Русская версия

Claude-Codex Review Skill

Two AI agents review your plans and code iteratively — Claude Code as the initiator, Codex as an independent reviewer — catching issues that a single agent misses. Up to 5 rounds.

Why review with a different model

A single AI agent tends to be consistent with itself: it makes the same assumptions throughout a task and rarely questions its own decisions. A second, independent agent breaks this pattern:

• Blind spots — each agent has different biases; cross-checking surfaces issues neither would find alone
• Confirmation resistance — when one agent must defend a decision to another, weak reasoning gets exposed before it reaches production
• Higher signal — two independent reviewers agreeing on an issue is a much stronger signal than one agent's self-assessment

Environment

This skill is optimized and tested for Claude Code and Codex running as VS Code extensions in the same workspace. Both agents share the project directory and read/write files directly. Recommended pairing as of April 2026 — Opus or Sonnet 4.6 vs. GPT 5.4 extra high. Using Sonnet instead of Opus even for the plan is smoothed out by the presence of a strong model as the reviewer.

Implemented as a robust file-based protocol between two VS Code extension panels, not as an MCP-based Codex integration or a CLI-mode workflow.

How it works

A review session lives in .dual-review/<session_id>/ and progresses through rounds:

R1-01-round-start.md     <- Claude prepares review context
R1-02-codex-claimed.flg  <- Codex claims the round
R1-03-codex-review.md    <- Codex writes structured review
R1-04-claude-claimed.flg <- Claude processes the review

Each round, Claude analyzes the feedback — accepting, rejecting, or inline-fixing each issue with explicit reasoning. If the verdict is NEEDS_WORK, Claude applies changes and opens a new round. The cycle continues until APPROVED, REJECTED, or max_rounds is reached, producing a final.md with the session outcome.

Review scopes

Quality contract

Every review enforces:

• Only significant issues — no stylistic nitpicks or weak hypotheses
• Each issue backed by evidence (file path, method name, logic)
• Explicit checks for missed requirements, regression risks, and over-engineering

Design decisions

• Issues as hypotheses — the initiator treats every reviewer issue as a hypothesis to verify, not a command to follow. Each issue gets an explicit verdict (ACCEPTED, REJECTED, INLINE FIX) with reasoning. This prevents the common pattern where an agent blindly applies all suggestions, including wrong ones.

• Claim flags for crash recovery — .flg files mark that an agent has started processing a round. If the agent crashes mid-work, the next run detects the orphaned flag and resumes from the correct state instead of creating a duplicate or skipping the round.

• Append-only immutability — no protocol file is ever overwritten or deleted during a session. This eliminates race conditions between agents and produces a complete audit trail by construction, not by discipline.

• Scope lock — the review scope is fixed in round 1 and cannot change between rounds. This prevents scope creep where later rounds drift into unrelated areas and the review never converges.

• Round-start as contract — R*-01-round-start.md is not just context but a binding contract: it lists what changed since the last round, which issues were accepted/rejected, and what the reviewer should focus on. The reviewer must verify claims against actual files.

• Background handoff — a lightweight file watcher (wait-for-review.sh) runs in the background and notifies Claude Code when Codex completes its review. This makes round transitions seamless without polling loops in the agent's main thread.

Installation

Copy the claude-code/ directory into your project's skills:

mkdir -p .claude/skills/codex-dual-review-file-based
cp claude-code/* .claude/skills/codex-dual-review-file-based/

Usage

In Claude Code:

/codex-dual-review-file-based <task description> [path/to/plan.md] [max_rounds=5]

Claude prepares the first round and prompts you to pass the reviewer instructions to Codex. A background file watcher detects when Codex completes its review and automatically triggers the next step.

File structure

claude-code/
  SKILL.md              # Initiator protocol (Claude Code skill definition)
  reviewer-prompt.txt   # Reviewer protocol (prompt for Codex)
  wait-for-review.sh    # Background watcher script
docs/
  review-findings.md    # Protocol self-review (design decisions, edge cases)

Requirements

• Claude Code as a VS Code extension
• Codex as a VS Code extension (in the same workspace)
• Bash (for wait-for-review.sh)

Limitations

• Codex must have read/write access to the .dual-review/ directory in the project
• The wait-for-review.sh watcher polls every 10 seconds; not instant
• Session files are append-only by design — no file is ever overwritten or deleted during a session