LLMSkillHub
mirroredCopied from carlos-algms/agentic.nvim under MIT. carlos-algms has not claimed this page, and is not affiliated with LLM SkillHub.

carlos-algms / agentic-pr-workflow

Draft-first PR flow and single-commit CodeRabbit review rounds. Use before opening a PR, marking one ready for review, or handling a CodeRabbit review round and its findings.

0
SKILL.md
3.2 KB

Pull request workflow

CodeRabbit reviews every push to a non-draft PR and has a rate limit. Each rule
below exists to keep the number of pushes, and so the number of reviews, low.

Opening a PR

  1. Open every PR as draft. Iteration pushes on a non-draft PR each trigger a
    review and hit the rate limit

  2. The PR title follows Conventional Commits. This repo only squash-merges: the
    title becomes the commit subject and the description becomes the commit body

  3. Never stage docs/plans/, docs/superpowers/, or rules-report.md
  4. Flip to "ready for review" only when the user asks

A review round is one batch

  1. Read every finding. Inline threads are not all of them: read the full review
    body too, including every collapsed <details> block ("Outside diff range",
    nitpicks). The "Actionable comments posted: N" header counts inline threads
    only, and some passes with findings print no header at all

  2. Fix each valid finding. Skip findings that are false or not worth the
    change. Take the diagnosis, re-derive the patch: CodeRabbit's suggested code
    has broken this repo's banned-call rules, so check it against the routed
    docs before you apply it

  3. For each fixed finding, ask whether a written rule would have prevented it.
    If yes, apply skill agentic-learn in the same change

  4. One commit and one push, at the end of the round. The commit follows its
    normal rules. Answering a review never skips them

  5. Only after the push, reply to the skipped findings. CodeRabbit then reviews
    the new commit once

  6. A resolved or "addressed" thread is not proof the fix landed. CodeRabbit
    has marked a finding addressed before the fix commit existed. Trust it only
    when a later review pass covers the fix commit

Merging

Do not merge until CodeRabbit's review of the last push has completed: its
summary comment exists, with or without findings. A PR merged within minutes of
flipping to ready can carry no review at all.

Out-of-diff findings

A finding in the review body has no thread and no resolved state, so nothing in
the GitHub UI tracks it. Fix it or skip it, then post a top-level comment that
starts with @coderabbitai and states what you did and why. The comment is the
only visibility and history: CodeRabbit may never raise an out-of-diff finding
again in later reviews. Do not edit the PR description for it.

Rejecting a finding

  • Reply in its thread, or in a top-level comment for an out-of-diff finding
  • Start with @coderabbitai. Without the tag, a top-level reply goes nowhere
  • State the rule for the whole codebase, not for this line. CodeRabbit stores
    your wording as a learning, and narrow wording lets the same claim fire again
    on another file

  • A finding that correctly cites a repo rule is still false when only a test
    double can reach the scenario. Name the production path, or reject it

  • A rejection on scope alone ("pre-existing") does not answer whether the
    finding is correct. Fix it, or open a tracked follow-up in the same reply