菜单
Use OpenAI Codex from inside Claude Code — code reviews, adversarial reviews, and delegating tasks to Codex. Use when user asks to /codex:review, /codex:rescue, /codex:adversarial-review, delegate to Codex, run Codex review, or install the Codex plugin.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | codex-in-claude-code |
| description | Use OpenAI Codex from inside Claude Code — code reviews, adversarial reviews, and delegating tasks to Codex. Use when user asks to /codex:review, /codex:rescue, /codex:adversarial-review, delegate to Codex, run Codex review, or install the Codex plugin. |
Plugin oficjalny OpenAI: openai/codex-plugin-cc
Repo: https://github.com/openai/codex-plugin-cc
Pozwala wywołać Codex bezpośrednio z Claude Code — review kodu, adversarial review i delegowanie tasków do Codexa jako background job.
@openai/codex) zainstalowany globalnieWAZNE: Te /plugin* komendy musi wpisac uzytkownik w prompt (nie wolaj ich z Bash, nie pisz /plugin ... jako komendy do siebie). Claude nie moze zainstalowac plugina sam — to akcja po stronie harnessu.
Pelna sekwencja (4 kroki, KAZDY w osobnym promptcie):
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup
Kolejnosc jest obowiazkowa. Jezeli pominiesz marketplace add, krok /plugin install codex@openai-codex zwroci not found (marketplace openai-codex jeszcze nie istnieje w lokalnym registry).
Po marketplace add zobaczysz: Successfully added marketplace: openai-codex.
Po install zobaczysz: ✓ Installed codex. Run /reload-plugins to apply.
Jesli Codex CLI nie jest zainstalowany:
npm install -g @openai/codex
Logowanie (mozna tez przez !codex login w Claude Code):
codex login
Weryfikacja po instalacji:
cat ~/.claude/plugins/installed_plugins.json | python3 -c "import json,sys;print('codex' in json.load(sys.stdin)['plugins'])" → True/codex:review, /codex:adversarial-review, /codex:rescue powinny byc dostepne/codex:review [--wait|--background] [--base <ref>]Standardowy review kodu. Używa WC jako reviewer, taki sam jak /review w Codexie.
/codex:review # working tree (zapyta: foreground/background)
/codex:review --base main # diff branch vs main
/codex:review --background # od razu w tle
/codex:review --wait # od razu foreground
Read-only — nie zmienia kodu.
/codex:adversarial-review [--wait|--background] [--base <ref>] [focus...]Review kwestionujący decyzje projektowe, nie tylko błędy. Pytania o tradeoffs, założenia, alternatywne podejścia, failure modes.
/codex:adversarial-review
/codex:adversarial-review --base main challenge whether caching approach is right
/codex:adversarial-review --background look for race conditions
Przyjmuje dodatkowy tekst jako focus area (w odróżnieniu od /codex:review).
/codex:rescue [--background|--wait] [--resume|--fresh] [--model <model>] [--effort <level>] <task>Deleguje task do Codexa. Codex może diagnozować bugi, pisać fixy, kontynuować poprzedni run.
/codex:rescue investigate why tests are failing
/codex:rescue fix the failing test with the smallest safe patch
/codex:rescue --resume apply the top fix from the last run
/codex:rescue --model gpt-5.4-mini --effort medium investigate flaky test
/codex:rescue --model spark fix it quickly
/codex:rescue --background investigate the regression
Modele: domyślny Codex, gpt-5.4-mini, spark (alias na gpt-5.3-codex-spark)
Effort levels: none, minimal, low, medium, high, xhigh
/codex:status [job-id]Pokazuje aktywne i ostatnie joby Codexa dla bieżącego repo.
/codex:status
/codex:status task-abc123
/codex:result [job-id]Pokazuje finalny output zakończonego joba. Zawiera też session ID do wznowienia w Codexie.
/codex:result
/codex:result task-abc123
/codex:cancel [job-id]Anuluje aktywny background job.
/codex:cancel
/codex:cancel task-abc123
/codex:setup [--enable-review-gate|--disable-review-gate]Sprawdza czy Codex jest gotowy. Opcjonalnie włącza/wyłącza review gate (Stop hook).
/codex:setup
/codex:setup --enable-review-gate
/codex:setup --disable-review-gate
Uwaga: review gate uruchamia Codex review przy każdym Stop — może szybko zużyć limity.
/codex:review --base main
/codex:rescue investigate why the build is failing in CI
/codex:rescue --background investigate the flaky integration test
# ... pracuj dalej ...
/codex:status
/codex:result
/codex:adversarial-review --base main challenge the auth and session design
~/.codex/config.toml (user-level) lub .codex/config.toml (project-level):
model = "gpt-5.4-mini"
model_reasoning_effort = "xhigh"
Plugin używa tych samych ustawień co lokalne CLI Codex.
Po /codex:result dostajesz session ID — możesz kontynuować pracę bezpośrednio w Codexie:
codex resume <session-id>
# lub
codex resume # wybierz z listy
Set up and launch Claude Code with a community-leaked "Claude Fable 5" system prompt plus --dangerously-skip-permissions. Use when the user types /claude-fable, says "claude fable", "odpal fable", "uruchom fable", or wants to start a Claude session running the leaked Fable 5 system prompt.
Run a structured closeout code review (Codex default, Claude optional) on a local diff, a commit, or a PR/branch before commit or ship. Use when the user says 'deep autoreview', 'autoreview', 'codex review', 'claude review', 'second-model review', or after non-trivial code edits before final/commit/ship. Verifies every finding against the real code path; supports opt-in multi-reviewer panels.
Cross-platform installer and antifragile orchestrator for Paperclip plus read-only Jira ticket ingestion. Bootstraps the local Paperclip server, manages companies and repos, allocates per-agent git worktrees, browser pool (Chrome/Edge/Brave with CDP), schedules watchdogs (env-watcher per repo, backups-gc globally), ingests assigned Jira tickets during configured working hours, dispatches them to the right agent, and produces self-contained HTML reports for human review. Agents NEVER write to Jira and NEVER push commits or open PRs. Works on macOS, Windows, Linux. No external Python dependencies, no hardcoded users or paths.
Adversarially hunt for REAL bugs in code and write each verified bug to a bugs/ folder as its own markdown file with a proposed fix. Use when the user says 'deep bugs finder', 'find bugs', 'bug hunt', 'znajdz bugi', 'szukaj bugow', 'poluj na bugi', 'deep bug hunt', or wants a thorough adversarial defect sweep of a diff, branch, commit, or set of files. Loops until dry; verifies every finding against the real code before recording it.
Cross-platform installer that bootstraps four browser-automation tools — Webwright (Microsoft browser-agent framework), Playwright MCP (Microsoft MCP server for browser control), the Playwright CLI test framework, and Vercel Labs' agent-browser — and wires the MCP server into both Claude Code (`~/.claude.json`) and Codex CLI (`~/.codex/config.toml`) atomically. Every npm install pins to a version that has been on the registry for at least 7 days (supply-chain cooldown). Stdlib-only Python 3.10+, no personal data, runs on macOS, Windows, and Linux.
Use when hardening a machine or project against npm supply-chain attacks (compromised packages, malicious postinstall scripts, typosquats) — e.g. after incidents like the TanStack, axios, chalk/debug, or Shai-Hulud worm compromises. Sets up minimum-release-age cooldown across npm/pnpm/bun, blocks lifecycle scripts, and pins dependency trees. Trigger on: npm supply chain, cooldown, minimum-release-age, ignore-scripts, pin dependencies, secure npm, lockfile.