@godv61/dsh-task-engine 0.24.0 → 0.26.0

package/docs/testing/0.23.2/R03/344/270/232/345/212/241/346/265/213/350/257/225/346/230/216/347/273/206.md

@@ -1,38 +1,38 @@
-# 自动化测试逐项记录
-
-来源：r03-business-tests.log；共 32 项。
-
-| ID | 用例 | 结果 | 耗时 ms |
-| --- | --- | --- | ---: |
-| 1 | 单元 aggregate：同批次累计、码元升序并输出 defectRate | 通过 | 1.695 |
-| 2 | 单元 aggregate：空数组得到空 batches 与全零 totals | 通过 | 0.2052 |
-| 3 | 单元 aggregate：inspected 为 0 时 defectRate 为 0 | 通过 | 0.1467 |
-| 4 | 单元 aggregate：defectRate 至多两位小数且为数值类型 | 通过 | 0.2306 |
-| 5 | 单元 aggregate：totals.defectRate 基于累计值而非各批次比率平均 | 通过 | 0.111 |
-| 6 | 单元 aggregate：按 batchId 字符串码元升序（大写先于小写） | 通过 | 0.0989 |
-| 7 | 单元 validateRecords：batchId 去除前后空格后作为同一批次 | 通过 | 0.1417 |
-| 8 | 单元 validateRecords：接受非负整数与零值边界 | 通过 | 0.1628 |
-| 9 | 单元 validateRecords：拒绝非数组、非法元素与非法字段 | 通过 | 0.683 |
-| 10 | AC1 同批次累计，stdout 仅一行 JSON，totals 正确 | 通过 | 74.2086 |
-| 11 | AC2 空数组合法 | 通过 | 75.882 |
-| 12 | AC3 零数合法 | 通过 | 85.2254 |
-| 13 | AC4 batchId 去空格归并并按码元升序 | 通过 | 82.9206 |
-| 14 | AC5 非法 JSON 失败且 stdout 为空 | 通过 | 90.9043 |
-| 15 | AC6 顶层非数组失败且 stdout 为空 | 通过 | 379.1634 |
-| 16 | AC7 batchId 为空字符串 → 非零退出、stdout 为空、stderr 具体 | 通过 | 80.0845 |
-| 17 | AC7 batchId 为纯空白 → 非零退出、stdout 为空、stderr 具体 | 通过 | 74.659 |
-| 18 | AC7 batchId 缺失 → 非零退出、stdout 为空、stderr 具体 | 通过 | 78.5134 |
-| 19 | AC8 inspected 为负数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 121.0524 |
-| 20 | AC8 defective 为负数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 77.35 |
-| 21 | AC9 inspected 为小数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 66.7267 |
-| 22 | AC9 defective 为小数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 68.4103 |
-| 23 | AC10 inspected 为字符串数字 → 非零退出、stdout 为空、stderr 具体 | 通过 | 71.1564 |
-| 24 | AC10 defective 为字符串数字 → 非零退出、stdout 为空、stderr 具体 | 通过 | 71.2365 |
-| 25 | AC11 defective 大于 inspected → 非零退出、stdout 为空、stderr 具体 | 通过 | 76.4629 |
-| 26 | 非对象元素 → 非零退出、stdout 为空、stderr 具体 | 通过 | 72.6897 |
-| 27 | AC12 无参数失败并提示用法 | 通过 | 75.6079 |
-| 28 | 多余参数失败并提示用法 | 通过 | 73.248 |
-| 29 | AC13 文件不存在失败且 stdout 为空 | 通过 | 71.4465 |
-| 30 | AC14 --help 退出 0 并输出用法 | 通过 | 141.8224 |
-| 31 | AC14 --help 优先于其他参数 | 通过 | 70.1995 |
-| 32 | AC15 examples/batches.json 可直接运行并得到文档所述结果 | 通过 | 73.5514 |
+# 自动化测试逐项记录
+
+来源：r03-business-tests.log；共 32 项。
+
+| ID | 用例 | 结果 | 耗时 ms |
+| --- | --- | --- | ---: |
+| 1 | 单元 aggregate：同批次累计、码元升序并输出 defectRate | 通过 | 1.695 |
+| 2 | 单元 aggregate：空数组得到空 batches 与全零 totals | 通过 | 0.2052 |
+| 3 | 单元 aggregate：inspected 为 0 时 defectRate 为 0 | 通过 | 0.1467 |
+| 4 | 单元 aggregate：defectRate 至多两位小数且为数值类型 | 通过 | 0.2306 |
+| 5 | 单元 aggregate：totals.defectRate 基于累计值而非各批次比率平均 | 通过 | 0.111 |
+| 6 | 单元 aggregate：按 batchId 字符串码元升序（大写先于小写） | 通过 | 0.0989 |
+| 7 | 单元 validateRecords：batchId 去除前后空格后作为同一批次 | 通过 | 0.1417 |
+| 8 | 单元 validateRecords：接受非负整数与零值边界 | 通过 | 0.1628 |
+| 9 | 单元 validateRecords：拒绝非数组、非法元素与非法字段 | 通过 | 0.683 |
+| 10 | AC1 同批次累计，stdout 仅一行 JSON，totals 正确 | 通过 | 74.2086 |
+| 11 | AC2 空数组合法 | 通过 | 75.882 |
+| 12 | AC3 零数合法 | 通过 | 85.2254 |
+| 13 | AC4 batchId 去空格归并并按码元升序 | 通过 | 82.9206 |
+| 14 | AC5 非法 JSON 失败且 stdout 为空 | 通过 | 90.9043 |
+| 15 | AC6 顶层非数组失败且 stdout 为空 | 通过 | 379.1634 |
+| 16 | AC7 batchId 为空字符串 → 非零退出、stdout 为空、stderr 具体 | 通过 | 80.0845 |
+| 17 | AC7 batchId 为纯空白 → 非零退出、stdout 为空、stderr 具体 | 通过 | 74.659 |
+| 18 | AC7 batchId 缺失 → 非零退出、stdout 为空、stderr 具体 | 通过 | 78.5134 |
+| 19 | AC8 inspected 为负数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 121.0524 |
+| 20 | AC8 defective 为负数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 77.35 |
+| 21 | AC9 inspected 为小数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 66.7267 |
+| 22 | AC9 defective 为小数 → 非零退出、stdout 为空、stderr 具体 | 通过 | 68.4103 |
+| 23 | AC10 inspected 为字符串数字 → 非零退出、stdout 为空、stderr 具体 | 通过 | 71.1564 |
+| 24 | AC10 defective 为字符串数字 → 非零退出、stdout 为空、stderr 具体 | 通过 | 71.2365 |
+| 25 | AC11 defective 大于 inspected → 非零退出、stdout 为空、stderr 具体 | 通过 | 76.4629 |
+| 26 | 非对象元素 → 非零退出、stdout 为空、stderr 具体 | 通过 | 72.6897 |
+| 27 | AC12 无参数失败并提示用法 | 通过 | 75.6079 |
+| 28 | 多余参数失败并提示用法 | 通过 | 73.248 |
+| 29 | AC13 文件不存在失败且 stdout 为空 | 通过 | 71.4465 |
+| 30 | AC14 --help 退出 0 并输出用法 | 通过 | 141.8224 |
+| 31 | AC14 --help 优先于其他参数 | 通过 | 70.1995 |
+| 32 | AC15 examples/batches.json 可直接运行并得到文档所述结果 | 通过 | 73.5514 |

package/docs/testing/0.23.2//346/265/213/350/257/225/346/212/245/345/221/212.md

@@ -1,65 +1,65 @@
-# 0.23.2 实机循环回归报告
-
-[← 发布说明](../../releases/0.23.2.md)
-
-日期：2026-09-17。结论：权限缺陷修复后，两个新会话从同一空业务基线完成标准研发流程、独立验收、本地提交和归档回退；门禁专项通过。验证对象是本机 Harness 上的独立 CLI 项目，不是 QMS 页面、数据库或生产业务。
-
-## 环境与方法
-
-Harness 提交 `d347e703908d0406b7a7ef80e3a0e594d86b2215`，Windows、Node 22.22.3、web profile，模型 `deepseek-v4.1-flash`，工程化开发引擎预设。独立仓库 `D:\dshtest`，分支 `codex/plugin-live-regression`，基线 `aae3bf8994ef803cf8370f5f53840b508539499c` 仅有 `.gitignore`。
-
-冻结需求为零依赖 ESM 质检批次统计 CLI：严格校验 JSON 记录，按 trim 后的 batchId 累计并按码元排序，计算百分数缺陷率，正确区分成功 JSON 与失败 stderr，提供单元测试、真实 CLI 子进程测试、样例及 README。所有业务代码由 DSH 编写；观察者只审阅、执行独立验收、归档及修复插件。每轮使用新会话，不读取前轮实现；QMS 未改动。
-
-## 轮次结果
-
-| 轮次 | 插件 | 业务测试 / 独立黑盒 | 流程结果 | 本地提交 |
-| --- | --- | --- | --- | --- |
-| R01 | 0.23.1 | 24 / 19 通过 | 插件验证权限缺陷阻塞交付；停止、归档，未冒充完成 | 无 |
-| R02 | 0.23.2-rc.1 | 50 / 19 通过 | 审批、3项实施审核、验证、代码审核、提交回写、完成、归档回退通过 | `15ffd9c230363520a934586107ce9f4023e828ed` |
-| R03 | 0.23.2-rc.3 | 32 / 19 通过 | 设计退回修订、4项实施审核、验证、代码审核、提交回写、完成、归档回退通过 | `4e476a514d3109417bac602e818460b09dd57289` |
-
-R02 与 R03 的权限核心构建文件 SHA256 均为 `FD116354D20D870E8981B4C098E9F0BDBE9B213397918D7181FE07E0E1713688`；期间只修正提交规则及文档。rc.2 安装核对时发现新指南误称 status 返回格式字段，已改为实际的 `commit.label` 并重打 rc.3，rc.2 不计作业务轮次。
-
-R02 为27项单元、23项CLI子进程测试；R03 为9项单元、23项CLI子进程测试。DSH 的 R03 最终文字误写8+24，本报告按实际测试文件与TAP用例核对为9+23。逐项结果和耗时见[R02明细](R02业务测试明细.md)、[R03明细](R03业务测试明细.md)。独立黑盒19项覆盖主用例、空数组、零值、trim与取整、4类非法顶层输入、7类非法字段、非法JSON、缺参数、缺文件和帮助。
-
-## 门禁专项
-
-| 场景 | 实际结果 | 结论 |
-| --- | --- | --- |
-| 未加载绑定技能即推进 | 返回 load skill requirement-analysis，未推进 | 通过 |
-| 缺少需求产物与确认即推进 | 返回 requirement_confirmation / artifacts_present 未满足 | 通过 |
-| 使用合法消息提前提交 | 返回当前阶段无提交检查点，revision仍为1 | 通过 |
-| 普通权限运行 node --test | 真实exit1，子进程spawn EPERM，回执mode=workspace-write | 按预期失败 |
-| 拒绝该命令的权限升级 | 工具报rejected，台账逐字节哈希不变 | 通过 |
-| 明确重新授权后批准原命令 | 50项测试通过，exit0，回执mode=danger-full-access | 通过 |
-| 修改唯一范围文件 | status出现验证过期的evidence_blockers，冻结旧回执保留 | 通过 |
-| 同会话后续普通shell调用 | 原命令重新EPERM，未继承单次权限 | 通过 |
-
-拒绝前后台账 SHA256 均为 `99F13BB166B4D0C9D7EC4EAE00F0471E7CEABD97541ADD85820A64A0D2008306`。专项任务 `DSH-GATE-PROBE` 停在需求阶段，不计作完整业务轮次。过期提交的检查点路径由自动化回归覆盖；本次专项实机验证到动态过期提示，未声称在该检查点另外执行过过期提交。
-
-R03完整流程另通过正式 `dev_task verify` 执行 `npm test`：单次审批后32项通过、exit0、实际模式danger-full-access。R02使用可在受限环境运行的真实文件描述符子进程测试，其workspace-write成功回执不冒充权限修复证据。R03修正后的提交消息首次预检即获批准，随后核对真实Git提交并登记。
-
-## 插件回归与缺陷
-
-P1：0.23.1在执行验证命令之后才审批写台账，导致命令先运行且未收到批准模式。回归测试先复现“审批服务不可用却已调用一次shell”，修复后拒绝、取消、审批不可用、非法参数均不运行命令或改台账；verify与skill_result批准顺序、请求上下文、单次权限和回执写入不重复审批均受断言保护。
-
-P2：随包提交规则要求模块名，而引擎要求任务id，R02实际出现两次格式拒绝。规则及手册改为当前任务id和status.commit.label；R03未再出现该格式错误。
-
-源码构建、host/client类型检查、126项P0和37项node:test回归通过；候选包干净安装、入口导入、客户端注册、包内测试及CLI语法验证通过。源码测试明细见[自动化测试明细](自动化测试明细.md)。最终0.23.2的 `npm run verify:package` 由用户在插件源码目录执行并确认通过；此项依据用户执行反馈，观察者没有独立保存该次终端日志。发布版本与远端提交另外核对，不以本报告代替发布成功回执。
-
-## 归档与干预
-
-两个完成轮次的父提交均为B0，文件分别严格限制为10个、5个；归档前工作区干净，任务均为完成、review=pass，最终登记hash等于HEAD。业务ZIP、任务JSON、会话日志及独立验收记录保存在本地 `live-project-test/regression-20260917`；备份分支为 `codex/archive-dsh-reg-r02` 和 `codex/archive-dsh-reg-r03`。业务分支最终恢复B0，保留被忽略的历史台账。
-
-| 归档 | 业务ZIP SHA256 |
-| --- | --- |
-| R02 | `2545B54A201D70D536B85BDB52F2356CE2364310FCAF32E99F759408BFC98D82` |
-| R03 | `F7BB52228137BA0A37722097A26AF6BBE1AEFB09C416792F4EE94EA29E71DD56` |
-
-观察者处理了正式需求/设计/权限审批；R03退回已知Node22测试目录入口问题，并要求DSH通过正式审批区分EPERM与入口错误，保留默认测试隔离和真实子进程。观察者未代写业务代码。测试包含这些干预，不能称为完全自主执行。
-
-## 能力边界与未覆盖范围
-
-`dev_task`对阶段、证据和提交登记实施门禁，通用文件写入不会逐次经过状态机；R03实际在实施项仍为todo时写入文件，随后才逐项审核和标记done。没有安装提交钩子时，不能据此宣称插件能阻止直接shell绕过提交预检。本轮未测试强行篡改台账、直接Git绕过、多人并发、QMS集成、全部Harness/模型版本或生产部署。
-
-scope_hash只覆盖任务files的内容，不能证明命令或范围外业务文件未变；观察者已纠正DSH一次过度解读。模型报告也可能出现测试分类计数错误，必须复核实际日志。自动化中的模拟宿主与真实DSH会话分开计证，不互相替代。
+# 0.23.2 实机循环回归报告
+
+[← 发布说明](../../releases/0.23.2.md)
+
+日期：2026-09-17。结论：权限缺陷修复后，两个新会话从同一空业务基线完成标准研发流程、独立验收、本地提交和归档回退；门禁专项通过。验证对象是本机 Harness 上的独立 CLI 项目，不是 QMS 页面、数据库或生产业务。
+
+## 环境与方法
+
+Harness 提交 `d347e703908d0406b7a7ef80e3a0e594d86b2215`，Windows、Node 22.22.3、web profile，模型 `deepseek-v4.1-flash`，工程化开发引擎预设。独立仓库 `D:\dshtest`，分支 `codex/plugin-live-regression`，基线 `aae3bf8994ef803cf8370f5f53840b508539499c` 仅有 `.gitignore`。
+
+冻结需求为零依赖 ESM 质检批次统计 CLI：严格校验 JSON 记录，按 trim 后的 batchId 累计并按码元排序，计算百分数缺陷率，正确区分成功 JSON 与失败 stderr，提供单元测试、真实 CLI 子进程测试、样例及 README。所有业务代码由 DSH 编写；观察者只审阅、执行独立验收、归档及修复插件。每轮使用新会话，不读取前轮实现；QMS 未改动。
+
+## 轮次结果
+
+| 轮次 | 插件 | 业务测试 / 独立黑盒 | 流程结果 | 本地提交 |
+| --- | --- | --- | --- | --- |
+| R01 | 0.23.1 | 24 / 19 通过 | 插件验证权限缺陷阻塞交付；停止、归档，未冒充完成 | 无 |
+| R02 | 0.23.2-rc.1 | 50 / 19 通过 | 审批、3项实施审核、验证、代码审核、提交回写、完成、归档回退通过 | `15ffd9c230363520a934586107ce9f4023e828ed` |
+| R03 | 0.23.2-rc.3 | 32 / 19 通过 | 设计退回修订、4项实施审核、验证、代码审核、提交回写、完成、归档回退通过 | `4e476a514d3109417bac602e818460b09dd57289` |
+
+R02 与 R03 的权限核心构建文件 SHA256 均为 `FD116354D20D870E8981B4C098E9F0BDBE9B213397918D7181FE07E0E1713688`；期间只修正提交规则及文档。rc.2 安装核对时发现新指南误称 status 返回格式字段，已改为实际的 `commit.label` 并重打 rc.3，rc.2 不计作业务轮次。
+
+R02 为27项单元、23项CLI子进程测试；R03 为9项单元、23项CLI子进程测试。DSH 的 R03 最终文字误写8+24，本报告按实际测试文件与TAP用例核对为9+23。逐项结果和耗时见[R02明细](R02业务测试明细.md)、[R03明细](R03业务测试明细.md)。独立黑盒19项覆盖主用例、空数组、零值、trim与取整、4类非法顶层输入、7类非法字段、非法JSON、缺参数、缺文件和帮助。
+
+## 门禁专项
+
+| 场景 | 实际结果 | 结论 |
+| --- | --- | --- |
+| 未加载绑定技能即推进 | 返回 load skill requirement-analysis，未推进 | 通过 |
+| 缺少需求产物与确认即推进 | 返回 requirement_confirmation / artifacts_present 未满足 | 通过 |
+| 使用合法消息提前提交 | 返回当前阶段无提交检查点，revision仍为1 | 通过 |
+| 普通权限运行 node --test | 真实exit1，子进程spawn EPERM，回执mode=workspace-write | 按预期失败 |
+| 拒绝该命令的权限升级 | 工具报rejected，台账逐字节哈希不变 | 通过 |
+| 明确重新授权后批准原命令 | 50项测试通过，exit0，回执mode=danger-full-access | 通过 |
+| 修改唯一范围文件 | status出现验证过期的evidence_blockers，冻结旧回执保留 | 通过 |
+| 同会话后续普通shell调用 | 原命令重新EPERM，未继承单次权限 | 通过 |
+
+拒绝前后台账 SHA256 均为 `99F13BB166B4D0C9D7EC4EAE00F0471E7CEABD97541ADD85820A64A0D2008306`。专项任务 `DSH-GATE-PROBE` 停在需求阶段，不计作完整业务轮次。过期提交的检查点路径由自动化回归覆盖；本次专项实机验证到动态过期提示，未声称在该检查点另外执行过过期提交。
+
+R03完整流程另通过正式 `dev_task verify` 执行 `npm test`：单次审批后32项通过、exit0、实际模式danger-full-access。R02使用可在受限环境运行的真实文件描述符子进程测试，其workspace-write成功回执不冒充权限修复证据。R03修正后的提交消息首次预检即获批准，随后核对真实Git提交并登记。
+
+## 插件回归与缺陷
+
+P1：0.23.1在执行验证命令之后才审批写台账，导致命令先运行且未收到批准模式。回归测试先复现“审批服务不可用却已调用一次shell”，修复后拒绝、取消、审批不可用、非法参数均不运行命令或改台账；verify与skill_result批准顺序、请求上下文、单次权限和回执写入不重复审批均受断言保护。
+
+P2：随包提交规则要求模块名，而引擎要求任务id，R02实际出现两次格式拒绝。规则及手册改为当前任务id和status.commit.label；R03未再出现该格式错误。
+
+源码构建、host/client类型检查、126项P0和37项node:test回归通过；候选包干净安装、入口导入、客户端注册、包内测试及CLI语法验证通过。源码测试明细见[自动化测试明细](自动化测试明细.md)。最终0.23.2的 `npm run verify:package` 由用户在插件源码目录执行并确认通过；此项依据用户执行反馈，观察者没有独立保存该次终端日志。发布版本与远端提交另外核对，不以本报告代替发布成功回执。
+
+## 归档与干预
+
+两个完成轮次的父提交均为B0，文件分别严格限制为10个、5个；归档前工作区干净，任务均为完成、review=pass，最终登记hash等于HEAD。业务ZIP、任务JSON、会话日志及独立验收记录保存在本地 `live-project-test/regression-20260917`；备份分支为 `codex/archive-dsh-reg-r02` 和 `codex/archive-dsh-reg-r03`。业务分支最终恢复B0，保留被忽略的历史台账。
+
+| 归档 | 业务ZIP SHA256 |
+| --- | --- |
+| R02 | `2545B54A201D70D536B85BDB52F2356CE2364310FCAF32E99F759408BFC98D82` |
+| R03 | `F7BB52228137BA0A37722097A26AF6BBE1AEFB09C416792F4EE94EA29E71DD56` |
+
+观察者处理了正式需求/设计/权限审批；R03退回已知Node22测试目录入口问题，并要求DSH通过正式审批区分EPERM与入口错误，保留默认测试隔离和真实子进程。观察者未代写业务代码。测试包含这些干预，不能称为完全自主执行。
+
+## 能力边界与未覆盖范围
+
+`dev_task`对阶段、证据和提交登记实施门禁，通用文件写入不会逐次经过状态机；R03实际在实施项仍为todo时写入文件，随后才逐项审核和标记done。没有安装提交钩子时，不能据此宣称插件能阻止直接shell绕过提交预检。本轮未测试强行篡改台账、直接Git绕过、多人并发、QMS集成、全部Harness/模型版本或生产部署。
+
+scope_hash只覆盖任务files的内容，不能证明命令或范围外业务文件未变；观察者已纠正DSH一次过度解读。模型报告也可能出现测试分类计数错误，必须复核实际日志。自动化中的模拟宿主与真实DSH会话分开计证，不互相替代。

package/hooks/commit-msg

@@ -172,19 +172,23 @@ var STANDARD = {
     { from: "\u4EA4\u4ED8", to: "\u4EE3\u7801\u5BA1\u6838", requires: ["verified"] },
     { from: "\u4EE3\u7801\u5BA1\u6838", to: "\u5B8C\u6210", requires: ["review_passed", "artifacts_present"] }
   ],
+  // No artifacts, no commit policy, no skills: the skeleton carries none of them.
+  artifacts: [],
+  commit: { policy: "task", message_pattern: "", message_hint: "\u65E0\u683C\u5F0F\u8981\u6C42", checkpoints: ["\u4EE3\u7801\u5BA1\u6838"], file_scope: true },
+  high_risk_requires_verification: true
+};
+var STANDARD_RECOMMENDATION = {
+  commit: {
+    // Only the TEXT conventions. Policy, checkpoints and scope stay on the
+    // skeleton: they decide when a gate fires, not how a message reads.
+    message_pattern: "^\u3010(\\S+)\u3011\u3010(?:TASK|T\\d+)\u3011.+",
+    message_hint: "\u3010<task_id>\u3011\u3010TASK/T1\u3011\u8BF4\u660E \u2014\u2014 \u7B2C\u4E00\u6BB5\u586B\u672C\u4EFB\u52A1 id\uFF08\u5982 GREET-001\uFF09\uFF0C\u5199\u7ED3\u679C\u4E0D\u5199\u7A7A\u6CDB\u52A8\u4F5C"
+  },
   artifacts: [
     { stage: "\u9700\u6C42\u8BC4\u5BA1", id: "requirement", name: "\u9700\u6C42\u8BF4\u660E", fields: ["scope", "acceptance_criteria"] },
     { stage: "\u8BBE\u8BA1", id: "design", name: "\u8BBE\u8BA1\u6587\u6863", fields: ["approach", "risks", "impact"] },
     { stage: "\u4EE3\u7801\u5BA1\u6838", id: "review", name: "\u8BC4\u5BA1\u8BB0\u5F55", fields: ["conclusion", "issues"] }
   ],
-  commit: {
-    policy: "task",
-    message_pattern: "^\u3010(\\S+)\u3011\u3010(?:TASK|T\\d+)\u3011.+",
-    message_hint: "\u3010<task_id>\u3011\u3010TASK/T1\u3011\u8BF4\u660E \u2014\u2014 \u7B2C\u4E00\u6BB5\u586B\u672C\u4EFB\u52A1 id\uFF08\u5982 GREET-001\uFF09\uFF0C\u5199\u7ED3\u679C\u4E0D\u5199\u7A7A\u6CDB\u52A8\u4F5C",
-    checkpoints: ["\u4EE3\u7801\u5BA1\u6838"],
-    file_scope: true
-  },
-  high_risk_requires_verification: true,
   stage_bindings: {
     "\u9700\u6C42\u8BC4\u5BA1": { skills: [bundled("requirement-analysis", "security-redlines")] },
     "\u8BBE\u8BA1": { skills: [bundled("solution-design")] },
@@ -204,30 +208,38 @@ var AGILE = {
   transitions: [
     { from: "\u9700\u6C42", to: "\u5F00\u53D1", requires: ["requirement_confirmation"] },
     { from: "\u5F00\u53D1", to: "\u4EA4\u4ED8", requires: ["todos_done"] },
+    // Entering 审查 is unconditional. A review guard on this edge would be
+    // circular — it would demand the verdict BEFORE the stage that produces it —
+    // and it would also block the commit at 交付, since a commit requires its
+    // stage's outgoing guards. What 审查 requires is declared as completion_guards
+    // instead, because a terminal stage has no outgoing edge to carry it.
     { from: "\u4EA4\u4ED8", to: "\u5BA1\u67E5", requires: [] }
   ],
+  artifacts: [],
+  commit: { policy: "item", message_pattern: "", message_hint: "\u65E0\u683C\u5F0F\u8981\u6C42", checkpoints: ["\u4EA4\u4ED8"], file_scope: true },
+  high_risk_requires_verification: false,
+  // 审查 is terminal and IS the review, so there is no outgoing edge to carry
+  // the guard. Declaring it here is what makes a blocked review prevent
+  // completion; without it the stage could be reached and finished with the
+  // review failing, which is what the assessment found.
+  completion_guards: ["review_passed"]
+};
+var AGILE_RECOMMENDATION = {
+  commit: {
+    // `item` policy emits two label shapes: `T<n>` mid-flow and `TASK` at the
+    // closing checkpoint. The pattern accepts BOTH, so the label the engine hands
+    // the model always satisfies the rule that validates it.
+    message_pattern: "^\u3010(\\S+)\u3011\u3010(?:TASK|T\\d+)\u3011.+",
+    message_hint: "\u3010<task_id>\u3011\u3010TASK/T1\u3011\u8BF4\u660E \u2014\u2014 \u7B2C\u4E00\u6BB5\u586B\u672C\u4EFB\u52A1 id\uFF1B\u5B9E\u65BD\u9879\u63D0\u4EA4\u7528 T1\u3001T2\uFF0C\u6536\u5C3E\u63D0\u4EA4\u7528 TASK"
+  },
   artifacts: [
     { stage: "\u9700\u6C42", id: "requirement", name: "\u9700\u6C42\u8BF4\u660E", fields: ["scope"] },
     // The 审查 stage binds code-review, whose instructions write this artifact with
-    // `record artifact=review`. Declaring it here keeps that binding executable:
-    // the engine refuses an undeclared artifact, so a bound skill pointing at one
-    // would be a contract the preset itself made impossible to satisfy.
+    // `record artifact=review`. Declaring it keeps that binding executable: the
+    // engine refuses an undeclared artifact, so a bound skill pointing at one would
+    // be a contract the recommendation made impossible to satisfy.
     { stage: "\u5BA1\u67E5", id: "review", name: "\u8BC4\u5BA1\u8BB0\u5F55", fields: ["conclusion", "issues"] }
   ],
-  commit: {
-    // `item` policy emits two label shapes: `T<n>` for a mid-flow item commit and
-    // `TASK` for the closing commit at the checkpoint. The pattern below accepts
-    // BOTH, so the label the engine hands the model always satisfies the preset's
-    // own rule. Previously it accepted only `T\d+` while the checkpoint stage
-    // returned `TASK`, which made the closing commit impossible to phrase: the
-    // status line and the validator disagreed about the contract.
-    policy: "item",
-    message_pattern: "^\u3010(\\S+)\u3011\u3010(?:TASK|T\\d+)\u3011.+",
-    message_hint: "\u3010<task_id>\u3011\u3010TASK/T1\u3011\u8BF4\u660E \u2014\u2014 \u7B2C\u4E00\u6BB5\u586B\u672C\u4EFB\u52A1 id\uFF1B\u5B9E\u65BD\u9879\u63D0\u4EA4\u7528 T1\u3001T2\uFF0C\u6536\u5C3E\u63D0\u4EA4\u7528 TASK",
-    checkpoints: ["\u4EA4\u4ED8"],
-    file_scope: true
-  },
-  high_risk_requires_verification: false,
   stage_bindings: {
     "\u9700\u6C42": { skills: [bundled("requirement-analysis", "security-redlines")] },
     "\u5F00\u53D1": { skills: [bundled("code-implement", "coding-conventions", "security-redlines")] },
@@ -242,18 +254,13 @@ var MINIMAL = {
     { from: "\u5F00\u53D1", to: "\u4EA4\u4ED8", requires: ["todos_done"] }
   ],
   artifacts: [],
-  commit: {
-    policy: "task",
-    message_pattern: "",
-    message_hint: "\u65E0\u683C\u5F0F\u8981\u6C42",
-    checkpoints: ["\u4EA4\u4ED8"],
-    file_scope: false
-  },
-  high_risk_requires_verification: false,
+  commit: { policy: "task", message_pattern: "", message_hint: "\u65E0\u683C\u5F0F\u8981\u6C42", checkpoints: ["\u4EA4\u4ED8"], file_scope: false },
+  high_risk_requires_verification: false
+};
+var MINIMAL_RECOMMENDATION = {
   // A fast-change flow checks each change once, not twice under two headings a
   // short change rarely distinguishes. The item still has to be reviewed — what
-  // drops is the duplicated verdict, which is where the weight actually was: this
-  // flow used to carry the same per-item audit burden as the full one.
+  // drops is the duplicated verdict, which is where the weight actually was.
   review_depth: "single",
   stage_bindings: {
     "\u5F00\u53D1": { skills: [bundled("code-implement", "coding-conventions", "security-redlines")] },
@@ -269,54 +276,126 @@ function deriveCapabilities(config2) {
   if (config2.commit.file_scope) caps.push("file_scope");
   return caps;
 }
-function preset(id, version, label, description, config2) {
-  return { id, version, label, description, capabilities: deriveCapabilities(config2), config: config2 };
+function preset(id, version, label, description, config2, recommendation) {
+  return {
+    id,
+    version,
+    label,
+    description,
+    capabilities: deriveCapabilities(config2),
+    config: config2,
+    // exactOptionalPropertyTypes: omit rather than assign undefined.
+    ...recommendation !== void 0 ? { recommendation } : {}
+  };
 }
 var FLOW_PRESETS = {
-  standard: preset("standard", 2, "\u5B8C\u6574\u7814\u53D1", "\u65B0\u529F\u80FD\u3001\u67B6\u6784\u6216\u8DE8\u6A21\u5757\u6539\u52A8\u3001\u9AD8\u98CE\u9669\u4EFB\u52A1\uFF1A\u9700\u6C42\u786E\u8BA4 \u2192 \u65B9\u6848\u786E\u8BA4 \u2192 \u5B9E\u73B0 \u2192 \u9A8C\u8BC1 \u2192 \u5BA1\u6838 \u2192 \u63D0\u4EA4", STANDARD),
-  // Version 2: the closing-commit label shape was unified with the message pattern,
-  // and the `review` artifact the bound code-review skill writes was declared. Tasks
-  // created before this keep their frozen version 1 config and are unaffected.
-  agile: preset("agile", 2, "\u65E5\u5E38\u8FED\u4EE3", "\u76EE\u6807\u660E\u786E\u7684\u5E38\u89C4\u529F\u80FD\u4E0E\u7F3A\u9677\u4FEE\u590D\uFF1A\u76EE\u6807\u4E0E\u9A8C\u6536 \u2192 \u5B9E\u73B0 \u2192 \u9A8C\u6536\u4E0E\u5BA1\u67E5 \u2192 \u63D0\u4EA4", AGILE),
-  // Version 2: per-item review depth is declared as `single` instead of inheriting
-  // the full flow's two verdicts per item. Tasks created before this keep their
-  // frozen version 1 config, which has no `review_depth` and therefore reads as
-  // two-stage exactly as it behaved.
-  minimal: preset("minimal", 2, "\u5FEB\u901F\u4FEE\u6539", "\u5C40\u90E8\u3001\u4F4E\u98CE\u9669\u3001\u65B9\u6848\u660E\u786E\u7684\u6539\u52A8\uFF1A\u4FEE\u6539 \u2192 \u68C0\u67E5\u4E0E\u63D0\u4EA4\uFF0C\u6BCF\u9879\u4E00\u6B21\u68C0\u67E5", MINIMAL)
+  standard: preset("standard", 3, "\u5B8C\u6574\u7814\u53D1", "\u65B0\u529F\u80FD\u3001\u67B6\u6784\u6216\u8DE8\u6A21\u5757\u6539\u52A8\u3001\u9AD8\u98CE\u9669\u4EFB\u52A1\uFF1A\u9700\u6C42\u786E\u8BA4 \u2192 \u65B9\u6848\u786E\u8BA4 \u2192 \u5B9E\u73B0 \u2192 \u9A8C\u8BC1 \u2192 \u5BA1\u6838 \u2192 \u63D0\u4EA4", STANDARD, STANDARD_RECOMMENDATION),
+  // Version 4 for agile: 交付 → 审查 now requires review_passed, where it previously
+  // required nothing. That is a gate-semantics change, so a task created before it
+  // keeps its frozen version 3 config and behaves exactly as it did.
+  //
+  // Version 3 across all three: the preset is now a bare skeleton and its former
+  // built-in bindings, commit rule and artifacts are a separate recommendation a
+  // user adopts explicitly. Tasks created before this keep their frozen config and
+  // are unaffected; a project's existing config is likewise left exactly as it is.
+  agile: preset("agile", 4, "\u65E5\u5E38\u8FED\u4EE3", "\u76EE\u6807\u660E\u786E\u7684\u5E38\u89C4\u529F\u80FD\u4E0E\u7F3A\u9677\u4FEE\u590D\uFF1A\u76EE\u6807\u4E0E\u9A8C\u6536 \u2192 \u5B9E\u73B0 \u2192 \u9A8C\u6536\u4E0E\u5BA1\u67E5 \u2192 \u63D0\u4EA4", AGILE, AGILE_RECOMMENDATION),
+  minimal: preset("minimal", 3, "\u5FEB\u901F\u4FEE\u6539", "\u5C40\u90E8\u3001\u4F4E\u98CE\u9669\u3001\u65B9\u6848\u660E\u786E\u7684\u6539\u52A8\uFF1A\u4FEE\u6539 \u2192 \u68C0\u67E5\u4E0E\u63D0\u4EA4\uFF0C\u6BCF\u9879\u4E00\u6B21\u68C0\u67E5", MINIMAL, MINIMAL_RECOMMENDATION)
 };
 function dedupe(names) {
   return [...new Set(names)];
 }
-function mergeBindings(base, override) {
-  if (override === void 0) return base;
-  let changed = false;
-  const merged = { ...base.stage_bindings };
-  for (const [stage, addition] of Object.entries(override)) {
-    const existing = merged[stage];
-    const skills = [...existing?.skills ?? []];
-    const seen = new Set(skills.map((skill) => formatResourceRef(skill.skill)));
-    for (const skill of addition.skills ?? []) {
-      const key = formatResourceRef(skill.skill);
-      if (seen.has(key)) continue;
-      seen.add(key);
-      skills.push(skill);
+function applyProjectConfig(base, project) {
+  if (project === void 0) return base;
+  const next = { ...base };
+  if (project.stage_bindings !== void 0) {
+    const profiles = { ...project.skill_profiles ?? {} };
+    const explicitProfiles = new Set(Object.keys(project.skill_profiles ?? {}));
+    const problems = [];
+    const bindings = {};
+    for (const [stage, binding] of Object.entries(project.stage_bindings)) {
+      const entry = {};
+      if (binding === null || typeof binding !== "object") {
+        problems.push(`stage_bindings "${stage}" must be an object`);
+        bindings[stage] = entry;
+        continue;
+      }
+      if (binding.skill_refs !== void 0 && binding.skills !== void 0) {
+        problems.push(`stage_bindings "${stage}" has both skill_refs and legacy skills`);
+      }
+      if (binding.skill_refs !== void 0) {
+        if (!Array.isArray(binding.skill_refs)) {
+          problems.push(`stage_bindings "${stage}" skill_refs must be an array`);
+        }
+        entry.skills = (Array.isArray(binding.skill_refs) ? binding.skill_refs : []).flatMap((ref) => {
+          if (ref === null || typeof ref !== "object") {
+            problems.push(`stage_bindings "${stage}" has an invalid skill reference`);
+            return [];
+          }
+          const key = formatResourceRef(ref);
+          const profile = profiles[key];
+          if (profile === void 0) {
+            problems.push(`skill profile "${key}" is missing`);
+            return [];
+          }
+          if (profile === null || !Array.isArray(profile.rules)) {
+            problems.push(`skill profile "${key}" must declare a rules array`);
+            return [];
+          }
+          return [{ skill: ref, rules: profile.rules, ...profile.evidence !== void 0 ? { evidence: profile.evidence } : {} }];
+        });
+      } else if (binding.skills !== void 0) {
+        if (!Array.isArray(binding.skills)) {
+          problems.push(`stage_bindings "${stage}" skills must be an array`);
+        }
+        entry.skills = (Array.isArray(binding.skills) ? binding.skills : []).flatMap((skill) => {
+          if (skill === null || typeof skill !== "object" || skill.skill === null || typeof skill.skill !== "object") {
+            problems.push(`stage_bindings "${stage}" has an invalid skill binding`);
+            return [];
+          }
+          const key = formatResourceRef(skill.skill);
+          if (explicitProfiles.has(key)) {
+            const profile = profiles[key];
+            if (profile === null || !Array.isArray(profile?.rules)) {
+              problems.push(`skill profile "${key}" must declare a rules array`);
+              return [];
+            }
+            return [{
+              skill: skill.skill,
+              rules: profile.rules,
+              ...profile.evidence !== void 0 ? { evidence: profile.evidence } : {}
+            }];
+          }
+          const candidate = { rules: skill.rules ?? [], ...skill.evidence !== void 0 ? { evidence: skill.evidence } : {} };
+          const existing = profiles[key];
+          if (existing !== void 0 && JSON.stringify(existing) !== JSON.stringify(candidate)) {
+            problems.push(`skill "${key}" has different rules or evidence in multiple stages; copy it as a distinct skill`);
+          } else profiles[key] = candidate;
+          return [skill];
+        });
+      }
+      const legacy = dedupe(binding.legacy_rules ?? []);
+      if (legacy.length > 0) entry.legacy_rules = legacy;
+      bindings[stage] = entry;
     }
-    const legacy = dedupe([...existing?.legacy_rules ?? [], ...addition.legacy_rules ?? []]);
-    const binding = {};
-    if (skills.length > 0) binding.skills = skills;
-    if (legacy.length > 0) binding.legacy_rules = legacy;
-    merged[stage] = binding;
-    changed = true;
+    next.stage_bindings = bindings;
+    next.skill_profiles = profiles;
+    if (problems.length > 0) next.configuration_errors = problems;
+  } else if (project.skill_profiles !== void 0) {
+    next.skill_profiles = project.skill_profiles;
   }
-  if (!changed) return base;
-  return { ...base, stage_bindings: merged };
+  if (project.commit !== void 0) next.commit = project.commit;
+  if (project.artifacts !== void 0) next.artifacts = project.artifacts;
+  if (project.review_depth !== void 0) next.review_depth = project.review_depth;
+  if (project.commit_required !== void 0) next.commit_required = project.commit_required;
+  return next;
 }
-function resolveFlow(flow, override) {
+function resolveFlow(flow, project) {
   const preset2 = FLOW_PRESETS[flow];
   if (preset2 === void 0) {
     return { ok: false, code: "UNKNOWN_FLOW", flow, knownFlows: Object.keys(FLOW_PRESETS) };
   }
-  return { ok: true, config: mergeBindings(preset2.config, override?.stage_bindings), preset: preset2 };
+  const { flow: _ignored, ...rest } = project ?? { flow };
+  return { ok: true, config: applyProjectConfig(preset2.config, { flow, ...rest }), preset: preset2 };
 }
 
 // src/snapshot.ts
@@ -471,8 +550,16 @@ function loadWorkflow(parsed) {
   if (typeof flow !== "string" || flow.trim() === "") {
     refuse('\u9879\u76EE .dsh/eng.json \u7F3A "flow" \u5B57\u6BB5\u2014\u2014\u8BBE\u4E3A standard|agile|minimal\uFF0C\u6216\u5220\u9664\u8BE5\u6587\u4EF6');
   }
-  const override = typeof parsed.stage_bindings === "object" && parsed.stage_bindings !== null && !Array.isArray(parsed.stage_bindings) ? { stage_bindings: parsed.stage_bindings } : void 0;
-  const resolved = resolveFlow(flow, override);
+  const project = {
+    flow,
+    ...typeof parsed.stage_bindings === "object" && parsed.stage_bindings !== null && !Array.isArray(parsed.stage_bindings) ? { stage_bindings: parsed.stage_bindings } : {},
+    ...typeof parsed.skill_profiles === "object" && parsed.skill_profiles !== null && !Array.isArray(parsed.skill_profiles) ? { skill_profiles: parsed.skill_profiles } : {},
+    ...typeof parsed.commit === "object" && parsed.commit !== null && !Array.isArray(parsed.commit) ? { commit: parsed.commit } : {},
+    ...Array.isArray(parsed.artifacts) ? { artifacts: parsed.artifacts } : {},
+    ...typeof parsed.review_depth === "string" ? { review_depth: parsed.review_depth } : {},
+    ...typeof parsed.commit_required === "boolean" ? { commit_required: parsed.commit_required } : {}
+  };
+  const resolved = resolveFlow(flow, project);
   if (!resolved.ok) {
     refuse(`\u672A\u77E5\u6D41\u7A0B "${resolved.flow}"\uFF08\u5DF2\u77E5\u6D41\u7A0B\uFF1A${resolved.knownFlows.join("\u3001")}\uFF09`);
   }