Delegate coding work to the OpenAI Codex CLI for repository changes, reviews, and focused fixes.
Install
mkdir -p .claude/skills/codex-open-curiosity && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/16131" && unzip -o skill.zip -d .claude/skills/codex-open-curiosity && rm skill.zipInstalls to .claude/skills/codex-open-curiosity
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Delegate coding work to the OpenAI Codex CLI for repository changes, reviews, and focused fixes.About this skill
Codex CLI
Use the OpenAI Codex CLI when the user wants an autonomous coding pass in a git repository: feature work, bug fixes, refactors, review passes, or isolated experiments.
Prerequisites
- Install:
npm install -g @openai/codex - Authenticate with the Codex CLI login flow, or provide a valid OpenAI API key.
- Run inside a git repository. For scratch work, create a temporary repo first.
- Use
pty=truefor interactive Codex commands.
Codex CLI OAuth usually lives under ~/.codex/auth.json. Do not assume Codex
is unauthenticated only because OPENAI_API_KEY is missing.
When to Use
- User asks for a second coding agent to implement, review, or investigate.
- Work can be scoped to a repository, branch, or worktree.
- A task benefits from a separate autonomous pass while Gini keeps the main approval, audit, and trace path.
When NOT to Use
- User only needs a small direct edit that Gini can do faster.
- The working directory is not a git repo and the user did not ask for scratch prototyping.
- The request involves secrets, credential files, or destructive commands without explicit approval boundaries.
Quick Tasks
Run a focused one-shot from the target repo:
codex exec "Find why the dashboard test is flaky, patch the smallest fix, and run the targeted test."
Review local changes:
codex exec "Review git diff --stat and git diff for bugs, regressions, and missing tests. Return findings only."
Scratch prototype in a disposable repo:
tmp=$(mktemp -d)
cd "$tmp"
git init
codex exec "Create a tiny Bun CLI that parses a JSON file and prints a summary."
Long-Running Work
For long tasks, launch in the background with a PTY, then poll the process rather than starting duplicate Codex runs.
codex exec --full-auto "Refactor the settings loader. Keep the diff narrow and commit when tests pass."
Use --full-auto only when repository writes are expected and the user has
accepted the risk. Avoid --yolo unless the user explicitly asks for it.
Worktree Pattern
Use worktrees for parallel issue fixes so each Codex instance has its own branch and filesystem.
git worktree add -b fix/runtime-health /tmp/gini-runtime-health main
git worktree add -b fix/skills-copy /tmp/gini-skills-copy main
cd /tmp/gini-runtime-health
codex exec --full-auto "Fix the runtime health regression and run the related tests."
cd /tmp/gini-skills-copy
codex exec --full-auto "Add coverage for bundled skill loading changes."
After each run, inspect the diff, run the expected checks, and remove completed
worktrees with git worktree remove <path>.
PR Review Pattern
git fetch origin main
codex exec "Review the current branch against origin/main. Prioritize correctness, security, and missing tests."
For a GitHub PR:
gh pr checkout 42
codex exec "Review this PR against origin/main. Return only actionable findings with file and line notes."
Rules
- Keep Codex scoped to a git repo or throwaway initialized repo.
- Prefer
codex exec "prompt"for one-shot tasks. - Use
pty=truefor interactive Codex sessions. - Prefer
--full-autoover--yolofor write-heavy work. - Monitor long runs before launching another agent on the same branch.
- Inspect Codex changes before trusting or committing them.