AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
SKILL verified MIT Self-run

Review Contract

skill-jerrylalala-compound-engineering-review-contract · by Jerrylalala

Fork Overlay:Review Contract 三档 Tier 分类 + Anti-Leniency 注入。在 ce:review 基础上增加 Blocking/Analytical/Advisory 分层和结论类型系统。使用时机:运行 ce:review 前阅读此协议以理解审查标准。

No reviews yet
0 installs
0 views
view→install

Install

$ agentstack add skill-jerrylalala-compound-engineering-review-contract

✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

View the full security report →

Verified badge

Passed review? Show it. Paste this badge into your README, it links to the public security report.

AgentStack Verified badge Links to your public security report.
[![AgentStack Verified](https://agentstack.voostack.com/badges/verified.svg)](https://agentstack.voostack.com/security/report/skill-jerrylalala-compound-engineering-review-contract)

Reliability & compatibility

Security review passed
0 installs to date
no reviews yet
2mo ago

Declared compatibility

Claude CodeClaude Desktop

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Review Contract? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

Review Contract — Fork Overlay

> 基础层:上游 ce:review 已实现 P0-P3 严重级别、confidence 阈值(≥0.60 才报告)、autofix_class 路由。 > > 本 Overlay 增量:三档 Tier 分类 + 对应的 anti-leniency 强度 + conclusion_type 覆盖层。


Tier 分类

> 命名规范[上游] = 在 skills/ce-review/references/persona-catalog.md 中有记录,由 ce:review 编排器实际派发。 > [本地] = 仅在 agents/review/ 中存在但不在 persona-catalog,需手动调用,不被编排器自动派发。

| Tier | Agents | Anti-Leniency 强度 | |------|--------|-------------------| | Blocking | security-reviewer [上游], data-migrations-reviewer [上游], deployment-verification-agent [上游] | 零容忍:confidence ≥ 0.50 且 severity P0 必须报告,不确定时用 needs-human-check | | Analytical | architecture-strategist [本地], performance-reviewer [上游], kieran-rails-reviewer [上游], kieran-typescript-reviewer [上游], kieran-python-reviewer [上游], julik-frontend-races-reviewer [上游], dhh-rails-reviewer [上游], pattern-recognition-specialist [本地] | 严格务实:必须提供触发条件/复现场景;没有场景的担忧降级为 question | | Advisory | code-simplicity-reviewer [本地], agent-native-reviewer [上游], schema-drift-detector [上游] | 建议为主:默认 autofix_class: advisory,除非与明确 spec/CLAUDE.md 冲突 |

已更正的名称对照(原名 → persona-catalog 名):

  • security-sentinelsecurity-reviewer
  • data-integrity-guardian + data-migration-expertdata-migrations-reviewer(合并,上游统一入口)
  • performance-oracleperformance-reviewer

结论类型系统(conclusion_type 覆盖层)

> 注意conclusion_type 是本 overlay 新增的本地语义字段,不存在于上游 findings-schema.json。 > 上游 schema 使用 autofix_classseverityconfidence 等字段。 > 本字段仅在此 overlay 内部使用,不写入上游 schema。

上游 schema 已有 autofix_class 路由。本 overlay 增加一个更语义化的 conclusion_type 字段,与上游字段并存:

| conclusion_type | 映射关系 | 何时使用 | |-------------------|----------|----------| | finding | 上游 P0/P1/P2 + confidence ≥ 0.70 | 有确凿证据的问题,有代码引用 | | question | 上游 residual_risks / confidence 0.60-0.69 | 不确定,需要讨论 | | needs-human-check | 上游 requires_verification: true | 超出 AI 判断能力 | | clear | 上游 findings 为空 | 此项无问题 |


Anti-Leniency 参考原则(仅供分类框架参考)

> 注意:以下原则是读取本 skill 时的行为指南,不是通过 system_prompt_suffix 自动注入到 agent 中。 > ce:review 不支持动态 prompt 注入机制。加载本 skill 后,Claude 在调用各 Tier agent 时应主动遵循对应原则。

Blocking Tier — 零容忍

你是 [agent-name],属于 Blocking Tier。你的审查结论直接影响生产环境安全。

铁律(在上游 confidence_thresholds 基础上加强):
1. P0 问题:confidence ≥ 0.50 即必须报告(上游默认 0.60,Blocking Tier 降低阈值)
2. 不确定时必须输出 requires_verification: true,不能直接给 safe_auto
3. 每个 finding 必须有 evidence[],包含文件路径 + 行号 + 代码片段
4. 禁止使用「应该没问题」「可能不影响」等措辞
5. 自问:「如果这段代码导致数据泄露,我的 clear 结论能站住脚吗?」

Analytical Tier — 严格务实

你是 [agent-name],属于 Analytical Tier。

规则(在上游 confidence_thresholds 基础上):
1. 每个 finding 必须提供触发条件或复现场景
2. 没有具体场景的担忧 → 放入 residual_risks,不作为 findings
3. 性能问题必须估算影响量级(O(n) vs O(n²)、毫秒 vs 秒级)
4. 「理论上可能有问题」→ residual_risks,不是 findings

Advisory Tier — 建议为主

你是 [agent-name],属于 Advisory Tier。

规则:
1. 默认 autofix_class: advisory,除非与明确 spec/CLAUDE.md 冲突
2. 代码品味问题 → P3 + advisory,不是 P0-P2
3. 不要因为「代码可以更好」就报 finding,放入 residual_risks 更合适

与上游 ce:review 的集成方式

不修改上游 ce:review SKILL.md。在调用 ce:review 前:

  1. 加载本 skill(review-contract
  2. 确认当前 PR 涉及哪些 Tier 的 agent(根据 diff 类型)
  3. ce:review 运行时,对 Blocking Tier agent 的结果验证 evidence[] 是否完整
  4. 对 Advisory Tier 的 P0-P2 findings 降级为 P3(代码品味不应 blocking)

Eval Set 测试:本协议的正确性由 docs/eval/ 中的案例验证:

  • fn-01-sql-injection → Blocking Tier 漏报测试
  • fp-01-safe-env-var → Blocking Tier 误报测试
  • al-d-opinion-vs-finding → Advisory Tier 降级测试

与上游 findings-schema.json 的对齐

| 本 overlay 字段 | 上游 schema 对应字段 | 说明 | |----------------|---------------------|------| | conclusion_type: finding | severity: P0-P2 + confidence ≥ 0.70 | 高置信度有证据问题 | | conclusion_type: question | residual_risks[] + confidence 0.60-0.69 | 不确定的担忧 | | conclusion_type: needs-human-check | requires_verification: true | 需要人工判断 | | conclusion_type: clear | findings: [] (空数组) | 此项无问题 | | Blocking Tier | 无上游对应 | 本 overlay 新增分类 | | confidence ≥ 0.50 (P0) | 上游默认 0.60 | Blocking Tier 降低阈值 |


遗留问题(待 P2 完成后评估)

  1. conclusion_type 是否需要写入 findings-schema.json?(当前策略:不修改上游 schema,在 overlay 层处理)
  2. Anti-leniency prompt suffix 的具体注入机制(systempromptsuffix vs overlay skill 说明)
  3. 多模型仲裁权重:Claude 1.0 / Codex 0.85 / Gemini 0.80(待 P7 实现时启用)

Integration with [T] Mode

ce:review 使用 [T] flag 时,本 skill 的 Tier 分类由 Stage 5 Deterministic Patch Gate 自动消费。用户不再需要手动加载此 skill — Patch Gate 内联引用 Tier 逻辑。

Tier → Patch Gate 行为映射

| Tier | 代理 | Patch Gate 行为 | |------|------|----------------| | Blocking | security-reviewer, data-migrations-reviewer, deployment-verification-agent | 所有 findings 强制 downgrade 为 gated_auto(无论原始 autofixclass) | | Analytical | architecture-strategist, performance-reviewer, kieran-/julik-/dhh-rails-* | 维持原 autofixclass,但自动添加 requires_verification: true | | Advisory | code-simplicity-reviewer, agent-native-reviewer, schema-drift-detector | 维持原 autofix_class,不触发额外验证 |

解决的遗留问题

本集成解决了此前 skill 中记录的"虚拟字段无消费者"问题:

  • conclusion_type 字段 — Patch Gate 通过 Tier 分类间接消费(Blocking Tier = gated_auto,对应 conclusion_type: finding 的最高置信度处理)
  • confidence ≥ 0.50 (P0) Blocking Tier 规则 — Patch Gate 的 Rule 1 强制 gated_auto 确保 Blocking Tier 高风险发现不会被自动应用
  • Anti-leniency 原则 — 通过 Patch Gate 的 Blocking Tier 强制路由实现,不依赖 prompt suffix 注入

激活条件

Patch Gate 集成仅在以下条件同时满足时生效:

  • ce:review 收到 [T] flag
  • mode:autofix 被选中
  • .team-contract.md 存在于 repo 根目录

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.