Setup Harness

Set up harness

/implement carries a ticket through review to merge. Three things around it are properties of the repository rather than of the skill:

• Repository hooks — behavior invoked by agent and Git events
• Merge automation — what merges a pull request once its conditions hold, so the agent never merges its own work
• Worktree cleanup — what a repository releases when a merged worktree goes away

This is a prompt-driven skill, not a deterministic script. Explore, present what you found, confirm with the user, then write. The configuration is the working files themselves; this skill writes no documentation describing them.

Process

1. Explore

Read what exists; assume nothing:

• git remote -v — a GitHub repository? Which one? Which default branch?
• .codex/hooks.json and .cursor/hooks.json — hooks this repository already runs, harness-owned or not.
• .mergify.yml — present? What does it already merge on?
• Who merged the last few pull requests: gh pr list --state merged --limit 5 --json number,mergedBy. A bot login means merge automation is already installed and running; only your account means nothing merges on its own.
• .github/workflows/ — the job names that report as checks. These are the candidates for named merge conditions.
• .agents/hooks/cleanup — this skill's prior output.
• git worktree list and the repository's own scripts — does a linked worktree receive resources of its own, such as a database, a port, or a generated environment file? A checkout hook under core.hooksPath that provisions anything is the strongest signal.
• How the repository runs its checks, from AGENTS.md, CLAUDE.md, package.json scripts, a Makefile, or scripts/. /implement opens on this command, so confirm it is findable rather than recording it anywhere.

2. Wire the hooks

Run this skill's scripts/install_hooks.py. The script owns the complete agent and Git hook installation and is idempotent, so ask nothing here and run it on every invocation. Keep hook membership and wiring in the installer; this step stays unchanged as hooks evolve.

3. Present findings and ask

Summarise what is present and what is missing. Then take the sections in order — one section, one answer, then the next. Lead each section with the recommended answer so the user can accept it in a word, and skip a section outright when exploration already settled it.

On a repeat run the recommended answer is whatever the repository already carries: propose keeping it, and write a file only where the user asks for a change. Running this skill again leaves a hand-tuned configuration as it is.

Section A — Merge automation.

Explainer: /implement makes every merge condition hold and waits. Something else does the merging. Mergify is the default here because GitHub's own required checks and auto-merge need a paid plan on a private repository.

Recommended: Mergify, gated on the checks this repository actually requires. Take the job names found in .github/workflows/ and propose the ones that must pass — usually the test or integration job, not a job that only runs for some changes. Write mergify.yml, substituting the default branch and one check-success line per required check.

Offer these alternatives:

• Count checks instead of naming them — replace the check-success lines with #check-failure = 0 and #check-pending = 0. It survives a workflow being renamed, at the cost of a race: a pull request whose checks have not been created yet satisfies both counts. Propose it only when the required set genuinely varies.
• No automation — the reviewer merges by hand while /implement waits.

When .mergify.yml already exists, change only the rule this skill owns and leave the rest of the file as it stands.

Mergify merges nothing until its GitHub App is installed on the repository. If exploration found no bot merges, say so and point the user at https://github.com/apps/mergify.

Section B — Worktree cleanup. Skip this section entirely when exploration found no per-worktree resources — /implement already removes the merged worktree and its branch, and a repository that owns nothing else needs no hook.

Otherwise ask one question:

Which command releases what a worktree owned, once the worktree is gone?

Write it as an executable .agents/hooks/cleanup, using cleanup as the starting point with its RELEASE_COMMAND placeholder replaced by the answer. /implement runs it after removing a merged worktree and reports its output, so have the command name what it released rather than working silently.

4. Confirm and write

Show the user each file you are about to write and let them edit it first. Then write them, and make .agents/hooks/cleanup executable.

5. Done

Tell the user to commit the generated files so the tooling stays a property of the repository, and that Codex needs /hooks to trust the project hook once while Cursor reloads .cursor/hooks.json on save. Updating the tooling itself is npx skills@latest update; this skill is worth running again to change how the repository merges or what it releases, and safe to run again for any reason.