pull-requests

Branch naming

Branch prefixes: feature/, fix/, refactor/, perf/, docs/, then a lowercase kebab-case title. Never codex/.

Title

The title names the one main change and why it matters, in words a person reads at a glance. The rest goes in the description.

BAD

fix (preview): set mermaid svg width to max-content

GOOD

fix (preview): draw diagrams at their natural size instead of stretching them

Description

The diff already says what changed. The description is for what the diff can't show: why the bug happens, why this fix over the alternatives, what you tried that didn't work.

• Summarise, don't enumerate. Don't account for every change, trace code paths by function name, or list each case you checked — that inventory is the diff. Give the one-level-up story: what was broken, what it does now, what to watch out for. Two or three short paragraphs, about 150 words, even for a large change. A dependency bump or a typo fix gets one line.
• Prose, not bullets. Reasoning is causal, and bullets break it into disconnected facts. Use a list only for genuinely parallel items. No section headers on anything under a few paragraphs.
• Name a symbol only when the reviewer needs it to find the change. Compress the verification to a clause ("checked in the app and Quick Look").
• Flag what the reviewer can't see in the diff: a behavior change, a trade-off you took on purpose, a decision you want a second opinion on.
• Agent-written descriptions run long and bury the point. Trim before opening.

BAD

Changed MarkdownHTML+Stylesheet.swift to remove width: 100% from .mermaid svg and added max-width: 100%. Updated EditorHTML.swift to match. Verified with flowchart, sequence, gantt, pie, and class diagrams in light and dark mode, in the app and Quick Look, at three window widths.

GOOD

Small diagrams were stretched to fill the column, so a three-box flowchart rendered with huge text. They now draw at their natural size and only shrink when wider than the column. Checked in the app and Quick Look.

Creating a PR

Do these without being asked, in order:

1. Run /simplify on the diff. Fix what holds up; skip a finding only with a reason you can state.
2. Run the checks CI runs (.github/workflows/swift-tests.yml), plus an app build:

   swift test --package-path tests/swift-tests --parallel
   python3 tests/test-mode-transitions.py
   python3 tests/test-checkbox-saves.py
   xcodebuild -project md-preview.xcodeproj -scheme md-preview -configuration Debug build

   Fix failures before opening the PR. Do not open a PR on red checks. If a test also fails on main on this Mac, it is not from your change (some WebKit tests depend on the macOS version; CI runs macOS 15): say so in the PR description instead of changing the test.
3. Check git diff for a DEVELOPMENT_TEAM hunk in project.pbxproj and revert it (see "Signing & secrets" in AGENTS.md), and grep docs for claims the change made false.
4. Open the PR ready for review (gh pr create), not as a draft. Say in the PR description what you skipped from step 1, and why.
5. Wait for the bot reviews and handle them (below). The PR is not handed off until both bots have reviewed the latest commit.

Bot reviews: CodeRabbit and Greptile

Both bots (coderabbitai[bot], greptile-apps[bot]) review every PR on their own, usually 3–5 minutes after it opens. Wait for them; don't hand off the PR or end the turn first. Poll in the background (a Monitor or a background shell), not with foreground sleeps. Start it only after git push returns — started alongside the push, it reads the old head commit and waits on that:

pr=<N>; repo=pluk-inc/markdown-preview
head=$(gh pr view $pr --json headRefOid -q .headRefOid)
for i in $(seq 1 10); do
  cr=$( { gh api --paginate repos/$repo/issues/$pr/comments -q '.[] | select(.user.login=="coderabbitai[bot]") | .body'
          gh api --paginate repos/$repo/pulls/$pr/reviews -q '.[] | select(.user.login=="coderabbitai[bot]") | .body'
        } | grep -c "and $head")
  busy=$(gh pr view $pr --json comments -q '[.comments[] | select(.author.login=="coderabbitai")
         | select(.body | test("Come back again in a few minutes"))] | length')
  roots=$(gh api --paginate repos/$repo/pulls/$pr/comments \
    -q '.[] | select(.in_reply_to_id == null) | .pull_request_review_id')
  gr=$(gh api --paginate repos/$repo/pulls/$pr/reviews \
      -q ".[] | select(.commit_id==\"$head\" and .user.login==\"greptile-apps[bot]\") | \"\\(.id) \\(.body | length)\"" |
    while read id len; do { [ "$len" -gt 0 ] || grep -qx "$id" <<<"$roots"; } && echo x; done | wc -l | tr -d ' ')
  [ "$cr" -gt 0 ] && [ "$busy" -eq 0 ] && [ "$gr" -gt 0 ] && { echo reviewed; exit 0; }
  sleep 60
done; echo "timeout: coderabbit=$cr greptile=$gr"

The two bots show a finished review in different ways, so the loop checks each one on its own terms:

• CodeRabbit reviews incrementally. A round with findings files a review; a clean round files nothing and only updates its summary comment ("No actionable comments were generated in the recent review"). Both name the range they covered as "between <old> and <head>", so the loop looks for the head commit there. While a review runs, the summary says "Come back again in a few minutes".
• Greptile counts only with a review on the head commit that has a body or opens a new thread. A bot's reply in a thread also files a review on the head commit, with an empty body and no new thread — that is not a review of the new code.

On a timeout, ask for the missing review in a PR comment: @coderabbitai review or @greptileai review, then run the loop one more time. If it times out again, tell the user which bot is missing rather than waiting on. Greptile often skips later pushes and may ignore the request; CodeRabbit answers "Already reviewed the last commit" when its clean round already counted.

When the reviews are in, handle them with the resolve skill — it has the thread query and the rules for reading bot findings:

• Verify each finding against the current code. Fix the valid ones with the smallest change; skip the rest with a reason. Bots are often wrong about WebKit, AppKit, and sandbox behavior — check before you change anything.
• Reply to every bot thread with what you did or why you skipped it, then mark it resolved.
• A push starts a new review round. Wait again and repeat until a round has no new valid findings, then hand off with one line per finding: fixed or skipped, and why.