peer-review-loop（同行评审循环）

Peer Review Loop — Cavekit + Ralph Loop + Codex Peer reviewer

Run a Cavekit cavekit through a Ralph Loop where Claude builds and Codex adversarially reviews. This is the most rigorous automated quality process available: every few iterations, a completely different model (different training data, different biases, different blind spots) challenges your implementation.

Why This Works

Architecture

┌─────────────────────────────────────────────────────┐
│                   Ralph Loop                         │
│  (Stop hook feeds same prompt each iteration)        │
│                                                      │
│  ┌──────────┐    ┌──────────────┐    ┌────────────┐ │
│  │  Claude   │───▶│ Build from   │───▶│  Commit    │ │
│  │  (Build)  │    │ cavekit +  │    │  changes   │ │
│  └──────────┘    └──────────────┘    └──────┬─────┘ │
│       ▲                                      │       │
│       │                                      ▼       │
│  ┌──────────┐    ┌──────────────┐    ┌────────────┐ │
│  │  Fix      │◀──│ Parse        │◀──│  Codex CLI │ │
│  │  findings │    │ findings     │    │  (Review)  │ │
│  └──────────┘    └──────────────┘    └────────────┘ │
│                                                      │
│  Completion: all cavekit requirements met +         │
│              no CRITICAL/HIGH findings               │
└─────────────────────────────────────────────────────┘

Review Invocation: Codex CLI (primary) vs MCP (legacy)

The peer review loop supports two invocation paths:

1. Codex CLI delegation (primary) — Uses scripts/codex-review.sh which calls codex directly in --approval-mode full-auto with a structured review prompt. Faster, no MCP server overhead, findings are parsed and appended to context/impl/impl-review-findings.md automatically.
2. MCP server (legacy fallback) — Configures Codex as an MCP server in .mcp.json. Claude calls the MCP tool on review iterations. Used only when Codex CLI delegation is unavailable (e.g., older Codex versions).

The build script (setup-build.sh) auto-detects which path to use: if codex-review.sh is present and codex CLI is available, it uses CLI delegation. Otherwise it falls back to MCP configuration.

Quick Start

# Basic: implement a cavekit with peer review
/ck:peer-review-loop context/kits/cavekit-auth.md

# With options
/ck:peer-review-loop context/kits/cavekit-api.md --max-iterations 20 --codex-model gpt-5.4-mini

# Review-only mode (review existing code, don't build new)
/ck:peer-review-loop context/kits/cavekit-api.md --review-only

# Review every iteration instead of every 2nd
/ck:peer-review-loop context/kits/cavekit-auth.md --review-interval 1

What the Command Does

1. Validates the cavekit file exists and Codex CLI is installed
2. Configures Codex as an MCP server in .mcp.json (if not already configured)
3. Builds a Ralph Loop prompt that embeds:

- The cavekit path and related plan/impl files - Instructions to alternate between build and review iterations - The peer review prompt template for Codex - Completion criteria tied to cavekit acceptance criteria

1. Starts the Ralph Loop via the stop hook mechanism

Codex Review Invocation

Primary: Codex CLI via codex-review.sh

When codex CLI is available, the loop delegates review to scripts/codex-review.sh which exposes the bp_codex_review function. This runs Codex in full-auto mode with a structured adversarial review prompt, parses findings into a standardized table, and appends them to context/impl/impl-review-findings.md.

# What the build loop runs on review iterations:
source scripts/codex-review.sh
bp_codex_review --base main

The CLI path is faster (no MCP server startup), produces structured findings with severity levels (P0-P3), and handles fallback gracefully if Codex is unavailable.

Legacy fallback: Codex MCP Server

When Codex CLI delegation is not available, the command configures Codex as an MCP server automatically:

{
  "mcpServers": {
    "codex-reviewer": {
      "command": "codex",
      "args": ["mcp-server", "-c", "model=\"gpt-5.4\""]
    }
  }
}

Claude calls this MCP server on review iterations to get peer review feedback. The MCP server exposes Codex as a tool that accepts prompts and returns responses — Claude sends the cavekit + code diff, Codex returns findings.

Changing the Codex Model

Use --codex-model to specify which OpenAI model Codex should use:

/ck:peer-review-loop cavekit.md --codex-model gpt-5.4-mini    # faster, cheaper
/ck:peer-review-loop cavekit.md --codex-model gpt-5.4          # default, most capable

Iteration Pattern

Iteration 1: BUILD  — Read cavekit, implement first requirement
Iteration 2: REVIEW — Call Codex CLI (or MCP fallback), get findings, fix CRITICAL/HIGH
Iteration 3: BUILD  — Continue implementing, address remaining findings
Iteration 4: REVIEW — Call Codex CLI (or MCP fallback) again, new findings on new code
...
Iteration N: BUILD  — All requirements met, all findings fixed
             → outputs <promise>SPEC COMPLETE</promise>

The review interval is configurable. Default is every 2nd iteration. Use --review-interval 1 for maximum rigor (review every iteration).

Peer Review Findings File

Review findings are tracked in context/peer-review-findings.md:

# Peer Review Findings

## Latest Review: Iteration 4 — 2026-03-14T10:30:00Z
### Reviewer: Codex (gpt-5.4)

| # | Severity | File | Issue | Status |
|---|----------|------|-------|--------|
| 1 | CRITICAL | src/auth.ts:L42 | Missing input validation on token | FIXED |
| 2 | HIGH | src/auth.ts:L67 | Race condition in session refresh | FIXED |
| 3 | MEDIUM | src/auth.ts:L15 | Unused import | NEW |
| 4 | LOW | src/auth.ts:L3 | Comment typo | WONTFIX |

## History
### Iteration 2
| # | Severity | File | Issue | Status |
|---|----------|------|-------|--------|
| 1 | CRITICAL | src/auth.ts:L20 | SQL injection in login query | FIXED |

Completion Criteria

The loop exits when the completion promise is output. The prompt instructs Claude to ONLY output it when ALL of these are true:

• All cavekit requirements (R-numbers) have been implemented
• All acceptance criteria pass
• No CRITICAL or HIGH peer review findings remain unfixed
• Build passes
• Tests pass
• At least one review iteration completed with no new CRITICAL/HIGH findings

Modes

Build + Review (default)

Alternates between implementing cavekit requirements and calling Codex for review. Use for greenfield implementation from a cavekit.

Review Only (--review-only)

Skips building. Each iteration calls Codex to review existing code against the cavekit, then fixes issues found. Use when code already exists and you want peer review QA.

Prerequisites

1. Codex CLI installed: npm install -g @openai/codex
2. OpenAI API key configured: Codex needs authentication (via codex login or env var)
3. Cavekit context directory: Cavekit file must exist at the given path
4. Ralph Loop plugin: The ralph-loop plugin must be installed (provides the stop hook)

Convergence Signals

The peer review loop has converged when:

• Codex's findings drop to zero or only LOW/MEDIUM severity
• Code diffs between iterations are minimal
• All cavekit requirements confirmed as met by both Claude and Codex

If the loop hits max iterations without converging:

• Check context/peer-review-findings.md for persistent issues
• Consider whether the cavekit needs clarification
• Run /ck:revise to trace issues back to kits

Cross-References

• peer-review — The underlying peer review patterns and prompt templates
• convergence-monitoring — How to detect convergence vs ceiling
• validation-first — Validation gates that run on every build iteration
• impl-tracking — How implementation progress is tracked across iterations
• Ralph Loop — The underlying Ralph Loop mechanism