.task/reviews/iterative-delivery.json

@@ -0,0 +1,257 @@
+{
+  "status": "complete",
+  "verdict": "Ready with fixes",
+  "verdict_reason": "R1-R10 are addressed. One confirmed P3 needs a localized template/reference fix; there are no P0-P2 findings. This source review does not establish platform merge readiness: no GATE was verified.",
+  "scope": {
+    "mode": "branch-remote",
+    "base": "2147d5cb6566c301eaea3c65448bdda873093aea",
+    "branch": "feature/iterative-delivery",
+    "head_sha": "125e159924d90954c3bafc8572cc260e1e12eadc",
+    "pr_url": null,
+    "files_changed": 9
+  },
+  "intent": "Make Dev delivery iterative with ten stages and a requirements/fit table; default to stage-by-stage human review while honoring explicit scoped continuous authorization. Align generated task cards, entrypoint docs, and capability matrix; remove implicit task closure writes from post-merge without weakening safe cleanup or claiming manual approvals are machine verified.",
+  "intent_confidence": "explicit",
+  "reviewers": [
+    "correctness",
+    "project-standards",
+    "testing",
+    "agent-native",
+    "adversarial"
+  ],
+  "findings": [
+    {
+      "#": 1,
+      "autofix_class": "gated_auto",
+      "confidence": 100,
+      "evidence": [
+        "scripts/task.sh:184 -- - 流程依据：[policy/delivery.md](../policy/delivery.md)",
+        "scripts/task.sh:322-340 -- card_backfill updates status and main record, then appends closure notes; it does not rewrite relative links.",
+        "scripts/task.sh:391 and 401 -- git mv \"$c\" \".task/archive/${c##*/}\"",
+        "git ls-tree 125e159924d90954c3bafc8572cc260e1e12eadc:.task contains no policy subtree; policy/delivery.md exists at repository root."
+      ],
+      "file": "scripts/task.sh",
+      "first_evidence": "scripts/task.sh:184 -- - 流程依据：[policy/delivery.md](../policy/delivery.md)",
+      "independent_reviewers": [],
+      "line": 184,
+      "owner": "downstream-resolver",
+      "pre_existing": false,
+      "requires_verification": true,
+      "reviewers": [
+        "scope-audit"
+      ],
+      "severity": "P3",
+      "suggested_fix": "Use a plain-text repository-root reference, such as 流程依据：仓库根目录 policy/delivery.md, in the generated card and update its template assertion. This keeps the reference meaningful in both active and archived cards without modifying the generic archive mechanism.",
+      "title": "Keep the delivery reference valid after task archival",
+      "why_it_matters": "Every newly generated card has a working link while stored in .task/, but normal task-close moves it into .task/archive/ and leaves the link unchanged. The link then resolves to the nonexistent .task/policy/delivery.md, so historical task cards lose their clickable policy reference. A plain-text repository-root reference avoids dependence on the card directory.",
+      "validation_status": "confirmed",
+      "validation_reason": "The diff newly adds scripts/task.sh:184's ../policy/delivery.md link, while scripts/task.sh:391 and :401 move every closed card from .task/ to .task/archive/ without rewriting it, so it resolves to the absent .task/policy/delivery.md at reviewed head (the root policy/delivery.md exists); this occurs whenever a new task card is closed, and tests/test_task.py:605 only checks the pre-archive form.",
+      "protected_subject": null
+    }
+  ],
+  "actionable_findings": [
+    {
+      "#": 1,
+      "autofix_class": "gated_auto",
+      "confidence": 100,
+      "evidence": [
+        "scripts/task.sh:184 -- - 流程依据：[policy/delivery.md](../policy/delivery.md)",
+        "scripts/task.sh:322-340 -- card_backfill updates status and main record, then appends closure notes; it does not rewrite relative links.",
+        "scripts/task.sh:391 and 401 -- git mv \"$c\" \".task/archive/${c##*/}\"",
+        "git ls-tree 125e159924d90954c3bafc8572cc260e1e12eadc:.task contains no policy subtree; policy/delivery.md exists at repository root."
+      ],
+      "file": "scripts/task.sh",
+      "first_evidence": "scripts/task.sh:184 -- - 流程依据：[policy/delivery.md](../policy/delivery.md)",
+      "independent_reviewers": [],
+      "line": 184,
+      "owner": "downstream-resolver",
+      "pre_existing": false,
+      "requires_verification": true,
+      "reviewers": [
+        "scope-audit"
+      ],
+      "severity": "P3",
+      "suggested_fix": "Use a plain-text repository-root reference, such as 流程依据：仓库根目录 policy/delivery.md, in the generated card and update its template assertion. This keeps the reference meaningful in both active and archived cards without modifying the generic archive mechanism.",
+      "title": "Keep the delivery reference valid after task archival",
+      "why_it_matters": "Every newly generated card has a working link while stored in .task/, but normal task-close moves it into .task/archive/ and leaves the link unchanged. The link then resolves to the nonexistent .task/policy/delivery.md, so historical task cards lose their clickable policy reference. A plain-text repository-root reference avoids dependence on the card directory.",
+      "validation_status": "confirmed",
+      "validation_reason": "The diff newly adds scripts/task.sh:184's ../policy/delivery.md link, while scripts/task.sh:391 and :401 move every closed card from .task/ to .task/archive/ without rewriting it, so it resolves to the absent .task/policy/delivery.md at reviewed head (the root policy/delivery.md exists); this occurs whenever a new task card is closed, and tests/test_task.py:605 only checks the pre-archive form.",
+      "protected_subject": null
+    }
+  ],
+  "triage_groups": [],
+  "pre_existing_findings": [],
+  "requirements_completeness": {
+    "plan_source": "explicit",
+    "plan": ".task/feature-iterative-delivery.md",
+    "requirements": [
+      {
+        "id": "R1",
+        "status": "met",
+        "assessment": "Ten stages name their artifacts and exits.",
+        "evidence": "policy/delivery.md: stages and exits"
+      },
+      {
+        "id": "R2",
+        "status": "met",
+        "assessment": "Generated cards include stable requirement IDs, fit evidence, acceptance criteria and disposition.",
+        "evidence": "scripts/task.sh: requirement table"
+      },
+      {
+        "id": "R3",
+        "status": "met",
+        "assessment": "Manual review defaults and scoped continuous authorization are explicit; templates start unreviewed.",
+        "evidence": "policy/delivery.md: authorization; scripts/task.sh: stage review table"
+      },
+      {
+        "id": "R4",
+        "status": "met",
+        "assessment": "Design, implementation and test plans retain versioned approval objects, assumptions and verification methods.",
+        "evidence": "policy/delivery.md: stages, exploration and feedback loops"
+      },
+      {
+        "id": "R5",
+        "status": "met",
+        "assessment": "Four impact classes determine continuation, revision and renewed review.",
+        "evidence": "policy/delivery.md: feedback-loop impact table"
+      },
+      {
+        "id": "R6",
+        "status": "met",
+        "assessment": "Post-merge defaults to reminders; governance tests exercise real merge/pull without closure writes.",
+        "evidence": "scripts/hooks/post-merge; tests/test_task.py; independent local replay"
+      },
+      {
+        "id": "R7",
+        "status": "met",
+        "assessment": "Merge and cleanup exits preserve registered-workspace, writer and unique-file checks, with receipts outside deleted workspaces.",
+        "evidence": "policy/delivery.md: merge, cleanup and recovery; policy/governance.md"
+      },
+      {
+        "id": "R8",
+        "status": "met",
+        "assessment": "Entrypoint, adoption records, generated cards and capability classifications align; governance replay passed. Finding #1 is a narrow archival-reference defect.",
+        "evidence": "AGENTS.md; policy/adoption.md; policy/tool-adaptation-matrix.md; scripts/task.sh"
+      },
+      {
+        "id": "R9",
+        "status": "met",
+        "assessment": "Bounded experiments can precede complete design, with explicit approval and separate acceptance.",
+        "evidence": "policy/delivery.md: design exploration loop"
+      },
+      {
+        "id": "R10",
+        "status": "met",
+        "assessment": "Incremental work records discoveries, adjustments, renewed review/tests, decisions, outcomes and stopping conditions.",
+        "evidence": "policy/delivery.md: feedback loops; scripts/task.sh: exploration and iteration table"
+      }
+    ],
+    "implementation_units": []
+  },
+  "learnings": [],
+  "agent_native_gaps": [],
+  "deployment_notes": [],
+  "residual_risks": [],
+  "testing_gaps": [
+    "Full application suite and build were not rerun at the reviewed head; the reported earlier author run is not independently verified.",
+    "Remote PR CI and GATE identity/SHA were not verified.",
+    "The later uncommitted archival-reference fix and assertion were not reviewed or verified."
+  ],
+  "coverage": {
+    "notes": [
+      "Reviewed only the nine-file diff between fixed base and head; main checkout source was not used for changed files. Full roster selected because instruction prose and changed tests rule out lite coverage.",
+      "Project standards used root AGENTS.md as instruction fallback; no CODING_STANDARDS.md applies. No matching solutions corpus or declared local pack was selected.",
+      "Independent replay by /root/delivery_review at fixed head 125e159924d90954c3bafc8572cc260e1e12eadc: make test-governance && git diff --check; exit 0, 104 tests passed. Receipt /tmp/iterative-delivery-review-replay-20261007.json remains source SELF; local evidence does not authenticate CI or reviewer identity.",
+      "Full application suite and build were not rerun by this review. Author reported they passed on 9454819 with only task metadata changed afterward; remote PR CI is not yet available and is not validated by this review.",
+      "Supplemental scope-audit is a non-independent source inspection; the five dispatched specialist reviewers returned no findings.",
+      "The root caller is preparing the minimal archival-reference fix outside the fixed review head. Review does not claim that uncommitted fix verified."
+    ],
+    "reviewer_selection": [
+      {
+        "reviewer": "correctness",
+        "tier": "session",
+        "reason": "always-on"
+      },
+      {
+        "reviewer": "project-standards",
+        "tier": "mid",
+        "reason": "AGENTS.md governs all changed files"
+      },
+      {
+        "reviewer": "testing",
+        "tier": "mid",
+        "reason": "real-Git hook tests and generated task-card assertions changed"
+      },
+      {
+        "reviewer": "agent-native",
+        "tier": "mid",
+        "reason": "agent instructions, commands and task-card context changed"
+      },
+      {
+        "reviewer": "adversarial",
+        "tier": "session",
+        "reason": "post-merge persistence side effects and command composition changed"
+      }
+    ],
+    "lite": false,
+    "cross_model": {
+      "outcome": "in-process-fallback",
+      "note": "Independent external cross-model pass not run: branch-remote scope uses the local adversarial reviewer against fixed refs. No code sent externally."
+    },
+    "validator": {
+      "outcome": "verdicts",
+      "batches": 1,
+      "confirmed": 1,
+      "rejected": 0,
+      "unresolved": 0,
+      "malformed": 0,
+      "failed": 0,
+      "shortcut_skipped": 0,
+      "shortcut_basis": "No verified independent cross-model corroboration.",
+      "bound_exceeded": false,
+      "results": [
+        {
+          "#": 1,
+          "status": "confirmed",
+          "protected_subject": null,
+          "reason": "The diff newly adds scripts/task.sh:184's ../policy/delivery.md link, while scripts/task.sh:391 and :401 move every closed card from .task/ to .task/archive/ without rewriting it, so it resolves to the absent .task/policy/delivery.md at reviewed head (the root policy/delivery.md exists); this occurs whenever a new task card is closed, and tests/test_task.py:605 only checks the pre-archive form."
+        }
+      ],
+      "protected_subject_reclassifications": 0,
+      "degraded_blockers": []
+    },
+    "suppressed_by_confidence": {},
+    "malformed_findings": 0,
+    "malformed_returns": 0,
+    "mode_aware_demotions": 0,
+    "missing_first_evidence_demotions": 0,
+    "first_evidence_backfilled": 0,
+    "settlement_evaluated": true,
+    "settlement_suppressed": 0,
+    "non_independent_attributed_sources": [
+      "scope-audit"
+    ],
+    "confidence_basis": "Direct fixed-head source evidence; no confidence promotion from scope-audit attribution or validator agreement.",
+    "independent_replay": {
+      "source": "SELF",
+      "reviewer": "/root/delivery_review",
+      "reviewer_identity_authenticated": false,
+      "head_sha": "125e159924d90954c3bafc8572cc260e1e12eadc",
+      "command": "make test-governance && git diff --check",
+      "exit_code": 0,
+      "tests_passed": 104,
+      "receipt": "/tmp/iterative-delivery-review-replay-20261007.json"
+    },
+    "gate_verified": false,
+    "failed_reviewers": [],
+    "bound_exceeded": [],
+    "compound_packs": {
+      "roots": [],
+      "errors": [],
+      "warnings": []
+    },
+    "apply": "Skipped: agent report-only; later author edits are outside the fixed review head."
+  },
+  "artifact_path": "/tmp/compound-engineering-1000/ce-code-review/20261007-025812-7b611869",
+  "run_id": "20261007-025812-7b611869"
+}

AGENTS.md

@@ -1,7 +1,9 @@
 # Agent 工作说明（Agent-System-Dev）
 
 本仓开发遵循 Development-Guidelines（见 `deps/development-guidelines`，gitlink 固定版本）：
-核心规则与术语见其 `core/`，按任务触发阅读 `standards/` 专题；日常交付走 `workflows/delivery.md`。
+核心规则与术语见其 `core/`，按任务触发阅读 `standards/` 专题；通用交付依据为 `workflows/delivery.md`。
+本项目的需求、设计试验、实施调整及交付**必读 [policy/delivery.md](policy/delivery.md)**：
+默认逐阶段人工评审，阶段内迭代、按影响回退；仅在用户明确授权范围内连续推进。
 
 ## UI 设计门禁（强制）
 
@@ -37,9 +39,8 @@
   `make task-finish ABANDON="--abandon <原因>"` 并记录原因；放弃不豁免现场保护。
 - **任务卡闭合（防欠账）**：分支合入 main 后，其任务卡须回填"已完成"+主记录 PR 链接并
   归档至 `.task/archive/`。`task-finish` 回收后提示所有者运行 `make task-close`，不借回收修改
-  共享主检出或推送另一任务。自动路径保留主检出 pull 触发 post-merge
-  钩子（实测 ff-only pull 也触发）；钩子内 close 只扫主检出本地 `.task/`，远端刚合入的卡片
-  要等 pull 完成才能被看到，扫不到时由手动 `make task-close` 兜底。手动闭合在主检出运行→建
+  共享主检出或推送另一任务。主检出 pull/merge 的 post-merge 钩子仅提示待办，不自动写入或推送。
+  `make task-close` 会提交、推送并可能创建 PR，须先核对闭合范围和对应授权。闭合在主检出运行→建
   `feature/card-close-*` 分支+PR，任务分支内加 `FOLD=--fold` 随当前 PR 合入。
   precheck 的 `[cards]` 行报告欠账，验收前应清零；闭合 PR 由所有者合入。
 
@@ -102,5 +103,5 @@
 - **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。
-- 任务卡闭合 PR（`make task-close` / `task-finish` 自动闭合）经 API 建 PR 时以
+- 任务卡闭合 PR（获授权后运行 `make task-close`）经 API 建 PR 时以
   `GITEA_TOKEN` 环境变量传入同上的 token。

policy/adoption.md

@@ -16,6 +16,7 @@
 | standards/verification.md | 验收 | 本地 make test-governance + 组合测试；当前证据与历史 scratch 记录分开 |
 | standards/multi-repo.md | 跨仓/gitlink | manual；受控历史迁移通道已于 2026-09-30 行使一次（documents/records/2026-09-30-history-migration.md） |
 | 工作区回收 | 交付终点 | advisory（task-finish）；负例：合入前 finish 拒绝输出 |
+| policy/delivery.md（项目收紧） | 需求、试验、实施迭代及交付 | manual：人工评审和有范围的连续推进授权；任务卡提供记录，post-merge 默认只提示，行为回归见 tests/test_task.py |
 | standards/tool-adaptation.md（draft.8） | 为工具创建规则/钩子实例 | 本组合 CLI/git 级实例已登记：policy/tool-adaptation-matrix.md（A1–A7 逐行定级+正负例证据） |
 | 同上（zcode 实例） | zcode 客户端写路径 | policy/tool-adaptation-zcode.md：PreToolUse 写守卫；脚本负例与客户端接线分开，客户端 pending-live |
 | core/baseline.md「规范层级」+ documentation.md「记忆分层/机制化」（draft.9） | 层级效力、记忆权威、经验机制化 | manual：本仓 policy/ 仅收紧不放宽；记忆四层按条文执行（任务卡=任务记忆，documents/records=项目记忆，Dev 仓 AGENTS.md 指针已就位） |
@@ -25,6 +26,7 @@
 | 原规则/章节 | 项目采用方式 | 原因与风险 | 负责人/复查条件 |
 |---|---|---|---|
 | baseline 硬边界 1（授权范围） | 高后果双钥匙：改写历史/解除分支保护/删归档/发布——Agent 只准备，用户执行（AGENTS.md 条文） | 单人+Agent 场景下消除单点；风险=用户不在场时高后果操作排队 | weichen；首次触发时复查 |
+| delivery.md §2、design-verification.md 的轻量流程默认做法 | 按 [policy/delivery.md](delivery.md) 默认逐阶段人工评审；小任务简化篇幅，阶段内可迭代；明确授权时在指定范围连续推进 | 保留需求、方案和实施决策权；按变化影响重审，避免每条命令审批及全流程重复返工 | weichen；阶段等待或试验边界影响交付时复查 |
 
 - 历史记录（2026-09-30）报告 11 个组合仓 main 禁直推、拒绝陈旧候选；当前平台能力须独立重验，本地脚本测试不替代远端证据。
 - 当前规范内容以 gitlink 及其 VERSION 为准。治理可靠性修复的命令与证据约定见 [开发门禁](governance.md)，本地与客户端/平台能力分别登记；本轮不改变分支保护、发布授权或通用库版本。

policy/delivery.md

@@ -0,0 +1,110 @@
+# 需求、试验与迭代交付
+
+适用：本项目的需求、修复、规范修改和其他交付任务。依据固定版本的
+[通用交付流程](../deps/development-guidelines/workflows/delivery.md)，本项目收紧为
+**默认逐阶段人工评审；阶段内迭代，按影响回退，按增量交付**。
+本文件是阶段与评审的权威表述；启动、同步、命令证据和回收操作见
+[开发门禁](governance.md)。规则的执行能力见 [能力矩阵](tool-adaptation-matrix.md)。
+
+## 授权与阶段评审
+
+任务卡默认记为 `manual`。完成当前阶段产物并取得人工评审结论后，才进入下一阶段。
+评审记录包含阶段、对象版本、评审人、时间、结论、意见及可追溯依据；结论使用
+未评审、通过、需修改、拒绝、不适用。人工结论与命令执行证据分开记录；Agent 自审、
+测试通过、工具写入或用户沉默均不能代替人工批准，也不能由作者代签独立审核。
+
+只有用户明确要求自动或连续推进时，才将对应范围记为 `authorized-continuous`，记录
+授权来源、覆盖阶段、允许动作、交付终点、资源约束和停止条件。未覆盖阶段仍逐阶段评审。
+授权连续推进不免除需求分析、试验判定、测试、代码 Review、证据和平台准入；
+既有独立审核要求及高后果操作双钥匙仍有效。单独的合入授权不补齐尚不存在的前序产物。
+
+阶段内已授权的调查、编辑、自检、缺陷修复和等价实现可以持续进行，无需逐命令批准。
+需求拆解、评估、需求表可以共用一份任务卡，但分别保留阶段结论。小任务缩短篇幅；
+确实不适用的内容说明原因，经人工确认或在明确授权的裁剪范围内处理，不能静默跳步。
+任务启动检查、独立工作区和收据用于准备现场，不表示任何阶段评审已通过。
+
+## 阶段与出口
+
+| 阶段 | 必需产物 | 进入下一阶段的条件 |
+|---|---|---|
+| 1 需求拆解 | 目标、用户场景、子需求、边界、非目标及待澄清问题 | 人工确认拆解与原始意图一致 |
+| 2 需求评估 | 价值、可行性、成本、风险、依赖、现有能力及在途任务重叠；与项目定位、模块职责和公开契约的符合度 | 人工确认取舍；重大未知转入下述受控试验 |
+| 3 需求表确认 | 稳定 ID、描述、适配依据、优先级、可判定验收条件、依赖和处理结论 | 人工确认本增量范围；影响实施的关键待核实项已处理 |
+| 4 方案设计 | 推荐方案与主要备选、职责和接口、数据与权限影响、失败恢复、待验证假设及试验结论 | 人工批准当前方案版本及剩余风险处理方式 |
+| 5 实施计划 | 关联需求的工作项、依赖顺序、责任人、文件或模块范围、测试方法与环境、完成条件及停止条件 | 人工批准当前实施和测试计划；远期工作可保留粗粒度 |
+| 6 实施与自检 | 可审查变更、局部测试、计划偏差与迭代记录 | 人工确认当前增量实现产物，必要设计调整已重新获批 |
+| 7 正式测试与验收 | 需求到证据映射；正常、异常、边界、回归结果及缺项 | 人工审阅当前候选的验证结果；正式验收前同步并固定被测对象 |
+| 8 代码 Review | 需求符合性、正确性、维护性、测试充分性及问题闭合记录 | 完成必要重放、整改、补测和复审，由有权评审者放行当前版本 |
+| 9 合入 | 最新 base/head、必要检查及评审、冲突处置、合入授权 | 合入前再次同步核验；经 PR 合入后取得真实合入结果 |
+| 10 合入核验与收口 | 实际合入内容及组件版本、验证依据、成果保存位置、写者状态、清理清单与任务归档安排 | 清理前取得人工批准或对应连续推进授权；安全回收后核对结果、闭合任务卡 |
+
+表中的人工节点仅在明确的连续推进授权范围内按授权执行；阶段产物及验证条件保留。
+正式测试和最终 Review 的顺序不限制设计阶段试验、实施期测试或提前审阅。
+文本合并无冲突不证明组合兼容；合入后仍核验真实内容与被测候选是否一致。
+
+## 需求表与追溯
+
+需求表以任务卡为主记录。项目符合度使用“符合 / 需调整后符合 / 不符合 / 待核实”，
+附当前源码、项目约束或实测依据。处理结论使用“纳入本增量 / 后续处理 / 拒绝 / 待定”，
+与符合度分开；关键依据缺失时保留待核实，不凭推测判定通过。
+
+每个纳入实施的需求必须能追溯到：
+
+`需求 ID 与验收条件 → 方案及工作项 → 测试或审阅方法 → 当前版本结果与证据`。
+
+没有可判定验收方法的需求先澄清；待定或不符合的需求未经重新评审不得进入正式实施。
+跨模块、权限、持久化、异步和恢复需求按实际风险补充真实链路及故障场景。
+规范和文档任务采用差异、链接、规则一致性及相关工具行为检查，不制造无关业务测试。
+
+## 设计探索循环
+
+需求分析和设计阶段允许先试验再完善方案，无须先批准尚未能确定的完整设计。
+执行试验前提交最小试验计划：待回答的问题或假设、对象和隔离环境、方法、预期判据、
+允许副作用、时间或成本边界、停止条件及产物保存或处置方式，并取得人工批准；
+已有连续推进授权明确覆盖该试验范围时，引用该授权。
+
+在已批准范围内可以反复尝试：
+
+`初步方案 → 批准试验范围 → 原型或测试 → 分析证据 → 修订方案与计划 → 结果评审`。
+
+记录成功、失败及不确定结果，以及下一步继续探索、采用、换方案或暂停的理由。
+超出范围、触及未授权共享环境、达到预算或持续无有效进展时，暂停受影响试验并提交决策。
+无需强求全部未知一次消除；影响下一步的关键假设必须有结论，其余风险有责任人与处理安排。
+
+原型标明用途和限制，不自动成为正式实现。决定复用时纳入方案、计划、代码 Review 与正式验证。
+试验成功仅证明相应假设，不能替代验收；失败结果用于调整方案，不能静默降低验收标准。
+
+## 实施、测试与 Review 的反馈循环
+
+按可独立验证的增量推进；近期工作细化到可执行，后续工作根据证据逐步细化。
+每轮形成“发现与证据 → 原假设或困难 → 调整及影响 → 重审和补测范围 → 决策 → 结果”。
+记录放在同一任务卡，较长材料只引用文件或 PR，避免多份状态。
+
+| 变化 | 执行动作与评审边界 |
+|---|---|
+| 已批准范围内的内部组织、等价实现、局部缺陷修正 | 当前阶段继续调整、自测并记录；阶段出口统一人工评审 |
+| 技术路线、模块职责、公开契约、数据或权限设计、外部影响，或计划的关键依赖、交付顺序和资源约束变化 | 暂停受影响工作，修订方案或计划，经人工重审后继续；已有明确授权覆盖该类调整时按其边界执行 |
+| 目标、范围、优先级或验收条件变化 | 返回需求评估和需求表，取得对应批准后更新下游方案、计划和验证 |
+| 影响不明 | 先调查；仍无法判定时提交人工决定，不自行归入细节调整 |
+
+未受影响且已有授权的工作可以继续。测试或 Review 发现实现缺陷，返回实施修复；
+发现设计缺陷返回方案；发现需求理解错误返回需求阶段。只重走受影响环节，并说明
+未重走部分为何仍有效。新版本不能沿用已受影响的批准、测试或 Review 结论。
+
+批准对象使用文档修订或提交标识；正式验证固定源码、组件组合与相关配置。
+对 head/base、依赖、需求或契约变化重新判断证据及评审的有效性，按影响补测和复审。
+阶段状态、测试状态、代码审核状态及合入状态分别记录，不以一个“完成”互相替代。
+
+## 合入、清理与恢复
+
+合入与清理分别核对授权及条件。仅回收本任务登记工作区，按开发门禁核对当前权威目标、
+父子仓成果、未跟踪和 ignored 文件、独有提交与引用、活动 Git 操作以及所有写者退出。
+无法证明成果已保存时保留现场；本地删除与远端分支删除分别报告。
+
+清理前保存任务主记录与证据。执行回收时将命令输出用 evidence 的 `RECEIPT` 模式保存到
+待删目录之外；回收后的结果写入后续闭合记录，不能向已删除的任务卡补写收据。
+任务卡须回填真实主 PR 和完成状态并归档；闭合 PR 的创建、合入均依当前授权执行。
+post-merge 默认只提示，不因 pull 或 merge 自动提交、推送或创建闭合 PR。
+
+中断或交接时记录当前有效版本、批准范围、未决假设、最新试验结论、剩余事项及下一动作。
+恢复后核对现场和授权，从最早未满足的环节续办；不重做仍有效的工作，不省略失效环节。

policy/governance.md

@@ -2,6 +2,8 @@
 
 本文件细化根 AGENTS.md 的启动、验证和回收规则。通用语义以固定的
 `deps/development-guidelines` 为准；门禁实现及回归用例随本仓版本管理。
+需求表、设计试验、迭代调整、阶段产物与人工评审按 [迭代交付](delivery.md) 执行。
+任务卡记录批准来源与对象版本；本地 evidence 仅记录命令，不能证明人工批准。
 
 ## 启动与恢复
 
@@ -64,6 +66,9 @@ Git 对含子模块的工作区需要显式删除选项时，只有全部现场
 owner 须先独立确认其保存与归属，工具不把缓存的远端引用当作唯一数据已经备份的证据。
 远端分支删除与本地回收分别报告，失败不能声明两者全部完成。
 回收后由所有者执行 `make task-close` 闭合卡片；回收命令不顺带改写共享主检出。
+post-merge 仅提示，不自动调用闭合。`make task-close` 会提交、推送并可能创建 PR，
+执行者先按迭代交付规则核对该动作的授权及所选任务，保留未授权的其他任务现场。
+回收命令使用待删目录之外的 `RECEIPT` 保存输出，清理结果由后续闭合记录引用。
 
 ## CI 凭据边界
 

policy/tool-adaptation-matrix.md

@@ -14,6 +14,7 @@
 | A5 回收 | 本地命令检查；写者退出由 owner 确认 | 登记根目录、锁/操作状态、父子仓与唯一文件、合入核验 | `tests/test_task.py` |
 | A6 秘密 | 本地脱敏及 CI 获取边界；权限与宿主隔离另验 | 命令/输出/摘要脱敏，凭据不持久写 Git 配置 | `tests/test_evidence.py`、`tests/test_ci_credentials.py` |
 | A7 证据 | 本地 SELF 收据；外部来源独立核验 | 首次捕获、完整脱敏输出/摘要、对象记录、原样导入 | `tests/test_evidence.py`；CI/审阅链接另核 |
+| A8 阶段评审与迭代 | manual；模板和提示不构成批准身份校验或全流程阻断 | policy/delivery.md；任务卡记录需求、方案/计划、评审与试验迭代；post-merge 默认仅提示 | `tests/test_task.py` 核对模板及钩子不自动写入；人工核对批准范围和版本 |
 
 ## 失败处理
 

scripts/hooks/post-merge

@@ -1,15 +1,12 @@
 #!/usr/bin/env bash
-# 本地钩子（post-merge）：pull/merge 落地后尝试闭合已合入任务卡；ff-only 也会触发。
-# precheck [cards] 报告欠账，所有者可用 make task-close 手动兜底。
+# 本地钩子（post-merge）：pull/merge 落地后提示 owner 按授权闭合任务卡；ff-only 也会触发。
+# precheck [cards] 报告欠账；此钩子只提示，不执行闭合或其他写入。
 # 触发条件收紧到"主检出 + 集成分支 main"——任务 worktree 里 task-sync 的 merge 不触发，
-# 防止把卡片闭合混进任务分支现场。失败只警告，绝不阻断 pull。
-# 注意：task.sh close 在主检出会另建 feature/card-close-* 分支提交（集成分支不直推），
-# 结束后切回 main；PR 由所有者合入。
+# 防止把卡片闭合提示混进任务分支现场。
 common=$(git rev-parse --git-common-dir) && common=$(cd "$common" 2>/dev/null && pwd) || exit 0
 mainrepo=$(dirname "$common")
 [ "$mainrepo" = "$PWD" ] || exit 0
 [ "$(git rev-parse --abbrev-ref HEAD)" = "main" ] || exit 0
 [ -f "$mainrepo/scripts/task.sh" ] || exit 0
-bash "$mainrepo/scripts/task.sh" close \
-  || echo "[warn] post-merge 卡片自动闭合失败——在主检出手动运行 make task-close" >&2
+echo "[next] owner 按已有授权在主检出运行 make task-close 闭合已合入任务卡；未授权时先取得授权。"
 exit 0

scripts/task.sh

@@ -99,7 +99,7 @@ cmd_precheck() {
   else
     echo "[facts]   WARN policy/facts.md 不存在——外部事实登记表缺失（AGENTS.md 证据与验证门禁）"
   fi
-  echo "[required-reading] core/baseline.md, workflows/delivery.md §1$([ "$dirty" -gt 0 ] && echo ", standards/collaboration.md（树脏=他人或未收尾现场，不得写入）")"
+  echo "[required-reading] core/baseline.md, workflows/delivery.md §1$([ -f policy/delivery.md ] && echo ", policy/delivery.md")$([ "$dirty" -gt 0 ] && echo ", standards/collaboration.md（树脏=他人或未收尾现场，不得写入）")"
   echo "[R0] $(echo "$rmain" | cut -c1-10)（完整：$rmain）"
   if [ "$dirty" -gt 0 ] || [ "${sub_drift:-0}" -gt 0 ]; then
     echo "[verdict] SHARED_WORKSPACE_BUSY — 工作树不干净，不得写入；先收尾或经 owner 确认"
@@ -178,11 +178,48 @@ cmd_start() {
 - 工作区：${wt}（worktree）；分支：${name}
 ${parent_note}
 - 授权与交付终点：<本次允许的动作与终点；未授权事项写明>
+- 交付模式：manual（逐阶段人工评审）；明确授权后可记为 authorized-continuous
+- 自动推进授权：无；授权范围：无；授权来源：无；授权终点：无；停止条件：无（未授权）
+- 当前阶段：需求拆解；当前增量：I1（待定义）
+- 流程依据：仓库根目录 policy/delivery.md（归档后仍按仓库根目录定位）
 
 ## 目标与非目标
 
 <预期结果、受影响入口/模块/契约；明确不做什么。>
 
+## 需求分析
+
+| ID | 描述 | 项目符合度及依据 | 优先级 | 验收条件 | 依赖 | 结论 |
+|---|---|---|---|---|---|---|
+| R1 | 待填写 | 待评估 | 待定 | 待定义 | 待确认 | 待评审 |
+
+## 方案与计划
+
+- 方案引用：待填写（含对象版本）
+- 实施计划引用：待填写（含增量与需求 ID）
+- 测试计划引用：待填写（含需求 ID 与验收条件）
+
+## 阶段评审
+
+| 阶段 | 对象版本 | 结论 | 评审人 | 时间 | 依据 |
+|---|---|---|---|---|---|
+| 需求拆解 | 待定 | 未评审 | — | — | — |
+| 需求评估 | 待定 | 未评审 | — | — | — |
+| 需求表确认 | 待定 | 未评审 | — | — | — |
+| 方案设计 | 待定 | 未评审 | — | — | — |
+| 实施计划 | 待定 | 未评审 | — | — | — |
+| 实施与自检 | 待定 | 未评审 | — | — | — |
+| 正式测试与验收 | 待定 | 未评审 | — | — | — |
+| 代码 Review | 待定 | 未评审 | — | — | — |
+| 合入 | 待定 | 未评审 | — | — | — |
+| 合入核验与收口 | 待定 | 未评审 | — | — | — |
+
+## 探索试验与迭代
+
+| 增量/试验 | 假设 | 范围 | 判断和停止条件 | 批准依据 | 结果 | 变更影响 | 需重审补测 |
+|---|---|---|---|---|---|---|---|
+| 待定义 | — | — | — | 未批准 | NOT_RUN | 待评估 | 待评估 |
+
 ## 同步记录
 
 | 节点 | sync_base_sha | tested_head_sha | checked_at |
@@ -200,6 +237,11 @@ ${parent_note}
 - 结果：
 - 验证：<结论摘要 PASS/FAIL/SKIP/NOT_RUN 逐项；逐条证据由 make evidence 写入上方证据表>
 - 影响与剩余项：
+
+## 合入与清理
+
+- 合入记录：NOT_RUN（授权依据、PR、合入 SHA、验证来源）
+- 清理记录：NOT_RUN（授权依据、写者退出确认、worktree/分支/任务卡处置）
 EOF
   (cd "$wt" && IMPORT_RECEIPT="$PRECHECK_RECEIPT" bash "${SCRIPT_DIR}/evidence.sh")
   git -C "$wt" add .task

tests/test_task.py

@@ -89,6 +89,18 @@ class TaskLifecycleTests(unittest.TestCase):
         self.assertNotEqual(result.returncode, 0, result.stdout)
         self.assertIn("SYNC_REQUIRED", result.stdout)
 
+    def test_precheck_routes_existing_project_delivery_policy(self):
+        result = self.task(self.repo, "precheck")
+        self.assertNotIn("policy/delivery.md", result.stdout)
+        policy = self.repo / "policy/delivery.md"
+        policy.parent.mkdir()
+        policy.write_text("# Project delivery policy\n")
+        self.git(self.repo, "add", "policy")
+        self.git(self.repo, "commit", "-m", "Adopt project delivery policy")
+        result = self.task(self.repo, "precheck")
+        reading = next(line for line in result.stdout.splitlines() if line.startswith("[required-reading]"))
+        self.assertIn("policy/delivery.md", reading)
+
     def test_sync_materializes_updated_component(self):
         child = self.add_submodule()
         wt = self.worktree()
@@ -586,11 +598,99 @@ class TaskLifecycleTests(unittest.TestCase):
         self.task(self.repo, "start", "feature/new", env=self.env | {"PRECHECK_RECEIPT": str(receipt)})
         wt = self.root / "repo-wt-feature-new"
         self.assertTrue((wt / ".task/feature-new.md").is_file())
-        self.assertIn("make precheck", (wt / ".task/feature-new.md").read_text())
+        card = (wt / ".task/feature-new.md").read_text()
+        self.assertIn("make precheck", card)
+        self.assertIn("交付模式：manual（逐阶段人工评审）", card)
+        self.assertIn("authorized-continuous", card)
+        self.assertIn("流程依据：仓库根目录 policy/delivery.md", card)
+        for field in ("授权范围", "授权来源", "授权终点", "停止条件"):
+            self.assertIn(f"{field}：无", card)
+        for field in ("当前阶段", "当前增量", "方案引用", "实施计划引用", "测试计划引用"):
+            self.assertIn(field + "：", card)
+        self.assertIn("| ID | 描述 | 项目符合度及依据 | 优先级 | 验收条件 | 依赖 | 结论 |", card)
+        reviews = card.split("## 阶段评审\n", 1)[1].split("\n## ", 1)[0]
+        self.assertIn("| 阶段 | 对象版本 | 结论 | 评审人 | 时间 | 依据 |", reviews)
+        for stage in ("需求拆解", "需求评估", "需求表确认", "方案设计", "实施计划", "实施与自检",
+                      "正式测试与验收", "代码 Review", "合入", "合入核验与收口"):
+            row = next(line for line in reviews.splitlines() if line.startswith(f"| {stage} |"))
+            self.assertEqual(row.split("|")[3].strip(), "未评审")
+        self.assertNotIn("通过", reviews)
+        self.assertIn("## 探索试验与迭代", card)
+        for field in ("假设", "范围", "判断和停止条件", "批准依据", "结果", "变更影响", "需重审补测"):
+            self.assertIn(field, card.split("## 探索试验与迭代\n", 1)[1].split("\n## ", 1)[0])
+        self.assertIn("## 合入与清理", card)
+        for field in ("合入记录", "清理记录"):
+            self.assertIn(field + "：NOT_RUN", card)
         body = self.git(wt, "log", "-1", "--format=%B").stdout
         for field in ("Why:", "Evidence:", "Impact:", "Refs:"):
             self.assertIn(field, body)
 
+    def install_post_merge_hook(self):
+        hook = self.repo / ".git/hooks/post-merge"
+        shutil.copyfile(SOURCE / "scripts/hooks/post-merge", hook)
+        hook.chmod(0o755)
+
+    def publish_merged_task(self):
+        publisher = self.root / "publisher"
+        self.git(self.root, "clone", str(self.remote), str(publisher))
+        self.git(publisher, "switch", "-c", "feature/merged")
+        card = publisher / ".task/feature-merged.md"
+        card.parent.mkdir()
+        card.write_text("# Merged task\n任务主记录：待回填；状态：进行中；owner：test\n"
+                        "- 工作区：fixture；分支：feature/merged\n")
+        self.git(publisher, "add", ".task")
+        self.git(publisher, "commit", "-m", "Record merged task")
+        self.git(publisher, "switch", "main")
+        self.git(publisher, "merge", "--no-ff", "feature/merged", "-m",
+                 "Merge pull request 'Task' (#7) from feature/merged into main")
+        self.git(publisher, "push", "origin", "main")
+        return self.git(publisher, "rev-parse", "HEAD").stdout.strip(), card.read_text()
+
+    def assert_post_merge_only_reminds_owner(self, operation):
+        expected_head, expected_card = self.publish_merged_task()
+        self.git(self.repo, "fetch", "origin")
+        old_head = self.git(self.repo, "rev-parse", "HEAD").stdout.strip()
+        local_heads = self.git(self.repo, "show-ref", "--heads").stdout
+        worktrees = self.git(self.repo, "worktree", "list", "--porcelain").stdout
+        remote_refs = self.git(self.remote, "show-ref").stdout
+        self.install_post_merge_hook()
+        result = self.git(self.repo, *operation)
+        self.assertIn("make task-close", result.stdout)
+        self.assertIn("授权", result.stdout)
+        self.assertNotIn("[close]", result.stdout)
+        self.assertEqual(self.git(self.repo, "branch", "--show-current").stdout.strip(), "main")
+        self.assertEqual(self.git(self.repo, "rev-parse", "HEAD").stdout.strip(), expected_head)
+        self.assertEqual(self.git(self.repo, "show-ref", "--heads").stdout,
+                         local_heads.replace(old_head, expected_head))
+        self.assertEqual(self.git(self.repo, "worktree", "list", "--porcelain").stdout,
+                         worktrees.replace(old_head, expected_head))
+        self.assertEqual(self.git(self.remote, "show-ref").stdout, remote_refs)
+        self.assertEqual(self.git(self.repo, "status", "--porcelain").stdout, "")
+        self.assertEqual((self.repo / ".task/feature-merged.md").read_text(), expected_card)
+        self.assertFalse((self.repo / ".task/archive").exists())
+
+    def test_post_merge_main_merge_only_reminds_owner(self):
+        self.assert_post_merge_only_reminds_owner(("merge", "--ff-only", "origin/main"))
+
+    def test_post_merge_main_pull_only_reminds_owner(self):
+        self.assert_post_merge_only_reminds_owner(("pull", "--ff-only"))
+
+    def test_post_merge_is_silent_outside_main_checkout(self):
+        wt = self.worktree()
+        self.install_post_merge_hook()
+        self.advance_remote()
+        result = self.git(wt, "merge", "--ff-only", "origin/main")
+        self.assertNotIn("make task-close", result.stdout)
+        self.assertNotIn("[close]", result.stdout)
+        self.git(self.repo, "switch", "-c", "feature/primary")
+        self.git(wt, "switch", "-c", "feature/advance")
+        (wt / "feature-work").write_text("another task\n")
+        self.git(wt, "add", "feature-work")
+        self.git(wt, "commit", "-m", "Advance feature")
+        result = self.git(self.repo, "merge", "--ff-only", "feature/advance")
+        self.assertNotIn("make task-close", result.stdout)
+        self.assertNotIn("[close]", result.stdout)
+
     def test_start_stacked_branch_records_parent_and_main_baseline(self):
         base = self.git(self.repo, "rev-parse", "origin/main").stdout.strip()
         parent = self.worktree("feature/parent")