← 返回 Skills 市场

adversarial-coach

Name: adversarial-coach
Author: killerapp

作者 Vaskin Kissoyan · GitHub ↗ · v0.9.0

cross-platform ✓ 安全检测通过

1709

总下载

当前安装

版本数

在 OpenClaw 中安装

/install adversarial-coach

功能描述

Adversarial implementation review based on Block's g3 dialectical autocoding research. Use when validating implementation completeness against requirements with fresh objectivity.

使用说明 (SKILL.md)

/coach - Adversarial Implementation Review

Usage

/coach [requirements-file]

/coach - Infer requirements from context
/coach requirements.md - Validate against specific file

Coach-Player Loop

You orchestrate this dialectical loop between implementing agent (player) and reviewer (coach):

You (player) implement features
/coach invokes adversarial review with independent evaluation of compliance to requirements
Coach returns: IMPLEMENTATION_APPROVED or specific fixes
Address feedback, loop until approved

Review Process

Step 1: Identify Requirements

Check (in order):

Specified requirements file or issue/ticket mentioned
requirements.md, REQUIREMENTS.md, SPEC.md, TODO.md
Conversation context; ask user if nothing found

Step 2: Adversarial Review

Review with fresh objectivity - discard prior knowledge, don't rationalize shortcuts.

Check Category	Items
Requirements	Each item: implemented or missing with specific gap
Compilation	Compiles? Tests pass? Runs?
Common Gaps	Auth on endpoints, token refresh endpoint, HTTPS, bcrypt for passwords, error handling, input validation
Functional	Test actual flows (not just compilation), verify edge cases work
Test Coverage	Auth error cases (401/403), token expiry, invalid inputs, rate limits

Step 3: Return Verdict

If approved (>95% complete):

IMPLEMENTATION_APPROVED

- [Requirement 1]: Verified
- [Requirement 2]: Verified
- Compilation: Success
- Tests: All passing

If fixes needed:

REQUIREMENTS COMPLIANCE:
- [Requirement]: Implemented
- [Requirement]: Missing - [specific gap]

IMMEDIATE ACTIONS NEEDED:
1. [Specific fix with file/line if known]
2. [Specific fix]

Key Principles

Rigorous but fair:

Catch real gaps (security, logic, data flow), not style preferences
Functionality over aesthetics; always flag security issues (auth, crypto, validation)

Concise:

Bullets, not essays; specific issues, not vague concerns
No file contents or verbose analysis in output

Fresh context is your superpower:

Review as if you've never seen this code; validate against requirements, not intentions

Magic Signal

IMPLEMENTATION_APPROVED = termination signal

Only use when: all requirements met, code compiles/runs, tests pass, no significant gaps.

If in doubt, don't approve.

Minimal Example

User: /coach SPEC.md

Claude:
REQUIREMENTS COMPLIANCE:
- User registration endpoint (bcrypt hashing)
- Login with JWT tokens
- Logout endpoint - not implemented
- Token refresh - missing /auth/refresh endpoint

IMMEDIATE ACTIONS NEEDED:
1. Add POST /auth/logout endpoint
2. Add POST /auth/refresh for token renewal
3. Add tests for 401 on expired tokens

---

User: [fixes] /coach SPEC.md

Claude:
IMPLEMENTATION_APPROVED
- All auth endpoints verified (register, login, logout, refresh)
- 18 tests passing including auth error cases

Research

Paper: Adversarial Cooperation in Code Synthesis
Implementation: g3
Key insight: Discard implementing agent's self-report; perform independent evaluation against requirements.

安全使用建议

This skill is an instruction-only code-review helper and appears internally consistent. Before installing: 1) Be aware the coach will ask to inspect your repository/workspace and may ask you to run builds/tests — grant only the minimal access required. 2) If you plan to let the agent run commands or CI, ensure build environments don't expose secrets (API keys, private tokens). 3) Review any file contents or code the skill outputs before sharing externally. 4) Because the skill can be invoked autonomously by agents, consider who can grant it access to your codebase; revoke access if you see unexpected behavior. Overall, the skill itself does not request credentials or install software and looks safe for use as a review assistant.

功能分析

Type: OpenClaw Skill Name: adversarial-coach Version: 0.9.0 The OpenClaw skill 'adversarial-coach' is designed for an AI agent to perform code reviews against specified requirements. The `SKILL.md` file contains instructions for the agent to identify requirements from common files like `requirements.md` or `SPEC.md` and then conduct an 'adversarial review' focusing on security, compilation, and functional gaps. There is no evidence of malicious intent, data exfiltration, unauthorized execution, persistence mechanisms, or prompt injection attempting to subvert the agent's core purpose. All instructions align with the stated goal of a rigorous code review.

能力评估

✓ Purpose & Capability

Name/description (adversarial implementation review) align with the SKILL.md instructions: locate requirements, perform independent checks (compilation, tests, edge cases), and return a verdict. The skill requests no unrelated binaries, env vars, or config paths.

✓ Instruction Scope

Instructions tell the agent to look for requirements files, use conversation context, run compilation/tests, and report specific gaps. They do not instruct reading unrelated system files, exfiltrating secrets, or posting results to external endpoints. The guidance is appropriately scoped to code-review tasks.

✓ Install Mechanism

No install spec and no code files — instruction-only. This minimizes disk-write and supply-chain risk.

✓ Credentials

The skill requires no environment variables, credentials, or config paths. The checks it describes (compilation, tests) are proportional to a review task and do not imply access to unrelated services or secrets.

✓ Persistence & Privilege

always is false and the skill does not request persistent system modifications or cross-skill configuration changes. Autonomous invocation is allowed by default but not combined with other high-risk behaviors.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install adversarial-coach
安装完成后，直接呼叫该 Skill 的名称或使用 /adversarial-coach 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v0.9.0

From Foundry: Adversarial implementation review based on Block's g3 dialectical autocoding res

元数据

Slug adversarial-coach

版本 0.9.0

许可证 —

累计安装 4

当前安装数 3

历史版本数 1

常见问题