@abelxiaoxing/cadence 1.0.14 → 1.1.0

package/prompts/abel-design.md

@@ -1,243 +1,165 @@
 ---
 name: abel-design
-description: 在不写产品代码的前提下，产出经双闸门批准、可直接实施且完全可追溯的 OpenSpec 变更。
+description: Produce a validated, approved, directly executable OpenSpec delivery without implementing product code
 category: abel
-tags: [abel, design, openspec, PBT, subagents, agents]
+tags: [abel, design, openspec, PBT, evidence]
 argument-hint: "<requirement> | --change <change_name>"
 ---
 
+Load the bundled `abel-workflow` Skill and read the complete value inside `<abel-request>` without tokenizing it a second time.
+
 <abel-request>
 $ARGUMENTS
 </abel-request>
 
 <!-- ABEL:PROMPT:abel-design -->
 <!-- ABEL:START -->
-# 不可协商规则（最高优先级）
-
-1. 仅设计：绝不生成或修改产品代码。
-2. 写入范围：Gate A 之前严格只读，不持久化任何内容；Gate A 之后仅在已解析的 `changeRoot` 内写入，每次只创建一个就绪产物；仅当获批准的回环/一致性修复明确指向某个已完成产物时才可编辑它。
-3. 绝不假设或猜测：每个阻塞性决策都必须提交用户（见决策模型）。
-4. **Design 阶段 AGENTS 权限仅在本阶段生效**：仓库 `AGENTS.md` 保持只读，只能审计当前索引是否过期并把影响写入任务契约，本阶段绝不编辑。
-   此只读规则不继承、不适用于 Implement；进入 Implement 后按 Gate B 已批准的结构化 AGENTS 影响契约执行父代理检查点。
-   绝不触碰 OpenSpec 管理的 `openspec/` 生成物（schemas、`config.yaml`、skills 与 slash commands）。
-5. 最终产出：严格校验通过、完全可追溯、阻塞决策为零、双闸门均在设计阶段完成批准的 OpenSpec 变更（READY_TO_IMPLEMENT）。
-
-技能集成：见全局 AGENTS《阶段技能矩阵》Design 列。
-
-## 决策模型
-
-- 在会话内维护决策台账：id、类别（`behavior | technical`）、问题、证据、选项、建议、结论、状态、受影响产物。
-  `BLOCKING_DECISIONS` 为其中未决的非机械决策数。
-- 行为决策回答 WHAT：可观察结果、范围/非目标、场景、失败行为、数据/安全/隐私/兼容策略、成功标准。
-- 技术决策回答 HOW：接口、数据流、依赖、存储/算法、实现错误机制、关键技术参数。
-- 必须由用户批准：目标、范围/非目标与可观察成功行为；数据/安全/隐私/兼容/迁移规则；新依赖、跨模块架构、不可逆变更；任何有实质性权衡的技术选择（含关键参数）。
-- 可由 agent 机械决定：仓库惯例唯一确定的命名/位置/局部结构；易逆且无外部行为变化的细节；由已批准设计直接推出的测试放置与执行顺序。
-- 机械决策记录后不再重问；存在两个以上有实质差异的可行选项 → 升级为阻塞决策。
-- 绝不创建运行时用户/会话状态文件。
-  跨上下文只允许在 `changeRoot` 内生成 `gate-a.yaml` 与 `ready.yaml` 作为可复算的审计凭据；它们不是身份签名或工具权限，不得写入 AGENTS 索引。
-
-## 阶段 0 —— 进入、模式与就绪（只读）
-
-- 从 `<abel-request>` 读取完整需求或 `--change <name>`；不得用 frontmatter hint 或正文示例替代实际参数。
-  参数为空或歧义阻塞时提问。
-- 要求 OpenSpec 根已初始化且 CLI 能力齐全：`new change`、`list --json`、`schemas --json`、`schema which --json`、`schema validate --json`、`templates --json`、`status --json`、`instructions --json`、`validate --strict`。
-  缺失则停止并给出 `/abel-init` 补救，本命令不执行初始化。
-- 要求根 `AGENTS.md` 存在且托管区块结构有效；阅读相关嵌套索引；逐条对照代码、清单和可执行配置核实目标范围内声明。
-  缺失/结构无效则停止并给出 `/abel-init`；过期声明记录进任务影响，本阶段不编辑。
-- 模式解析：显式 `--change` → 恢复，不存在则停止并请用户纠正名称或选择新建；精确匹配已有变更 → 恢复；否则新建。
-  绝不把疑似笔误静默当作新需求。
-- schema 解析优先级：显式选择 > 已有变更 metadata > 项目配置 > `spec-driven`，且须出现在 `openspec schemas --json` 中。
-  创建前运行 `schema which --json`、`schema validate --json` 并检查模板，做行为/技术/混合依赖预检；实施兼容性要求具体非空的 `apply.tracks` 与恰好一个产物的 `generates` 匹配，不兼容则在创建前关闭。
-- 新建模式在探索前的最低输入：问题/目标陈述 + 范围锚点（涉及哪个模块/目录）；缺失则先简要提问。
-- 生成暂定 kebab-case 变更名并用 `openspec list --changes --json` 查重；Gate A 定稿范围后重算并确认。
-  此前不持久化任何内容。
-
-## 阶段 1 —— 证据探索（只读）
-
-- 使用私有工具 `abel_dispatch` 的 `action: "run"` 派发探索包。
-  每个包对应一个独立 tool call，`request` 必须是单个 envelope；禁止使用 `requests` 数组或其他批量 wrapper。
-- 用户要求两个或更多并行包时，下一次 assistant response 必须只包含全部 sibling `abel_dispatch` tool calls，并在该同一 assistant turn（同一响应轮次）一次发齐；少发、先等一个结果、在后续 turn 补发均为调度失败，不得用串行重试或父代理并行读取冒充 Subagent。
-- Design 单包 envelope 的精确形状如下；字段名、固定值和数组形状不得猜测或改写：
-
-  ```json
-  {
-    "action": "run",
-    "request": {
-      "stage": "abel-design",
-      "role": "design-explorer",
-      "id": "unique-packet-id",
-      "phase": "evidence",
-      "objective": "bounded evidence objective for this packet",
-      "roots": ["."],
-      "context": {
-        "agents": "applicable AGENTS.md instructions",
-        "contract": "exact scope, retrieval, read-only, output, and budget contract"
-      },
-      "declared": {
-        "read": ["exact/relative/path"],
-        "write": [],
-        "conflicts": [],
-        "resources": []
-      },
-      "output": "evidence"
-    }
+
+# Design outcome
+
+Design resolves product and technical decisions, collects cited evidence, and compiles one trusted delivery.
+It never launches an implementation Worker, creates a product-code candidate, applies a product diff, runs Red/Green/Refactor, or mutates the main workspace.
+
+If the requirement/change is missing, absent, ambiguous, or not unique, stop before exploration and ask one focused question.
+An explicit `--change <name>` resumes only that existing change; never reinterpret a misspelling as a new requirement.
+
+## Write boundary
+
+- Before explicit Gate A approval, all repository work is read-only.
+  A provisional Design run may exist only in the repository-external private run store.
+- After Gate A, write only inside the canonical `openspec/changes/<change>/` root and only the next schema-ready artifact or a mechanically invalidated receipt.
+- Every repository `AGENTS.md` remains read-only throughout Design.
+  Audit it and seal any needed `none | update-existing | create-index | remove-index` operation into the implementation plan; never edit the index here.
+- Never edit OpenSpec schemas, `openspec/config.yaml`, generated OpenSpec skills/commands, product code, tests, package manifests, or repository state outside the change root.
+
+## Durable Design identity
+
+Use `start` once to create or recover the Design run identity and `status` to inspect it.
+A new requirement may begin with a code-generated SHA-256 `provisionalKey`; after Gate A fixes the change name, a named `start` binds the same provisional run.
+These commands record lifecycle only and do not execute implementation work.
+
+```json
+{"version":2,"command":"start","stage":"abel-design","provisionalKey":"<sha256>","operationId":"<unique-operation>"}
+{"version":2,"command":"start","stage":"abel-design","change":"<change>","provisionalKey":"<same-sha256>","operationId":"<unique-operation>"}
+{"version":2,"command":"status","stage":"abel-design","change":"<change>"}
+```
+
+## Read-only evidence packets
+
+Use `abel_dispatch` with `action: "run"` only for bounded Design evidence.
+Each packet is one tool call, has `stage: "abel-design"`, role `design-explorer`, phase `evidence`, an exact read scope, an empty write set, and output `evidence`.
+Never send an implementation phase to a Design packet.
+
+```json
+{
+  "action":"run",
+  "request":{
+    "stage":"abel-design",
+    "role":"design-explorer",
+    "id":"<unique-packet-id>",
+    "phase":"evidence",
+    "objective":"<bounded evidence question>",
+    "roots":["."],
+    "context":{"agents":"<applicable index context>","contract":"<exact read-only contract>"},
+    "declared":{"read":["<safe relative path>"],"write":[],"conflicts":[],"resources":[]},
+    "output":"evidence"
   }
-  ```
-
-- 主 agent 只负责核实相关 AGENTS、拆分探索包、派发、校验证据、综合、澄清与 Gate 呈现；广泛代码库探索必须委派。
-  主 agent 仅可为范围划分、引用抽查和冲突裁决做最小定向读取。
-- 默认并行派发多个只读探索子代理：优先按独立上下文边界拆分；只有一个边界时，按互不重叠的路径或符号证据面拆分。
-  平台不支持并行或范围客观不可拆分时，记录原因并降级为一个只读子代理，不得静默改由主 agent 广泛探索。
-- 每个探索包包含唯一 `packet_id`、精确路径/符号/问题边界、相关 AGENTS 内容、强制检索策略、禁止写入约束与输出预算；包集合覆盖目标范围，重叠仅用于核对跨边界依赖。
-- 子代理强制返回 JSON：`packet_id`、`module_name`、`scope`、`files_read`、`evidence`（每项含 claim、path、line_start、line_end）、`existing_structures`、`existing_conventions`、`constraints_discovered`、`open_questions`、`dependencies`、`write_set_hints`、`validation_hints`、`agents_impact_hints`、`risks`、`success_criteria_hints`。
-- 聚合前逐包校验 JSON、范围和引用；无引用、越界或互相冲突的结论必须定向重派或最小抽查。
-  把约束、依赖、风险、潜在写集冲突与开放问题并入决策台账。
-- 父代理负责验证与综合；子代理绝不批准 Gate、绝不写入。
-- 按需 `/context7-auto-research` 核对候选库/API 的官方契约；按需 `/grok-search` 做架构模式调研。
-- PBT 边界筛查：空输入、幂等性、顺序、大小/取值边界、状态转换合法性 → 形成阶段 2 的问题清单。
-
-## 阶段 2 —— 行为澄清循环（允许多轮）
-
-- 只覆盖 WHAT；库、协议、算法、存储、拓扑、实现参数一律路由到阶段 4。
-- 每轮只提当前影响最大的阻塞问题，简明分组，各附证据、影响与建议默认值。
-- 反模式（标记并拒绝）：可观察行为推迟到实现期决定；技术机制伪装成产品需求。
-- 答案扩大模块/场景/数据边界 → 仅增量地回到阶段 1。
-- 循环至未决行为决策为零。
-
-## ⛔ Gate A —— 批准行为契约
-
-- 呈现行为契约与相关台账条目；用户显式批准目标、范围/非目标、场景/成功标准与各项策略。
-- 按批准的范围重算变更名并重新查重。
-- 新建模式：仅此一时点运行 `openspec new change <name>`（仅显式非默认选择时加 `--schema`）。
-- 构建产物计划，然后只物化安全且就绪的行为类产物。
-- 最后写 `changeRoot/gate-a.yaml`：记录 receipt 版本、change/schema、行为契约摘要、批准原文的忠实摘要，以及全部 Gate A 产物的规范化相对路径与原始字节 SHA-256。
-  禁止身份、时间戳、会话 id、模型 id 与工具授权字段。
-
-## 产物计划与写入协议
-
-- 创建前依据 schema 定义/模板做兼容性预判；创建后或恢复时依据 `status --json` 与 `instructions --json` 构建产物 DAG：id、输出路径、依赖/状态、决策类别（`behavior | technical | mixed`）、写入闸门、受影响决策；记录 `apply.tracks`。
-  仅携带机械信息不使产物变为 mixed。
-  `gate-a.yaml`/`ready.yaml` 是工作流审计凭据，不冒充 schema artifact，也不改变 DAG/status。
-- 按产物承载的决策分类，绝不按硬编码产物名：behavior → Gate A；technical/mixed → Gate B；依赖 Gate B 产物的行为类 → 等 Gate B 后沿 DAG 推进。
-- schema 若要求 Gate A 前写入、写入 `changeRoot` 之外、或以未批准的技术决策解锁行为 → 创建前停止，请用户选择兼容的 schema/映射。
-  schema 顺序绝不凌驾于批准之上。
-- 每次写入的强制循环：
-  1. 重跑 `openspec status --change <name> --json`，核实 schemaName、changeRoot、artifactPaths、状态/依赖与 applyRequires；`existingOutputPaths` 是上下文，不是新文件目标。
-  2. 运行 `openspec instructions <artifact-id> --change <name> --json` 并遵循其模板/规则/依赖。
-  3. 阅读依赖与已有输出，做双向一致性检查。
-  4. 在内存备好内容，于对应闸门前展示决策摘要或统一 diff；物化暴露新的实质决策 → 回到相应循环重新批准。
-  5. 批准后只写一个就绪产物（或编辑一个被明确指向的已完成产物）；重跑 status，拓扑推进新解锁的产物。
-- Receipt 写入仅免除不存在的 `instructions <artifact-id>`，其余边界、一致性与逐文件写入规则不变；生成前必须重算覆盖文件的 hash。
-  任何 receipt 覆盖文件发生变化都立即使该 receipt 失效：回到最早受影响 Gate，实质变化须重新批准，机械一致性修复也须重算 receipt；`ready.yaml` 必须最后生成。
-
-## 阶段 3 —— 技术推导
-
-- 从 Gate A 契约、既有代码模式与官方 API 契约推导技术设计；机械决策直接记录进设计，实质性权衡进入阶段 4。
-
-## 阶段 4 —— 技术决策与验证循环
-
-- 覆盖 HOW：接口、数据流、实现错误机制、依赖/算法与关键参数（如 JWT vs session、经批准的 bcrypt cost factor）。
-- PBT 适用性规则（六类筛查：交换/结合、幂等、往返、不变量保持、单调、边界）：具备不变量/往返/幂等/顺序/边界/状态转换的行为必须提取性质与证伪策略；不适合的记录原因并采用示例/E2E/静态验证，不强行套类。
-- 每个场景获得稳定引用 `<spec-path>#<requirement-heading>/<scenario-heading>`，标题在 spec 内唯一；全程维护 Requirement → Scenario → Verification → Task。
-- 先构建实施任务 DAG：所有任务引用存在且无环；依赖边标记 `artifact | contract | conflict-order`、原因与所需输出。
-  由 DAG、写集和资源锁派生建议执行波次；波次只用于审阅，Implement 必须从可信契约重新计算。
-- 每个任务恰好一个 schema 复选框，验证契约使用普通缩进符号列表（绝不嵌套 `- [ ]`/`- [x]`）：
-  - 任务 ID；Requirement + 稳定 Scenario 引用
-  - 直接前置任务（无则 `[]`）、依赖类型、所需输出与原因
-  - 非任务前置条件及其可执行或静态检查
-  - 派发上下文：完成任务所需的最小 spec、AGENTS、代码路径与符号
-  - 建议执行波次；并行资格 `eligible | serial` 及原因
-  - 计划写集：修改/新建/删除的精确路径或最窄可判定 glob；预期交付文件/符号
-  - 冲突集：任务 ID + 原因；共享状态、生成物、清单、数据库、端口、fixture 或验证资源及所需资源锁
-  - 验证类型：`property | example | E2E | static`
-  - 每个 phase 的 Runtime verification contract：只使用 `kind: "vitest"`、`kind: "package-script"`、`kind: "static-check"` 或 `kind: "steps"`；Red 命令 + 预期失败原因；Green 预期行为
-  - `vitest` 明确 runner、安全相对 `testFiles`、`args`、`minTests`；`package-script` 明确 `bun | npm | pnpm | yarn`、真实 script、Gate B 固定的完整 script command 与 `args`；`static-check` 明确本地 executable、`npx` + `noInstall: true` 或安全相对 Node script
-  - 复合验证建模为有序 `steps`：前置检查为 `expected-green`，最后一步匹配 phase classification；不得使用 `&&` 或其他 shell operator
-  - 每个 phase 的 `verificationInputs`：每个直接输入恰好绑定为 Gate B 时已存在的 `workspace` 安全普通文件，或唯一 `outputId`
-  - 受影响套件命令；目标范围；验证命令的并行资格及资源锁
-  - 影响闭包（impact closure）：`changedSurfaces`、检索命令/命中证据、全部相关既有测试及其 `current-task | regression-task | unaffected` 归类、affected suite 中的既有测试证据
-  - 已批准依赖变化清单（无则 `[]`）
-  - AGENTS 影响：`none | update-existing | create-index | remove-index`
-  - AGENTS 目标索引（`none` 时省略）+ 基于证据的原因；`agentsManagedOnly: true`；AGENTS 验证命令（可执行或静态）
-- 将全部任务汇总成唯一、机器可读的 `ImplementGraphBoundary`：graph 持有 `changeId`、完整 tasks 与 outputs；每个 task 显式包含 `dependsOn` 和固定的 objective/context、逐 phase 精确 read/write、verification/verificationInputs、scheduling、impact closure 与 AGENTS 合同。
-  每个 output 显式声明 `id`、安全相对 `path`、producer `taskId/phase` 与 `regular-file` postcondition；同一路径只能有一个 producer，且路径必须在 producer phase write set 中。
-  write set 只授予写权限，不证明 artifact 已产生，provenance 不得从 write set 推断。
-  Runtime 首先接收一次 `admit-graph`；后续每个阶段只接收 `task-attempt`，包含 change/task/request/phase identity 和新的动态 snapshot，不得重述或改变 graph/task 稳定事实。
-- **影响闭包审计**：任务修改路由授权、页面状态、API 响应或公共 HTML 结构时，必须用 URL、路由名、handler/template 符号检索所有现有入口和测试。
-  相关 E2E、theme/layout、authorization、HTML/template 与 API contract 测试必须纳入当前写集、绑定明确后续回归任务，或以引用证据证明不受影响。
-  `affected suite` 不得只包含新增测试；例如 `/videos`、`/api/videos` 变化必须覆盖上述全部测试面。
-  此审计只校验契约与检索证据，不引入业务扫描器。
-- Gate B 前审计调度契约：前置条件可判定，波次满足拓扑顺序；同波次任务的写集、共享资源和验证锁均不冲突；每个冲突均有 `conflict-order` 串行边；每个任务可由最小派发上下文独立执行。
-- 写集未知、过宽或可能重叠的任务不得标为并行；先重塑任务，或用 `conflict-order` 串行化。
-- 非行为变更任务的 Red 是变更前即可执行的静态验证。
-  仅人工验证的任务不具备实施就绪性，绝不通过 Gate B 或退出；重塑任务直至具备可执行验证。
-- Gate B 呈现前，使用 Cadence 与 Implement Runtime 共用的 `src/implement-graph.ts::assessImplementGraphReadiness` 在 consumer root 校验完整 canonical graph；初始 facts 为 completed/blocked tasks 与 applied phases 均为空，且无 candidate overlay。
-  Gate B 要求静态 `closure` 为 `{ executable: true, diagnostics: [] }`；尚待前置任务完成是动态 readiness，不使合法 closure 失败。
-  调用方不得自行构造 missing-input 集合或在 graph 外另算可执行性。
-  `workspace` binding 在 Gate B 时必须已经是安全普通文件。
-  缺失输入必须绑定唯一 output；跨任务 consumer 必须在 producer 的传递 DAG 后继中；suggested wave、conflict、resource 与 verification lock 均不能代替 `dependsOn`。
-  同一 phase candidate 可以创建并使用自己的 output，更早 phase output 必须先成功应用；Red 不得依赖仅由 Green/Refactor 生产的 output。
-  仅允许内容与合同完全一致的 package script、可用的显式 package manager、本地 `node_modules` executable，以及 `npx` 的 `noInstall: true`/`--no-install` 语义；绝不允许隐式联网下载、任意 shell 或 shell operator。
-  绝对路径、非规范路径、`..`、NUL、root escape、任一已存在组件 symlink、非目录父项及非普通最终输入/output 均关闭失败。
-  不支持的合同以 `design-readiness/verification-contract-unsupported` 诊断关闭；脚本缺失以 `verification-adapter/script-missing` 关闭。
-  任何 closure 诊断存在时不得呈现 Gate B 为可批准，也不得写 ready receipt。
-- `vitest` 才允许 Runtime 注入 JSON reporter、要求 `minTests` 并用 assertion identity 判定 Red；`package-script` 与 `static-check` 只按退出码及批准的稳定输出 identity 判定，绝不附加 Vitest 参数。
-  `bun run check`、`bun run test:target`、precheck 与 affected suite 都不得由 Runtime 隐式补充；需要时必须成为 Gate B 明确批准的 contract/step。
-  Design 只生成上述结构化 contract；argv-only shape 无效。
-- 范围内每个索引过期发现分配给一个任务，或报告为无关的既有过期问题。
-- 循环至未决技术决策为零；在内存中备好剩余产物内容/统一 diff。
-
-## ⛔ Gate B —— 批准实施契约
-
-- 核实阶段 3/4 忠实展开 Gate A 契约，未引入未批准的新决策。
-- 对将写入 receipt 的同一个 `ImplementGraphBoundary` 重新运行 `assessImplementGraphReadiness`；只有静态 closure 为 `{ executable: true, diagnostics: [] }` 才可请求 Gate B 批准。
-  否则以精确 readiness/adapter 诊断返回阶段 3/4 修正，不把缺失 consumer capability 延迟到 Implement Red。
-- 呈现实质技术决策、任务 DAG/建议波次、前置条件、计划写集/冲突/资源锁、验证映射、AGENTS 影响矩阵与产物物化预览；用户显式批准。
-- 按写入协议逐个写入剩余就绪产物。
-- 最后写 `changeRoot/ready.yaml`：记录当前 receipt 版本、change/schema、Gate B 批准原文的忠实摘要（含任务 DAG、建议波次、前置条件、计划写集、冲突与资源锁）、唯一 `implementGraph`、`hashImplementGraphBoundary(implementGraph)` 产生的 `implementGraphHash`、`verificationClosure: { executable: true, diagnostics: [] }`、任务验证与 AGENTS 影响摘要、`gate-a.yaml` 的 SHA-256，以及全部规划产物的规范化相对路径与 SHA-256；排除 `ready.yaml` 自身。
-  receipt 不保存另一份 task readiness 摘要、Runtime task/session 状态或 retry budget。
-  任务跟踪文件计算 hash 前只将任务完成标记 `- [x]`/`- [X]` 规范化为 `- [ ]`，其余字节不得忽略。
-
-## 回环规则
-
-- 答案扩大边界 → 阶段 1；技术分析推翻行为契约 → 阶段 2（只重批受影响决策并同步所有受影响产物）；Gate B 发现物化不忠实 → 阶段 3/4（Gate A 未受影响的决策保持批准）；严格校验/验证契约/可追溯性失败 → 引入不一致的最早阶段。
-  绝不隐瞒迟发现的阻塞问题。
-
-## 恢复规则
-
-- 仅当相应 receipt、artifact hash、追溯链与 strict validation 全部通过时，才将设计阶段已完成的 Gate 视为可信交付；仅切换上下文不触发重复批准。
-- 恢复以当前 schema/status 的 `changeRoot` 与 artifactPaths 为准；receipt 使用上述固定根文件名，但文件存在本身不证明批准。
-- 算法：
-  1. `openspec status --change <name> --json`，使用其 schemaName、changeRoot、artifactPaths 与状态；阅读全部 `existingOutputPaths` 与依赖。
-  2. 校验 `gate-a.yaml`/`ready.yaml` 的格式、change/schema 绑定、规范化相对路径、hash 与覆盖范围；拒绝绝对路径、`..`、越界及符号链接逃逸。
-  3. 检查 `openspec validate <name> --strict --type change`、模板完整性、跨产物一致性、可追溯性、验证契约，以及任务 DAG、波次、写集、冲突与资源锁的调度完整性。
-  4. 重建产物计划，从最早未完成或不一致的阶段继续，不重新请求仍有有效 receipt 的 Gate。
-  5. 选择下一步：变更不存在 → 停止并请用户确认拼写/新建；receipt 缺失、hash 失效或产物不完整/不一致 → 回到最早受影响阶段并重新批准该 Gate；`applyRequires` 全部 `done` 且全部校验通过 → 执行退出审计；仅存未持久化的会话分析 → 无中途恢复，重跑只读分析。
-
-## 退出条件
-
-- [ ] `openspec validate <name> --strict --type change` 问题为零
-- [ ] `applyRequires` 中每个产物 id 状态均为 `done`
-- [ ] `apply.tracks` 解析到 `changeRoot` 内生成的任务产物
-- [ ] 产物一致且可追溯；canonical graph/hash 一致，静态 verification closure 无诊断；任务 DAG 无环、前置条件可判定、波次及冲突排序有效、计划写集与资源锁完整；每个任务验证契约完整且可执行，无仅人工任务
-- [ ] 每个任务的结构化 AGENTS 影响、批准依赖与影响闭包契约完整；仓库索引在 Design 阶段未被修改
-- [ ] BLOCKING_DECISIONS = 0
-- [ ] `gate-a.yaml` 与 `ready.yaml` 格式、覆盖范围及全部 hash 校验通过
-- [ ] Gate A 与 Gate B 已在可信实施交付前于设计阶段完成
-- [ ] 状态：READY_TO_IMPLEMENT
-
-## 参考命令
-
-- `openspec context --json` / `openspec schemas --json`
-- `openspec view` / `openspec list --changes --json` / `openspec list --specs`
-- `openspec status --change <name> --json` / `openspec instructions <artifact-id> --change <name> --json`
-- `openspec new change <name>`（仅 Gate A）
-- 校验失败时 `openspec show <name> --json --deltas-only`
-- 定义新约束前 `rg -n "Constraint:|MUST|MUST NOT|INVARIANT:|PROPERTY:" openspec/`
-<!-- ABEL:END -->
+}
+```
+
+When two or more independent packets are needed, issue all sibling calls in the same assistant turn.
+Do not wrap them in a `requests` array.
+Validate packet id, scope, citations, and structure before synthesis; uncited, conflicting, or out-of-scope claims require a bounded follow-up or a minimal parent check.
+
+Evidence packets may read code and tests but may not execute validation.
+Design may inspect manifests and verification capability statically; it must not run product tests as a substitute for Implement evidence.
+
+## Decision ledger and Gate A
+
+Maintain a decision ledger with id, `behavior | technical`, question, cited evidence, alternatives, recommendation, user decision, status, and affected artifacts.
+
+Ask the user only for substantive choices: observable behavior, scope/non-goals, data/security/privacy/compatibility/migration policy, new dependencies, architecture/policy, irreversible changes, and technical choices with real trade-offs.
+Resolve reversible details mechanically from one established repository convention and do not ask them repeatedly.
+
+Gate A approves the complete WHAT: goal, observable scenarios, failures, scope/non-goals, compatibility/migration/security policy, and success criteria.
+Before approval, present the unresolved behavior decisions together.
+After explicit approval:
+
+1. resolve the final kebab-case change name and schema;
+2. create the change only if it does not already exist;
+3. materialize schema-ready behavior artifacts one at a time;
+4. compile `gate-a.yaml` with code-owned canonical serialization and raw SHA-256 bindings.
+
+Gate A is not tool permission and carries no session/model/timestamp identity.
+
+## Technical contract and ImplementPlan
+
+Derive HOW from Gate A plus repository evidence.
+Ask only unresolved substantive technical decisions, then build a complete task DAG with stable Requirement → Scenario → Verification → Task references.
+
+Each task must seal:
+
+- stable id, objective, exact dependencies and edge reasons;
+- exact phase-local read/write/delete paths and output producers;
+- Red, Green, optional Refactor structured verification and exact verification inputs;
+- task-affected and repair verification contracts;
+- conflicts, resources, verification locks, dependency changes, impact closure, and existing-test evidence;
+- AGENTS impact/target and managed-only ownership;
+- precise context sufficient for a fresh Worker without a conversation transcript.
+
+Impact closure must name affected existing tests and fixtures; a suite made only from newly added tests is insufficient.
+For public UI or API work, seal the relevant route authorization, page state, API response and contract, public HTML/template/theme behavior, and an approved browser E2E check when those surfaces are affected.
 
-If the required input is missing or absent, or the request is ambiguous and not unique, stop before any work and ask the user for the missing or clarified input.
+The plan admits only shell-free `vitest`, `package-script`, `static-check`, or ordered `steps` contracts.
+Pin local runners, package scripts, arguments, `minTests`, classifications, and `noInstall` behavior.
+Reject shell operators, implicit downloads, absolute/escaping paths, missing local capability, and unsupported verification shapes during Design readiness.
 
-Design work is read-only before an explicit Gate A approval: do not write product code, and wait for the user when a blocking decision or ambiguity remains.
-Only after Gate A approval may design artifacts be written inside the change root.
-Resume an existing change with --change <change_name>: validate the Gate receipt, artifact hash, and trace strictly; when an artifact is inconsistent or invalid, report readiness only when the delivery is complete and READY_TO_IMPLEMENT holds.
+Compile one canonical `ImplementPlan` with `compileImplementPlan`.
+It contains:
+
+- tasks and regular-file outputs;
+- target/affected/full-suite baselines and normalized failure identity policy;
+- change-level affected, full-suite, and post-apply verification;
+- an explicit `artifactCorrection.maxAttempts` of 2-3 total candidate launches per task phase and operation, including the initial launch;
+- bounded in-boundary repair policy and attribution classes `pre-existing | introduced | unresolved | environment`;
+- parent-owned `tasks.md` tracking;
+- when required, one sealed managed-block AGENTS operation per approved target, including impact, owning task ids, complete marker-bounded content, and verification.
+
+Use the shared readiness proof to require an executable static closure with no diagnostics.
+A write set grants authority but never proves that an output exists.
+Every absent future input must be a unique declared output from a transitive dependency.
+
+## Gate B and trusted delivery
+
+Gate B approves the complete HOW: substantive technical choices, task DAG, exact boundaries, verification/repair contracts, output postconditions, dependency changes, scheduling declarations, tracking, and AGENTS operations.
+Do not request approval while any capability, closure, traceability, or blocking decision is unresolved.
+
+After explicit approval, write remaining schema artifacts one at a time, run OpenSpec strict validation, then generate in this order:
+
+1. canonical `implement-plan.json`;
+2. canonical Gate A/plan/artifact/traceability bindings;
+3. `ready.yaml` last, using `compileReadyReceipt`.
+
+`ready.yaml` references the plan path, schema version, raw-byte hash, canonical plan hash, executable verification closure, Gate A hash, artifact hashes, and traceability hash.
+It does not embed caller-owned runtime task state, completed/blocked arrays, snapshots, launch identities, or retry budgets.
+
+Only the sealed task checkboxes may be normalized from `[x]`/`[X]` to `[ ]` when validating the tracked `tasks.md` hash.
+Every other byte remains bound.
+
+## Recovery without approval loops
+
+On resume, validate current artifacts and receipts first.
+Preserve decisions whose bound content did not change.
+A mechanical hash, formatting, checkbox, stale path, or traceability repair regenerates only invalidated artifacts and does not reopen a Gate.
+Reopen Gate A only for changed behavior authority; reopen Gate B only for changed technical authority.
+Report the exact invalidated decision/artifact instead of restarting the whole Design process.
+
+Endpoint/transport failure in an evidence packet pauses that packet and retains other evidence.
+Missing package scripts/runners or a non-executable verification contract must be corrected before Gate B; they must not be deferred as an Implement surprise.
+
+Report `READY_TO_IMPLEMENT` only when strict validation passes, both Gates have zero unresolved decisions, all artifacts/hashes/traceability resolve, and the canonical plan is executable.
+Otherwise report the current Design state, retained evidence, and the one unresolved decision or artifact that prevents readiness.
+
+Do not implement product code and do not archive, publish, release, stage, or commit implicitly.
+
+<!-- ABEL:END -->

package/prompts/abel-diagnose.md

@@ -3,33 +3,52 @@ description: Diagnose and minimally repair one or more existing bugs
 argument-hint: "<problem-description>"
 ---
 
-Load the bundled `abel-workflow` Skill before doing any work.
-Read the complete value inside `<abel-request>` without a second whitespace-tokenization pass.
+Load the bundled `abel-workflow` Skill before work and read the complete value inside `<abel-request>` without tokenizing it a second time.
 
 <abel-request>
 $ARGUMENTS
 </abel-request>
 
 <!-- ABEL:PROMPT:abel-diagnose -->
-The request must uniquely identify one or more existing bugs.
-If required input is missing or ambiguous, stop and ask for the missing or disambiguating information instead of guessing.
 
-## Diagnose procedure
+The request must uniquely identify one or more existing defects.
+If required input is missing or ambiguous, stop before work and ask one focused question.
 
-Record target, affected-suite, and full-suite baselines before writes, keeping pre-existing failures separate from the reported defect.
-Collect evidence for each report, reproduce it, propose candidate root causes, and actively falsify each candidate before accepting a root cause.
-If the defect cannot be reproduced or the root cause remains unverified, mark it blocked and do not generate a repair.
+## Independent Diagnose contract
 
-For each verified root cause, first add and run an executable regression verification that fails for the target defect.
-Only then apply the minimum repair, rerunning the target after every code or test edit and the affected suite after refactoring.
-At a stable checkpoint, classify AGENTS impact, apply only verified managed-index changes, and validate routes.
-Finish only when regression and affected verification pass and the full suite has no new failure relative to baseline.
+Diagnose is not an Implement recovery route and does not inherit Implement task states, retry budgets, approval codes, delivery revisions, or Gate routing.
+A transport, environment, artifact, or verification failure in Diagnose remains a typed Diagnose pause with retained evidence; it never becomes an instruction to change workflow stage.
 
-If resolution requires new behavior, substantive architecture, or a choice not uniquely determined by an existing contract, stop and return the requirement to Design.
-Do not archive, publish, or commit implicitly.
+Use bounded `abel_dispatch` packets with `action: "run"`, `stage: "abel-diagnose"`, and role `diagnosis-worker` only when evidence collection benefits from isolation.
+Each packet covers one defect or one falsification question.
+Do not send an Implement control command from Diagnose.
 
-If the required input is missing or absent, or the request is ambiguous and not unique, stop before any work and ask the user for the missing or clarified input.
+The parent runs reproduction and verification commands and applies accepted diffs.
+A diagnosis Worker is read-only: it proposes cited evidence or one complete candidate diff and never claims that it executed a command or observed a passing result.
 
-Reproduce and falsify the root cause first: complete the failing regression verification before the minimum repair.
-The fix must not introduce new behavior or architecture; a substantive decision returns to Design.
-When dev-browser is missing or unavailable, block only approved browser E2E tasks, must not block other work, and provide executable remediation.
+## Evidence-first algorithm
+
+For each defect, keep this order:
+
+1. Record target, affected-suite, and full-suite baselines.
+   Keep unrelated and pre-existing failures separate.
+2. Reproduce the reported symptom with exact inputs and an observable failure identity.
+3. List plausible root causes and actively falsify each one.
+   Accept a root cause only when evidence rules out material alternatives.
+4. Add the smallest executable regression verification and run it.
+   It must fail for the reproduced defect, not for syntax, setup, environment, or another reason.
+5. Apply the minimum repair within the existing behavior and architecture contract.
+   Run the regression after every edit and the affected suite after refactoring.
+6. Compare the full suite with baseline and require no introduced failure.
+   Classify any AGENTS impact and apply only a verified managed-region update at a stable parent checkpoint.
+
+If reproduction is unavailable, root-cause evidence is insufficient, the regression cannot witness the defect, or an external capability is unavailable, return `paused` with the exact safe evidence and a concrete resume condition.
+Do not invent a repair.
+
+If the requested result actually requires new observable behavior, a new dependency, wider paths, substantive architecture/policy, or another product decision, return `scope-decision-required` with the evidence and decision that the user must make.
+Do not silently widen Diagnose and do not transform an ordinary repair failure into that result.
+
+When an approved browser E2E check requires `dev-browser` and it is unavailable, pause only that check and provide executable remediation; continue independent non-browser evidence when safe.
+
+Finish only when reproduction, falsification, failing regression, minimum repair, affected verification, and baseline comparison are all evidenced.
+Never archive, publish, or commit implicitly.