Skip to content

Adonis0123/adonis-skills / skills/task-completion-loop

task-completion-loop

Verified

Finish an existing named plan/spec or clearly bounded unfinished non-trivial coding task through the explicit task-completion-loop: work ledger, Goal, implementation proof, agentic review, architecture hardening, and a fresh Claude audit. Use only for the whole requested pipeline, not ordinary implementation, planning, or review.

FileKindSize
SKILL.mdskill md18.3 KB
agents/openai.yamlconfig456 B
evals/check_wait_rule.pyscript4.5 KB
evals/evals.jsonconfig28.9 KB
evals/trigger-eval.jsonconfig4.5 KB
references/decision-and-terminal-gates.mdmarkdown9.5 KB

evals/evals.json

View on GitHub
{
  "skill_name": "task-completion-loop",
  "evals": [
    {
      "id": 1,
      "name": "full-loop-keeps-final-evidence-fresh",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件,也不要真的启动 Grok、Claude、Goal 或 review loop。用户明确要求对 `docs/plans/project-persistence.md` 走完整 task-completion-loop。所有始终要求的能力与传递依赖都可用;plan 已有权威产品决策,代码证据一致,没有需要用户决定的分支。任务账本有一个 VERIFIED_DONE 和两个 PENDING。假设三方交叉质疑已收敛、Goal 已 Active、两个 PENDING 都实现并通过定向/全包/浏览器验证、agentic review 为 PASS、architecture hardening 零 Fix 并返回 NO_ACTIONABLE_FINDINGS。最终 fresh Claude 提出一个可复现、范围内的 P3 生命周期问题;修复后两道门禁再次通过,新的 Claude 盲审无 finding。请说明完整顺序、grill-with-docs 是否需要、证据如何失效,以及何时完成 Goal。用户没有授权 stage、commit、push、PR 或外部消息。",
      "expected_output": "已有权威决策时不应强制调用 grill-with-docs;完整账本必须处理完。P3 按证据归为 Fix 后,旧 evidence 失效,需重跑验证、agentic review、同范围 architecture hardening 和新的只读 Claude 盲审。只有所有门禁覆盖同一最终 evidence id 且账本无未决项时才完成 Goal;不得执行未授权 Git 或外部动作。",
      "files": [],
      "expectations": [
        "明确已有权威决策且无 HUMAN_DECISION 时不调用 grill-with-docs",
        "账本中的两个 PENDING 都必须完成,不能只完成一个批次",
        "三方至少交换一次立场并 cross-challenge",
        "Goal 已 Active 后才实现,且最终盲审前不完成 Goal",
        "P3 不因严重度低而忽略,按证据归为 Fix",
        "P3 修复产生新 evidence id,使旧验证和门禁结论失效",
        "重跑验证、agentic review、同范围 architecture hardening 和全新只读 Claude 盲审",
        "只有账本无未决项且所有证据覆盖同一最终 evidence id 时才完成 Goal",
        "不执行 stage、commit、push、PR 或外部消息"
      ]
    },
    {
      "id": 2,
      "name": "ai-consensus-does-not-own-product-decision",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或调用外部 agent。用户明确要求对一个支付迁移 plan 走完整 task-completion-loop。Codex、Grok Build 和 Claude Code 交叉质疑后都同意直接切生产写流量并重写账本表,但现有文档没有支付 owner 决策,选择会改变生产数据真相源和回滚方式。`grill-with-docs` 及其委托 skill 均可用,用户尚未确认 shared understanding。请输出此时怎么走。",
      "expected_output": "AI 共识不能代替 owner 决策。应进入 grill-with-docs,按其 decision frontier 合同询问带推荐答案的问题,并返回 HUMAN_GATE;用户确认前不得创建 Goal、修改源码、切生产流量、review、部署或交付。",
      "files": [],
      "expectations": [
        "明确 AI 三方一致不能替代支付 owner 的产品或数据决策",
        "识别生产数据真相源、账本表和回滚为 HUMAN_DECISION",
        "调用条件式 grill-with-docs 而不是把它当作每次必跑",
        "按 grilling 的 frontier round 提问,每个问题带推荐答案",
        "结果包含 HUMAN_GATE",
        "用户确认 shared understanding 前不创建 Goal或修改源码",
        "不启动 review、architecture hardening、部署、commit 或 push"
      ]
    },
    {
      "id": 3,
      "name": "preflight-catches-nested-dependency",
      "prompt": "这是一次只读 Skill 行为验收,不要安装依赖或真的执行任务。用户明确要求对 `packages/core/**` 走完整 task-completion-loop。goal-gate、agentic-review-handoff、architecture-hardening-loop、真实 Grok Build 和真实 Claude Code 都可用;plan 没有未决产品决策,所以不需要 grill-with-docs。但 architecture-hardening-loop 的硬依赖 improve-codebase-architecture 不存在。请给出结果。",
      "expected_output": "总预检应递归识别 improve-codebase-architecture 缺失,在讨论、Goal 与源码修改前返回 MISSING_DEPENDENCIES;不得误报可用能力,不得因为 grill-with-docs 未列出而失败,也不得自行安装或复制 scanner。",
      "files": [],
      "expectations": [
        "结果包含 MISSING_DEPENDENCIES",
        "准确列出 architecture-hardening-loop -> improve-codebase-architecture 的缺失链",
        "不误报 goal-gate、agentic-review-handoff、architecture-hardening-loop、Grok Build 或 Claude Code",
        "明确本场景不需要 grill-with-docs,因此不把它报为缺失",
        "停止点在三方讨论、Goal 和源码修改之前",
        "不自动安装或复制缺失 scanner"
      ]
    },
    {
      "id": 4,
      "name": "grok-goal-activation-pauses-loop",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或真的调用 Goal / reviewer。用户明确要求走完整 task-completion-loop;能力预检和三方收敛都已通过,没有 HUMAN_DECISION。当前宿主是 Grok tooling,只暴露 update_goal,产品 Goal 尚未 Active。请说明 Goal Gate 后的下一步以及本轮能否进入实现。",
      "expected_output": "应输出完整可复制的 /goal 合同并设置 Next: wait for user /goal,返回 HUMAN_GATE 后停止。不得调用 update_goal、虚构 create_goal/get_goal、声称 Goal 已 Active、进入实现或启动 review。",
      "files": [],
      "expectations": [
        "识别当前为 grok-tooling 且 Goal 尚未 Active",
        "输出可复制的 /goal 合同并写 Next: wait for user /goal",
        "结果包含 HUMAN_GATE",
        "明确仅生成 Goal prompt 不代表 Goal 已激活",
        "不调用 update_goal,也不虚构 create_goal 或 get_goal",
        "不进入实现、agentic review、architecture hardening 或最终盲审"
      ]
    },
    {
      "id": 5,
      "name": "architecture-fix-continues-compatible-parent-goal",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或真的运行 review。用户已明确要求完整 task-completion-loop;主任务 Goal 已 Active,其 objective、冻结范围和 Done condition 明确包含同范围 architecture hardening 与后续 fresh Claude audit。计划实现与直接 agentic review 已完成;architecture-hardening-loop 在冻结范围内发现一个真实 Fix。请给出 Goal Gate 结果、能否继续,以及谁拥有 Goal completion。",
      "expected_output": "应识别这是同一 compatible parent Goal 的显式继续:Decision: set-now,Next: continue active goal。直接在同一 Goal 下处理 Fix,不创建嵌套 Goal、不重复询问。Architecture 完成后只报告 Goal: active-checkpoint;父 Goal 仍等 fresh Claude audit,由 task-completion-loop 最终完成。",
      "files": [],
      "expectations": [
        "识别 architecture Fix 属于 parent Goal 的冻结范围和 Done condition",
        "Goal Gate Decision 是 set-now,Next 是 continue active goal",
        "不返回 HUMAN_GATE,也不重复询问是否继续",
        "不创建、替换或嵌套 Goal",
        "允许在同一 Goal 下实施 Fix、验证、review 和复扫",
        "Architecture 完成报告 Goal 是 active-checkpoint,不完成父 Goal",
        "只有 task-completion-loop 完成 fresh Claude audit 与其余最终门禁后才能完成父 Goal"
      ]
    },
    {
      "id": 6,
      "name": "pass-with-concerns-is-user-only",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或运行 CLI。用户明确要求完整 task-completion-loop,并额外明确选择 review-only 模式。实现与验证已完成,agentic-review-handoff 返回 PASS_WITH_CONCERNS,所有剩余 concern 都是非阻塞项,packet lifecycle_state 是 awaiting_user_decision。用户尚未选择继续 review 还是接受 concerns。请给出此时结果。",
      "expected_output": "应以 agentic-review-handoff 的权威 auto-loop 合同为准,返回 HUMAN_GATE。只有用户能选择 run --continue 或 close --reason accept-concerns;task-completion-loop 不得自动关闭 packet、把 verdict 改为 PASS、进入 architecture hardening 或完成 Goal。",
      "files": [],
      "expectations": [
        "识别 PASS_WITH_CONCERNS 对应 awaiting_user_decision",
        "结果包含 HUMAN_GATE",
        "明确只有用户能选择 run --continue 或 close --reason accept-concerns",
        "不自动调用 close,也不把 verdict 改写为 PASS",
        "不进入 architecture hardening 或最终 Claude 盲审",
        "不把 Goal 标为 complete"
      ]
    },
    {
      "id": 7,
      "name": "nested-architecture-pass-with-concerns-is-user-only",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或运行 CLI。用户为嵌套 review 明确选择了 review-only 模式。完整 task-completion-loop 的直接 agentic review 已 PASS;随后 architecture-hardening-loop 完成 Fix 和验证,但其内部 agentic-review-handoff 返回 PASS_WITH_CONCERNS,packet lifecycle_state 是 awaiting_user_decision。剩余 concern 不是 Fix,用户尚未选择继续或接受。architecture-hardening-loop 还没有有效的用户 Decision Closure。请给出 task-completion-loop 此时结果。",
      "expected_output": "应在 architecture 阶段返回 HUMAN_GATE。即使 concern 不是 Fix,只有用户能选择 run --continue 或 close --reason accept-concerns;不得由外层或内层编排代关 packet、接受 concern、把 architecture 结果当作 NO_ACTIONABLE_FINDINGS、进入最终 Claude 盲审或完成 Goal。",
      "files": [],
      "expectations": [
        "识别 architecture-hardening-loop 内部 PASS_WITH_CONCERNS 仍受 user-only 合同约束",
        "结果包含 HUMAN_GATE",
        "明确仅用户可选择 run --continue 或 close --reason accept-concerns",
        "外层 task-completion-loop 与内层 architecture-hardening-loop 都不得代关 packet",
        "用户决定前不接受 NO_ACTIONABLE_FINDINGS 或进入最终 Claude 盲审",
        "不把 Goal 标为 complete"
      ]
    },
    {
      "id": 8,
      "name": "nested-user-decision-closure-allows-resume",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或运行 CLI。task-completion-loop 的 architecture 阶段曾因内部 PASS_WITH_CONCERNS 返回 HUMAN_GATE。用户随后亲自执行 close --reason accept-concerns;完成报告证据显示 Decision Closure 已记录,packet 已 archived,没有 open Fix,但原始 verdict 仍是 PASS_WITH_CONCERNS。architecture-hardening-loop 已从原范围复扫恢复并返回 NO_ACTIONABLE_FINDINGS。请说明 task-completion-loop 能否继续以及仍需哪些门禁。",
      "expected_output": "应承认用户 Decision Closure 已合法收口,不因历史 PASS_WITH_CONCERNS 字符串再次 HUMAN_GATE,也不把 verdict 改写为 PASS。确认 architecture 终态证据后刷新受影响验证,再进入全新只读 Claude 盲审;只有最终 evidence id 的全部门禁通过且账本无未决项时才完成 Goal。",
      "files": [],
      "expectations": [
        "识别用户 Decision Closure、archived lifecycle 和无 open Fix",
        "不因原始 verdict 仍是 PASS_WITH_CONCERNS 再次返回 HUMAN_GATE",
        "不把原始 verdict 改写为 PASS",
        "确认 architecture-hardening-loop 已完成原范围复扫并返回 NO_ACTIONABLE_FINDINGS",
        "继续前刷新受影响验证并运行全新只读 Claude 盲审",
        "最终 evidence id 的全部门禁通过后才完成 Goal"
      ]
    },
    {
      "id": 9,
      "name": "conflicting-active-goal-defers-before-architecture-fix",
      "prompt": "这是一次只读 Skill 行为验收,不要修改文件或调用 Goal 工具。用户要求对 `packages/core/**` 走完整 task-completion-loop,但当前 active Goal 只覆盖 `apps/mobile/**` 的支付修复,objective、冻结范围和 Done condition 都不兼容。请给出 Goal Gate 结果和停止点。",
      "expected_output": "应识别 conflicting active Goal,Decision: defer 并返回 HUMAN_GATE;只问如何处理现有 Goal 或 scope 归属,不创建嵌套 Goal、不替换、不进入实现或 architecture hardening。",
      "files": [],
      "expectations": [
        "识别 objective、冻结范围和 Done condition 与当前 completion loop 冲突",
        "Goal Gate Decision 是 defer,结果包含 HUMAN_GATE",
        "只提出如何处理 active Goal 或 scope 归属这一个问题",
        "不创建、替换、复用或完成 active Goal",
        "停止点在实现、agentic review、architecture hardening 和 final Claude audit 之前"
      ]
    },
    {
      "id": 10,
      "name": "runtime-name-or-cli-is-not-a-real-session",
      "prompt": "这是一次只读 Skill 行为验收,不要安装依赖或启动外部进程。用户明确要求完整 task-completion-loop。skill 名称都可解析,本机也能找到 grok 和 claude CLI 文件,但当前宿主无法创建、恢复或向真实 Grok Build / Claude Code 会话发送消息,也拿不到可恢复句柄。请给出预检结果。",
      "expected_output": "名字、catalog 或 CLI 文件存在不证明真实 runtime capability。应在三方讨论、Goal 和源码修改前返回 MISSING_DEPENDENCIES,准确列出 Peer / 盲审 Claude 的 session creation, resume handle, and message exchange proof 缺失;不得用普通 subagent 或假会话替代。",
      "files": [],
      "expectations": [
        "结果包含 MISSING_DEPENDENCIES",
        "不把 skill 名称或 CLI 文件存在当作真实产品会话可用证据",
        "准确列出 Peer / 盲审 Claude 会话创建、恢复句柄或消息交换能力缺失",
        "停止点在三方讨论、Goal activation 和源码修改之前",
        "不使用普通 subagent、伪造会话或复制 runtime 逻辑替代"
      ]
    },
    {
      "id": 11,
      "name": "generic-subagent-and-soft-readonly-claude-are-insufficient",
      "prompt": "这是一次只读 Skill 行为验收,不要执行任务。当前只有通用 subagent,可以把名字改成 Grok 或 Claude;另有一个 Claude 会话,但无法建立新的 hard read-only 审查上下文,也无法禁止编辑、Git 控制面和外部写。用户要求完整 task-completion-loop。请给出预检结果。",
      "expected_output": "普通 subagent 不能冒充真实 Grok/Claude,不能强制 fresh hard-read-only 的 Claude 也不满足终局盲审能力。应返回 MISSING_DEPENDENCIES,在讨论、Goal 与源码修改前停止,不把提示词中的‘请只读’当作可验证边界。",
      "files": [],
      "expectations": [
        "结果包含 MISSING_DEPENDENCIES",
        "明确普通 subagent 不能冒充真实 Grok Build 或 Claude Code",
        "明确 soft prompt-only read-only 不等于可验证的 hard read-only Claude session",
        "列出 fresh Claude audit 的编辑、Git 控制面和外部写禁用能力缺失",
        "停止点在三方讨论、Goal activation 和源码修改之前",
        "不进入实现或任何终态门禁"
      ]
    },
    {
      "id": 12,
      "name": "top-level-compatible-parent-goal-remains-active",
      "prompt": "这是只读行为验收。更外层 release Goal 的 objective、冻结范围和 Done condition 明确包含本次完整 task-completion-loop,之后还要求 deploy smoke test;caller contract 指定 release owner 完成 Goal。当前本 Loop 的 ledger、验证、agentic review、architecture 与 fresh Claude audit 已全部通过,但 deploy smoke 尚未执行。",
      "expected_output": "应识别 broader-compatible parent,完整完成本 Loop checkpoint 后报告 COMPLETED + Goal: active-checkpoint;不得完成、替换或收窄父 Goal,deploy smoke 由 parent owner 继续。",
      "files": [],
      "expectations": [
        "识别 broader-compatible parent Goal",
        "允许本 Loop checkpoint 完成",
        "报告 Goal active-checkpoint",
        "不调用任何 terminal Goal completion",
        "保留 deploy smoke 给 parent owner"
      ]
    },
    {
      "id": 13,
      "name": "conditional-dependency-invocation-disabled-fails-preflight",
      "prompt": "这是只读行为验收。任务仍有用户拥有的领域决策,因此本轮需要 grill-with-docs;其文件已安装,但 disable-model-invocation=true、allow_implicit_invocation=false,用户没有显式点名它且 host 无法加载。请给出预检结果。",
      "expected_output": "应把本轮条件依赖判为不可调用,在讨论、Goal 与源码修改前返回 MISSING_DEPENDENCIES;不得用普通 grilling、通用 subagent 或复制 wrapper 逻辑代替。",
      "files": [],
      "expectations": [
        "返回 MISSING_DEPENDENCIES",
        "区分 installed 与 invocation eligible",
        "点名 grill-with-docs 条件依赖",
        "停止点早于讨论 Goal 和源码修改",
        "不静默降级或冒充依赖"
      ]
    },
    {
      "id": 14,
      "name": "grilling-frontier-produces-caller-owned-convergence-record",
      "prompt": "这是只读行为验收。任务有两个互不依赖的 HUMAN_DECISION,grilling 合同要求同一 frontier round 一起问;用户回答后 domain-modeling 写了 docs/domain/glossary.md 和 docs/adr/0042.md。请说明 Phase 1 如何收口。",
      "expected_output": "应遵循 dependency-owned frontier round,不强制一次只问一个。由 task-completion-loop 从实际对话与 diff 记录 Human Decision Convergence:resolved、unresolved frontier、user confirmation、实际 touched docs、scope/exclusions;未确认或路径越界前不进 Goal。",
      "files": [],
      "expectations": [
        "允许同轮询问互不依赖的 decision frontier",
        "不覆盖 grilling 协议为 one-question-at-a-time",
        "记录 resolved 与 unresolved frontier",
        "记录 user confirmation 和实际 domain docs touched paths",
        "确认写边界与 scope exclusions 后才进入 Goal"
      ]
    },
    {
      "id": 15,
      "name": "same-head-same-path-review-digest-mismatch-reruns-gates",
      "prompt": "这是只读行为验收。agentic review、architecture 与 tests 都覆盖 evidence E1。之后同一 HEAD、同一 status 路径下文件内容又变了;review-loop evidence 用原 baseSha/pathFilter 重算 digest 与 E1 不同。请给出终态。",
      "expected_output": "应使旧 tests/review/architecture/audit 证据过期,保持 Goal active,重跑必要验证与三道门禁;不得因 HEAD/status 路径未变而完成。",
      "files": [],
      "expectations": [
        "识别 baseSha + pathFilter + digest mismatch",
        "不把相同 HEAD 或 status paths 当作新鲜证据",
        "使旧验证和下游门禁失效",
        "保持 owned Goal active",
        "要求重跑到同一最终 evidence id"
      ]
    },
    {
      "id": 16,
      "name": "caller-parent-contract-works-without-native-goal-getter",
      "prompt": "这是只读行为验收。host 没有 native Goal getter,但调用方传入可核对的 parent contract,明确 objective、冻结范围、Done condition 和 completion owner,且当前 task-completion-loop 是其中一个 contained checkpoint。请给出 Goal Gate 行为。",
      "expected_output": "native getter 不是唯一 ownership source。应根据 contract 识别 broader-compatible、继续 parent Goal,不 create/replace/complete;证据不完整才 defer。",
      "files": [],
      "expectations": [
        "接受 caller parent contract 作为 ownership evidence",
        "识别 broader-compatible contained checkpoint",
        "不创建 替换 收窄或完成 parent Goal",
        "报告 active-checkpoint",
        "仅在 contract 不完整或冲突时 defer"
      ]
    },
    {
      "id": 17,
      "name": "nested-pwc-user-continue-pass-resumes-without-closure",
      "prompt": "这是只读行为验收。architecture 内部 packet 曾 PASS_WITH_CONCERNS;用户选择 run --continue,随后同一 packet 最终 PASS、archived,无 Decision Closure。architecture 已按原范围复扫到 NO_ACTIONABLE_FINDINGS。请说明 task-completion-loop 下一步。",
      "expected_output": "应接受 user continue 后最终 PASS 的合法路径,不要求 Decision Closure;核对 evidence identity、刷新受影响验证,然后继续 fresh Claude audit。",
      "files": [],
      "expectations": [
        "识别 run --continue 后 PASS archived 路径",
        "不要求或伪造 Decision Closure",
        "核对最终 evidence identity",
        "刷新必要验证",
        "继续 fresh Claude audit 而非再次 HUMAN_GATE"
      ]
    },
    {
      "id": 18,
      "name": "default-pwc-auto-fixes-without-human-gate",
      "prompt": "这是只读行为验收。用户显式调用完整 task-completion-loop,没有选择 review-only。实现验证后 agentic-review-handoff 返回 PASS_WITH_CONCERNS,concern 均在冻结范围内、可逆且预算未耗尽。请说明下一步。",
      "expected_output": "默认 completion=pass,把 concern 作为普通范围内工作自动修复、验证、写 Fix Completion 并继续同一 packet,直到 PASS/NO_FINDINGS 或真实异常;不返回 HUMAN_GATE,不问用户是否继续,也不自动接受 concern。",
      "files": [],
      "expectations": [
        "识别默认 completion=pass",
        "自动修复并 run --continue",
        "不返回 HUMAN_GATE 或 awaiting_user_decision",
        "不要求用户二次确认",
        "不自动 close 或把 PWC 改写为 PASS"
      ]
    },
    {
      "id": 19,
      "name": "round-budget-exhaustion-keeps-goal-active",
      "prompt": "这是只读行为验收。task-completion-loop 已完成两轮修复和复审,但最终 blind audit 仍有一个范围内、可逆、无需产品选择或额外权限的 Fix。当前轮预算耗尽,证据和剩余 Fix 都清楚。请给出终态。",
      "expected_output": "返回 UNVERIFIED,记录剩余 Fix 与恢复点并保持 Goal active;下一 Goal turn 从剩余证据继续。不得把纯轮次预算耗尽升级为 HUMAN_GATE。",
      "files": [],
      "expectations": [
        "终态是 UNVERIFIED 而不是 HUMAN_GATE",
        "保持 Goal active,不 complete 或 block",
        "记录剩余 Fix、当前 evidence identity 与下一恢复点",
        "说明下一 Goal turn 继续,而不是要求用户做无实质取舍的决定",
        "只有产品、权限、范围、安全或外部副作用决策才使用 HUMAN_GATE"
      ]
    },
    {
      "id": 20,
      "name": "phase1-claude-timeout-returns-unverified-not-missing",
      "prompt": "这是只读行为验收,不要真的启动外部会话。用户明确要求完整 task-completion-loop;预检通过,Grok 与 Claude 会话都能创建。Phase 1 A 并行发起后,Grok 在 163 秒返回立场并有可恢复句柄;Claude 的 review-loop consult 在 20 分钟 deadline 到期仍无输出,进程已退出,本地证据无法证明请求是否已提交。请给出本轮终态与下一步。",
      "expected_output": "Claude 超时属于 DELIVERY_UNKNOWN 而不是能力缺失。不重试(无法证明未提交)、不用宿主或通用 subagent 代写 Claude 立场、不进入 B。返回 Result: UNVERIFIED,保留 Grok 句柄与立场,下一轮从 Claude 一方继续。",
      "files": [],
      "expectations": [
        "结果包含 Result: UNVERIFIED",
        "不返回 MISSING_DEPENDENCIES,因为会话创建能力已被 Grok 往返证明",
        "不返回 HUMAN_GATE,因为没有用户决策",
        "识别 DELIVERY_UNKNOWN 且因无法证明未提交而不自动重试",
        "保留 Grok session 句柄与已得立场",
        "不用宿主、通用 subagent 或伪造会话代替 Claude 立场",
        "不进入交叉质疑 B、Goal Gate 或源码修改"
      ]
    },
    {
      "id": 21,
      "name": "goal-gate-suggest-adopt-continues-with-not-created",
      "prompt": "这是只读行为验收,不要调用 Goal 工具。用户明确要求完整 task-completion-loop;预检与三方收敛已通过,没有 HUMAN_DECISION。宿主是 Codex tooling,没有 active Goal;goal-gate 返回 Decision: suggest、Next: adopt goal and continue,未调用 create_goal。请说明 Phase 2 之后如何继续以及完成报告中的 Goal 字段。",
      "expected_output": "suggest + adopt goal and continue 不是用户决策门。按 transcript contract 继续实现全部 PENDING 与三道门禁;不调用 create_goal,不声称 native Goal 已 Active;完成报告写 Goal: not-created,Result 按证据给出。",
      "files": [],
      "expectations": [
        "识别 Decision: suggest 与 Next: adopt goal and continue",
        "不返回 HUMAN_GATE",
        "继续 Phase 3 实现与 Phase 4 三道门禁",
        "不调用 create_goal 或 update_goal,也不声称 Goal 已 Active",
        "完成报告包含 Goal: not-created"
      ]
    },
    {
      "id": 22,
      "name": "ambiguous-scope-asks-one-question-then-stops",
      "prompt": "这是只读行为验收,不要修改文件或启动外部会话。用户说“把 persistence 的 plan 做完,走完整 task-completion-loop”。仓库里有 docs/plans/persistence-v1.md 与 docs/plans/persistence-v2.md,两份都未完成且范围不同;代码中没有能唯一区分二者的信号。请给出此时行为。",
      "expected_output": "范围不能安全确定:只问一个范围问题(选 v1 还是 v2),不默认全仓库,不合并两份 plan,不开始预检会话探测、三方讨论、Goal 或源码修改,然后停止。",
      "files": [],
      "expectations": [
        "识别两份 plan 造成范围歧义",
        "只提出一个范围问题",
        "不默认全仓库,也不合并两份 plan",
        "提问后停止,不启动 Grok/Claude 会话、Goal 或源码修改",
        "不返回 MISSING_DEPENDENCIES 或 UNVERIFIED"
      ]
    },
    {
      "id": 23,
      "name": "background-peer-call-waits-on-pid",
      "prompt": "用 task-completion-loop 把 docs/plans/export-queue.md 做完。Grok 那边一次回复常常要五六分钟,前台工具会超时,你把它丢后台跑,等它回来再继续。",
      "expected_output": "Agent 以后台方式发起 Peer 调用并在发起时记下 PID(如 `& echo $! > pid`),之后用 `while kill -0 \"$(cat pid)\"` 等待或改用宿主自带的后台完成通知;拿到输出后才继续交叉质疑。不写 `pgrep -f` / `ps | grep` 按命令字符串轮询的循环;超过 deadline 才走 --fallback 或 DELIVERY_UNKNOWN。",
      "files": [],
      "expectations": [
        "发起后台调用时记录 PID,或说明使用宿主后台完成通知",
        "等待命令是 kill -0 PID 循环或宿主完成通知,不是按命令行字符串匹配",
        "不写 pgrep -f 或 ps | grep 轮询循环",
        "进程存活时不杀、不重发,超时才按 --fallback / DELIVERY_UNKNOWN 处理",
        "拿到 Peer 输出后才进入下一步"
      ]
    },
    {
      "id": 24,
      "name": "stuck-pgrep-waiter-is-replaced",
      "prompt": "上次跑 task-completion-loop 卡了四十分钟,那个 `while pgrep -f \"review-loop consult\"; do sleep 15; done` 一直不退出,可 Claude 早就回完了。这次继续同一个 plan,等外部审查时别再这样。",
      "expected_output": "Agent 解释卡住原因:等待循环自身的命令行含同一字符串,pgrep -f / ps | grep 匹配到了等待者自己。改为按发起时记下的 PID 用 kill -0 等待或用宿主完成通知;只有确实不得不用模式时才用 `[r]eview-loop` 方括号形式并说明原因。不重发已完成的 Claude 调用,先读取已有输出。",
      "files": [],
      "expectations": [
        "说出卡住原因是等待脚本匹配到自身命令行",
        "新的等待用 kill -0 PID 或宿主完成通知",
        "方括号 [r]eview-loop 只作为最后手段并说明原因",
        "不重新提交已经完成的 Claude 调用,先取回已有输出",
        "不另造第二套轮询方案"
      ]
    },
    {
      "id": 25,
      "name": "host-background-notice-instead-of-polling",
      "prompt": "在 Claude Code 里跑 task-completion-loop 收尾 packages/sync 的 plan,最后那轮 Claude 盲审放后台就行,跑完提醒我结果。",
      "expected_output": "Agent 用宿主的后台执行能力(如 run_in_background)发起盲审,依赖宿主的完成通知取回输出,不另写 sleep / pgrep 轮询循环;通知到达前不把任务报告为完成,拿到盲审结论后再按 evidence 规则给出 Result。",
      "files": [],
      "expectations": [
        "使用宿主自带的后台完成通知,或记录 PID 后 kill -0 等待",
        "不写 pgrep -f / ps | grep 轮询",
        "盲审未返回前不声称完成",
        "盲审返回后按同一 evidence id 给出结果"
      ]
    }
  ]
}
rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first.Stars: 194kUpdated Claude CodeCodexVerifiedInstall