claude-code-review

name	claude-code-review
description	Launch a bounded, read-only Claude Code session through the Claude Agent SDK to review the current issue-branch diff and advise Codex. Use when the user requests a Claude or cross-model code review, or when flow-auto remains stuck after two materially distinct attempts and needs implementation support.

Claude Code Review

Request the Codex-native claude:code_review capability. Claude advises; Codex retains ownership of edits, tests, and lifecycle decisions.

Run the review

1. Confirm the current directory is the issue worktree, not the default branch.

2. Summarize the blocker without including credentials or raw environment values.

3. Resolve SKILL_DIR as the directory containing this SKILL.md.

4. Run:

   <SKILL_DIR>/scripts/claude-code-review \
     --cwd "$PWD" \
     --issue "<issue number and concise acceptance criteria>" \
     --blocker "<failed approaches, exact safe error summary, and question>"
   

   Add --base-ref <ref> only when the automatic base selection is wrong.

5. Relay Claude's verdict and findings with attribution. Verify each finding against the code before changing anything.

6. Apply accepted fixes with Codex, rerun the relevant test, and continue the parent workflow. Do not ask Claude to edit or commit.

Guardrails

• Treat this as a bounded escalation, not an autonomous retry loop. Invoke it at most once for the same blocker fingerprint unless new evidence materially changes the question.
• The runner uses the system claude executable and its existing login. Never read, print, copy, or pass stored OAuth credentials.
• The SDK session receives only Read, Glob, and Grep; Bash, writes, web access, MCP servers, plugins, skills, and subagents are disabled.
• The runner computes the diff locally and sends it as untrusted review data.
• If Claude Code, uv, the SDK, or authentication is unavailable, report that the escalation is unavailable and continue with Codex's normal blocker path.