Menu
Run a structured, read-only Claude Code review over the current git diff or a branch diff.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Compatibility alias for stricter Claude review; prefer claude-review with --adversarial for new use.
Diagnose whether Claude Code is installed and authenticated for this local Codex plugin without making a billable Claude call.
Delegate a foreground read-only or write-capable task to Claude Code from Codex using the local companion script.
Configure or explain the optional Codex Stop hook that can run a read-only Claude review when a turn ends.
Cancel a running Claude Code companion job by sending process signals through the plugin job index.
Analyze plugin-owned Claude Code job logs for permission prompts and export reviewed allowedTools arguments.
| name | claude-review |
| description | Run a structured, read-only Claude Code review over the current git diff or a branch diff. |
Use this skill when the user wants Claude Code to review code changes without mutating files.
Use the companion script from the installed plugin root. If the current working
directory is not this plugin checkout, resolve the script relative to this skill
file, for example ../../scripts/claude-companion.mjs, and pass --cwd <target-repo> for the repository being reviewed.
Real review calls send diffs and review prompts to Claude Code and may spend quota. If the Codex host offers persistent approvals, ask the user to approve the narrow prefix node scripts/claude-companion.mjs review instead of broad commands like node. If host policy blocks external disclosure, do not bypass it.
Review uncommitted changes:
node scripts/claude-companion.mjs review
Review a branch diff:
node scripts/claude-companion.mjs review --base main
Run a stricter review over the same schema:
node scripts/claude-companion.mjs review --adversarial --base main
For machine-readable output:
node scripts/claude-companion.mjs review --base main --json
Useful options:
--cwd <path> to run from a specific repository.--base <ref> to review <ref>...HEAD.--model <model> to choose a Claude model. The default is sonnet; use Claude model names or full Claude model IDs.--effort <level> to pass Claude Code effort: low, medium, high, xhigh, or max.--schema <path> to override the review JSON schema.--max-diff-bytes <n> to raise or lower the single-review diff limit. The default is 200000 bytes.--timeout-ms <n> to cap the Claude review subprocess runtime. The default is 600000 milliseconds.--adversarial to use a stricter prompt that challenges assumptions and looks for subtle concrete failure modes.--include-untracked to explicitly include untracked files in the default review diff.Default review includes tracked staged and unstaged changes. It does not include untracked files unless --include-untracked is set. Review always uses Claude Code --permission-mode plan and one-shot JSON output. It should not edit files. Foreground review commands create plugin-owned job state and log files before invoking Claude, so use claude-status, claude-result, or claude-cancel if a review is still running or needs to be inspected from another shell.
Model guidance:
--model opus for serious reviews, merge readiness, hook gate validation, or high-risk changes.--effort low for smoke tests, cheap sanity checks, or explicitly low-effort requests.sonnet and opus matching their requested families, but a haiku request initialized as Haiku and still reported claude-sonnet-4-6 in actual usage. After real Claude calls, inspect and report actual model usage from JSON output when available.