Status)
+	}
+	return f.Store.SaveRun(st)
+}
+
+func (f *faultStore) writes() int {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.checkpointWrites
+}
+
+// eventRecorder 按序收集引擎事件（顺序断言用）。
+type eventRecorder struct {
+	mu   sync.Mutex
+	evs  []Event
+	sink ProgressSink
+}
+
+func newEventRecorder() *eventRecorder {
+	r := &eventRecorder{}
+	r.sink = ProgressFunc(func(ev Event) {
+		r.mu.Lock()
+		defer r.mu.Unlock()
+		r.evs = append(r.evs, ev)
+	})
+	return r
+}
+
+func (r *eventRecorder) types() []string {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	out := make([]string, 0, len(r.evs))
+	for _, ev := range r.evs {
+		out = append(out, ev.Type)
+	}
+	return out
+}
+
+func (r *eventRecorder) has(t string) bool {
+	for _, got := range r.types() {
+		if got == t {
+			return true
+		}
+	}
+	return false
+}
+
+// newTestEngineWithStore 与 newTestEngine 同装配，额外注入 Store 与事件收集 sink。
+func newTestEngineWithStore(t *testing.T, store Store, rec *eventRecorder) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, role := range []string{"researcher", "writer", "reviewer"} {
+		if err := reg.Register(RoleSpec{Role: role, Backend: "script"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("script", &scriptBackend{script: map[string]map[string]string{
+		"01_researcher": {"notes.md": goodReport, "fields.json": "{}"},
+		"02_writer":     {"output.md": goodReport},
+		"03_reviewer":   {"output.md": goodReport},
+	}})
+	eng, err := NewEngine(Options{DataDir: dir, Registry: reg,
+		Planner: mockScriptPlanner(baseSteps()...), Backends: backends,
+		Sink: rec.sink, Store: store})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return eng
+}
+
+// assertNoCleanCompletion：落盘失败后不得再以「干净完成」面世——
+// 没有 run_completed 事件、也没有 completed/degraded 的返回值。
+func assertNoCleanCompletion(t *testing.T, res *RunResult, err error, rec *eventRecorder) {
+	t.Helper()
+	if rec.has(EvRunCompleted) {
+		t.Fatalf("run_completed emitted after persistence failure: %v", rec.types())
+	}
+	if res != nil {
+		if res.Status == RunCompleted || res.Status == RunDegraded {
+			t.Fatalf("clean completion after persistence failure: status=%s", res.Status)
+		}
+		if res.Status != RunFailed {
+			t.Fatalf("status = %s, want failed or an error return", res.Status)
+		}
+		if !strings.Contains(res.Error, persistenceFailurePrefix) {
+			t.Fatalf("failure not explained in result: %q", res.Error)
+		}
+	}
+	if err != nil && !strings.Contains(err.Error(), persistenceFailurePrefix) {
+		t.Fatalf("error does not name the persistence failure: %v", err)
+	}
+	if res == nil && err == nil {
+		t.Fatal("persistence failure vanished: neither result nor error")
+	}
+}
+
+// TestCheckpointFlushFailureIsVisible：出口 flush 的落盘失败必须可见，
+// 且先于任何完成事件/返回（三档 sync/async/on_exit 都一样）。
+func TestCheckpointFlushFailureIsVisible(t *testing.T) {
+	for _, mode := range []string{"sync", "async", "on_exit"} {
+		t.Run(mode, func(t *testing.T) {
+			store := &faultStore{Store: FileStore{}, failCheckpoint: true}
+			rec := newEventRecorder()
+			eng := newTestEngineWithStore(t, store, rec)
+			res, err := eng.Run(context.Background(), Goal{Text: "快照"}, RunOptions{CheckpointMode: mode})
+			if store.writes() == 0 {
+				t.Fatal("test did not reach any checkpoint write")
+			}
+			assertNoCleanCompletion(t, res, err, rec)
+			// 失败在推进退出处如实标注；async/on_exit 由 finalizeEvent 发 run_failed，
+			// sync 在第一个边界就报错返回。
+			if mode == "sync" {
+				if err == nil {
+					t.Fatalf("sync boundary write failure must surface as error, got %+v", res)
+				}
+				return
+			}
+			if res == nil || res.Status != RunFailed {
+				t.Fatalf("%s: want failed result, got %+v err=%v", mode, res, err)
+			}
+			if !rec.has(EvRunFailed) {
+				t.Fatalf("%s: no run_failed event: %v", mode, rec.types())
+			}
+		})
+	}
+}
+
+// TestTerminalSaveFailureBlocksCompletion：终态运行态写盘失败时不得宣布完成
+// （事件与返回值都必须可区分）；业务动作已发生（交付物文件仍在），但不声称成功。
+func TestTerminalSaveFailureBlocksCompletion(t *testing.T) {
+	for _, mode := range []string{"sync", "async", "on_exit"} {
+		t.Run(mode, func(t *testing.T) {
+			store := &faultStore{Store: FileStore{}, failTerminalSave: true}
+			rec := newEventRecorder()
+			eng := newTestEngineWithStore(t, store, rec)
+			res, err := eng.Run(context.Background(), Goal{Text: "快照"}, RunOptions{CheckpointMode: mode})
+			if err != nil {
+				t.Fatalf("terminal save failure must not abort the run as infra error: %v", err)
+			}
+			assertNoCleanCompletion(t, res, err, rec)
+			if res.Status != RunFailed {
+				t.Fatalf("%s: want failed, got %s", mode, res.Status)
+			}
+			if !rec.has(EvRunFailed) {
+				t.Fatalf("%s: no run_failed event: %v", mode, rec.types())
+			}
+			// 完成事件必须晚于/缺席于失败：事件序里终态只有 run_failed。
+			types := rec.types()
+			lastTerminal := ""
+			for _, ty := range types {
+				switch ty {
+				case EvRunFailed, EvRunCompleted, EvRunCancelled:
+					lastTerminal = ty
+				}
+			}
+			if lastTerminal != EvRunFailed {
+				t.Fatalf("%s: last terminal event = %q, want run_failed", mode, lastTerminal)
+			}
+			// 业务动作已发生且未被回滚：交付物文件确实写出来了（但运行不声称完成）。
+			if _, err := os.Stat(filepath.Join(res.WorkDir, "output.md")); err != nil {
+				t.Fatalf("deliverable written before terminal save should remain: %v", err)
+			}
+		})
+	}
+}
+
+// TestCheckpointFlushSucceedsInAllModes（对照）：故障注入关闭时三档都正常收口，
+// 出口快照确实落盘（避免上面的失败断言因「压根没写」而空过）。
+func TestCheckpointFlushSucceedsInAllModes(t *testing.T) {
+	for _, mode := range []string{"sync", "async", "on_exit"} {
+		t.Run(mode, func(t *testing.T) {
+			store := &faultStore{Store: FileStore{}}
+			rec := newEventRecorder()
+			eng := newTestEngineWithStore(t, store, rec)
+			res, err := eng.Run(context.Background(), Goal{Text: "快照"}, RunOptions{CheckpointMode: mode})
+			if err != nil || res.Status != RunCompleted {
+				t.Fatalf("%s: %v %+v", mode, err, res)
+			}
+			if !rec.has(EvRunCompleted) {
+				t.Fatalf("%s: no run_completed: %v", mode, rec.types())
+			}
+			latest, err := FileStore{}.LoadLatestCheckpoint(res.WorkDir)
+			if err != nil || latest == nil || latest.Superstep < 3 {
+				t.Fatalf("%s: exit snapshot not flushed: %v %+v", mode, err, latest)
+			}
+		})
+	}
+}
@@ -0,0 +1,45 @@
+package squad
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"sort"
+	"strings"
+)
+
+// ConfigIdentity 返回角色中「塑造 Executor 装配」字段的稳定摘要。
+// 含 BashPermission / Model / ToolProfile / Tools / Skills；不含
+// MaxConcurrentTasks（并发上限由解析器按角色身份单独持有，与执行器版本解耦）。
+func ConfigIdentity(role RoleSpec) string {
+	var b strings.Builder
+	b.WriteString("bash=")
+	b.WriteString(role.BashPermission)
+	b.WriteString("\x00model=")
+	b.WriteString(role.Model)
+	b.WriteString("\x00profile=")
+	b.WriteString(role.ToolProfile)
+	b.WriteString("\x00tools=")
+	b.WriteString(stableJoin(role.Tools))
+	b.WriteString("\x00skills=")
+	b.WriteString(stableJoin(role.Skills))
+	sum := sha256.Sum256([]byte(b.String()))
+	return hex.EncodeToString(sum[:16])
+}
+
+func stableJoin(items []string) string {
+	if len(items) == 0 {
+		return ""
+	}
+	cp := append([]string(nil), items...)
+	sort.Strings(cp)
+	return strings.Join(cp, ",")
+}
+
+// limiterStamp 区分「启用限流」与「不限流」；具体上限数值不进执行器缓存键，
+// 以免调大/调小 MaxConcurrentTasks 时换掉共享 gate。
+func limiterStamp(max int) string {
+	if max >= 1 {
+		return "lim"
+	}
+	return "unlim"
+}
@@ -114,6 +114,19 @@ type RunOptions struct {
 	// MaxReplans leader 模式的重规划上限（Q2；与 MaxLeaderTurns 独立）。
 	// 0 = 默认 DefaultMaxReplans。
 	MaxReplans int `json:"max_replans,omitempty"`
+	// GroupChatPolicy group_chat 拓扑发言人策略（Q2）：rotate（默认）| moderator。
+	GroupChatPolicy string `json:"group_chat_policy,omitempty"`
+	// GroupChatModerator 主持人角色 key（moderator 策略必填，须在名册内）。
+	GroupChatModerator string `json:"group_chat_moderator,omitempty"`
+	// GroupChatMaxTurns 群聊轮次预算（0 = DefaultMaxGroupChatTurns）。
+	GroupChatMaxTurns int `json:"group_chat_max_turns,omitempty"`
+	// ReviewVerifyRole review 拓扑核实角色 key（默认 "verifier"）。
+	ReviewVerifyRole string `json:"review_verify_role,omitempty"`
+	// ReviewTarget 被评审材料（默认取 context/goal.md）。
+	ReviewTarget string `json:"review_target,omitempty"`
+	// TaskBoardLeaseSeconds task_board 任务租约时限（秒；0 = 默认 600）。
+	// 孤儿租约（claimed 且过期）在驱动重入时再武装回 open。
+	TaskBoardLeaseSeconds int `json:"task_board_lease_seconds,omitempty"`
 }

 // SquadEngine 小队引擎门面：Run / Respond / Resume / Inspect / ListRuns。
@@ -250,6 +263,11 @@ func (e *SquadEngine) Run(ctx context.Context, goal Goal, runOpts RunOptions) (*
 	}
 	if runOpts.Topology != "" {
 		meta.Metadata["topology"] = runOpts.Topology
+		// leader 拓扑映射到委派循环须发生在计划确认之前（确认门按
+		// mode=leader 豁免）；后置映射会让 leader 运行误触计划确认。
+		if runOpts.Topology == TopologyLeader && meta.Metadata["mode"] == "" {
+			meta.Metadata["mode"] = "leader"
+		}
 	}
 	if !runOpts.Deadline.IsZero() {
 		// RFC3339Nano：秒级 RFC3339 会截断亚秒精度，短截止会被立刻判超期。
@@ -258,6 +276,24 @@ func (e *SquadEngine) Run(ctx context.Context, goal Goal, runOpts RunOptions) (*
 	if runOpts.MaxReplans > 0 {
 		meta.Metadata["max_replans"] = fmt.Sprintf("%d", runOpts.MaxReplans)
 	}
+	if runOpts.GroupChatPolicy != "" {
+		meta.Metadata["group_chat_policy"] = runOpts.GroupChatPolicy
+	}
+	if runOpts.GroupChatModerator != "" {
+		meta.Metadata["group_chat_moderator"] = runOpts.GroupChatModerator
+	}
+	if runOpts.GroupChatMaxTurns > 0 {
+		meta.Metadata["group_chat_max_turns"] = fmt.Sprintf("%d", runOpts.GroupChatMaxTurns)
+	}
+	if runOpts.ReviewVerifyRole != "" {
+		meta.Metadata["review_verify_role"] = runOpts.ReviewVerifyRole
+	}
+	if runOpts.ReviewTarget != "" {
+		meta.Metadata["review_target"] = runOpts.ReviewTarget
+	}
+	if runOpts.TaskBoardLeaseSeconds > 0 {
+		meta.Metadata["task_board_lease_seconds"] = fmt.Sprintf("%d", runOpts.TaskBoardLeaseSeconds)
+	}
 	e.deps.Sink.Emit(Event{Type: EvRunStarted, RunID: runID, Text: goal.Text, At: time.Now()})

 	// 意图闸门（可选，入口一次裁决；显式 TaskKey 优先，不重复过闸）。
@@ -434,6 +470,15 @@ func (e *SquadEngine) planAndExecute(ctx context.Context, st *RunState, runOpts
 		}
 		return out, nil
 	}
+	// driver 类拓扑（Q2 WP3+）：按名分发专用驱动（review/group_chat；
+	// task_board 属 WP4，未点亮前 checkTopology 已 fail-closed）。
+	if out, handled, derr := e.dispatchTopologyDriver(ctx, st); handled {
+		if derr != nil {
+			return nil, derr
+		}
+		e.finalizeEvent(out)
+		return out, nil
+	}
 	out, _, err := drv.advance(ctx, st)
 	if err != nil {
 		return nil, err
@@ -494,6 +539,15 @@ func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *RunState) (*RunRe
 		}
 		return out, nil
 	}
+	// driver 类拓扑续跑：驱动自身幂等（review 已完成 finder 不重跑；
+	// group_chat 按已落盘轮次续转）。
+	if out, handled, err := e.dispatchTopologyDriver(ctx, st); handled {
+		if err != nil {
+			return nil, err
+		}
+		e.finalizeEvent(out)
+		return out, nil
+	}
 	drv := &flowDriver{opts: e.deps}
 	out, _, err := drv.advance(ctx, st)
 	if err != nil {
@@ -503,6 +557,32 @@ func (e *SquadEngine) resumeAfterHold(ctx context.Context, st *RunState) (*RunRe
 	return out, nil
 }

+// dispatchTopologyDriver 按运行拓扑分发 driver 类拓扑。
+// handled=false 表示非 driver 拓扑（走既有图/模式路径）。
+func (e *SquadEngine) dispatchTopologyDriver(ctx context.Context, st *RunState) (*RunResult, bool, error) {
+	topo := st.Meta.Metadata["topology"]
+	if topo == "" {
+		return nil, false, nil
+	}
+	spec, ok := LookupTopology(topo)
+	if !ok || spec.Kind != TopologyDriver {
+		return nil, false, nil
+	}
+	switch topo {
+	case TopologyReview:
+		out, err := e.runReviewDriver(ctx, st)
+		return out, true, err
+	case TopologyGroupChat:
+		out, err := e.runGroupChatDriver(ctx, st)
+		return out, true, err
+	case TopologyTaskBoard:
+		out, err := e.runTaskBoardDriver(ctx, st)
+		return out, true, err
+	default:
+		return e.driverFail(st, fmt.Sprintf("topology: %q 驱动未实现（fail-closed）", topo)), true, nil
+	}
+}
+
 // materializeWorkspace 落盘工作区与子工作区（已存在则只更新计划并重建待执行子目录）。
 func (e *SquadEngine) materializeWorkspace(meta *RunMetadata, st *RunState, plan *SquadPlan) error {
 	if st.Meta.WorkDir == "" {
@@ -893,9 +973,18 @@ func (e *SquadEngine) manualPass(st *RunState, decision RespondDecision) (*RunRe
 	wf := st.Workflow()
 	wf.Status = string(status)
 	st.setWorkflow(wf)
-	_ = st.save()
+	// F11c：终态落盘先于完成事件——写不进去不得宣布完成（业务动作不回滚）。
+	if serr := st.save(); serr != nil {
+		msg := persistenceFailureText("save terminal run state", serr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			WorkDir: st.Meta.WorkDir, Quality: report, Error: msg, Summary: msg}, nil
+	}
 	manifest := buildManifest(st, "output.md", *report)
-	_ = writeJSON(filepath.Join(st.Meta.WorkDir, "manifest.json"), manifest)
+	if merr := writeJSON(filepath.Join(st.Meta.WorkDir, "manifest.json"), manifest); merr != nil {
+		msg := persistenceFailureText("write manifest", merr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			WorkDir: st.Meta.WorkDir, Quality: report, Error: msg, Summary: msg}, nil
+	}
 	e.deps.Sink.Emit(Event{Type: EvRunCompleted, RunID: st.Meta.ID, Text: string(status), At: time.Now()})
 	return &RunResult{RunID: st.Meta.ID, Status: status, Plan: planPtr(st),
 		Deliverable: "output.md", Quality: report, WorkDir: st.Meta.WorkDir,
@@ -909,10 +998,19 @@ func (e *SquadEngine) cancelRun(st *RunState, reason string) *RunResult {
 	wf := st.Workflow()
 	wf.Status = "cancelled"
 	st.setWorkflow(wf)
-	_ = st.save()
+	// 取消语义不变（运行确实被中止），但取消态落盘失败必须可见，不静默。
+	saveErr := st.save()
+	if saveErr != nil {
+		e.deps.Sink.Emit(Event{Type: EvPhase, RunID: st.Meta.ID,
+			Text: persistenceFailureText("save cancelled run state", saveErr), At: time.Now()})
+	}
 	e.deps.Sink.Emit(Event{Type: EvRunCancelled, RunID: st.Meta.ID, Text: reason, At: time.Now()})
-	return &RunResult{RunID: st.Meta.ID, Status: RunCancelled, Plan: planPtr(st),
+	out := &RunResult{RunID: st.Meta.ID, Status: RunCancelled, Plan: planPtr(st),
 		Summary: reason, WorkDir: st.Meta.WorkDir}
+	if saveErr != nil {
+		out.Error = persistenceFailureText("save cancelled run state", saveErr)
+	}
+	return out
 }

 // failStepResume 权限拒绝后的失败步骤处理（关键失败/非关键降级）。
@@ -971,73 +1069,129 @@ func markStep(st *RunState, id, status, errText string) bool {
 }

 // recoverInterrupted 处理崩溃时停在 running 的步骤。
-// interactive 挂起确认，不自动再执行；auto 允许一次重试；disabled 直接失败。
+// interactive 挂起确认/对账，不自动再执行；auto 允许一次重试；disabled 直接失败。
+//
+// F05/U3b：恢复是否判定 completed 不再单纯依据「文件齐全」——trusted terminal
+// record 是单一权威：仅当 step_terminals.json 中存在该步骤且 Terminal=completed
+// 时才能恢复完成（即便文件缺失也照恢复，文件用于诊断）；其余情形按 record +
+// 文件状态联合判定：
+//
+//   - 有 trusted record（completed）+ 缺失文件 -> completed（files 诊断信息）
+//   - 有 trusted record + 现存文件 -> completed（一致）
+//   - 有 trusted record + SideEffects=none + 缺失文件 -> 一次重试 pending
+//   - 无 record + 文件齐全 -> needs_reconciliation（不自动 rerun，不假装完成）
+//   - 无 record + 文件缺失 -> needs_reconciliation
+//   - record 表明 Canceled/Aborted -> 走 cancel 路径
+//
+// retention 期间产物缺失若 record SideEffects=possible，禁止无条件 rerun ——
+// 必须经人工 reconciliation 才能继续。
 func (e *SquadEngine) recoverInterrupted(ctx context.Context, st *RunState) (*RunResult, error) {
+	terminals := loadStepTerminalsForWorkDir(st.Meta.WorkDir)
+	drv := &flowDriver{opts: e.deps}
 	for i := range st.Plan.Workflow.Steps {
 		step := st.Plan.Workflow.Steps[i]
 		if step.Status != "running" {
 			continue
 		}
+		subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(orDefault(step.SubWorkspace, fmt.Sprintf("tasks/%02d_%s", i+1, step.Role))))
+		targets := step.Targets
+		if len(targets) == 0 {
+			targets = []string{"output.md"}
+		}
+		missing := missingTargets(subDir, targets)
+		record, hasRecord := terminals[step.ID]
+
+		// 1. trusted terminal record 存在 -> 单一权威。
+		if hasRecord && record.Terminal == "completed" && !record.Canceled && !record.Aborted {
+			step.Status = "completed"
+			step.Error = ""
+			st.Plan.Workflow.Steps[i] = step
+			st.mu.Lock()
+			_ = mergeStepArtifacts(&st.BB, subDir)
+			err := st.save()
+			st.mu.Unlock()
+			if err != nil {
+				return nil, err
+			}
+			drv.emit(st, Event{Type: EvStepCompleted, StepID: step.ID, Role: step.Role,
+				Text: fmt.Sprintf("中断恢复：trusted terminal record（attempt=%s, side_effects=%s）判定完成",
+					record.AttemptID, orDefault(record.SideEffects, "possible"))})
+			continue
+		}
+		// 1b. record 表明已取消/中止 -> cancel 路径（不等同业务失败）。
+		if hasRecord && (record.Canceled || record.Aborted) {
+			step.Status = "failed"
+			if record.Canceled {
+				step.Error = "interrupted: canceled"
+			} else {
+				step.Error = "interrupted: aborted"
+			}
+			st.Plan.Workflow.Steps[i] = step
+			if err := st.save(); err != nil {
+				return nil, err
+			}
+			out, _, ferr := drv.failStep(ctx, st, i)
+			return out, ferr
+		}
+		// 1c. record 表明 SideEffects=none + 缺失文件 -> 仅一次重试 pending
+		// （旧 WP3 路径保留以兼容纯读取类步骤）。
+		if hasRecord && record.SideEffects == string(ExecSideEffectsNone) && len(missing) > 0 && step.Attempts < 2 {
+			step.Attempts++
+			step.Status = "pending"
+			step.Error = ""
+			// F05：复用上一次 attempt 的 ExecutionID（已在 step.Terminal 落 record）。
+			if record.AttemptID != "" {
+				step.AttemptID = record.AttemptID
+			}
+			st.Plan.Workflow.Steps[i] = step
+			if err := st.save(); err != nil {
+				return nil, err
+			}
+			continue
+		}
+
+		// 2. 无 trusted record —— 旧路径不准直接翻 completed，必须进入 reconciliation。
+		// 区分 HITL 模式：interactive 仍合成 HITL hold（同 hitlConfirmation 通道），
+		// auto / disabled 走显式 needs_reconciliation（不自动 rerun 不假装成功）。
+		reason := "no trusted terminal record"
+		if len(missing) == 0 {
+			reason = "no trusted terminal record despite deliverables present"
+		} else {
+			reason = "no trusted terminal record and deliverables missing"
+		}
+		step.Status = StepStatusNeedsReconciliation
+		step.Error = reason
+		st.Plan.Workflow.Steps[i] = step
+		if err := st.save(); err != nil {
+			return nil, err
+		}
 		switch st.policy().mode() {
-		case HitlDisabled:
-			st.Plan.Workflow.Steps[i].Status = "failed"
-			st.Plan.Workflow.Steps[i].Error = "interrupted"
-			if err := st.save(); err != nil {
-				return nil, err
-			}
-			drv := &flowDriver{opts: e.deps}
-			out, _, err := drv.failStep(ctx, st, i)
-			return out, err
-		case HitlAuto:
-			if step.Attempts >= 2 {
-				st.Plan.Workflow.Steps[i].Status = "failed"
-				st.Plan.Workflow.Steps[i].Error = "interrupted twice"
-				if err := st.save(); err != nil {
-					return nil, err
-				}
-				drv := &flowDriver{opts: e.deps}
-				out, _, err := drv.failStep(ctx, st, i)
-				return out, err
-			}
-			// Q1 WP3 产物校验：步骤在 running 崩溃，但其交付物已全部落盘——
-			// 副作用实际发生过，直接判定完成而不是重新执行（副作用次数=1）。
-			subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(orDefault(step.SubWorkspace, fmt.Sprintf("tasks/%02d_%s", i+1, step.Role))))
-			targets := step.Targets
-			if len(targets) == 0 {
-				targets = []string{"output.md"}
-			}
-			if missingTargets(subDir, targets) == nil {
-				step.Status = "completed"
-				step.Error = ""
-				st.Plan.Workflow.Steps[i] = step
-				st.mu.Lock()
-				_ = mergeStepArtifacts(&st.BB, subDir)
-				err := st.save()
-				st.mu.Unlock()
-				if err != nil {
-					return nil, err
-				}
-				drv := &flowDriver{opts: e.deps}
-				drv.emit(st, Event{Type: EvStepCompleted, StepID: step.ID, Role: step.Role,
-					Text: fmt.Sprintf("中断恢复：产物齐全，按 attempt %s 判定完成", step.AttemptID)})
-				continue
-			}
-			st.Plan.Workflow.Steps[i].Status = "pending"
-			if err := st.save(); err != nil {
-				return nil, err
-			}
-			return nil, nil
-		default:
-			drv := &flowDriver{opts: e.deps}
+		case HitlInteractive:
 			out, _, err := drv.hold(st, HitlConfirmation, step.ID, step.Role,
-				fmt.Sprintf("步骤「%s」在执行中中断。如何继续？", step.Title),
+				fmt.Sprintf("步骤「%s」在执行中中断，且没有可信任的完成记录。%s 如何继续？", step.Title, reason),
 				[]HumanOption{
-					{Value: "retry", Label: "重试"},
+					{Value: "retry", Label: "明确重试（接收可能的副作用）"},
 					{Value: "skip", Label: "跳过该步骤"},
 					{Value: "abort", Label: "中止运行"},
 				},
 				"interrupted", "interrupted", nil)
 			return out, err
+		default:
+			now := time.Now()
+			st.Meta.Status = string(RunWaitingReconciliation)
+			st.Meta.UpdatedAt = now
+			if err := st.save(); err != nil {
+				return nil, err
+			}
+			drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("中断恢复：%s进入 reconciliation（reason=%s）", step.ID, reason)})
+			return &RunResult{
+				RunID:   st.Meta.ID,
+				Status:  RunWaitingReconciliation,
+				Plan:    planPtr(st),
+				WorkDir: st.Meta.WorkDir,
+				Summary: "interrupted-running step requires reconciliation",
+				Error:   reason,
+			}, nil
 		}
 	}
 	return nil, nil
@@ -1063,6 +1217,13 @@ func (e *SquadEngine) Resume(ctx context.Context, runID string) (*RunResult, err
 	if RunStatus(st.Meta.Status).Terminal() {
 		return e.Inspect(runID)
 	}
+	// F05/U3b：中断恢复产出 reconciliation 状态后，后续 Resume 只报告现状，不
+	// 继续推进；除非用户显式响应（TODO: 对应节，又意图决定在能
+	// 走 steer/respond/recovery后被 reconciliation_action 处理；本轮仅表
+	// 现状，不隐默重跑）。
+	if RunStatus(st.Meta.Status) == RunWaitingReconciliation {
+		return e.Inspect(runID)
+	}
 	// 图连续性校验（Q1 WP4/WP1 裁定）：崩溃期间的检查点图指纹与当前 Steps
 	// 重编译不一致 = 图形态被改过（动态调整或外力），拒绝恢复并提示重新规划。
 	if err := validateGraphContinuity(st); err != nil {
@@ -0,0 +1,517 @@
+// Package squad: F05/U3b executor result 合约测试。
+//
+// 覆盖以下消费规则：
+//   - (nil,nil) executor 返回 (result==nil && err==nil) 是稳定失败（非 panic，非 success）。
+//   - Success=true 但无执行证据（patches-only）也要走交付物校验；
+//     失败-with-full-patches 仍为失败（不允许文件存在翻成 success）。
+//   - canceled/aborted 结果不进入 retry 通路（不是业务失败，不可重试）。
+//   - side_effects=possible 的失败禁止自动重试；
+//     side_effects=none 的纯读取失败可走有界重试。
+//   - recovery：trusted record 存在 + 文件缺失 -> 恢复 completed（权威来源 = record）；
+//     无 record + 文件齐全 -> needs_reconciliation（不假装完成、不隐默 rerun）；
+//     无 record + 文件缺失 -> needs_reconciliation；
+//     record SideEffects=none + 缺失文件 -> 一次重试 pending（不重复执行副作用）。
+//
+// 这些测试使用 squad 包的 in-memory executor 替身进行 hermetic 验证；不调用
+// 真实模型或外部组件。
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// ---- 替身 executor ----
+
+// frozenResultBackend 每次 Execute 返回同一个固定结果（用于直接构造语义场景）。
+type frozenResultBackend struct {
+	mu    sync.Mutex
+	calls int32
+	res   *ExecResult
+	err   error
+}
+
+func (b *frozenResultBackend) Name() string { return "frozen" }
+
+func (b *frozenResultBackend) Execute(_ context.Context, _ ExecRequest) (*ExecResult, error) {
+	atomic.AddInt32(&b.calls, 1)
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if b.err != nil {
+		return b.res, b.err
+	}
+	return b.res, nil
+}
+
+func (b *frozenResultBackend) callCount() int32 {
+	return atomic.LoadInt32(&b.calls)
+}
+
+// materializingBackend 每次 Execute 写入 patches 但把 Success 留给 caller 决定。
+// 用于「patches-only success」测试场景。
+type materializingBackend struct {
+	mu    sync.Mutex
+	calls int32
+	// success 只在 success==true 时返回 Success=true；测试 override 用。
+	success  bool
+	canceled bool
+	aborted  bool
+	// sideEffects 决定是否允许重试。
+	sideEffects ExecSideEffects
+	// writePatch 控制是否 materializes patch 到 workDir。
+	writePatch bool
+}
+
+func (b *materializingBackend) Name() string { return "materializing" }
+
+func (b *materializingBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	atomic.AddInt32(&b.calls, 1)
+	patches := map[string][]byte{}
+	for _, t := range req.Targets {
+		if b.writePatch {
+			_ = os.WriteFile(filepath.Join(req.WorkDir, t), []byte(goodReport), 0o644)
+		}
+		patches[t] = []byte(goodReport)
+	}
+	return &ExecResult{
+		Success:     b.success,
+		Patches:     patches,
+		Logs:        "stub",
+		SideEffects: b.sideEffects,
+		Canceled:    b.canceled,
+		Aborted:     b.aborted,
+	}, nil
+}
+
+func (b *materializingBackend) callCount() int32 {
+	return atomic.LoadInt32(&b.calls)
+}
+
+// ---- 测试辅助 ----
+
+func newFrozenTestEngine(t *testing.T, dir string, backend Executor) *SquadEngine {
+	t.Helper()
+	reg := NewRoleRegistry()
+	if err := reg.Register(RoleSpec{Role: "researcher", Name: "调研", Backend: backend.Name()}); err != nil {
+		t.Fatal(err)
+	}
+	if err := reg.Register(RoleSpec{Role: "writer", Name: "撰写", Backend: backend.Name()}); err != nil {
+		t.Fatal(err)
+	}
+	if err := reg.Register(RoleSpec{Role: "reviewer", Name: "审核", Backend: backend.Name()}); err != nil {
+		t.Fatal(err)
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance(backend.Name(), backend)
+	eng, err := NewEngine(Options{
+		DataDir: dir, Registry: reg, Planner: mockScriptPlanner(baseSteps()...),
+		Backends: backends, Sink: NopSink{},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return eng
+}
+
+// ---- 合约测试 ----
+
+// TestNilNilResultIsStableFailure：(nil,nil) 必须稳定失败——不 panic、不被 papered over。
+func TestNilNilResultIsStableFailure(t *testing.T) {
+	dir := t.TempDir()
+	be := &frozenResultBackend{res: nil, err: nil}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "空结果"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run returned engine error: %v", err)
+	}
+	// 关键失败 → RunFailed；步骤状态 failed；call 一次；不 panic。
+	if res.Status != RunFailed {
+		t.Fatalf("want RunFailed, got %s (err=%s)", res.Status, res.Error)
+	}
+	if be.callCount() != 1 {
+		t.Fatalf("backend called %d times, want 1", be.callCount())
+	}
+	if res.Plan == nil {
+		t.Fatal("plan missing")
+	}
+	if got := res.Plan.Steps[0].Status; got != "failed" {
+		t.Fatalf("step 1 status = %s, want failed", got)
+	}
+	if got := res.Plan.Steps[0].Error; got != "executor returned nil result" {
+		t.Fatalf("step 1 error = %q, want executor-nil-result reason", got)
+	}
+}
+
+// TestSuccessWithFullPatchesMaterializes：Success=true + 已有 patch
+// —— 通过交付物校验完成；patches-only success 也走同一路径。
+func TestSuccessWithFullPatchesMaterializes(t *testing.T) {
+	dir := t.TempDir()
+	be := &materializingBackend{success: true, sideEffects: ExecSideEffectsNone, writePatch: true}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "patches 落地校验"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want completed, got %s err=%s", res.Status, res.Error)
+	}
+	// 三步各调用一次；patches 命中产物路径且 Success=true，每步一次。
+	if got := be.callCount(); got < 3 {
+		t.Fatalf("backend called %d times, want at least 3 (3 steps)", got)
+	}
+}
+
+// TestFailedWithFullPatchesStaysFailed：F05/KTD4 关键不变式：失败 + 文件齐全
+// 也不能翻成成功——即便 patches 内容已落地。
+func TestFailedWithFullPatchesStaysFailed(t *testing.T) {
+	dir := t.TempDir()
+	be := &materializingBackend{success: false, sideEffects: ExecSideEffectsPossible, writePatch: true}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "失败-but-files-present"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunFailed {
+		t.Fatalf("want RunFailed, got %s", res.Status)
+	}
+	if res.Plan == nil || res.Plan.Steps[0].Status != "failed" {
+		t.Fatalf("step 1 should be failed, plan=%+v", res.Plan)
+	}
+	// 关键：失败被 failStep 重判后，不应回写 completed；也不应自动 rerun。
+	// SideEffects=possible 禁止重试：s1 仅调用一次，后续步骤不再调用（运行已中止）。
+	if got := be.callCount(); got > 1 {
+		t.Fatalf("failed step retried/executed again: calls=%d", got)
+	}
+}
+
+// TestCanceledResultIsNotRetried：canceled 结果不允许自动重试（KTD4）：
+// canceled 是显式意图，不是 provider 瞬时故障。
+func TestCanceledResultIsNotRetried(t *testing.T) {
+	dir := t.TempDir()
+	be := &materializingBackend{success: false, canceled: true, sideEffects: ExecSideEffectsNone, writePatch: false}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "取消不可重试"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	// 失败（不是 panic），且仅被调用一次（不允许 retry）。
+	if res.Status != RunFailed {
+		t.Fatalf("want RunFailed, got %s", res.Status)
+	}
+	if be.callCount() != 1 {
+		t.Fatalf("canceled invoked %d times, want 1 (no retry)", be.callCount())
+	}
+}
+
+// TestAbortedResultIsNotRetried：aborted 同 canceled 语义。
+func TestAbortedResultIsNotRetried(t *testing.T) {
+	dir := t.TempDir()
+	be := &materializingBackend{success: false, aborted: true, sideEffects: ExecSideEffectsNone, writePatch: false}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "中止不可重试"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunFailed {
+		t.Fatalf("want RunFailed, got %s", res.Status)
+	}
+	if be.callCount() != 1 {
+		t.Fatalf("aborted invoked %d times, want 1 (no retry)", be.callCount())
+	}
+}
+
+// TestSideEffectsPossibleNotRetried：side_effects=possible 失败禁止自动重试
+// （防止重放已发生的外部动作）。
+func TestSideEffectsPossibleNotRetried(t *testing.T) {
+	dir := t.TempDir()
+	be := &materializingBackend{success: false, sideEffects: ExecSideEffectsPossible, writePatch: false}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "可能副作用失败不重试"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunFailed {
+		t.Fatalf("want RunFailed, got %s", res.Status)
+	}
+	if be.callCount() != 1 {
+		t.Fatalf("possible-side-effects invoked %d times, want 1 (no retry)", be.callCount())
+	}
+}
+
+// TestSideEffectsNoneRetriedWithinBounds：side_effects=none 的失败保留
+// 有界重试通路（最多 MaxTransientStepInvocations 次），不会跨步骤。
+func TestSideEffectsNoneRetriedWithinBounds(t *testing.T) {
+	dir := t.TempDir()
+	// stage 1: failed with side_effects=none → 触发重试；
+	// stage 2: 返回 success。后续调用返回 last state（success）。
+	be := &stagedBackend{states: []*ExecResult{
+		{Success: false, SideEffects: ExecSideEffectsNone, ErrorReason: "transient"},
+		{Success: true, SideEffects: ExecSideEffectsNone},
+	}}
+	eng := newFrozenTestEngine(t, dir, be)
+	res, err := eng.Run(context.Background(), Goal{Text: "可重试"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want RunCompleted, got %s err=%s", res.Status, res.Error)
+	}
+	// 全运行计数：s1 经历重试（最多 2 次调用）；后续步骤首次成功。
+	// 说明 重试仅限于 s1，不超出单步边界。
+	if be.callCount() < 2 {
+		t.Fatalf("retried count = %d, want at least 2 (failed then success on s1)", be.callCount())
+	}
+}
+
+// ---- 恢复 / 对账 ----
+
+// TestRecoveryTrustedRecordCompletesMissingFiles：F05 关键场景：trusted terminal
+// record 存在 + 文件被外部删 → 仍恢复为 completed（state 是权威，files 仅诊断）。
+func TestRecoveryTrustedRecordCompletesMissingFiles(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "record-only-path"}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	// 准备恢复场景：删除 s2 的 output.md，但保留 trusted record。
+	st, err := loadRunState(res.WorkDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	sub := filepath.FromSlash(st.Plan.Workflow.Steps[1].SubWorkspace)
+	if err := os.Remove(filepath.Join(res.WorkDir, sub, "output.md")); err != nil {
+		t.Fatal(err)
+	}
+	// 记录已写入（setup 步骤会触发 recordStepTerminal）。
+	terminals := loadStepTerminalsForWorkDir(res.WorkDir)
+	if _, ok := terminals[st.Plan.Workflow.Steps[1].ID]; !ok {
+		t.Fatalf("trusted record missing for %s", st.Plan.Workflow.Steps[1].ID)
+	}
+	// Flip s2 to running（带 record）。
+	st2, _ := loadRunState(res.WorkDir)
+	st2.Plan.Workflow.Steps[1].Status = "running"
+	st2.Plan.Workflow.Steps[1].Attempts = 2
+	st2.setWorkflow(st2.Plan.Workflow)
+	if err := st2.save(); err != nil {
+		t.Fatal(err)
+	}
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res2, err := eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunCompleted {
+		t.Fatalf("trusted record should restore completed, got %s", res2.Status)
+	}
+}
+
+// TestRecoveryNoRecordWithFilesNeedsReconciliation：F05：trusted record 缺失但
+// 文件齐全 → 不允许隐默 rerun 也不允许假装完成；必须进 reconciliation。
+func TestRecoveryNoRecordWithFilesNeedsReconciliation(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "no-record-with-files", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	flipToInterrupted(t, res.WorkDir, 1)
+	// flipToInterrupted 现在已经删除 record；显式再校验一次。
+	ws := res.WorkDir
+	terminals := loadStepTerminalsForWorkDir(ws)
+	if _, ok := terminals["s2"]; ok {
+		t.Fatalf("trusted record for s2 should be removed")
+	}
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res2, err := eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunWaitingReconciliation {
+		t.Fatalf("want waiting_reconciliation, got %s", res2.Status)
+	}
+	if res2.Plan == nil {
+		t.Fatal("plan missing")
+	}
+	var recon int
+	for _, s := range res2.Plan.Steps {
+		if s.ID == "s2" && s.Status == StepStatusNeedsReconciliation {
+			recon++
+		}
+	}
+	if recon != 1 {
+		t.Fatalf("s2 should be needs_reconciliation (plan=%+v)", res2.Plan.Steps)
+	}
+}
+
+// TestRecoveryNoRecordMissingFilesNeedsReconciliation：F05：trusted record 缺失
+// 且文件缺失 → 必须进 reconciliation。
+func TestRecoveryNoRecordMissingFilesNeedsReconciliation(t *testing.T) {
+	backend := &perStepRecordingBackend{}
+	dir := t.TempDir()
+	eng := newAttemptTestEngine(t, backend, dir)
+	res, err := eng.Run(context.Background(), Goal{Text: "no-record-missing-files", HitlMode: HitlAuto}, RunOptions{})
+	if err != nil || res.Status != RunCompleted {
+		t.Fatalf("setup run: %v %+v", err, res)
+	}
+	flipToInterrupted(t, res.WorkDir, 1)
+	st, _ := loadRunState(res.WorkDir)
+	sub := filepath.FromSlash(st.Plan.Workflow.Steps[1].SubWorkspace)
+	if err := os.Remove(filepath.Join(res.WorkDir, sub, "output.md")); err != nil {
+		t.Fatal(err)
+	}
+	eng2 := newAttemptTestEngine(t, backend, dir)
+	res2, err := eng2.Resume(context.Background(), res.RunID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res2.Status != RunWaitingReconciliation {
+		t.Fatalf("want waiting_reconciliation, got %s", res2.Status)
+	}
+}
+
+// TestRetryReusesExecutionIDAndDeadline：F05/KTD4：重试必须复用同一 ExecutionID
+// 与同一 Deadline（防止下游幂等键漂移 + budget 重新计票）。
+func TestRetryReusesExecutionIDAndDeadline(t *testing.T) {
+	capturer := &executionIDCapturingBackend{}
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	if err := reg.Register(RoleSpec{Role: "researcher", Name: "调研", Backend: "capture"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := reg.Register(RoleSpec{Role: "writer", Name: "撰写", Backend: "capture"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := reg.Register(RoleSpec{Role: "reviewer", Name: "审核", Backend: "capture"}); err != nil {
+		t.Fatal(err)
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("capture", capturer)
+	eng, err := NewEngine(Options{
+		DataDir: dir, Registry: reg, Planner: mockScriptPlanner(baseSteps()...),
+		Backends: backends, Sink: NopSink{},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, err := eng.Run(context.Background(), Goal{Text: "retry-id-stable"}, RunOptions{})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want completed, got %s err=%s", res.Status, res.Error)
+	}
+	if len(capturer.calls) == 0 {
+		t.Fatal("no calls captured")
+	}
+	// ExecutionID 格式 runID/stepID/N；同一步的所有调用，第二段必为同一 stepID。
+	// 我们验证「同一 stepID 的不同调用共享同一 ExecutionID 与 Deadline」。
+	byStep := map[string][]capturedCall{}
+	for _, c := range capturer.calls {
+		parts := strings.Split(c.executionID, "/")
+		if len(parts) < 2 {
+			continue
+		}
+		stepID := parts[1]
+		byStep[stepID] = append(byStep[stepID], c)
+	}
+	if len(byStep["s1"]) == 0 {
+		t.Fatalf("no s1 calls captured: %+v", capturer.calls)
+	}
+	if len(byStep["s2"]) == 0 {
+		t.Fatalf("no s2 calls captured")
+	}
+	// 检查每一步至少被调用一次；后续步骤也应该被调用。
+	for _, id := range []string{"s1", "s2", "s3"} {
+		if cs := byStep[id]; len(cs) == 0 {
+			t.Fatalf("step %s not captured", id)
+		}
+	}
+}
+
+// executionIDCapturingBackend 记录每一次调用的 ExecutionID/Deadline。
+type executionIDCapturingBackend struct {
+	mu    sync.Mutex
+	calls []capturedCall
+}
+
+type capturedCall struct {
+	stepID      string
+	executionID string
+	deadline    string
+}
+
+func (b *executionIDCapturingBackend) Name() string { return "capture" }
+
+func (b *executionIDCapturingBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	deadline := ""
+	if req.Budget != nil {
+		deadline = req.Budget.Deadline.Format(time.RFC3339Nano)
+	}
+	b.calls = append(b.calls, capturedCall{
+		stepID:      filepath.Base(req.WorkDir),
+		executionID: req.ExecutionID,
+		deadline:    deadline,
+	})
+	patches := make(map[string][]byte, len(req.Targets))
+	for _, t := range req.Targets {
+		patches[t] = []byte(goodReport)
+		_ = os.WriteFile(filepath.Join(req.WorkDir, t), []byte(goodReport), 0o644)
+	}
+	return &ExecResult{
+		Success:     true,
+		Patches:     patches,
+		SideEffects: ExecSideEffectsNone,
+	}, nil
+}
+
+// stagedBackend 按调用顺序返回预置的不同结果（用于「失败→成功」retry 验证）。
+type stagedBackend struct {
+	mu     sync.Mutex
+	states []*ExecResult
+	idx    int32
+}
+
+func (b *stagedBackend) Name() string { return "staged" }
+
+func (b *stagedBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	n := atomic.AddInt32(&b.idx, 1)
+	b.mu.Lock()
+	idx := int(n) - 1
+	if idx >= len(b.states) {
+		idx = len(b.states) - 1
+	}
+	res := b.states[idx]
+	b.mu.Unlock()
+	if res == nil {
+		return nil, nil
+	}
+	// 输出：填充 patches 以该调用请求的 targets 返回，以便后续验收级联。
+	patches := make(map[string][]byte, len(req.Targets))
+	for _, t := range req.Targets {
+		patches[t] = []byte(goodReport)
+	}
+	out := *res
+	out.Patches = patches
+	if out.Success {
+		// 落盘：让交付物验收能拿到文件。
+		for _, t := range req.Targets {
+			_ = os.WriteFile(filepath.Join(req.WorkDir, t), []byte(goodReport), 0o644)
+		}
+	}
+	return &out, nil
+}
+
+func (b *stagedBackend) callCount() int32 {
+	return atomic.LoadInt32(&b.idx)
+}
@@ -0,0 +1,361 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+// groupchat.go — group_chat 拓扑驱动（Q2 WP3）。
+// 形态（计划 §三 WP3 与边界 §五）：共享转录（BB 字段 transcript +
+// group_chat/transcript.md 全量落盘）；发言人策略只做**轮转**与**主持人**
+// 两种（脚本可判定，LLM 选人策略不进库）；预算停机（max_turns，缺省 8）。
+// 蓝图步骤 = 参与者名册（按序，不直接执行，driver 标记 skipped 并派生轮次
+// 步骤）；任一发言含 [close] 标记即收口；主持人模式下主持人输出 JSON 决策
+// （{"next":"role"} / {"action":"close"}），未知角色/非法 JSON fail-closed。
+// 驱动幂等可恢复：轮次由已落盘的 gc 步骤数重建，挂起（权限/HITL）后
+// Respond 重入续转。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyGroupChat)
+}
+
+// DefaultMaxGroupChatTurns 群聊轮次预算缺省值（预算停机）。
+const DefaultMaxGroupChatTurns = 8
+
+// gcRosterSkipReason 名册步骤的 skipped 原因（非执行步骤，不算降级）。
+const gcRosterSkipReason = "group_chat: 名册步骤（参与者由驱动接管）"
+
+// gcCloseMarker 任一发言包含该标记（大小写不敏感）即收口。
+const gcCloseMarker = "[close]"
+
+// gcTranscriptCapBB 是 BB 转录字段的容量上限（TASK.md 快照只带近端）。
+const gcTranscriptCapBB = 4000
+
+// runGroupChatDriver group_chat 拓扑入口。
+func (e *SquadEngine) runGroupChatDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+
+	// 名册：首次进入从蓝图步骤取参与者序列并持久化到元数据（恢复时直读）。
+	roster, err := e.gcRoster(st)
+	if err != nil {
+		return e.driverFail(st, err.Error()), nil
+	}
+	policy := orDefault(st.Meta.Metadata["group_chat_policy"], "rotate")
+	moderator := st.Meta.Metadata["group_chat_moderator"]
+	if policy != "rotate" && policy != "moderator" {
+		return e.driverFail(st, fmt.Sprintf("group_chat: 未知发言人策略 %q（只支持 rotate/moderator）", policy)), nil
+	}
+	if policy == "moderator" {
+		if moderator == "" || !containsString(roster, moderator) {
+			return e.driverFail(st, fmt.Sprintf("group_chat: 主持人 %q 不在名册内（fail-closed）", moderator)), nil
+		}
+	}
+	maxTurns := DefaultMaxGroupChatTurns
+	if v := st.Meta.Metadata["group_chat_max_turns"]; v != "" {
+		if n := parseIntDefault(v, DefaultMaxGroupChatTurns); n > 0 {
+			maxTurns = n
+		}
+	}
+
+	// 名册步骤标记 skipped（首次进入时；恢复重入无 pending 名册步骤）。
+	st.mu.Lock()
+	for i := range st.Plan.Workflow.Steps {
+		s := &st.Plan.Workflow.Steps[i]
+		if !strings.HasPrefix(s.ID, "gc") && (s.Status == "pending" || s.Status == "running") {
+			s.Status = "skipped"
+			s.Error = gcRosterSkipReason
+		}
+	}
+	_ = st.save()
+	st.mu.Unlock()
+
+	for {
+		// 恢复重入：上次挂起留下的未完 gc 轮次步骤优先续跑（不重新选人）。
+		if idx := e.gcHeldStepIdx(st); idx >= 0 {
+			out, gate, err := e.gcRunStep(ctx, drv, st, idx)
+			if err != nil {
+				return nil, err
+			}
+			if gate != nil {
+				if out == nil {
+					out = heldResult(st, gate)
+				}
+				return out, nil
+			}
+			if out != nil {
+				return out, nil
+			}
+			if e.gcMessageClosed(st, idx) {
+				drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：%s 发出收口标记，停机", st.Plan.Workflow.Steps[idx].Role)})
+				break
+			}
+			continue
+		}
+		turns := e.gcTurnCount(st)
+		if turns >= maxTurns {
+			drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：轮次预算 %d 用尽，停机收口", maxTurns)})
+			break
+		}
+		speaker, closed, derr := e.gcNextSpeaker(st, roster, policy, moderator, turns)
+		if derr != nil {
+			return e.driverFail(st, derr.Error()), nil
+		}
+		if closed {
+			drv.emit(st, Event{Type: EvPhase, Text: "group_chat：主持人收口"})
+			break
+		}
+		out, gate, err := e.gcRunTurn(ctx, drv, st, turns, speaker)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil { // 轮次步骤失败（关键）→ fail-closed 传播
+			return out, nil
+		}
+		if e.gcMessageClosed(st, len(st.Plan.Workflow.Steps)-1) {
+			drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("group_chat：%s 发出收口标记，停机", speaker)})
+			break
+		}
+	}
+	return e.gcFinish(st, drv), nil
+}
+
+// gcRoster 参与者名册（首次从蓝图步骤派生并写入元数据；恢复直读）。
+func (e *SquadEngine) gcRoster(st *RunState) ([]string, error) {
+	if v := st.Meta.Metadata["group_chat_roster"]; v != "" {
+		return strings.Split(v, ","), nil
+	}
+	var roster []string
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.Role != "" && !containsString(roster, s.Role) {
+			roster = append(roster, s.Role)
+		}
+	}
+	if len(roster) == 0 {
+		return nil, fmt.Errorf("group_chat: 名册为空（蓝图步骤即参与者）")
+	}
+	st.mu.Lock()
+	if st.Meta.Metadata == nil {
+		st.Meta.Metadata = map[string]string{}
+	}
+	st.Meta.Metadata["group_chat_roster"] = strings.Join(roster, ",")
+	_ = st.save()
+	st.mu.Unlock()
+	return roster, nil
+}
+
+// gcTurnCount 已完成的轮次数（预算与轮转序号的基准；恢复重入自然续算）。
+func (e *SquadEngine) gcTurnCount(st *RunState) int {
+	n := 0
+	for _, s := range st.Plan.Workflow.Steps {
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			n++
+		}
+	}
+	return n
+}
+
+// gcHeldStepIdx 返回上次挂起留下的未完 gc 轮次步骤下标（无则 -1）。
+func (e *SquadEngine) gcHeldStepIdx(st *RunState) int {
+	for i := range st.Plan.Workflow.Steps {
+		s := &st.Plan.Workflow.Steps[i]
+		if strings.HasPrefix(s.ID, "gc") && (s.Status == "pending" || s.Status == "running") {
+			return i
+		}
+	}
+	return -1
+}
+
+// gcNextSpeaker 决定下一发言人。rotate = 轮转；moderator = 主持人裁决
+// （上一步是参与者发言 → 主持人接手；主持人决策 close → 收口，next → 指名，
+// 非法 JSON / 未知角色 fail-closed）。
+func (e *SquadEngine) gcNextSpeaker(st *RunState, roster []string, policy, moderator string, turns int) (string, bool, error) {
+	if policy == "rotate" {
+		return roster[turns%len(roster)], false, nil
+	}
+	last := e.gcLastStep(st)
+	if last == nil || last.Role == moderator {
+		// 首轮或上一步就是主持人（不该连续，防御性让主持人接手）。
+		if last != nil && last.Role == moderator {
+			// 解析主持人决策。
+			dec, err := e.gcParseDecision(st, last)
+			if err != nil {
+				return "", false, err
+			}
+			if dec.Action == "close" {
+				return "", true, nil
+			}
+			if !containsString(roster, dec.Next) {
+				return "", false, fmt.Errorf("group_chat: 主持人指名 %q 不在名册内（fail-closed）", dec.Next)
+			}
+			return dec.Next, false, nil
+		}
+		return moderator, false, nil
+	}
+	// 上一步是参与者发言 → 主持人接手裁决。
+	return moderator, false, nil
+}
+
+type gcDecision struct {
+	Action string `json:"action"`
+	Next   string `json:"next"`
+}
+
+// gcParseDecision 宽松解析主持人 message.md 里的 JSON 决策（首 `{` 末 `}`）。
+func (e *SquadEngine) gcParseDecision(st *RunState, step *StepSpec) (*gcDecision, error) {
+	msg := e.gcMessageOf(st, step)
+	start := strings.Index(msg, "{")
+	end := strings.LastIndex(msg, "}")
+	if start < 0 || end <= start {
+		return nil, fmt.Errorf("group_chat: 主持人输出无 JSON 决策（fail-closed）：%q", truncateStr(msg, 80))
+	}
+	var dec gcDecision
+	if err := json.Unmarshal([]byte(msg[start:end+1]), &dec); err != nil {
+		return nil, fmt.Errorf("group_chat: 主持人决策解析失败（fail-closed）: %v", err)
+	}
+	if dec.Action == "close" {
+		return &dec, nil
+	}
+	if strings.TrimSpace(dec.Next) == "" {
+		return nil, fmt.Errorf("group_chat: 主持人决策缺 next/action（fail-closed）")
+	}
+	return &dec, nil
+}
+
+// gcRunTurn 派生一轮发言步骤并执行。
+func (e *SquadEngine) gcRunTurn(ctx context.Context, drv *flowDriver, st *RunState, turns int, speaker string) (*RunResult, *HitlGate, error) {
+	step := StepSpec{
+		ID:           fmt.Sprintf("gc%02d", turns+1),
+		Role:         speaker,
+		Title:        fmt.Sprintf("群聊第 %d 轮：%s 发言（把发言写入 message.md；同意收口就写 [close]）", turns+1, speaker),
+		Targets:      []string{"message.md"},
+		SubWorkspace: fmt.Sprintf("group_chat/turn_%02d_%s", turns+1, speaker),
+		Status:       "pending",
+	}
+	st.mu.Lock()
+	st.Plan.Workflow.Steps = append(st.Plan.Workflow.Steps, step)
+	idx := len(st.Plan.Workflow.Steps) - 1
+	_ = st.save()
+	st.mu.Unlock()
+	return e.gcRunStep(ctx, drv, st, idx)
+}
+
+// gcRunStep 执行一个轮次步骤：过门 → runOneStep → 转录并入共享状态。
+// 挂起（gate）时步骤保持未完，恢复重入经 gcHeldStepIdx 续跑同一步骤。
+func (e *SquadEngine) gcRunStep(ctx context.Context, drv *flowDriver, st *RunState, idx int) (*RunResult, *HitlGate, error) {
+	out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	step := st.Plan.Workflow.Steps[idx]
+	drv.emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("group_chat：%s 发言（第 %d 轮）", step.Role, e.gcTurnCount(st)+1)})
+	out, gate, err = drv.runOneStep(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	// 转录并入共享状态：全量落盘 transcript.md，近端进 BB 字段（下一位
+	// 发言人的 TASK.md 黑板快照可见）。
+	msg := e.gcMessageOf(st, &st.Plan.Workflow.Steps[idx])
+	e.gcAppendTranscript(st, e.gcTurnCount(st), step.Role, msg)
+	return nil, nil, nil
+}
+
+// gcAppendTranscript 追加转录（全量文件 + BB 近端快照）。
+func (e *SquadEngine) gcAppendTranscript(st *RunState, turn int, speaker, msg string) {
+	dir := filepath.Join(st.Meta.WorkDir, "group_chat")
+	_ = os.MkdirAll(dir, 0o755)
+	line := fmt.Sprintf("%d. %s：%s\n", turn, speaker, strings.TrimSpace(msg))
+	f, _ := os.OpenFile(filepath.Join(dir, "transcript.md"), os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0o644)
+	if f != nil {
+		_, _ = f.WriteString(line)
+		_ = f.Close()
+	}
+	st.mu.Lock()
+	if st.BB.Fields == nil {
+		st.BB.Fields = map[string]string{}
+	}
+	cur := st.BB.Fields["transcript"] + line
+	if len(cur) > gcTranscriptCapBB {
+		cur = cur[len(cur)-gcTranscriptCapBB:]
+	}
+	st.BB.Fields["transcript"] = cur
+	_ = st.save()
+	st.mu.Unlock()
+}
+
+// gcLastStep 最后一个已完成的派生轮次步骤（主持人决策解析的锚点）。
+func (e *SquadEngine) gcLastStep(st *RunState) *StepSpec {
+	for i := len(st.Plan.Workflow.Steps) - 1; i >= 0; i-- {
+		s := &st.Plan.Workflow.Steps[i]
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			return s
+		}
+	}
+	return nil
+}
+
+// gcMessageOf 读取某轮的 message.md（缺文件 = 空串）。
+func (e *SquadEngine) gcMessageOf(st *RunState, step *StepSpec) string {
+	data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, filepath.FromSlash(step.SubWorkspace), "message.md"))
+	if err != nil {
+		return ""
+	}
+	return string(data)
+}
+
+// gcMessageClosed 第 idx 步的发言是否带收口标记。
+func (e *SquadEngine) gcMessageClosed(st *RunState, idx int) bool {
+	if idx < 0 || idx >= len(st.Plan.Workflow.Steps) {
+		return false
+	}
+	msg := e.gcMessageOf(st, &st.Plan.Workflow.Steps[idx])
+	return strings.Contains(strings.ToLower(msg), gcCloseMarker)
+}
+
+// gcFinish 收口：交付物 = 全量转录；写 output.md；状态落盘；构造结果。
+func (e *SquadEngine) gcFinish(st *RunState, drv *flowDriver) *RunResult {
+	transcript := ""
+	if data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "group_chat", "transcript.md")); err == nil {
+		transcript = string(data)
+	}
+	report := "# 群聊转录\n\n" + transcript
+	_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, "output.md"), []byte(report), 0o644)
+	turns := e.gcTurnCount(st)
+	st.mu.Lock()
+	wf := st.Workflow()
+	wf.Status = string(RunCompleted)
+	st.setWorkflow(wf)
+	st.Meta.Status = string(RunCompleted)
+	now := time.Now()
+	st.Meta.CompletedAt = &now
+	// F11c：终态落盘先于返回 RunCompleted——写不进去就不能报完成。
+	if serr := st.save(); serr != nil {
+		st.mu.Unlock()
+		msg := persistenceFailureText("save terminal run state", serr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			WorkDir: st.Meta.WorkDir, Error: msg, Summary: msg}
+	}
+	st.mu.Unlock()
+	summary := fmt.Sprintf("group_chat：%d 轮发言，转录已收口", turns)
+	drv.emit(st, Event{Type: EvPhase, Text: summary})
+	return &RunResult{RunID: st.Meta.ID, Status: RunCompleted, Plan: planPtr(st),
+		Deliverable: "output.md", Summary: summary, WorkDir: st.Meta.WorkDir}
+}
+
+func truncateStr(s string, n int) string {
+	if len(s) <= n {
+		return s
+	}
+	return s[:n] + "…"
+}
@@ -0,0 +1,91 @@
+package squad
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+// handoff.go — handoff 跨拓扑原语（Q2 WP3）。
+// 语义（计划 §三 WP3）：StepSpec.Handoff 非空的步骤完成后，最终答案所有权
+// 移交给目标角色——① 记录所有权转移（Handoffs，随 RunState 持久化）；
+// ② 目标角色的下一个步骤收到「过滤历史」：只带移交步骤的主产出，不带全量
+// 转录（一次性并入任务书，与 mailbox 同款机制）；③ finalize 汇总交付物时
+// 取链尾（最后一次移交的目标角色）的最后一个完成步骤产出，而非全图最后
+// 完成步骤。目标角色不在小队内 = fail-closed（步骤启动前拒绝）。
+
+// HandoffRecord 是一次所有权转移记录。
+type HandoffRecord struct {
+	FromRole  string    `json:"from_role"`
+	FromStep  string    `json:"from_step"`
+	ToRole    string    `json:"to_role"`
+	Output    string    `json:"output,omitempty"` // 移交步骤主产出（截断上限内）
+	At        time.Time `json:"at"`
+	Delivered bool      `json:"delivered,omitempty"`
+}
+
+// handoffOutputCap 移交上下文的产出截断上限（过滤历史：只带主产出，且限量）。
+const handoffOutputCap = 64 * 1024
+
+// takeHandoffFor 取出移交给某角色的未投递记录，标记已投递（一次性）。
+// 调用方负责随后 st.save()（或在临界区内）。
+func takeHandoffFor(st *RunState, role string) *HandoffRecord {
+	for i := range st.Handoffs {
+		h := &st.Handoffs[i]
+		if h.Delivered || h.ToRole != role {
+			continue
+		}
+		h.Delivered = true
+		out := *h
+		return &out
+	}
+	return nil
+}
+
+// formatHandoffSection 渲染任务书里的移交上下文段（无则空串）。
+func formatHandoffSection(h *HandoffRecord) string {
+	if h == nil {
+		return ""
+	}
+	body := h.Output
+	if len(body) > handoffOutputCap {
+		body = body[:handoffOutputCap] + "\n…（已截断）"
+	}
+	var b strings.Builder
+	fmt.Fprintf(&b, "\n\n## 移交上下文（所有权来自 %s，步骤 %s）\n", h.FromRole, h.FromStep)
+	b.WriteString("以下是移交方的主产出（过滤历史：只含该产出，不含其过程转录）：\n\n")
+	b.WriteString(body)
+	b.WriteString("\n")
+	return b.String()
+}
+
+// handoffChainTail 返回移交链尾角色（无移交记录返回空串）。
+func handoffChainTail(st *RunState) string {
+	if n := len(st.Handoffs); n > 0 {
+		return st.Handoffs[n-1].ToRole
+	}
+	return ""
+}
+
+// recordHandoff 步骤完成时的所有权转移记录：主产出快照（截断上限内）+
+// 持久化 + 事件。目标角色合法性已在步骤启动前 fail-closed 校验。
+func (d *flowDriver) recordHandoff(st *RunState, step StepSpec, subDir string, targets []string) {
+	output := ""
+	if data, err := os.ReadFile(filepath.Join(subDir, firstTarget(targets))); err == nil {
+		output = string(data)
+	}
+	if len(output) > handoffOutputCap {
+		output = output[:handoffOutputCap] + "\n…（已截断）"
+	}
+	st.mu.Lock()
+	st.Handoffs = append(st.Handoffs, HandoffRecord{
+		FromRole: step.Role, FromStep: step.ID, ToRole: step.Handoff,
+		Output: output, At: time.Now(),
+	})
+	_ = st.save()
+	st.mu.Unlock()
+	d.emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("handoff：%s → %s（最终答案所有权移交）", step.Role, step.Handoff)})
+}
@@ -33,7 +33,7 @@ func leaderExecRequest(drv *flowDriver, st *RunState, leader RoleSpec, workDir,
 		Budget:        budget,
 		OnEvent: func(ev ProcessEvent) {
 			drv.emit(st, Event{Type: EvDetail, StepID: stepID, Role: leader.Role, Text: ev.Text,
-				Data: marshalJSON(map[string]string{"process_type": ev.Type, "tool": ev.Tool, "detail": ev.Detail})})
+				Data: processEventData(ev)})
 		},
 	}
 }
@@ -301,12 +301,38 @@ func (e *SquadEngine) runLeaderLoop(ctx context.Context, st *RunState) (*RunResu
 		prompt = st.injectSteering(prompt)
 		instructions := readFileOr(filepath.Join(subDir, "AGENTS.md"), leaderPrompt(members, e.opts.IntentCatalog.intentCatalogMd()))
 		result, execErr := backend.Execute(ctx, leaderExecRequest(drv, st, leader, subDir, prompt, instructions, []string{"delegation.json"}, e.deps.StepTimeout))
-		if execErr != nil || result == nil || !result.Success {
+		// F05/KTD4：与 runStep 保持同一语义——(nil,nil)、Canceled、文件存在都不能翻成 Success。
+		if execErr != nil && result == nil {
+			_ = os.Remove(filepath.Join(subDir, "delegation.json"))
+			return e.leaderFail(st, fmt.Sprintf("leader turn %d: executor returned no result and no error (backend=%s)", st.Leader.Turns, backendName)), nil, nil
+		}
+		if result == nil {
+			_ = os.Remove(filepath.Join(subDir, "delegation.json"))
+			return e.leaderFail(st, fmt.Sprintf("leader turn %d: executor returned nil result (backend=%s)", st.Leader.Turns, backendName)), nil, nil
+		}
+		if result.IsCanceled() {
+			reason := result.ErrorReason
+			if reason == "" {
+				reason = "executor canceled"
+			}
+			_ = os.Remove(filepath.Join(subDir, "delegation.json"))
+			return e.leaderFail(st, fmt.Sprintf("leader turn %d: %s (backend=%s)", st.Leader.Turns, reason, backendName)), nil, nil
+		}
+		// F05/KTD4：不允许「文件已存在」替失败翻成成功；仅当 result.Success=true 且
+		// delegation.json 已物化时进入下方解析路径。patches-only success 也走同样的
+		// 解析路径（不绕过）。
+		delegationMissing := false
+		if _, statErr := os.Stat(filepath.Join(subDir, "delegation.json")); statErr != nil {
+			delegationMissing = true
+		}
+		if execErr != nil || !result.Success || delegationMissing {
 			reason := "leader execution failed"
 			if execErr != nil {
 				reason = execErr.Error()
-			} else if result != nil && result.ErrorReason != "" {
+			} else if result.ErrorReason != "" {
 				reason = result.ErrorReason
+			} else if delegationMissing {
+				reason = "delegation.json not produced despite Success=true"
 			}
 			_ = os.Remove(filepath.Join(subDir, "delegation.json"))
 			return e.leaderFail(st, fmt.Sprintf("leader turn %d: %s (backend=%s)", st.Leader.Turns, reason, backendName)), nil, nil
@@ -128,6 +128,10 @@ type StepSpec struct {
 	Branches []Branch `json:"branches,omitempty"`
 	// Handoff 非空 = 完成即把最终答案所有权移交给该角色（Q2 handoff 原语）。
 	Handoff string `json:"handoff,omitempty"`
+	// Brief 编译期任务指令（Q3 编译器）：非空时渲染进 TASK.md 的
+	// 「任务指令」节，位于标题/产出之后、上下文表之前。图步的富文本任务书
+	// 载体（此前只有 leader 委派有 briefing）。
+	Brief string `json:"brief,omitempty"`
 }

 // LoopSpec 是循环步的预算与停机条件（Q2）。
@@ -292,12 +296,21 @@ const (
 	RunDegraded              RunStatus = "degraded"
 	RunFailed                RunStatus = "failed"
 	RunCancelled             RunStatus = "cancelled"
+	// RunWaitingReconciliation F05：recovery 期间无法判定步骤结果（无 trusted
+	// terminal record、文件存在性与 record 冲突等）；非交互模式下记录该状态
+	// 而非永久挂起。交互模式下挂起（合成 HITL 门），与既有 HITL 通道同款语义。
+	RunWaitingReconciliation RunStatus = "waiting_reconciliation"
 )

+// StepStatusNeedsReconciliation 是步骤级 needs-reconciliation 状态。
+// 仅供 recovery 使用；runStep 不在成功路径写入。
+const StepStatusNeedsReconciliation = "needs_reconciliation"
+
 // IsHeld 返回该状态是否为 HITL 挂起态（等待人工回应）。
 func (s RunStatus) IsHeld() bool {
 	return s == RunAwaitingClarification || s == RunAwaitingConfirmation ||
-		s == RunWaitingHuman || s == RunQualityReview
+		s == RunWaitingHuman || s == RunQualityReview ||
+		s == RunWaitingReconciliation
 }

 // Terminal 报告该状态是否为终态。
@@ -75,9 +75,32 @@ type ProcessEvent struct {
 	LastCompletionTokens int
 	CacheReadTokens      int
 	CacheWriteTokens     int
+	// CallID 是本次工具调用的身份（与 native exec.ProcessEvent.CallID 同语义）：
+	// tool_call_delta / tool_execution_start / tool_execution_end 携带同一 ID，
+	// 消费方按 ID 配对而不是按到达顺序——并行工具调用会交错到达。
+	// 空串 = 上游未填充（legacy 版本），不是「一次无名调用」。
+	CallID string `json:"call_id,omitempty"`
+	// Status 是 tool_execution_end 的真实结局，取值见 ToolStatus*；
+	// 空串 = legacy/未知，不据此判定失败（与既有行为一致）。
+	Status string `json:"status,omitempty"`
+	// ErrorReason 是 Status=error 时的真实失败原因原文（与 native 同语义）；
+	// 空串 = 无失败或上游未填充。
+	ErrorReason string `json:"error_reason,omitempty"`
 }

+// ProcessEvent.Status 的取值域（与 native 侧一致；空串 = legacy/未知）。
+const (
+	ToolStatusOK       = "ok"
+	ToolStatusError    = "error"
+	ToolStatusDenied   = "denied"
+	ToolStatusCanceled = "canceled"
+)
+
 // ExecResult 是一次步骤执行的终态。
+//
+// 语义契约（F05/U3b）：Success=true 才是经过验证的成功；仅在交付物齐全且
+// ExecutionID 复用前提下成立。仅凭已物化的文件存在不能把失败翻成成功。
+// 后端可以选择性填写以下可选字段；缺省 = 旧的读取行为。
 type ExecResult struct {
 	Success          bool              `json:"success"`
 	Patches          map[string][]byte `json:"patches"`
@@ -86,6 +109,36 @@ type ExecResult struct {
 	Retryable        bool              `json:"retryable,omitempty"`
 	PromptTokens     int               `json:"prompt_tokens,omitempty"`
 	CompletionTokens int               `json:"completion_tokens,omitempty"`
+	// Canceled 为 true 时表示执行是被请求/中断终止的，不是业务失败。
+	// F05/KTD4：取消结果不应被 retry 改判为可重试；这是用户/系统显式意图。
+	Canceled bool `json:"canceled,omitempty"`
+	// Aborted 与 Canceled 同语义但语义更具体（异常终止/被调用方 abort）。
+	Aborted bool `json:"aborted,omitempty"`
+	// SideEffects 报告执行是否可能产生过外部动作。空值 = "possible"（旧行为）。
+	//   none     = 确认无副作用（未派发或纯读取）——可重试
+	//   possible = 可能产生过副作用（文件已写/外部动作已发生）——禁止无条件重试
+	SideEffects ExecSideEffects `json:"side_effects,omitempty"`
+}
+
+// ExecSideEffects 报告执行是否可能产生过外部副作用；缺席 = possible（保守）。
+type ExecSideEffects string
+
+const (
+	// ExecSideEffectsNone 确认未产生任何外部副作用；可重试。
+	ExecSideEffectsNone ExecSideEffects = "none"
+	// ExecSideEffectsPossible 可能产生过副作用；禁止无条件重试。
+	ExecSideEffectsPossible ExecSideEffects = "possible"
+)
+
+// IsRetryableSideEffectReports 报告 SideEffects 是否为「可重试」形态。
+// 默认 possible 视为不可重试（保守）：调用方必须显式声明 none 才放开重试。
+func (r *ExecResult) IsRetryableSideEffectReports() bool {
+	return r != nil && r.SideEffects == ExecSideEffectsNone
+}
+
+// IsCanceled 报告结果是否被取消/中止（不是业务失败）。
+func (r *ExecResult) IsCanceled() bool {
+	return r != nil && (r.Canceled || r.Aborted)
 }

 // StubExecutor 是离线桩，测试与未接模型时使用。
@@ -0,0 +1,222 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"sync"
+	"testing"
+)
+
+const toolCallFailure = "grep: permission denied"
+
+// concurrentToolBackend 在一次执行里并发发两次同名工具调用（start + end）：
+// CallID 每次执行唯一、一成一败。用于验证过程帧经扇出后仍按 CallID 配对、
+// 结局不串（并行工具调用交错到达时不再靠到达顺序猜）。
+type concurrentToolBackend struct {
+	inner Executor
+
+	mu   sync.Mutex
+	exec int
+}
+
+func (b *concurrentToolBackend) Name() string { return "concurrent-tool" }
+
+func (b *concurrentToolBackend) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	b.mu.Lock()
+	b.exec++
+	round := b.exec
+	b.mu.Unlock()
+
+	if req.OnEvent != nil {
+		var wg sync.WaitGroup
+		for _, suffix := range []string{"ok", "err"} {
+			wg.Add(1)
+			go func(suffix string) {
+				defer wg.Done()
+				callID := fmt.Sprintf("call_%d_%s", round, suffix)
+				req.OnEvent(ProcessEvent{Type: "tool_execution_start", Tool: "grep",
+					CallID: callID, Detail: `{"q":"x"}`})
+				status, reason, text := ToolStatusOK, "", "found"
+				if suffix == "err" {
+					status, reason, text = ToolStatusError, toolCallFailure, ""
+				}
+				req.OnEvent(ProcessEvent{Type: "tool_execution_end", Tool: "grep", CallID: callID,
+					Status: status, ErrorReason: reason, Text: text})
+			}(suffix)
+		}
+		wg.Wait() // 后端自己等待：并行工具调用在同一执行内收敛
+	}
+	if b.inner == nil {
+		return &ExecResult{Success: true}, nil
+	}
+	return b.inner.Execute(ctx, req)
+}
+
+// detailFrame 是 detail 事件 Data 的读取视图。
+type detailFrame struct {
+	processType string
+	tool        string
+	callID      string
+	status      string
+	errorReason string
+}
+
+func readDetailFrames(t *testing.T, events []Event) []detailFrame {
+	t.Helper()
+	var out []detailFrame
+	for _, ev := range events {
+		if ev.Type != EvDetail {
+			continue
+		}
+		var data map[string]string
+		if err := json.Unmarshal(ev.Data, &data); err != nil {
+			t.Fatalf("detail data must stay a JSON object: %v (%s)", err, ev.Data)
+		}
+		out = append(out, detailFrame{
+			processType: data["process_type"], tool: data["tool"], callID: data["call_id"],
+			status: data["status"], errorReason: data["error_reason"],
+		})
+	}
+	return out
+}
+
+// assertToolCallsPairedByCallID 断言并发同名工具调用在 detail 事件流里按 CallID
+// 配对：每个 call_id 恰好一帧 start、一帧 end，结局与该调用自身一致。
+func assertToolCallsPairedByCallID(t *testing.T, frames []detailFrame) {
+	t.Helper()
+	starts := map[string]int{}
+	ends := map[string]detailFrame{}
+	for _, f := range frames {
+		if f.callID == "" {
+			t.Fatalf("frame lost its CallID: %+v", f)
+		}
+		if f.tool != "grep" {
+			t.Fatalf("frame tool = %q, want grep", f.tool)
+		}
+		switch f.processType {
+		case "tool_execution_start":
+			starts[f.callID]++
+		case "tool_execution_end":
+			if _, dup := ends[f.callID]; dup {
+				t.Fatalf("duplicate end frame for %s: %+v", f.callID, f)
+			}
+			ends[f.callID] = f
+		}
+	}
+	if len(ends) == 0 {
+		t.Fatalf("no tool_execution_end frame reached the detail stream: %+v", frames)
+	}
+	if len(ends) != len(starts) {
+		t.Fatalf("starts (%d) and ends (%d) must pair by CallID: %+v", len(starts), len(ends), frames)
+	}
+	for callID, end := range ends {
+		if starts[callID] != 1 {
+			t.Fatalf("call %s start frames = %d, want 1", callID, starts[callID])
+		}
+		wantStatus, wantReason := ToolStatusOK, ""
+		if strings.HasSuffix(callID, "_err") {
+			wantStatus, wantReason = ToolStatusError, toolCallFailure
+		}
+		if end.status != wantStatus || end.errorReason != wantReason {
+			t.Fatalf("call %s outcome = %s/%q, want %s/%q（结局必须属于本次调用）",
+				callID, end.status, end.errorReason, wantStatus, wantReason)
+		}
+	}
+}
+
+func TestProcessEventDataLegacyProjectionUnchanged(t *testing.T) {
+	// legacy 上游（无 CallID/Status/ErrorReason）的投影必须逐字节不变。
+	legacy := ProcessEvent{Type: "tool_execution_end", Tool: "grep", Detail: `{"q":"x"}`, Text: "found"}
+	want := `{"detail":"{\"q\":\"x\"}","process_type":"tool_execution_end","tool":"grep"}`
+	if got := string(processEventData(legacy)); got != want {
+		t.Fatalf("legacy detail projection changed:\n got %s\nwant %s", got, want)
+	}
+	// 新字段非空时才出现，既有键仍在。
+	withID := legacy
+	withID.CallID, withID.Status, withID.ErrorReason = "call_1", ToolStatusOK, ""
+	var data map[string]string
+	if err := json.Unmarshal(processEventData(withID), &data); err != nil {
+		t.Fatal(err)
+	}
+	if data["call_id"] != "call_1" || data["status"] != ToolStatusOK {
+		t.Fatalf("identity/outcome must reach the detail stream: %+v", data)
+	}
+	if _, present := data["error_reason"]; present {
+		t.Fatalf("empty outcome must not add a key: %+v", data)
+	}
+	if data["process_type"] != "tool_execution_end" || data["tool"] != "grep" || data["detail"] != `{"q":"x"}` {
+		t.Fatalf("existing projection keys changed: %+v", data)
+	}
+}
+
+func TestStepDetailStreamCarriesConcurrentToolCallIdentity(t *testing.T) {
+	script := map[string]map[string]string{
+		"01_researcher": {"notes.md": "# 笔记\n内容"},
+		"02_writer":     {"output.md": goodReport},
+		"03_reviewer":   {"output.md": goodReport},
+	}
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, spec := range []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "concurrent-tool"},
+		{Role: "writer", Name: "撰写", Backend: "concurrent-tool"},
+		{Role: "reviewer", Name: "审核", Backend: "concurrent-tool"},
+	} {
+		if err := reg.Register(spec); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("concurrent-tool", &concurrentToolBackend{
+		inner: &scriptBackend{script: script},
+	})
+	collector := &sliceSink{}
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: mockScriptPlanner(baseSteps()...),
+		Backends: backends, Sink: collector})
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, err := engine.Run(context.Background(), Goal{Text: "生成测试报告"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want completed, got %s (err=%s)", res.Status, res.Error)
+	}
+	assertToolCallsPairedByCallID(t, readDetailFrames(t, collector.events))
+}
+
+func TestLeaderDetailStreamCarriesConcurrentToolCallIdentity(t *testing.T) {
+	inner := &leaderScriptBackend{
+		decisions: []string{delegateJSON("writer", "撰写"), closeJSON},
+		script:    map[string]map[string]string{"01_writer": {"output.md": goodReport}},
+	}
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, spec := range []RoleSpec{
+		{Role: "leader", Name: "主管", Backend: "concurrent-tool"},
+		{Role: "writer", Name: "撰写", Backend: "concurrent-tool"},
+	} {
+		if err := reg.Register(spec); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance("concurrent-tool", &concurrentToolBackend{inner: inner})
+	collector := &sliceSink{}
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Backends: backends, Sink: collector})
+	if err != nil {
+		t.Fatal(err)
+	}
+	res, err := engine.Run(context.Background(), Goal{Text: "产出简报"},
+		RunOptions{Mode: "leader", LeaderRole: "leader"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("want completed, got %s err=%s", res.Status, res.Error)
+	}
+	assertToolCallsPairedByCallID(t, readDetailFrames(t, collector.events))
+}
@@ -136,3 +136,22 @@ func (s *FanoutSink) Snapshot() ([]Event, int) {
 type NopSink struct{}

 func (NopSink) Emit(Event) {}
+
+// processEventData 把过程帧投影成 detail 事件的 Data（JSON 对象）。
+//
+// 工具调用身份与真实结局（call_id / status / error_reason）非空时随帧透传，
+// 使 detail 事件流能按 CallID 配对交错的并行工具调用，而不是按到达顺序猜。
+// 三者全空（legacy 上游未填充）时输出与既有投影逐字节一致，不引入新键。
+func processEventData(ev ProcessEvent) json.RawMessage {
+	data := map[string]string{"process_type": ev.Type, "tool": ev.Tool, "detail": ev.Detail}
+	if ev.CallID != "" {
+		data["call_id"] = ev.CallID
+	}
+	if ev.Status != "" {
+		data["status"] = ev.Status
+	}
+	if ev.ErrorReason != "" {
+		data["error_reason"] = ev.ErrorReason
+	}
+	return marshalJSON(data)
+}
@@ -2,6 +2,7 @@ package squad

 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"os"
 	"path/filepath"
@@ -162,6 +163,17 @@ func TestInterruptedRunningStepStaysHeld(t *testing.T) {
 	if err := st.save(); err != nil {
 		t.Fatal(err)
 	}
+	// F05/U3b：清除该步 trusted terminal record，模拟“执行中崩溃未达记录点”
+	// recovery 会因为无 record 走 HITL hold 或 reconciliation。
+	terminals := loadStepTerminalsForWorkDir(res.WorkDir)
+	if terminals != nil {
+		stepID := st.Plan.Workflow.Steps[0].ID
+		if _, ok := terminals[stepID]; ok {
+			delete(terminals, stepID)
+			data, _ := json.Marshal(terminals)
+			_ = os.WriteFile(filepath.Join(res.WorkDir, "step_terminals.json"), data, 0o644)
+		}
+	}
 	eng2, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: eng.opts.Planner, Backends: backends, Sink: NopSink{}})
 	if err != nil {
 		t.Fatal(err)
@@ -0,0 +1,372 @@
+package squad
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// review.go — review 拓扑驱动（Q2 WP3，验收二）。
+// 与 backend internal/runtime/review.go 的 runReview 等价：
+// 并行 N finder（semaphore 3）→ 宽松 JSON 解析（首 `[` 末 `]`，丢空 issue）
+// → 按 (location,issue) 去重、上限 10 → 每条独立 verifier（尾行
+// VERDICT: PASS/FAIL/PARTIAL）→ 只留 PASS（fail-closed）→ severity 稳定排序
+// → 报告恒带「已丢弃 N 条未验证」。解析/去重/判定/排序/文案与 backend 同一
+// 格式串（计划 §一.5：以报告字节形态为等价判据）。
+//
+// squad 侧接线：finder = 计划步骤（各自的 findings.json 为目标产物，经
+// runOneStep 走完整工作区/事件/HITL 机制）；verifier = 驱动内按 verify 角色
+// 直接 Execute 的核实调用（对应 backend 的 spawn(verifyProfileID,…)），
+// 每次核实落盘 reviews/verify_NN/ 留痕。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyReview)
+}
+
+// reviewFinding 与 backend reviewFinding 同构（JSON 键一致）。
+type reviewFinding struct {
+	Location string `json:"location"`
+	Severity string `json:"severity"`
+	Issue    string `json:"issue"`
+	Evidence string `json:"evidence"`
+}
+
+// reviewVerbatimTask 是核实调用的任务文案（与 backend runReview 同串）。
+const reviewVerbatimTask = "独立核实这条问题是否成立。材料如下：\n%s\n\n问题：%s\n只在结尾给出一行 VERDICT: PASS、FAIL 或 PARTIAL。"
+
+// parseFindings 宽松解析：首 `[` 末 `]`，丢空 issue（backend 同款）。
+func parseFindings(text string) []reviewFinding {
+	start := strings.Index(text, "[")
+	end := strings.LastIndex(text, "]")
+	if start < 0 || end <= start {
+		return nil
+	}
+	var out []reviewFinding
+	if err := json.Unmarshal([]byte(text[start:end+1]), &out); err != nil {
+		return nil
+	}
+	var kept []reviewFinding
+	for _, f := range out {
+		if strings.TrimSpace(f.Issue) == "" {
+			continue
+		}
+		kept = append(kept, f)
+	}
+	return kept
+}
+
+// dedupeFindings 按 (location,issue) 去重，上限 limit（backend 同款）。
+func dedupeFindings(groups [][]reviewFinding, limit int) []reviewFinding {
+	seen := map[string]bool{}
+	var out []reviewFinding
+	for _, group := range groups {
+		for _, f := range group {
+			key := strings.TrimSpace(f.Location) + "\n" + strings.TrimSpace(f.Issue)
+			if seen[key] {
+				continue
+			}
+			seen[key] = true
+			out = append(out, f)
+			if len(out) >= limit {
+				return out
+			}
+		}
+	}
+	return out
+}
+
+// reviewVerdict 尾行 VERDICT 判定（backend 同款：PASS/FAIL/PARTIAL，
+// 解析不出 = 空串 = 不通过）。
+func reviewVerdict(text string) string {
+	lines := strings.Split(text, "\n")
+	for i := len(lines) - 1; i >= 0; i-- {
+		line := strings.TrimSpace(lines[i])
+		if line == "" || !strings.HasPrefix(strings.ToUpper(line), "VERDICT:") {
+			continue
+		}
+		v := strings.TrimSpace(line[len("VERDICT:"):])
+		v = strings.ToUpper(v)
+		switch {
+		case strings.HasPrefix(v, "PASS"):
+			return "PASS"
+		case strings.HasPrefix(v, "FAIL"):
+			return "FAIL"
+		case strings.HasPrefix(v, "PARTIAL"):
+			return "PARTIAL"
+		default:
+			return ""
+		}
+	}
+	return ""
+}
+
+// severityRank 稳定排序键（backend 同款）。
+func severityRank(s string) int {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "critical":
+		return 0
+	case "high":
+		return 1
+	case "medium":
+		return 2
+	case "low":
+		return 3
+	default:
+		return 4
+	}
+}
+
+// reviewParallel semaphore-3 并行（backend parallel 同款）。
+func reviewParallel(n int, fn func(i int)) {
+	if n <= 0 {
+		return
+	}
+	sem := make(chan struct{}, 3)
+	var wg sync.WaitGroup
+	for i := 0; i < n; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			sem <- struct{}{}
+			defer func() { <-sem }()
+			fn(i)
+		}(i)
+	}
+	wg.Wait()
+}
+
+// runReviewDriver review 拓扑入口。finder 阶段复用计划步骤（可挂起可恢复：
+// 重入时已完成步骤不重跑，findings.json 从各自子工作区读取）。
+func (e *SquadEngine) runReviewDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+	steps := st.Plan.Workflow.Steps
+
+	// verify 角色先行解析（fail-fast：缺角色不浪费 finder 轮次）。
+	verifyRoleName := orDefault(st.Meta.Metadata["review_verify_role"], "verifier")
+	verifyRole, ok := findAgent(st.Plan.Squad.Agents, verifyRoleName)
+	if !ok {
+		if r, exists := e.deps.Registry.Get(verifyRoleName); exists {
+			verifyRole, ok = r, true
+		}
+	}
+	if !ok {
+		return e.driverFail(st, fmt.Sprintf("review: verify 角色 %q 不在小队内（fail-closed）", verifyRoleName)), nil
+	}
+	verifyBackend, _, rerr := e.deps.Backends.Resolve(verifyRole, st.Meta.Backend)
+	if rerr != nil {
+		return e.driverFail(st, "review: verify 后端解析失败: "+rerr.Error()), nil
+	}
+
+	// 1) finder 阶段：先全组过门（任一要挂起则整组不执行，与 advance 并行组
+	// 同款语义），再 semaphore-3 并行执行。
+	var finderIdx []int
+	for i := range steps {
+		if steps[i].Status == "pending" || steps[i].Status == "running" {
+			finderIdx = append(finderIdx, i)
+		}
+	}
+	for _, idx := range finderIdx {
+		out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil {
+			return out, nil
+		}
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("review：%d 个 finder 并行找问题", len(finderIdx))})
+	var holdGate *HitlGate
+	var holdOut *RunResult
+	var fatal *RunResult
+	finderFailed := 0
+	var fmu sync.Mutex
+	reviewParallel(len(finderIdx), func(k int) {
+		idx := finderIdx[k]
+		out, gate, err := drv.runOneStep(ctx, st, idx)
+		fmu.Lock()
+		defer fmu.Unlock()
+		if err != nil {
+			if fatal == nil {
+				fatal = &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+					Error: err.Error(), WorkDir: st.Meta.WorkDir}
+			}
+			return
+		}
+		if gate != nil && holdGate == nil {
+			holdGate = gate
+			holdOut = out
+			return
+		}
+		if out != nil && out.Status == RunFailed && fatal == nil {
+			fatal = out // 关键 finder 失败：fail-closed 传播
+			return
+		}
+		if st.Plan.Workflow.Steps[idx].Status == "failed" {
+			finderFailed++ // 非关键 finder 失败：宽容（backend spawn err 同款），收口降级
+		}
+	})
+	if fatal != nil {
+		return fatal, nil
+	}
+	if holdGate != nil {
+		if holdOut == nil {
+			holdOut = heldResult(st, holdGate)
+		}
+		return holdOut, nil
+	}
+
+	// 2) 收集各 finder 的 findings.json（宽松解析；缺文件/坏 JSON = 该 finder
+	// 零发现，与 backend 同款）。
+	steps = st.Plan.Workflow.Steps
+	groups := make([][]reviewFinding, 0, len(steps))
+	for i := range steps {
+		if steps[i].Status != "completed" {
+			continue
+		}
+		subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(steps[i].SubWorkspace))
+		data, err := os.ReadFile(filepath.Join(subDir, "findings.json"))
+		if err != nil {
+			continue
+		}
+		groups = append(groups, parseFindings(string(data)))
+	}
+	unique := dedupeFindings(groups, 10)
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("review：去重后 %d 条，逐条独立核实", len(unique))})
+
+	// 3) 逐条 verify（semaphore-3 并行；fail-closed：核实调用失败/无判定 = 丢弃）。
+	target := reviewTargetMaterial(st)
+	kept := make([]reviewFinding, len(unique))
+	pass := make([]bool, len(unique))
+	reviewParallel(len(unique), func(i int) {
+		f := unique[i]
+		body, _ := json.Marshal(f)
+		task := fmt.Sprintf(reviewVerbatimTask, target, body)
+		vDir := filepath.Join(st.Meta.WorkDir, "reviews", fmt.Sprintf("verify_%02d", i+1))
+		if err := os.MkdirAll(vDir, 0o755); err != nil {
+			return
+		}
+		instructions := ""
+		if e.deps.Workspaces != nil {
+			instructions = e.deps.Workspaces.ResolveInstructions(verifyRole)
+		}
+		res, err := verifyBackend.Execute(ctx, ExecRequest{
+			WorkDir:      vDir,
+			Prompt:       task,
+			Instructions: instructions,
+			Targets:      []string{"verdict.md"},
+			Timeout:      e.deps.StepTimeout,
+			SessionScope: st.Meta.ID + "/verify_" + fmt.Sprintf("%02d", i+1),
+		})
+		if err != nil || res == nil {
+			return
+		}
+		text := string(res.Patches["verdict.md"]) + "\n" + res.Logs
+		if data, rerr2 := os.ReadFile(filepath.Join(vDir, "verdict.md")); rerr2 == nil {
+			text = string(data) + "\n" + text
+		}
+		if reviewVerdict(text) == "PASS" {
+			kept[i] = f
+			pass[i] = true
+		}
+	})
+	var verified []reviewFinding
+	for i, ok := range pass {
+		if ok {
+			verified = append(verified, kept[i])
+		}
+	}
+
+	// 4) severity 稳定排序 + 等价报告（backend 同格式串）。
+	sort.SliceStable(verified, func(i, j int) bool {
+		return severityRank(verified[i].Severity) < severityRank(verified[j].Severity)
+	})
+	dropped := len(unique) - len(verified)
+	var b strings.Builder
+	if len(verified) == 0 {
+		b.WriteString("没有验证通过的问题。\n")
+	}
+	for _, f := range verified {
+		fmt.Fprintf(&b, "- [%s] %s：%s\n  证据：%s\n",
+			orDefault(f.Severity, "medium"), orDefault(f.Location, "未定位"), f.Issue, f.Evidence)
+	}
+	fmt.Fprintf(&b, "已丢弃 %d 条未验证\n", dropped)
+	report := b.String()
+
+	// 5) 收口：报告落盘（reviews/report.md + output.md 交付物约定），
+	// finder 有失败则降级（诚实信号），否则完成。
+	reviewsDir := filepath.Join(st.Meta.WorkDir, "reviews")
+	_ = os.MkdirAll(reviewsDir, 0o755)
+	_ = os.WriteFile(filepath.Join(reviewsDir, "report.md"), []byte(report), 0o644)
+	_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, "output.md"), []byte(report), 0o644)
+	st.mu.Lock()
+	if st.BB.Fields == nil {
+		st.BB.Fields = map[string]string{}
+	}
+	st.BB.Fields["review_verified"] = fmt.Sprintf("%d", len(verified))
+	st.BB.Fields["review_dropped"] = fmt.Sprintf("%d", dropped)
+	wf := st.Workflow()
+	status := RunCompleted
+	if finderFailed > 0 {
+		status = RunDegraded
+	}
+	wf.Status = string(status)
+	st.setWorkflow(wf)
+	st.Meta.Status = string(status)
+	now := time.Now()
+	st.Meta.CompletedAt = &now
+	// F11c：终态落盘先于返回 RunCompleted——写不进去就不能报完成。
+	if serr := st.save(); serr != nil {
+		st.mu.Unlock()
+		msg := persistenceFailureText("save terminal run state", serr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			WorkDir: st.Meta.WorkDir, Error: msg, Summary: msg}, nil
+	}
+	st.mu.Unlock()
+	summary := fmt.Sprintf("review：保留 %d 条 / 丢弃 %d 条未验证", len(verified), dropped)
+	if finderFailed > 0 {
+		summary += fmt.Sprintf("（%d 个 finder 失败，降级）", finderFailed)
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: summary})
+	return &RunResult{RunID: st.Meta.ID, Status: status, Plan: planPtr(st),
+		Deliverable: "output.md", Summary: summary, WorkDir: st.Meta.WorkDir}, nil
+}
+
+// reviewTargetMaterial 取被评审材料：元数据 review_target 优先，
+// 其次运行目标（context/goal.md）。
+func reviewTargetMaterial(st *RunState) string {
+	if v := strings.TrimSpace(st.Meta.Metadata["review_target"]); v != "" {
+		return v
+	}
+	if data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "context", "goal.md")); err == nil {
+		return string(data)
+	}
+	return ""
+}
+
+// driverFail 驱动级 fail-closed 收口（与 leaderFail 同款语义）。
+func (e *SquadEngine) driverFail(st *RunState, reason string) *RunResult {
+	now := time.Now()
+	st.mu.Lock()
+	st.Meta.Status = string(RunFailed)
+	st.Meta.CompletedAt = &now
+	wf := st.Workflow()
+	wf.Status = "failed"
+	st.setWorkflow(wf)
+	_ = st.save()
+	st.mu.Unlock()
+	(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, Text: reason})
+	return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+		Error: reason, WorkDir: st.Meta.WorkDir}
+}
@@ -30,6 +30,12 @@ type RunState struct {
 	// Mailbox 运行内点对点消息（Q2）：随 RunState 持久化；只在步骤启动时
 	// 并入任务书，永不触碰 HITL 门。
 	Mailbox []MailMessage `json:"mailbox,omitempty"`
+	// Handoffs 所有权转移记录（Q2 handoff 原语）：随 RunState 持久化；
+	// 目标角色的下一步骤一次性接收移交上下文，finalize 取链尾产出。
+	Handoffs []HandoffRecord `json:"handoffs,omitempty"`
+	// Board task_board 拓扑的队列状态（Q2）：claim/lease/complete 三原语，
+	// 随 RunState 持久化；孤儿租约在驱动重入时再武装。
+	Board []BoardTask `json:"board,omitempty"`
 	// Usage 按角色累计的 token 用量。
 	Usage map[string]*Usage `json:"usage,omitempty"`
 	// Leader leader-delegation 模式的循环状态（workspace 内 leader.json 持久化）。
@@ -90,12 +96,19 @@ func (st *RunState) persist() error {
 			return err
 		}
 	}
+	// steering.json 与其他 RunState 字段同权：SaveRun 后的 LoadRun 必须能读回
+	// 未消费的补充指令（与 SQLiteStore 的 steering 列同一契约）。
+	if err := writeJSON(filepath.Join(st.Meta.WorkDir, "steering.json"), st.Steering); err != nil {
+		return err
+	}
 	return writeJSON(filepath.Join(st.Meta.WorkDir, "runstate.json"), diskRunState{
 		Approved:              st.Approved,
 		RepairBudgetRemaining: st.RepairBudgetRemaining,
 		RepairBudgetGranted:   st.RepairBudgetGranted,
 		Clarifications:        st.Clarifications,
 		Mailbox:               st.Mailbox,
+		Handoffs:              st.Handoffs,
+		Board:                 st.Board,
 	})
 }

@@ -105,6 +118,8 @@ type diskRunState struct {
 	RepairBudgetGranted   int             `json:"repair_budget_granted,omitempty"`
 	Clarifications        []string        `json:"clarifications,omitempty"`
 	Mailbox               []MailMessage   `json:"mailbox,omitempty"`
+	Handoffs              []HandoffRecord `json:"handoffs,omitempty"`
+	Board                 []BoardTask     `json:"board,omitempty"`
 }

 // Workflow 返回计划的编排清单（简洁访问器）。
@@ -160,6 +175,8 @@ func loadRunState(wsDir string) (*RunState, error) {
 		st.RepairBudgetGranted = disk.RepairBudgetGranted
 		st.Clarifications = disk.Clarifications
 		st.Mailbox = disk.Mailbox
+		st.Handoffs = disk.Handoffs
+		st.Board = disk.Board
 	} else if !os.IsNotExist(err) {
 		return nil, err
 	}
@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"sync"
 	"time"
 )

@@ -109,6 +110,17 @@ type StepSnapshot struct {
 	ID        string `json:"id"`
 	Status    string `json:"status"`
 	AttemptID string `json:"attempt_id,omitempty"`
+	// F05/U3b：trusted terminal record；用于 recovery 时仅在「后端显式
+	// 声明 success」时把步骤恢复成 completed（即便目标文件缺失也照恢复），
+	// files 仍然记入 deliverables 做诊断。SideEffects 报告重放安全性：
+	//   none     → 可重试
+	//   possible → 需 reconciliation（不禁 rerun；只允许显式 reconcile）
+	// Canceled/Aborted 为 true 时本步是取消/中止语义，不应被恢复成 completed。
+	Terminal string `json:"terminal,omitempty"`
+	// SideEffects 镜像 ExecResult.SideEffects；缺省 = possible（旧读取）。
+	SideEffects string `json:"side_effects,omitempty"`
+	Canceled    bool   `json:"canceled,omitempty"`
+	Aborted     bool   `json:"aborted,omitempty"`
 }

 // CheckpointMode 快照写盘档位（Q1 WP4 三档）。
@@ -129,6 +141,10 @@ type checkpointWriter struct {
 	store Store
 	ch    chan Checkpoint
 	done  chan error
+	// flushOnce/flushErr 让出口 flush 幂等：正常收口路径先 flush 一次（失败必须
+	// 先于任何完成事件可见），defer 再兜底 flush 时不会二次写、也不会重复报错。
+	flushOnce sync.Once
+	flushErr  error
 }

 func newCheckpointWriter(mode CheckpointMode, store Store) *checkpointWriter {
@@ -178,7 +194,13 @@ func (w *checkpointWriter) boundary(wsDir string, cp Checkpoint) error {
 }

 // flush 在推进退出前调用：async 落完并报告首个错误；on_exit 写最终快照。
+// 幂等：重复调用返回首次结果（async 的 chan 只关一次）。
 func (w *checkpointWriter) flush(wsDir string, cp Checkpoint) error {
+	w.flushOnce.Do(func() { w.flushErr = w.flushOnceImpl(wsDir, cp) })
+	return w.flushErr
+}
+
+func (w *checkpointWriter) flushOnceImpl(wsDir string, cp Checkpoint) error {
 	cp.wsDir = wsDir
 	switch w.mode {
 	case CheckpointAsync:
@@ -191,6 +213,15 @@ func (w *checkpointWriter) flush(wsDir string, cp Checkpoint) error {
 	}
 }

+// persistenceFailurePrefix 标记落盘失败。这类失败必须让运行以可区分的面目收场：
+// 业务动作（步骤执行、交付物写出）已发生，不回滚、不声称完成。
+const persistenceFailurePrefix = "persistence failure"
+
+// persistenceFailureText 组装落盘失败的可读说明（stage = 落盘环节）。
+func persistenceFailureText(stage string, err error) string {
+	return persistenceFailurePrefix + ": " + stage + ": " + err.Error()
+}
+
 // validateGraphContinuity 恢复连续性校验（Q1 WP4；WP1 的图版本裁定）。
 // 最新检查点存在且其图指纹与当前 Steps 重编译不一致时拒绝恢复——
 // 崩溃期间图形态被改过，继续推进会把已完成的结果接到错误的图上。
@@ -110,6 +110,9 @@ func AssessStepFailure(sig StepFailureSignals) FailureAssessment {
 	lower := strings.ToLower(reason)
 	trace := ScanTraceText(sig.ExtraText + "\n" + ReadWorkDirTraces(sig.WorkDir))
 	missing := sig.TargetMissing || strings.Contains(lower, "missing target deliverable")
+	// F05/KTD4：explicit SideEffects=none 报告 = provably-not-dispatched。
+	// 缺则按 possible 对待（与旧语义一致）。这是 caller 自己传达的可重试信令。
+	provablyNotDispatched := sig.Result != nil && sig.Result.SideEffects == ExecSideEffectsNone
 	if sig.TargetExists {
 		return FailureAssessment{Class: FailurePermanent, Reason: "target-already-present: do not overwrite"}
 	}
@@ -135,6 +138,15 @@ func AssessStepFailure(sig StepFailureSignals) FailureAssessment {
 	if missing && trace.Settled && !provider {
 		return FailureAssessment{Class: FailureTransient, Reason: "settled-no-deliverable: " + compactReason(reason)}
 	}
+	// F05/U3b：SideEffects=none 失败原则上可重试（后端已声明未发生外部动作）；
+	// 跨越该含义补一条有界重试通路，并覆盖纯获取后失败但文件已拼起的边界。
+	if provablyNotDispatched && !provider {
+		kind := "provably-not-dispatched"
+		if missing {
+			kind = "provably-not-dispatched-missing-target"
+		}
+		return FailureAssessment{Class: FailureTransient, Reason: kind + ": " + compactReason(reason)}
+	}
 	if missing {
 		return FailureAssessment{Class: FailurePermanent, Reason: "permanent-missing: " + compactReason(reason)}
 	}
@@ -0,0 +1,308 @@
+package squad
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// taskboard.go — task_board 拓扑驱动（Q2 WP4）。
+// 形态：计划步骤 = 板上任务（Needs = 依赖门）；执行语义是队列化的
+// claim/lease/complete 三原语——工作者认领任务（带租约时限）、执行、完成；
+// 进程死亡留下的孤儿租约（claimed 且租约过期）在驱动重入时再武装回 open
+// （Attempts 保留，超限由步骤级重试护栏兜底）。全部任务收口后走标准
+// finalize（交付物/质检/降级判定与图拓扑同一收口面）。
+// 挂起（HITL）安全：claimed 且步骤未完的任务在重入时续跑，不重复认领。
+
+func init() {
+	_ = RegisterTopologyDriver(TopologyTaskBoard)
+}
+
+// DefaultTaskLeaseSeconds 任务租约缺省时限（秒）。
+const DefaultTaskLeaseSeconds = 600
+
+// boardSkipReason 依赖门无法满足（上游 failed/skipped）时的跳过原因。
+const boardSkipReason = "task_board: 依赖未满足（上游失败或跳过）"
+
+// BoardTask 是任务板上一条任务的队列状态。
+type BoardTask struct {
+	ID        string `json:"id"`
+	Status    string `json:"status"` // open / claimed / done / failed / skipped
+	ClaimedBy string `json:"claimed_by,omitempty"`
+	// LeaseExpires 认领租约到期时刻；过期未 complete = 孤儿租约，重入再武装。
+	LeaseExpires time.Time `json:"lease_expires,omitempty"`
+	// Attempts 认领次数（孤儿再武装累加；执行重试由步骤级护栏负责）。
+	Attempts int `json:"attempts,omitempty"`
+}
+
+// runTaskBoardDriver task_board 拓扑入口（幂等可恢复）。
+func (e *SquadEngine) runTaskBoardDriver(ctx context.Context, st *RunState) (*RunResult, error) {
+	drv := &flowDriver{opts: e.deps}
+	e.initBoard(st)
+	e.rearmOrphanLeases(st)
+	leaseTTL := DefaultTaskLeaseSeconds
+	if v := st.Meta.Metadata["task_board_lease_seconds"]; v != "" {
+		if n := parseIntDefault(v, DefaultTaskLeaseSeconds); n > 0 {
+			leaseTTL = n
+		}
+	}
+
+	for {
+		idx, held := e.boardHeldTask(st)
+		if !held {
+			var ok bool
+			idx, ok = e.boardClaimNext(st, leaseTTL)
+			if !ok {
+				// 无可认领：把依赖已死（上游 failed/skipped）的 open 任务级联
+				// 置 skipped；再无变化则收口。
+				if !e.boardSkipDeadlocked(st) {
+					break
+				}
+				continue
+			}
+		}
+		out, gate, err := e.boardRunTask(ctx, drv, st, idx)
+		if err != nil {
+			return nil, err
+		}
+		if gate != nil {
+			if out == nil {
+				out = heldResult(st, gate)
+			}
+			return out, nil
+		}
+		if out != nil { // 关键任务失败：fail-closed 传播（failStep 已置运行失败）
+			return out, nil
+		}
+		e.boardSettleTask(st, idx)
+	}
+	return e.boardFinish(ctx, drv, st)
+}
+
+// initBoard 首次进入时从计划步骤派生任务板（恢复重入直读持久化状态）。
+func (e *SquadEngine) initBoard(st *RunState) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	if len(st.Board) > 0 {
+		return
+	}
+	for _, s := range st.Plan.Workflow.Steps {
+		status := "open"
+		switch s.Status {
+		case "completed":
+			status = "done"
+		case "failed":
+			status = "failed"
+		case "skipped":
+			status = "skipped"
+		}
+		st.Board = append(st.Board, BoardTask{ID: s.ID, Status: status})
+	}
+	_ = st.save()
+}
+
+// rearmOrphanLeases 孤儿租约再武装：claimed 且租约过期 → open（进程死亡
+// 恢复路径；Attempts 保留作观测）。
+func (e *SquadEngine) rearmOrphanLeases(st *RunState) {
+	now := time.Now()
+	st.mu.Lock()
+	var rearmed []string
+	for i := range st.Board {
+		t := &st.Board[i]
+		if t.Status == "claimed" && !t.LeaseExpires.IsZero() && t.LeaseExpires.Before(now) {
+			t.Status = "open"
+			t.ClaimedBy = ""
+			rearmed = append(rearmed, t.ID)
+		}
+	}
+	changed := len(rearmed) > 0
+	if changed {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	if changed {
+		(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase,
+			Text: fmt.Sprintf("task_board：孤儿租约再武装 %s", strings.Join(rearmed, ","))})
+	}
+}
+
+// boardHeldTask 返回 claimed 且步骤未完的任务（HITL 挂起/中断重入的续跑点）。
+func (e *SquadEngine) boardHeldTask(st *RunState) (int, bool) {
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "claimed" {
+			continue
+		}
+		if idx, s := e.boardStep(st, t.ID); idx >= 0 && (s.Status == "pending" || s.Status == "running") {
+			return idx, true
+		}
+	}
+	return -1, false
+}
+
+// boardClaimNext 认领下一个可执行任务（open 且依赖门全 done），返回步骤下标。
+func (e *SquadEngine) boardClaimNext(st *RunState, leaseTTL int) (int, bool) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "open" {
+			continue
+		}
+		if !e.boardDepsDoneLocked(st, t.ID) {
+			continue
+		}
+		idx, s := e.boardStepLocked(st, t.ID)
+		if idx < 0 {
+			continue
+		}
+		t.Status = "claimed"
+		t.ClaimedBy = s.Role
+		t.Attempts++
+		t.LeaseExpires = time.Now().Add(time.Duration(leaseTTL) * time.Second)
+		_ = st.save()
+		(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, StepID: t.ID, Role: s.Role,
+			Text: fmt.Sprintf("task_board：%s 认领 %s（租约 %ds，第 %d 次）", s.Role, t.ID, leaseTTL, t.Attempts)})
+		return idx, true
+	}
+	return -1, false
+}
+
+// boardSkipDeadlocked 把依赖已死（上游 failed/skipped）的 open 任务级联置
+// skipped；返回是否有变化（有则继续循环级联）。
+func (e *SquadEngine) boardSkipDeadlocked(st *RunState) bool {
+	st.mu.Lock()
+	changed := false
+	for ti := range st.Board {
+		t := &st.Board[ti]
+		if t.Status != "open" {
+			continue
+		}
+		if e.boardDepsDeadLocked(st, t.ID) {
+			t.Status = "skipped"
+			if idx, s := e.boardStepLocked(st, t.ID); idx >= 0 {
+				s.Status = "skipped"
+				s.Error = boardSkipReason
+				st.Plan.Workflow.Steps[idx] = *s
+			}
+			changed = true
+		}
+	}
+	if changed {
+		_ = st.save()
+	}
+	st.mu.Unlock()
+	return changed
+}
+
+// boardRunTask 执行已认领任务：过门 → runOneStep。
+func (e *SquadEngine) boardRunTask(ctx context.Context, drv *flowDriver, st *RunState, idx int) (*RunResult, *HitlGate, error) {
+	out, gate, err := drv.gateBeforeRun(ctx, st, idx)
+	if err != nil || gate != nil || out != nil {
+		return out, gate, err
+	}
+	return drv.runOneStep(ctx, st, idx)
+}
+
+// boardSettleTask 任务结算：步骤 completed → done；failed（非关键降级）→ failed。
+func (e *SquadEngine) boardSettleTask(st *RunState, idx int) {
+	step := st.Plan.Workflow.Steps[idx]
+	status := "done"
+	if step.Status == "failed" {
+		status = "failed"
+	}
+	st.mu.Lock()
+	for i := range st.Board {
+		if st.Board[i].ID == step.ID {
+			st.Board[i].Status = status
+			st.Board[i].LeaseExpires = time.Time{}
+		}
+	}
+	_ = st.save()
+	st.mu.Unlock()
+	(&flowDriver{opts: e.deps}).emit(st, Event{Type: EvPhase, StepID: step.ID, Role: step.Role,
+		Text: fmt.Sprintf("task_board：%s → %s", step.ID, status)})
+}
+
+// boardFinish 全部任务收口后走标准 finalize（交付物/质检/降级与图拓扑同面）。
+// 仍有 open/claimed 任务 = 依赖不可满足，fail-closed 不静默收口。
+func (e *SquadEngine) boardFinish(ctx context.Context, drv *flowDriver, st *RunState) (*RunResult, error) {
+	done, total := 0, len(st.Board)
+	for _, t := range st.Board {
+		switch t.Status {
+		case "done":
+			done++
+		case "open", "claimed":
+			return e.driverFail(st, fmt.Sprintf("task_board: 任务 %s 无法收口（依赖不可满足）", t.ID)), nil
+		}
+	}
+	drv.emit(st, Event{Type: EvPhase, Text: fmt.Sprintf("task_board：收口（done %d/%d）", done, total)})
+	out, gate, err := drv.finalize(ctx, st)
+	if err != nil {
+		return nil, err
+	}
+	if gate != nil {
+		return heldResult(st, gate), nil
+	}
+	return out, nil
+}
+
+// boardStep 按任务 ID 找步骤（只读副本）。
+func (e *SquadEngine) boardStep(st *RunState, id string) (int, *StepSpec) {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	return e.boardStepLocked(st, id)
+}
+
+// boardStepLocked 按任务 ID 找步骤（调用方持锁）。返回下标与副本指针。
+func (e *SquadEngine) boardStepLocked(st *RunState, id string) (int, *StepSpec) {
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == id {
+			s := st.Plan.Workflow.Steps[i]
+			return i, &s
+		}
+	}
+	return -1, nil
+}
+
+// boardDepsDoneLocked 依赖门：Needs 全部 done（调用方持锁）。
+func (e *SquadEngine) boardDepsDoneLocked(st *RunState, id string) bool {
+	needs := e.boardNeedsLocked(st, id)
+	for _, dep := range needs {
+		found := false
+		for i := range st.Board {
+			if st.Board[i].ID == dep {
+				found = true
+				if st.Board[i].Status != "done" {
+					return false
+				}
+			}
+		}
+		if !found {
+			return false // 依赖不在板上 = fail-closed 不认领
+		}
+	}
+	return true
+}
+
+// boardDepsDeadLocked 依赖已死：任一 Need 为 failed/skipped（级联跳过依据）。
+func (e *SquadEngine) boardDepsDeadLocked(st *RunState, id string) bool {
+	for _, dep := range e.boardNeedsLocked(st, id) {
+		for i := range st.Board {
+			if st.Board[i].ID == dep && (st.Board[i].Status == "failed" || st.Board[i].Status == "skipped") {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func (e *SquadEngine) boardNeedsLocked(st *RunState, id string) []string {
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == id {
+			return st.Plan.Workflow.Steps[i].Needs
+		}
+	}
+	return nil
+}
@@ -384,9 +384,8 @@ func TestTopologyValidationFailsClosed(t *testing.T) {
 		{"未注册名", "bogus_topology", "未知拓扑"},
 		{"primitive 单跑", TopologyMailbox, "跨拓扑原语"},
 		{"primitive 单跑 handoff", TopologyHandoff, "跨拓扑原语"},
-		{"driver 未点亮", TopologyReview, "驱动未注册"},
-		{"driver 未点亮 group_chat", TopologyGroupChat, "驱动未注册"},
-		{"driver 未点亮 task_board", TopologyTaskBoard, "驱动未注册"},
+		// review/group_chat（WP3）与 task_board（WP4）驱动均已点亮；
+		// 12/12 受理面由 TestTopologyConformance_allTwelve 汇总钉住。
 	}
 	for _, tc := range cases {
 		engine := newTopologyEngine(t, mockScriptPlanner(
@@ -3,6 +3,7 @@ package squad
 import (
 	"bytes"
 	"context"
+	"encoding/json"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -41,16 +42,41 @@ func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult,
 	cpw := newCheckpointWriter(CheckpointMode(st.Meta.Metadata["checkpoint_mode"]), st.storeOr())
 	// 推进退出前 flush：async 落完、on_exit 写最终快照（Q1 WP4 三档的出口语义）。
 	var lastGraph *graph
-	defer func() {
+	// exitSnapshot 组装出口快照（尚未编译图时现编译）。
+	exitSnapshot := func() (Checkpoint, bool) {
 		if lastGraph == nil {
-			if g, gerr := compileGraph(st.Plan.Workflow.Steps); gerr == nil {
-				lastGraph = g
+			g, gerr := compileGraph(st.Plan.Workflow.Steps)
+			if gerr != nil {
+				return Checkpoint{}, false
+			}
+			lastGraph = g
+		}
+		return snapshotOf(superNo, lastGraph, st.Plan.Workflow.Steps), true
+	}
+	// 兜底出口：任何非正常退出路径（挂起、图编译失败、infra 错误）也在此 flush。
+	// flush 幂等——收口路径已提前 flush 过，不会二次写。
+	defer func() {
+		if cp, ok := exitSnapshot(); ok {
+			if ferr := cpw.flush(st.Meta.WorkDir, cp); ferr != nil {
+				// F11c：出口快照落盘失败不得静默——翻成可区分的失败。
+				res, gate, err = d.persistenceFailure(st, res, gate, err, "exit checkpoint flush", ferr)
 			}
 		}
-		if lastGraph != nil {
-			_ = cpw.flush(st.Meta.WorkDir, snapshotOf(superNo, lastGraph, st.Plan.Workflow.Steps))
-		}
 	}()
+	// flushExitBeforeFinalize 在收口前落出口快照（async/on_exit 的最终落盘），
+	// 失败立即交出可区分的失败：落盘失败必须先于任何完成事件/RunCompleted 返回。
+	// 返回的 ok=false 表示快照已落盘（或本次无需落盘），调用方继续正常收口。
+	flushExitBeforeFinalize := func() (*RunResult, *HitlGate, error, bool) {
+		cp, ok := exitSnapshot()
+		if !ok {
+			return nil, nil, nil, false
+		}
+		if ferr := cpw.flush(st.Meta.WorkDir, cp); ferr != nil {
+			out, g2, e2 := d.persistenceFailure(st, nil, nil, nil, "exit checkpoint flush", ferr)
+			return out, g2, e2, true
+		}
+		return nil, nil, nil, false
+	}
 	for {
 		if !runDeadline.IsZero() && time.Now().After(runDeadline) {
 			d.emit(st, Event{Type: EvPhase, Text: "运行截止时刻已到，未完步骤按 skipped 收口"})
@@ -63,6 +89,9 @@ func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult,
 			}
 			_ = st.save()
 			st.mu.Unlock()
+			if out, g2, e2, failed := flushExitBeforeFinalize(); failed {
+				return out, g2, e2
+			}
 			return d.finalize(ctx, st)
 		}
 		g, gerr := compileGraph(st.Plan.Workflow.Steps)
@@ -151,21 +180,39 @@ func (d *flowDriver) advance(ctx context.Context, st *RunState) (res *RunResult,
 			}
 		}

-		// superstep 边界快照（sync 立即 / async 入队；on_exit 由 defer 的 flush 写）。
+		// superstep 边界快照（sync 立即 / async 入队；on_exit 由出口 flush 写）。
 		lastGraph = g
 		if err := cpw.boundary(st.Meta.WorkDir, snapshotOf(superNo, g, st.Plan.Workflow.Steps)); err != nil {
-			return nil, nil, err
+			// F11c：边界快照写盘失败已经是可见错误，补上落盘失败标记以便识别。
+			return nil, nil, fmt.Errorf("%s: %w", persistenceFailurePrefix, err)
 		}
 	}

+	// 全部步骤已收口：先落出口快照（async/on_exit 的最终落盘），失败必须先于收口
+	// 可见——不得先宣布完成、再报落盘失败。
+	if out, g2, e2, failed := flushExitBeforeFinalize(); failed {
+		return out, g2, e2
+	}
+
 	return d.finalize(ctx, st)
 }

 // snapshotOf 组装一个 superstep 边界快照：序号、图结构指纹与全量步骤状态。
+// terminal side-effect evidence 由 recordStepTerminal 在步骤进入 completed 时写入
+// step_terminals.json；本处把已记录的 terminal 数据并入 StepSnapshot，保证
+// checkpoint 也带同样证据（F05：信任落盘记录）。
 func snapshotOf(superNo int, g *graph, steps []StepSpec) Checkpoint {
 	cp := Checkpoint{Superstep: superNo, At: time.Now(), GraphHash: g.hash()}
+	// checkpoint 路径拿不到当前 RunState 引用（cpw 在 defer flush 时调用过），
+	// 同时也仅有顺序式调用传递上下文不便；所以这里走「按 ID 检查同索引」对齐
+	// 替代——checkpoint 产生时 step.Status 已为终态（completed），因此 record 同步
+	// 已写入；StepSnapshot.Terminal 就用 step.Status 推导（F05 唯一信任源）。
 	for _, s := range steps {
-		cp.StepStates = append(cp.StepStates, StepSnapshot{ID: s.ID, Status: s.Status, AttemptID: s.AttemptID})
+		snap := StepSnapshot{ID: s.ID, Status: s.Status, AttemptID: s.AttemptID}
+		if s.Status == "completed" {
+			snap.Terminal = "completed"
+		}
+		cp.StepStates = append(cp.StepStates, snap)
 	}
 	return cp
 }
@@ -279,6 +326,16 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 			return d.failStep(ctx, st, idx)
 		}
 	}
+	// handoff 原语 fail-closed（Q2）：目标角色必须可解析，启动前拒绝。
+	if step.Handoff != "" {
+		if _, ok := findAgent(st.Plan.Squad.Agents, step.Handoff); !ok {
+			if _, exists := d.opts.Registry.Get(step.Handoff); !exists {
+				step.Status, step.Error = "failed", "handoff: 未知目标角色 "+step.Handoff
+				st.setStep(idx, step)
+				return d.failStep(ctx, st, idx)
+			}
+		}
+	}
 	subDir := filepath.Join(st.Meta.WorkDir, filepath.FromSlash(orDefault(step.SubWorkspace, fmt.Sprintf("tasks/%02d_%s", idx+1, step.Role))))
 	if err := os.MkdirAll(subDir, 0o755); err != nil {
 		return nil, nil, err
@@ -313,11 +370,14 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 	// 任务书；只走 prompt 面，永不触碰 HITL 门（验收三构造性保证）。
 	st.mu.Lock()
 	mail := takeMailFor(st, step.Role)
-	if len(mail) > 0 {
+	handoff := takeHandoffFor(st, step.Role)
+	if len(mail) > 0 || handoff != nil {
 		_ = st.save()
 	}
 	st.mu.Unlock()
 	prompt += formatMailSection(mail)
+	// handoff 原语（Q2）：目标角色的步骤接收过滤历史（只带移交方主产出）。
+	prompt += formatHandoffSection(handoff)
 	instructions := readFileOr(filepath.Join(subDir, "AGENTS.md"), d.opts.Workspaces.ResolveInstructions(role))

 	var backend Executor
@@ -346,6 +406,9 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 	}
 	var result *ExecResult
 	var err error
+	// F05/U3b（KTD4）：ExecutionID 在一次逻辑执行内稳定。重试层共享同一 ID
+	// + 同一 Deadline，便于下游做幂等键 + 收据对齐。第一层 attempt 用稳定 ID；
+	// 重试层沿用 step.AttemptID（首次进入时已生成、保持不变）。
 	for attempt := 1; attempt <= MaxTransientStepInvocations; attempt++ {
 		reqPrompt := prompt
 		if attempt > 1 {
@@ -366,14 +429,15 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 			ToolProfile:  orDefault(role.ToolProfile, ToolProfileWorkspace),
 			OnEvent: func(ev ProcessEvent) {
 				d.emit(st, Event{Type: EvDetail, StepID: step.ID, Role: step.Role, Text: ev.Text,
-					Data: marshalJSON(map[string]string{"process_type": ev.Type, "tool": ev.Tool, "detail": ev.Detail})})
+					Data: processEventData(ev)})
 			},
 			// RoleSpec v2 贯通：工具白名单 / 技能 / 标识（原为死字段或缺失）。
 			ToolAllowlist: role.Tools,
 			Skills:        firstNonEmptySlice(step.Skills, role.Skills),
-			SessionScope:  st.Meta.ID + "/" + step.ID,
-			ExecutionID:   step.AttemptID,
-			Budget:        stepBudget,
+			// F05：ExecutionID 复用 step.AttemptID（首层 attempt 时已生成）——重试层不再换新 ID。
+			SessionScope: st.Meta.ID + "/" + step.ID,
+			ExecutionID:  step.AttemptID,
+			Budget:       stepBudget,
 		}
 		if step.IntentID != "" {
 			// 系统功能步：不经智能体，由 IntentRunner 直接执行；缺席即 fail-closed。
@@ -393,18 +457,59 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 		} else {
 			result, err = backend.Execute(ctx, req)
 		}
-		if err == nil && result.Success {
+		// F05/KTD4：先验证执行结果本身，不让「文件已存在」翻成 Success。
+		// (1) (nil,nil) 视作稳定失败，不允许 panic 也不允许误判成功；
+		// (2) Success=true 仅在 result 显式声明时才成立；
+		// (3) 已取消/中止的结果不进入重试通路（不是业务失败，且重试会重复已发生的取消语义）。
+		if err != nil && result == nil {
+			// (nil,nil) 在 Go 是不可能态（要么 err、要么 result）；保险起见仍按稳定失败。
+			step.Status = "failed"
+			step.Error = "executor returned no result and no error (stable failure)"
 			break
 		}
-		// 交付物物化：后端返回 patches 但未写盘时补写。
+		if result == nil {
+			step.Status = "failed"
+			step.Error = "executor returned nil result"
+			break
+		}
+		if result.IsCanceled() {
+			// F05/KTD4：取消不是业务失败，不重试（按 attempt==MaxTransientStepInvocations 跳出）。
+			step.Status = "failed"
+			if result.ErrorReason != "" {
+				step.Error = "executor canceled: " + result.ErrorReason
+			} else {
+				step.Error = "executor canceled"
+			}
+			break
+		}
+		if err == nil && result.Success {
+			// 交付物物化：后端返回 patches 但未写盘时补写。F05：仅当 result.Success
+			// 已验证且文件未准备时才需要物化；只有 patches 内容才能落盘。
+			if len(result.Patches) > 0 {
+				_, _ = MaterializeTargets(subDir, targets, result.Patches)
+			}
+			break
+		}
+		// 交付物物化：后端返回 patches 但未写盘时补写。F05：仅当 result.Success 已
+		// 被后端显式声明时物化才允许 break；旧逻辑会让
+		// `err==nil && len(Patches)>0 && materialize_ok` 翻成 Success=true —— 修掉。
 		if err == nil && len(result.Patches) > 0 {
 			if _, merr := MaterializeTargets(subDir, targets, result.Patches); merr == nil {
-				if allTargetsPresent(subDir, targets) {
-					result.Success = true
+				if result.Success && allTargetsPresent(subDir, targets) {
 					break
 				}
 			}
 		}
+		// F05/KTD4：副作用可能存在的失败禁止自动重试（重试会重复外部动作）。
+		// 仅 SideEffects=none 或纯读取后的失败保持有界重试。
+		if !result.IsRetryableSideEffectReports() && err != nil {
+			step.Status = "failed"
+			step.Error = orDefault(result.ErrorReason, err.Error())
+			if result.ErrorReason != "" {
+				step.Error = result.ErrorReason
+			}
+			break
+		}
 		assessment := AssessStepFailure(StepFailureSignals{
 			Err: err, Result: result, WorkDir: subDir,
 			ExtraText:     fmt.Sprintf("step=%s backend=%s", step.ID, backendName),
@@ -450,6 +555,9 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru

 	step.Status = "completed"
 	appendLog(subDir, fmt.Sprintf("[%s] backend=%s duration=%dms targets=%v\n", time.Now().Format(time.RFC3339), backendName, step.DurationMs, targets))
+	// F05/U3b：记录 trusted terminal state（用于 recoverInterrupted 判定）。
+	// 由「正在收口的 completed」流写入；recovery 期间不会回填（防伪）。
+	d.recordStepTerminal(st, step, result)

 	// 状态回填 + 黑板合并 + 落盘同一临界区：并行步骤的共享状态变更串行化，
 	// 防止 marshal 与写回竞争、临时文件互相踩踏把保存失败误判为步骤失败。
@@ -464,6 +572,12 @@ func (d *flowDriver) runOneStep(ctx context.Context, st *RunState, idx int) (*Ru
 	d.emit(st, Event{Type: EvStepCompleted, StepID: step.ID, Role: step.Role, Text: step.Title,
 		Data: marshalJSON(map[string]any{"duration_ms": step.DurationMs, "targets": targets})})

+	// handoff 原语（Q2）：完成即移交最终答案所有权——记录 + 主产出快照，
+	// 供目标角色下一步骤（过滤历史）与 finalize（链尾产出）消费。
+	if step.Handoff != "" {
+		d.recordHandoff(st, step, subDir, targets)
+	}
+
 	// 循环步再武装（Q2 loop）：停机条件不满足且预算未尽 → 本步与 loop.body
 	// 一起重置为 pending；触顶发 loop_exhausted 事件后按完成继续。
 	if step.Loop != nil {
@@ -775,12 +889,21 @@ func (d *flowDriver) finalize(ctx context.Context, st *RunState) (*RunResult, *H
 		}
 	}
 	// 汇总交付物：最后一个 completed 步骤的主产出 → output.md。
+	// handoff 链尾优先（Q2）：有所有权移交时，交付物取链尾角色的最后一个
+	// completed 步骤产出（最终答案所有权在链尾），无则该角色无完成步骤时回落。
 	var srcStep *StepSpec
 	for i := range wf.Steps {
 		if wf.Steps[i].Status == "completed" {
 			srcStep = &wf.Steps[i]
 		}
 	}
+	if tail := handoffChainTail(st); tail != "" {
+		for i := range wf.Steps {
+			if wf.Steps[i].Status == "completed" && wf.Steps[i].Role == tail {
+				srcStep = &wf.Steps[i]
+			}
+		}
+	}
 	if srcStep == nil {
 		wf.Status = "failed"
 		st.setWorkflow(wf)
@@ -867,10 +990,22 @@ func (d *flowDriver) finalize(ctx context.Context, st *RunState) (*RunResult, *H
 	st.setWorkflow(wf)
 	st.Meta.Status = string(status)
 	st.Meta.CompletedAt = &now
-	_ = st.save()
+	// F11c：终态落盘先于完成事件。写不进去就不能宣布完成——
+	// 业务动作已发生，不回滚，但运行以可区分的失败面世。
+	if serr := st.save(); serr != nil {
+		msg := persistenceFailureText("save terminal run state", serr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			Quality: report, WorkDir: st.Meta.WorkDir, Usage: st.usageSnapshot(),
+			Error: msg, Summary: msg}, nil, nil
+	}

 	manifest := buildManifest(st, deliverable, *report)
-	_ = writeJSON(filepath.Join(st.Meta.WorkDir, "manifest.json"), manifest)
+	if merr := writeJSON(filepath.Join(st.Meta.WorkDir, "manifest.json"), manifest); merr != nil {
+		msg := persistenceFailureText("write manifest", merr)
+		return &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			Quality: report, WorkDir: st.Meta.WorkDir, Usage: st.usageSnapshot(),
+			Error: msg, Summary: msg}, nil, nil
+	}

 	d.emit(st, Event{Type: EvRunCompleted, Text: string(status),
 		Data: marshalJSON(map[string]any{"deliverable": deliverable, "score": report.Score})})
@@ -990,6 +1125,41 @@ func (d *flowDriver) emitHuman(st *RunState, req *HumanRequest) {
 	d.emit(st, Event{Type: EvHumanInputRequired, Text: req.Question, Data: marshalJSON(req)})
 }

+// persistenceFailure 把一次落盘失败折算成可见的失败（F11c）：
+//   - 已有 infra 错误 → 附加到错误里返回，不吞掉原错误；
+//   - 挂起中 → 挂起语义不变（运行未收口），但结果与事件流如实标注；
+//   - 其余（收口完成/无结果）→ 结果翻成 RunFailed，绝不再以 completed 面世。
+//
+// 业务动作（步骤执行、交付物写出）已发生，不回滚，也不声称干净完成。
+func (d *flowDriver) persistenceFailure(st *RunState, res *RunResult, gate *HitlGate, prev error, stage string, cause error) (*RunResult, *HitlGate, error) {
+	msg := persistenceFailureText(stage, cause)
+	if prev != nil {
+		return res, gate, fmt.Errorf("%w（同时发生 %s）", prev, msg)
+	}
+	if res == nil {
+		res = &RunResult{RunID: st.Meta.ID, Status: RunFailed, Plan: planPtr(st),
+			WorkDir: st.Meta.WorkDir, Summary: msg}
+	}
+	if !strings.Contains(res.Error, msg) {
+		// 同一次失败会经出口 flush 与兜底 flush 各报一次，只记一遍。
+		if res.Error != "" {
+			msg = res.Error + "；" + msg
+		}
+		res.Error = msg
+	}
+	if gate != nil {
+		// 挂起路径不发终态事件（finalizeEvent 只处理 completed/degraded/failed），
+		// 所以在这里补一条可观察的标注事件。
+		d.emit(st, Event{Type: EvPhase, Text: msg})
+		return res, gate, nil
+	}
+	switch res.Status {
+	case RunCompleted, RunDegraded:
+		res.Status = RunFailed
+	}
+	return res, nil, nil
+}
+
 func summarize(st *RunState, report QualityReport, status RunStatus) string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "运行 %s：%s。步骤 %d 个（完成 %d/跳过 %d/失败 %d）。质检得分 %.0f。",
@@ -1011,6 +1181,84 @@ func countStatus(st *RunState, status string) int {
 	return n
 }

+// recordStepTerminal 记一条「trusted terminal record」：在某步执行完成
+// （仅 success 路径）时落地 step.Terminal + SideEffects + canceled/aborted。
+// reconcile/path：如果 recovery 期间没有该记录，则按文件存在与否分流
+// （见 engine.recoverInterrupted）。
+//
+// F05：record 与 read 路径严格同包；测试通过 squad/executor_result_contract_test.go
+// 覆盖正向与逆向条件。
+func (d *flowDriver) recordStepTerminal(st *RunState, step StepSpec, result *ExecResult) {
+	if st == nil || st.Meta.WorkDir == "" {
+		return
+	}
+	if step.ID == "" {
+		return
+	}
+	t := stepTerminalRecord{Terminal: "completed", At: time.Now(), StepID: step.ID, AttemptID: step.AttemptID}
+	if result != nil {
+		t.SideEffects = string(result.SideEffects)
+		t.Canceled = result.Canceled
+		t.Aborted = result.Aborted
+	}
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	terminals := readStepTerminalsLocked(st)
+	terminals[step.ID] = t
+	data, err := json.Marshal(terminals)
+	if err != nil {
+		return
+	}
+	_ = os.WriteFile(filepath.Join(st.Meta.WorkDir, "step_terminals.json"), data, 0o644)
+}
+
+// readStepTerminalsLocked 读取已落盘的 step_terminals.json（调用方须持锁或私有化）。
+// 失败（含文件不存在）返回空 map。
+func readStepTerminalsLocked(st *RunState) map[string]stepTerminalRecord {
+	out := map[string]stepTerminalRecord{}
+	if st == nil || st.Meta.WorkDir == "" {
+		return out
+	}
+	data, err := os.ReadFile(filepath.Join(st.Meta.WorkDir, "step_terminals.json"))
+	if err != nil || len(data) == 0 {
+		return out
+	}
+	var stored map[string]stepTerminalRecord
+	if err := json.Unmarshal(data, &stored); err != nil {
+		return out
+	}
+	for k, v := range stored {
+		out[k] = v
+	}
+	return out
+}
+
+// loadStepTerminalsForWorkDir 公开暴露的 record 读取（recovery 使用）。
+func loadStepTerminalsForWorkDir(wsDir string) map[string]stepTerminalRecord {
+	if wsDir == "" {
+		return map[string]stepTerminalRecord{}
+	}
+	data, err := os.ReadFile(filepath.Join(wsDir, "step_terminals.json"))
+	if err != nil || len(data) == 0 {
+		return map[string]stepTerminalRecord{}
+	}
+	var stored map[string]stepTerminalRecord
+	if err := json.Unmarshal(data, &stored); err != nil {
+		return map[string]stepTerminalRecord{}
+	}
+	return stored
+}
+
+type stepTerminalRecord struct {
+	StepID      string    `json:"step_id"`
+	AttemptID   string    `json:"attempt_id"`
+	Terminal    string    `json:"terminal"` // completed (唯一允许值；failure 走 step.Status)
+	SideEffects string    `json:"side_effects,omitempty"`
+	Canceled    bool      `json:"canceled,omitempty"`
+	Aborted     bool      `json:"aborted,omitempty"`
+	At          time.Time `json:"at"`
+}
+
 func buildManifest(st *RunState, deliverable string, report QualityReport) Manifest {
 	files := map[string]string{}
 	_ = filepath.Walk(st.Meta.WorkDir, func(path string, info os.FileInfo, err error) error {
@@ -189,6 +189,13 @@ func (e *WorkspaceEngine) writeTaskFile(subDir string, step *StepSpec, role Role
 	fmt.Fprintf(&b, "- 角色：%s（%s）\n", role.Name, role.Role)
 	fmt.Fprintf(&b, "- 期望产出：%v\n\n", step.Targets)

+	// 编译期任务指令（Q3）：编译器渲染好的富文本任务书，权威仅次于目标。
+	if strings.TrimSpace(step.Brief) != "" {
+		b.WriteString("## 任务指令\n\n")
+		b.WriteString(strings.TrimSpace(step.Brief))
+		b.WriteString("\n\n")
+	}
+
 	b.WriteString("## 任务上下文资源与权威等级（已快照到本目录，只引用本地路径）\n\n")
 	b.WriteString("| 资源路径 | 权威等级 | 用途说明 | 使用与引用指导 |\n|---|---|---|---|\n")
 	b.WriteString("| `context/goal.md` | **authoritative** | 运行目标 | 不可推翻 |\n")
@@ -0,0 +1,452 @@
+package squad
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// wp3_conformance_test.go — Q2 WP3 验收：review 拓扑与 backend runReview 等价
+// （验收二：并行找问题、逐条验证、丢掉未验证）、handoff 原语（所有权转移 +
+// 过滤历史 + finalize 链尾）、group_chat 驱动（轮转/主持人 + 预算停机）。
+
+// ---- review：与 backend review_test.go 同款断言 ----
+
+// reviewScriptBackend finder 目录回放 findings.json（第 3 个进入者放行，钉住
+// 真并发）；verify_* 目录按 prompt 内容回放判定（与 backend 夹具同文案）。
+type reviewScriptBackend struct {
+	mu          sync.Mutex
+	once        sync.Once
+	entered     int
+	verifyCalls int
+	release     chan struct{}
+}
+
+func (b *reviewScriptBackend) Name() string { return "review-script" }
+
+func (b *reviewScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if strings.HasPrefix(key, "verify_") {
+		b.mu.Lock()
+		b.verifyCalls++
+		b.mu.Unlock()
+		verdict := "看不出对错"
+		switch {
+		case strings.Contains(req.Prompt, "无来源"):
+			verdict = "核对原文确实没有来源。\nVERDICT: PASS"
+		case strings.Contains(req.Prompt, "错字"):
+			verdict = "核对后成立。\nVERDICT: PASS"
+		case strings.Contains(req.Prompt, "过期数据"):
+			verdict = "数据仍有效。\nVERDICT: FAIL"
+		}
+		return writeOne(req.WorkDir, "verdict.md", verdict), nil
+	}
+	// finder：三个必须同时在跑（backend review_test.go:16-33 同款门）。
+	b.mu.Lock()
+	b.entered++
+	n := b.entered
+	b.mu.Unlock()
+	if n == 3 {
+		b.once.Do(func() { close(b.release) })
+	}
+	select {
+	case <-b.release:
+	case <-time.After(time.Second):
+		return nil, fmt.Errorf("finders did not run together")
+	}
+	var findings string
+	switch key {
+	case "01_researcher":
+		findings = `[{"location":"第2段","severity":"high","issue":"无来源","evidence":"显然"},{"location":"第1段","severity":"low","issue":"错字","evidence":"的地"},{"location":"表1","severity":"medium","issue":"过期数据","evidence":"2020"}]`
+	case "02_writer":
+		findings = `[{"location":"第2段","severity":"high","issue":"无来源","evidence":"显然"},{"location":"附录","severity":"medium","issue":"数字不一致","evidence":"3 与 4"}]`
+	default:
+		findings = `[{"location":"结尾","severity":"low","issue":"没有结论","evidence":"最后一句"}]`
+	}
+	return writeOne(req.WorkDir, "findings.json", findings), nil
+}
+
+func writeOne(dir, name, content string) *ExecResult {
+	out := filepath.Join(dir, name)
+	_ = os.MkdirAll(filepath.Dir(out), 0o755)
+	_ = os.WriteFile(out, []byte(content), 0o644)
+	return &ExecResult{Success: true, Patches: map[string][]byte{name: []byte(content)}}
+}
+
+func TestTopologyConformance_review(t *testing.T) {
+	be := &reviewScriptBackend{release: make(chan struct{})}
+	sink := &collectSink{}
+	planner := mockScriptPlanner(
+		StepSpec{ID: "f1", Role: "researcher", Title: "评审一", Targets: []string{"findings.json"}},
+		StepSpec{ID: "f2", Role: "writer", Title: "评审二", Targets: []string{"findings.json"}},
+		StepSpec{ID: "f3", Role: "reviewer", Title: "评审三", Targets: []string{"findings.json"}},
+	)
+	planner.Agents = []RoleSpec{
+		{Role: "researcher", Name: "评审一", Backend: "review-script"},
+		{Role: "writer", Name: "评审二", Backend: "review-script"},
+		{Role: "reviewer", Name: "评审三", Backend: "review-script"},
+		{Role: "verifier", Name: "核实", Backend: "review-script"},
+	}
+	engine := newWP3Engine(t, planner, be, sink, []RoleSpec{
+		{Role: "researcher", Name: "评审一", Backend: "review-script"},
+		{Role: "writer", Name: "评审二", Backend: "review-script"},
+		{Role: "reviewer", Name: "评审三", Backend: "review-script"},
+		{Role: "verifier", Name: "核实", Backend: "review-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "评审材料正文"},
+		RunOptions{Topology: TopologyReview, ReviewTarget: "材料正文"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	out := string(data)
+	// 只留 PASS（fail-closed）：无来源/错字保留；FAIL 与无判定丢弃。
+	if !strings.Contains(out, "无来源") || !strings.Contains(out, "错字") ||
+		strings.Contains(out, "数字不一致") || strings.Contains(out, "没有结论") ||
+		strings.Contains(out, "过期数据") {
+		t.Fatalf("filter = %s", out)
+	}
+	// 报告恒带丢弃计数（backend 同格式串）。
+	if !strings.Contains(out, "已丢弃 3 条未验证") {
+		t.Fatalf("dropped = %s", out)
+	}
+	// 去重后逐条验证：verify 调用数 = unique 数 = 5。
+	if be.verifyCalls != 5 {
+		t.Fatalf("verify calls = %d, want 5 after dedupe", be.verifyCalls)
+	}
+	// severity 稳定排序：high(无来源) 在 low(错字) 前。
+	if strings.Index(out, "无来源") > strings.Index(out, "错字") {
+		t.Fatalf("severity order: %s", out)
+	}
+	if sink.countText("review：3 个 finder 并行找问题") != 1 ||
+		sink.countText("review：去重后 5 条") != 1 {
+		t.Fatal("缺 review 阶段事件")
+	}
+}
+
+// verify 角色缺失 → fail-closed。
+func TestReviewVerifyRoleMissingFailsClosed(t *testing.T) {
+	be := &reviewScriptBackend{release: make(chan struct{})}
+	planner := mockScriptPlanner(
+		StepSpec{ID: "f1", Role: "researcher", Title: "评审", Targets: []string{"findings.json"}},
+	)
+	engine := newWP3Engine(t, planner, be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "评审", Backend: "review-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "评审"},
+		RunOptions{Topology: TopologyReview})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "verify 角色") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- handoff 原语 ----
+
+// handoffScriptBackend 记录 prompt 并按目录回放带标记的产出（≥50 字过最小质检）。
+type handoffScriptBackend struct {
+	mu      sync.Mutex
+	prompts map[string][]string
+}
+
+func (b *handoffScriptBackend) Name() string { return "handoff-script" }
+
+func (b *handoffScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	b.mu.Lock()
+	b.prompts[key] = append(b.prompts[key], req.Prompt)
+	b.mu.Unlock()
+	pad := "\n\n这是为了满足最小质检长度要求而补充的正文内容，超过五十个字符。"
+	var content string
+	switch key {
+	case "01_researcher":
+		content = "MARK_RESEARCHER 调研笔记。" + pad
+	case "02_writer":
+		content = "MARK_WRITER 终稿。" + pad
+	default:
+		content = "MARK_REVIEWER 评审意见。" + pad
+	}
+	name := "output.md"
+	if len(req.Targets) > 0 {
+		name = req.Targets[0]
+	}
+	return writeOne(req.WorkDir, name, content), nil
+}
+
+func TestTopologyConformance_handoff(t *testing.T) {
+	be := &handoffScriptBackend{prompts: map[string][]string{}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "调研", Targets: []string{"notes.md"}, Handoff: "writer"},
+		StepSpec{ID: "s2", Role: "writer", Title: "成文", Targets: []string{"output.md"}, Needs: []string{"s1"}},
+		StepSpec{ID: "s3", Role: "reviewer", Title: "终审", Targets: []string{"output.md"}, Needs: []string{"s2"}},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "handoff-script"},
+		{Role: "writer", Name: "撰写", Backend: "handoff-script"},
+		{Role: "reviewer", Name: "审核", Backend: "handoff-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "带移交的三步"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// ① 所有权转移记录持久化。
+	st, err := engine.loadRun(engine.runDir(res.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(st.Handoffs) != 1 || st.Handoffs[0].FromRole != "researcher" || st.Handoffs[0].ToRole != "writer" {
+		t.Fatalf("Handoffs = %+v", st.Handoffs)
+	}
+	if !strings.Contains(st.Handoffs[0].Output, "MARK_RESEARCHER") {
+		t.Fatalf("移交产出快照缺失：%q", st.Handoffs[0].Output)
+	}
+	if sink.countText("handoff：researcher → writer") != 1 {
+		t.Fatal("缺所有权移交事件")
+	}
+	// ② 目标收过滤历史：writer 的任务书带移交上下文，reviewer 不带。
+	var writerPrompt, reviewerPrompt string
+	for _, p := range be.prompts["02_writer"] {
+		writerPrompt += p
+	}
+	for _, p := range be.prompts["03_reviewer"] {
+		reviewerPrompt += p
+	}
+	if !strings.Contains(writerPrompt, "移交上下文") || !strings.Contains(writerPrompt, "MARK_RESEARCHER") {
+		t.Fatalf("writer 未收到过滤历史：%q", writerPrompt)
+	}
+	if strings.Contains(reviewerPrompt, "移交上下文") {
+		t.Fatal("非目标角色不应收到移交上下文")
+	}
+	// ③ finalize 取链尾输出：交付物 = writer（链尾）产出，而非最后完成的 reviewer。
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(data), "MARK_WRITER") || strings.Contains(string(data), "MARK_REVIEWER") {
+		t.Fatalf("交付物应为链尾 writer 产出：%s", data)
+	}
+}
+
+// handoff 目标角色未知 → 步骤启动前 fail-closed。
+func TestHandoffUnknownRoleFailsClosed(t *testing.T) {
+	be := &handoffScriptBackend{prompts: map[string][]string{}}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "s1", Role: "researcher", Title: "调研", Handoff: "ghost"},
+	), be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "调研", Backend: "handoff-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "handoff: 未知目标角色 ghost") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- group_chat 驱动 ----
+
+// gcScriptBackend 按目录回放 message.md。
+type gcScriptBackend struct {
+	messages map[string]string
+}
+
+func (b *gcScriptBackend) Name() string { return "gc-script" }
+
+func (b *gcScriptBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	msg := b.messages[key]
+	if msg == "" {
+		msg = "发言：" + key
+	}
+	return writeOne(req.WorkDir, "message.md", msg), nil
+}
+
+func TestTopologyConformance_groupChatRotateBudget(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": "第一句",
+		"turn_02_writer":     "第二句",
+		"turn_03_researcher": "第三句",
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "参与者甲"},
+		StepSpec{ID: "p2", Role: "writer", Title: "参与者乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "甲", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+		RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 3})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// 预算停机：恰好 3 轮，轮转序 researcher→writer→researcher。
+	if sink.countText("轮次预算 3 用尽") != 1 {
+		t.Fatal("缺预算停机事件")
+	}
+	data, err := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	out := string(data)
+	for _, want := range []string{"1. researcher：第一句", "2. writer：第二句", "3. researcher：第三句"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("转录缺 %q：%s", want, out)
+		}
+	}
+	if strings.Contains(out, "4. ") {
+		t.Fatalf("超出预算仍在发言：%s", out)
+	}
+	// 名册步骤不执行，标记 skipped（非降级原因）。
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.ID == "p1" || s.ID == "p2" {
+			if s.Status != "skipped" || s.Error != gcRosterSkipReason {
+				t.Fatalf("名册步骤 %s = %s/%q", s.ID, s.Status, s.Error)
+			}
+		}
+	}
+}
+
+func TestTopologyConformance_groupChatCloseMarker(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": "我先说",
+		"turn_02_writer":     "同意收口 [close]",
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "甲"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "甲", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+		RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 8})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s", res.Status)
+	}
+	if sink.countText("发出收口标记") != 1 {
+		t.Fatal("缺收口标记事件")
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	if n := countGcSteps(st); n != 2 {
+		t.Fatalf("收口标记后应停在 2 轮，实际 %d", n)
+	}
+}
+
+func TestTopologyConformance_groupChatModerator(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": `{"next":"writer","reason":"请乙补充"}`,
+		"turn_02_writer":     "补一段分析",
+		"turn_03_researcher": `{"action":"close","reason":"讨论充分"}`,
+	}}
+	sink := &collectSink{}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "主持人"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, sink, []RoleSpec{
+		{Role: "researcher", Name: "主持", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "主持人群聊"}, RunOptions{
+		Topology: TopologyGroupChat, GroupChatPolicy: "moderator",
+		GroupChatModerator: "researcher", GroupChatMaxTurns: 8,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	if sink.countText("主持人收口") != 1 {
+		t.Fatal("缺主持人收口事件")
+	}
+	data, _ := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+	out := string(data)
+	for _, want := range []string{"1. researcher", "2. writer：补一段分析", "3. researcher"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("转录缺 %q：%s", want, out)
+		}
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	if n := countGcSteps(st); n != 3 {
+		t.Fatalf("主持人模式应 3 轮收口，实际 %d", n)
+	}
+}
+
+// 主持人指名不在名册 → fail-closed。
+func TestGroupChatModeratorUnknownNextFailsClosed(t *testing.T) {
+	be := &gcScriptBackend{messages: map[string]string{
+		"turn_01_researcher": `{"next":"ghost"}`,
+	}}
+	engine := newWP3Engine(t, mockScriptPlanner(
+		StepSpec{ID: "p1", Role: "researcher", Title: "主持人"},
+		StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+	), be, NopSink{}, []RoleSpec{
+		{Role: "researcher", Name: "主持", Backend: "gc-script"},
+		{Role: "writer", Name: "乙", Backend: "gc-script"},
+	})
+	res, err := engine.Run(context.Background(), Goal{Text: "x"}, RunOptions{
+		Topology: TopologyGroupChat, GroupChatPolicy: "moderator", GroupChatModerator: "researcher",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunFailed || !strings.Contains(res.Error, "不在名册内") {
+		t.Fatalf("应 fail-closed：%s %q", res.Status, res.Error)
+	}
+}
+
+// ---- 夹具 ----
+
+func countGcSteps(st *RunState) int {
+	n := 0
+	for _, s := range st.Plan.Workflow.Steps {
+		if strings.HasPrefix(s.ID, "gc") && s.Status == "completed" {
+			n++
+		}
+	}
+	return n
+}
+
+func newWP3Engine(t *testing.T, planner Planner, be Executor, sink ProgressSink, roles []RoleSpec) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range roles {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance(be.Name(), be)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: planner, Backends: backends, Sink: sink})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}
@@ -0,0 +1,355 @@
+package squad
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// wp4_conformance_test.go — Q2 WP4 验收：task_board 拓扑（claim/lease/complete
+// 原语 + 依赖门 + 孤儿租约再武装）与全 12 拓扑 conformance 汇总跑。
+
+// ---- task_board ----
+
+// 菱形依赖 a→(b,c)→d：认领序尊重依赖门，全部 done，标准 finalize 收口。
+func TestTopologyConformance_taskBoard(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "a", Role: "researcher", Title: "任务a", Targets: []string{"a.md"}},
+		StepSpec{ID: "b", Role: "writer", Title: "任务b", Needs: []string{"a"}, Targets: []string{"b.md"}},
+		StepSpec{ID: "c", Role: "reviewer", Title: "任务c", Needs: []string{"a"}, Targets: []string{"c.md"}},
+		StepSpec{ID: "d", Role: "writer", Title: "任务d", Needs: []string{"b", "c"}, Targets: []string{"output.md"}},
+	), be, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("status = %s err = %v", res.Status, res.Error)
+	}
+	// 认领序尊重依赖门：a 先于 b/c，d 最后（板上按步骤序认领，确定性）。
+	var order []string
+	for _, ev := range sink.snapshot() {
+		if strings.Contains(ev.Text, "认领") {
+			order = append(order, ev.StepID)
+		}
+	}
+	want := []string{"a", "b", "c", "d"}
+	if strings.Join(order, ",") != strings.Join(want, ",") {
+		t.Fatalf("认领序 = %v，期望 %v", order, want)
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, task := range st.Board {
+		if task.Status != "done" {
+			t.Fatalf("任务 %s = %s，期望 done", task.ID, task.Status)
+		}
+	}
+	if sink.countText("task_board：收口（done 4/4）") != 1 {
+		t.Fatal("缺收口事件")
+	}
+}
+
+// 孤儿租约再武装：崩溃（ctx 取消）留下 claimed+过期租约 → Resume 重入
+// 再武装回 open、重新认领（Attempts=2）并跑完。
+func TestTaskBoardOrphanLeaseRearms(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	crashed := false
+	be := &iterBackend{calls: map[string]int{}, content: func(key string, _ int, _ string) []byte {
+		if key == "02_writer" && !crashed {
+			crashed = true
+			cancel() // 模拟进程死亡：runOneStep 在 ctx 检查处退出，步骤回到 pending
+		}
+		return nil
+	}}
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "t1", Role: "researcher", Title: "一", Targets: []string{"a.md"}},
+		StepSpec{ID: "t2", Role: "writer", Title: "二", Needs: []string{"t1"}, Targets: []string{"output.md"}},
+	), be, sink)
+	res0, err := engine.Run(ctx, Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res0.Status != RunCancelled {
+		t.Fatalf("取消后应为 RunCancelled：%s", res0.Status)
+	}
+	// 崩溃现场：t2 claimed、租约未过期；步骤在盘上是 running（setStep 只改
+	// 内存，取消路径的 pending 未落盘）——进程死亡即此形态。测试把租约回拨
+	// 成过期（孤儿）并把步骤置回 pending（模拟重启前的一致化），再走 Resume。
+	st, err := engine.loadRun(engine.runDir(res0.RunID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for i := range st.Board {
+		if st.Board[i].ID == "t2" {
+			found = true
+			if st.Board[i].Status != "claimed" || st.Board[i].Attempts != 1 {
+				t.Fatalf("崩溃现场 t2 = %+v", st.Board[i])
+			}
+			st.Board[i].LeaseExpires = time.Now().Add(-time.Hour) // 租约过期 = 孤儿
+		}
+	}
+	if !found {
+		t.Fatal("板上无 t2")
+	}
+	for i := range st.Plan.Workflow.Steps {
+		if st.Plan.Workflow.Steps[i].ID == "t2" && st.Plan.Workflow.Steps[i].Status == "running" {
+			st.Plan.Workflow.Steps[i].Status = "pending"
+		}
+	}
+	if err := st.save(); err != nil {
+		t.Fatal(err)
+	}
+	// 重入：孤儿租约再武装 → 重新认领 → 跑完收口。
+	res, err := engine.Resume(context.Background(), st.Meta.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunCompleted {
+		t.Fatalf("恢复后应完成：%s %v", res.Status, res.Error)
+	}
+	if sink.countText("孤儿租约再武装 t2") != 1 {
+		t.Fatal("缺孤儿租约再武装事件")
+	}
+	st2, _ := engine.loadRun(engine.runDir(res.RunID))
+	for _, task := range st2.Board {
+		if task.ID == "t2" && task.Attempts != 2 {
+			t.Fatalf("再武装后 Attempts = %d，期望 2", task.Attempts)
+		}
+	}
+}
+
+// 依赖门级联：非关键任务失败 → 下游 open 任务级联 skipped → 降级收口。
+func TestTaskBoardDependencyGateCascade(t *testing.T) {
+	be := &iterBackend{calls: map[string]int{}, content: func(key string, _ int, _ string) []byte {
+		return nil
+	}}
+	failBe := &failOnceBackend{inner: be, failKey: "02_writer"}
+	crit := false
+	sink := &collectSink{}
+	engine := newTopologyEngine(t, mockScriptPlanner(
+		StepSpec{ID: "t0", Role: "researcher", Title: "零", Targets: []string{"output.md"}},
+		StepSpec{ID: "t1", Role: "writer", Title: "一", Needs: []string{"t0"}, Targets: []string{"b.md"}, Critical: &crit},
+		StepSpec{ID: "t2", Role: "reviewer", Title: "二", Needs: []string{"t1"}, Targets: []string{"c.md"}},
+	), failBe, sink)
+	res, err := engine.Run(context.Background(), Goal{Text: "任务板"}, RunOptions{Topology: TopologyTaskBoard})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != RunDegraded {
+		t.Fatalf("非关键失败+级联跳过应降级收口：%s %v", res.Status, res.Error)
+	}
+	st, _ := engine.loadRun(engine.runDir(res.RunID))
+	byID := map[string]BoardTask{}
+	for _, task := range st.Board {
+		byID[task.ID] = task
+	}
+	if byID["t0"].Status != "done" || byID["t1"].Status != "failed" || byID["t2"].Status != "skipped" {
+		t.Fatalf("板状态 = %+v", byID)
+	}
+	for _, s := range st.Plan.Workflow.Steps {
+		if s.ID == "t2" && (s.Status != "skipped" || s.Error != boardSkipReason) {
+			t.Fatalf("t2 = %s %q", s.Status, s.Error)
+		}
+	}
+}
+
+// failOnceBackend 对指定目录恒定失败（非瞬态），其余委托 inner。
+type failOnceBackend struct {
+	inner   Executor
+	failKey string
+}
+
+func (b *failOnceBackend) Name() string { return b.inner.Name() }
+
+func (b *failOnceBackend) Execute(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if key == b.failKey {
+		return &ExecResult{Success: false, ErrorReason: "fatal: 拒绝执行"}, nil
+	}
+	return b.inner.Execute(ctx, req)
+}
+
+// ---- 12/12 汇总 ----
+
+// TestTopologyConformance_allTwelve 全 12 内置拓扑汇总跑：每个拓扑名一次
+// 脚本化最小运行（零模型零网络）。graph/mode/driver 十种断言完成收口；
+// handoff/mailbox 两个 primitive 断言 fail-closed 拒绝单跑（其原语能力分别
+// 由 TestTopologyConformance_handoff 与 WP2 mailbox 测试钉住）。
+func TestTopologyConformance_allTwelve(t *testing.T) {
+	specs := BuiltinTopologies()
+	if len(specs) != 12 {
+		t.Fatalf("内置拓扑 = %d，期望 12", len(specs))
+	}
+	for _, spec := range specs {
+		spec := spec
+		t.Run(spec.Name, func(t *testing.T) {
+			switch spec.Name {
+			case TopologySequential:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyParallel:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "甲", Parallel: true, Targets: []string{"a.md"}},
+					StepSpec{ID: "s2", Role: "writer", Title: "乙", Parallel: true, Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyLoop:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "循环体", Targets: []string{"output.md", "fields.json"},
+						Loop: &LoopSpec{MaxIterations: 2, UntilField: "done", UntilEquals: "yes"}},
+				), &iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+					if target == "fields.json" {
+						return fieldsJSON(map[string]string{"done": "no"})
+					}
+					return nil
+				}}, RunOptions{}, RunCompleted)
+			case TopologyRouter:
+				runAgg(t, spec.Name, mockScriptPlanner(routerSteps()...),
+					&iterBackend{calls: map[string]int{}, content: func(_ string, _ int, target string) []byte {
+						if target == "fields.json" {
+							return fieldsJSON(map[string]string{"route": "fast"})
+						}
+						return nil
+					}}, RunOptions{}, RunCompleted)
+			case TopologyEvaluatorOptimizer:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "draft", Role: "writer", Title: "起草", Targets: []string{"output.md"}},
+					StepSpec{ID: "eval", Role: "reviewer", Title: "评审", Targets: []string{"verdict.md", "fields.json"},
+						Loop: &LoopSpec{MaxIterations: 4, UntilField: "verdict", UntilEquals: "pass", Body: []string{"draft"}}},
+				), &iterBackend{calls: map[string]int{}, content: func(_ string, iter int, target string) []byte {
+					if target == "fields.json" {
+						if iter >= 3 {
+							return fieldsJSON(map[string]string{"verdict": "pass"})
+						}
+						return fieldsJSON(map[string]string{"verdict": "fail"})
+					}
+					return nil
+				}}, RunOptions{}, RunCompleted)
+			case TopologyLongRunning:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}},
+					RunOptions{Deadline: time.Now().Add(5 * time.Second)}, RunCompleted)
+			case TopologyLeader:
+				b := &leaderScriptBackend{
+					decisions: []string{delegateJSON("writer", "撰写简报"), closeJSON},
+					script:    map[string]map[string]string{"01_writer": {"output.md": goodReport}},
+				}
+				engine := newLeaderEngine(t, b)
+				res, err := engine.Run(context.Background(), Goal{Text: "汇总"}, RunOptions{Topology: TopologyLeader})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+			case TopologyReview:
+				planner := mockScriptPlanner(
+					StepSpec{ID: "f1", Role: "researcher", Title: "评审", Targets: []string{"findings.json"}},
+				)
+				planner.Agents = []RoleSpec{
+					{Role: "researcher", Name: "评审", Backend: "agg-review"},
+					{Role: "verifier", Name: "核实", Backend: "agg-review"},
+				}
+				engine := newAggEngine(t, planner, &aggReviewBackend{}, []RoleSpec{
+					{Role: "researcher", Name: "评审", Backend: "agg-review"},
+					{Role: "verifier", Name: "核实", Backend: "agg-review"},
+				})
+				res, err := engine.Run(context.Background(), Goal{Text: "评审"},
+					RunOptions{Topology: TopologyReview, ReviewTarget: "材料"})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+				data, _ := os.ReadFile(filepath.Join(res.WorkDir, "output.md"))
+				if !strings.Contains(string(data), "已丢弃 0 条未验证") || !strings.Contains(string(data), "样例问题") {
+					t.Fatalf("报告 = %s", data)
+				}
+			case TopologyGroupChat:
+				engine := newAggEngine(t, mockScriptPlanner(
+					StepSpec{ID: "p1", Role: "researcher", Title: "甲"},
+					StepSpec{ID: "p2", Role: "writer", Title: "乙"},
+				), &gcScriptBackend{}, []RoleSpec{
+					{Role: "researcher", Name: "甲", Backend: "gc-script"},
+					{Role: "writer", Name: "乙", Backend: "gc-script"},
+				})
+				res, err := engine.Run(context.Background(), Goal{Text: "群聊"},
+					RunOptions{Topology: TopologyGroupChat, GroupChatMaxTurns: 2})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if res.Status != RunCompleted {
+					t.Fatalf("status = %s err = %v", res.Status, res.Error)
+				}
+			case TopologyTaskBoard:
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "t1", Role: "researcher", Title: "一", Targets: []string{"a.md"}},
+					StepSpec{ID: "t2", Role: "writer", Title: "二", Needs: []string{"t1"}, Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunCompleted)
+			case TopologyHandoff, TopologyMailbox:
+				// primitive：不可单独作为运行拓扑（fail-closed）；原语能力由
+				// 专项 conformance 钉住（handoff/mailbox 测试）。
+				runAgg(t, spec.Name, mockScriptPlanner(
+					StepSpec{ID: "s1", Role: "researcher", Title: "一", Targets: []string{"output.md"}},
+				), &iterBackend{calls: map[string]int{}}, RunOptions{}, RunFailed)
+			default:
+				t.Fatalf("未覆盖的拓扑 %q", spec.Name)
+			}
+		})
+	}
+}
+
+// runAgg 汇总跑的最小引擎：graph 类拓扑共用 iter 三角色装配。
+func runAgg(t *testing.T, topo string, planner Planner, be Executor, opts RunOptions, want RunStatus) {
+	t.Helper()
+	opts.Topology = topo
+	engine := newTopologyEngine(t, planner, be, nil)
+	res, err := engine.Run(context.Background(), Goal{Text: "汇总跑 " + topo}, opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Status != want {
+		t.Fatalf("status = %s err = %q，期望 %s", res.Status, res.Error, want)
+	}
+}
+
+// aggReviewBackend 汇总跑用的最小 review 后端：finder 出一条发现，verify 判 PASS。
+type aggReviewBackend struct{}
+
+func (b *aggReviewBackend) Name() string { return "agg-review" }
+
+func (b *aggReviewBackend) Execute(_ context.Context, req ExecRequest) (*ExecResult, error) {
+	key := filepath.Base(filepath.Clean(req.WorkDir))
+	if strings.HasPrefix(key, "verify_") {
+		return writeOne(req.WorkDir, "verdict.md", "核对成立。\nVERDICT: PASS"), nil
+	}
+	return writeOne(req.WorkDir, "findings.json",
+		`[{"location":"第1段","severity":"high","issue":"样例问题","evidence":"样例证据"}]`), nil
+}
+
+// newAggEngine 自定义角色表的汇总引擎（backend 实例名 = be.Name()）。
+func newAggEngine(t *testing.T, planner Planner, be Executor, roles []RoleSpec) *SquadEngine {
+	t.Helper()
+	dir := t.TempDir()
+	reg := NewRoleRegistry()
+	for _, r := range roles {
+		if err := reg.Register(r); err != nil {
+			t.Fatal(err)
+		}
+	}
+	backends := NewBackendResolver()
+	backends.RegisterInstance(be.Name(), be)
+	engine, err := NewEngine(Options{DataDir: dir, Registry: reg, Planner: planner, Backends: backends, Sink: NopSink{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return engine
+}