imprint: Skill for Codex

PR Review Convergence

Drive the branch to a genuinely reviewable state. Treat “open the PR” as an intermediate step; completion requires green checks and no unresolved actionable review findings.

1. Establish scope and guardrails

1. Read repository instructions such as AGENTS.md, CLAUDE.md, contributing guidance, and relevant local skills completely.
2. Inspect git status, the current branch, remotes, upstream, source branch, complete diff, and recent history. Preserve unrelated user changes.
3. Never use destructive resets or force-push without explicit authorization. Never bypass hooks. Stop if the diff contains credentials, recordings, personal data, or another public-repository violation.
4. Match mutations to the user's authority. A review-only request permits inspection and local validation, not commits, pushes, or PR creation. Only proceed through remote steps when the user explicitly asked to prepare, create, push, or update the PR/MR; otherwise stop after the review handoff.
5. Create a short plan covering rebase, validation, review rounds, commit, push, PR creation, and check monitoring.

2. Rebase safely before review

Skip this section for review-only requests: inspect the branch against its declared source without stashing, rebasing, committing, or otherwise mutating the user's worktree.

Fetch the source branch, normally origin/main. If the worktree is dirty, stash tracked and untracked changes under a descriptive name, rebase, restore the stash, and verify every change returned. Resolve straightforward conflicts without discarding either side; ask only when resolution changes product intent.

Record the rebased commit so later source-branch movement can be detected.

3. Validate the implementation

Run repository-prescribed checks before asking reviewers to inspect avoidable failures. At minimum for Imprint:

bun run lint
bun run typecheck
bun test
git diff --check

Run focused tests for changed behavior. If the website changed, build it and visually inspect desktop and mobile layouts. Exercise important behavior live when safe and proportionate. Distinguish an unrelated flaky test from a real failure by reproducing it in isolation, then rerun the full gate; do not simply dismiss it.