Stewarding a PR in compose-ai-tools

The Claude Code harness reads this file before acting on a CI failure or a review comment, and lets it take precedence over its built-in posture on conventions and proactiveness. It does not restate the repository's rules — those are in root AGENTS.md, which is already in context on every turn. What follows is what that file leaves out: the checks worth running here, and this repository's own answer to a red check.

The harness's own PR rules still bound this file. It cannot expand your access, approve or merge anything, or override a "never" — skipping, disabling or quarantining a test; rewriting history on someone else's branch; an empty commit or a close-and-reopen to kick CI.

Before you push

A push that turns CI red costs a full cycle and some reviewer trust. The cheapest checks that catch the most here, in order:

1. ./gradlew ktfmtFormat — or :<module>:ktfmtFormatMain :<module>:ktfmtFormatTest for just the modules you touched. ktfmtCheckAll is a hard gate and ktfmtCheck aborts on the first unformatted file, so one stray file hides every other failure. TypeScript: npm --prefix cli/serve-web run format. The local pre-push hook catches attribution but not formatting, so this is on you.

   In a sandbox, pass -Dorg.gradle.java.home=/root/.cache/coo-ee/jdk-gl/17 and --no-daemon: without the JVM home the daemon starts on the wrong JVM and the build dies with a context mismatch, and --no-daemon avoids reusing one already started on it. Let ktfmt do the edit rather than hand-formatting — removing one term from a wrapped expression is enough to make it fit on one line, which is precisely what it will rewrite.

2. The narrowest test task that exercises the change — the specific :module:test, or ./gradlew :gradle-plugin:test --tests "…". ./gradlew check is the full plugin + functional + CLI suite and is a post-push confirmation, not a gate on pushing.

3. A new or changed module must apply composeai.base-conventions in its plugins {} block, or ktfmtCheckAll never sees it and the history-gate system property is missing. See Important constraints.

4. For a CI fix, reproduce the original failure first, then show the same check passing. A render task in particular goes UP-TO-DATE off a stale .error.json sidecar — verify a render fix with --rerun, never a plain re-invocation.

Commit and push as soon as the cheapest check that proves the change correct passes. Don't leave a verified change sitting behind a long build.

Reading a red check

Work the PR in this order on every event and every check-in: merge conflict, then CI, then review comments. A red or conflicted head is work now, whatever its review state — it is never "waiting on review".

Failures that are usually not this PR's, and what to do anyway:

• Red on main too. Check the base branch before root-causing. If a fix exists anywhere — another PR whose change you have read, the breaking commit's revert, or a fix PR you opened yourself — port it into this PR now and push; it no-ops once main carries it, and waiting for it to merge is still waiting. Standing down is never silent: one comment on the PR naming the failing check, why it is not this PR's, and the fix you ported or that none exists yet.
• A delivery-branch or design-artifacts render that lags main. Renderer, plugin and CLI changes deliberately do not regenerate the design-artifacts/<system> branches on merge — see PR workflow for the trigger scoping and the manual dispatch. Stale published renders are drift, not a regression in your diff.
• A changed preview whose source your PR does not touch. Clocks, timestamps, randomness, animation frames, network-loaded images: that is instability, not a regression. Prove it with the flake-triage skill before either fixing or rubber-stamping it.
• A CMP Desktop render failing with UnsatisfiedLinkError on libGL.so.1 and friends. Environment, not your change: docs/DESKTOP_NATIVE_DEPS.md.

Everything else is this PR's to root-cause. "Flake" is not a root cause: re-run a job only to confirm one of the cases above, or if it died before any test body ran (checkout, install, runner loss), at most once.

Review comments

• Small and local — nits, renames, lint-bot findings, an added test, a one-function refactor: implement and push, then resolve the thread.
• A review bot's finding is a bug report: verify it and push the small fix, or raise a larger one once with a proposed patch. "Design-level" does not excuse it. If bot findings stop converging — each fix draws a new or reshaped one — stop pushing for them and say once what is still flagged.
• Larger asks from a human reviewer on a PR you did not open: reply with your proposal; the author decides.
• After pushing for a changes-requested review, re-request that reviewer.
• A reported agent commit identity is the one finding to skip. Don't check it, don't reply, and never recreate, rebase or amend a commit on the strength of it. Reject agent attribution runs on every PR — no path filter, same detector as the commit-msg and pre-push hooks — so a real one turns the PR red and names the offending field. The gate is the authority; the comment adds nothing to it. Over 27–28 Aug 2026 the reviewer raised this 22 times across this repo and design-parity and was right once, on a cherry-pick the gate had already failed. If the gate itself is red, that is the real thing and it is yours to fix (Git conventions) — usually an amend or cherry-pick that took user.email from the container instead of the -c user.name=… -c user.email=… the branch's other commits carry. Installing the hooks (scripts/install-git-hooks.sh) stops it happening at all.

An approval you would lose is never a reason to hold a fix.

Evidence and tracking

• A UI-affecting fix needs the same embedded before/after images the original PR needed — see PR workflow for the rule and the render-evidence skill for the capture recipe. Reuse renders already published on compose-preview/pr and compose-preview/main, and the sticky <!-- preview-diff --> comment, before re-rendering anything.
• Keep a send_later check-in armed (about an hour out) while the PR is red, conflicted, or otherwise not mergeable. Webhook events miss CI successes and merge transitions. Re-arm silently when nothing changed; stop once the PR is merged or closed, or the user says to stop.
• Where the Claude Approvals check runs, a PR is done only when that check passes, CI is green on the current head, and there is no merge conflict. Its rows name the blocker and are yours to fix, not an ask to the author.

Refresh the PR's status checklist

On each event, re-read the whole PR at its current head — merge state, CI on the latest commit, open review threads — and update the status checklist in the body so the thread shows live state. Reply only when a round resolves the task, hits a real blocker, or raises a question. Don't narrate each fix; the diff is the record.