Copient Skills

Public, project-agnostic developer skills from Copient AI.

Two so far, and they are one idea reviewed by two different models: a review → fix → re-review loop on your branch, run locally, that keeps the review's bulk out of your working context. No PR, no push, no GitHub round-trip.

Skill	Reviewer	Runs on
codex-review-loop	OpenAI Codex, via the codex CLI	Any agent that can run bash
pr-review-loop	Claude, in an isolated subagent	Claude Code only

Run both before a PR — they surface different classes of problem.

pr-review-loop's host requirement is not packaging trivia. Its isolation is the Task subagent; on an agent without subagents the review runs in your main context and the guarantee is gone, silently.

Install — pick one

npx skills@latest add Copient-ai/skills --global   # recommended

Editable copies, any agent, invoked as codex-review-loop / pr-review-loop. It asks which agents to install for; add -a claude-code -s '*' -y to skip that.

/plugin marketplace add Copient-ai/skills          # Claude Code only
/plugin install copient@copient-skills

Managed and read-only, invoked as copient:codex-review-loop — the plugin name becomes a prefix.

Do not install both. They duplicate rather than layer, and a project-level copy shadows a user-level one, so the copy you edit may not be the copy that runs.

Prerequisites

codex-review-loop needs the Codex CLI on your PATH, codex login done, and the repo trusted in ~/.codex/config.toml. Untrusted, the review stops for approval on every file it reads.

pr-review-loop needs Claude Code and nothing else.

Verify

These gate a push, so "it looks installed" is not enough.

bash ~/.claude/skills/codex-review-loop/scripts/codex-review.sh --version
# codex-review.sh 1.0.0   — compare against CHANGELOG.md

bash ~/.claude/skills/codex-review-loop/scripts/test-codex-review.sh
# ALL PASS                — offline, no Codex call needed

Adjust the path for your install: .agents/skills/… for agents other than Claude Code, ${CLAUDE_PLUGIN_ROOT}/skills/… under the plugin.

Using them

From a feature branch with your work committed:

/codex-review-loop
/pr-review-loop

Each reviews, you fix, it re-reviews, and it stops when the code stops changing. Blocking findings always get fixed. Nits are optional but ambitious, and one you decline is remembered so a later round cannot re-raise it. Three iterations, then a verification round that reviews without fixing. It pushes once, on convergence, never while escalating.

A clean verdict is not automatically convergence — a review can come back clean while listing nits you have neither fixed nor declined. And the rule underneath all of it: an iteration whose check did not run is not a completed iteration.

Telling the loops how to test your project

They assume no task runner. Commands resolve in this order:

1. .review-loop.json at the repo root wins:

   { "test": "just test-module", "lint": "just precommit" }

2. Otherwise, detection from the files your fix touched — justfile, package.json, Makefile, pyproject.toml/pytest.ini/tox.ini, Cargo.toml, go.mod, .pre-commit-config.yaml. A fix spanning two runs both, re-derived after each round. Picking one root marker and keeping it would run the JavaScript tests for a Go-only change and call it verified.

3. Nothing resolves → it stops and asks you, rather than skipping the check.

Pin a config when detection would guess wrong — a justfile whose recipes aren't named test/test-module/precommit/check is the common case.

Permissions

Two prompts are deliberate.

The helper script prompts on first use. Only the plugin copy is pre-approved, because a project-level install puts the script inside the repo being reviewed, where the branch under review could rewrite it and have it run before anything reviewed it. Silence it only for a copy outside any checkout:

{ "permissions": { "allow": [
  "Bash(bash ~/.claude/skills/codex-review-loop/scripts/:*)",
  "Bash(just test-module:*)"
] } }

Your test command prompts too — there's no way to know it in advance. Add it alongside, as above.

Approving a prompt is fine. Skipping the check to avoid one is not.

Staying current

An old copy is a safety problem, not a missing-features one: the bugs fixed in codex-review.sh are false-CLEAN bugs, where the tool reports approval for a branch it never parsed.

npx skills@latest update --global      # or: /plugin marketplace update copient-skills

Updates overwrite your edits silently — no warning, no prompt, no diff, just ✓ Updated. Keep anything worth keeping in .review-loop.json, your own settings, or a PR here.

If ~/.claude/skills is a symlink, check your links after any real update. The installer keeps files in ~/.agents/skills/ and links them in relatively; against a symlinked directory those resolve nowhere and the skills stop loading with no error.

test -e ~/.claude/skills/codex-review-loop/SKILL.md && echo ok || echo BROKEN
cd ~/.claude/skills && for s in codex-review-loop pr-review-loop; do
  rm -f "$s" && ln -s "$HOME/.agents/skills/$s" "$s"
done

Removing them

npx skills@latest remove codex-review-loop pr-review-loop
# or: /plugin uninstall copient@copient-skills

Nothing else to undo — no config is written outside the skill directories.

Rolling out to a team

Delete any copies vendored into your repos first, or they shadow whatever people install and nothing says which one ran. Pick one install path for everyone, since the invocation name differs between them. Pin .review-loop.json in the repos you review most.

Contributing

Everything here is public. A skill needing Copient context — repo knowledge, production data, credentials — does not belong in this repo.

Do not simplify codex-review.sh. It looks over-engineered. Each guard traces to a reproduced false-CLEAN, and together they mean: never report approval for something you did not parse. UNPARSED exists so an unrecognized transcript fails loudly. Every parser change needs a fixture proven red against the previous version before it goes green.

CI runs the parser suite on Linux, macOS, and stock macOS bash 3.2, plus shellcheck, on every push.

Layout

skills/            agent-neutral source of truth
.claude-plugin/    Claude Code adapter

MIT — see LICENSE.