steward
Skill127 starsChanged 37 days agoApache-2.0View on GitHub
Drive a pull request on this repository to green — which fast checks to run before pushing, how to read a red check, and what to do about a review comment. Use when a CI failure, a review comment, a merge conflict, or a scheduled check-in arrives on a PR you opened or were asked to drive.
19 other projects use a version of this.
Adopted by
Most recently taken up by
- btraceio/btracetheir own version, first seen Nov 16, 2026
- fworks-tech/agenthoodtheir own version, first seen Nov 16, 2026
- ForeactLabs/OpenExecutivetheir own version, first seen Nov 15, 2026
- 5thlayer/skillworkstheir own version, first seen Nov 8, 2026
- ElcanoTek/fleettheir own version, first seen Nov 8, 2026
- jmagly/kintsugi-usbtheir own version, first seen Nov 8, 2026
Used with
Projects with it also use release.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
MicrosoftSkillrelease
Use ONLY when the user explicitly asks to cut/publish a Huabu desktop release or tag a version. Performs an outward action — bumps the version in apps/desktop/package.json, commits + annotated-tags vX.Y.Z, and pushes the branch + tag. That tag push auto-builds the macOS/Windows installers and publishes a GitHub Release. Provides `status`, `bump`, `tag-push`, and a one-command `run` (bump → tag-push).
Also in 3 projects that use this
Other versions19 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 19
- btraceio/btrace★ 6ksame name
- ForeactLabs/OpenExecutive★ 5.8ksame name
- SenteLabsAI/OpenExecutive★ 5.8ksame name
- richlander/dotnet-inspect★ 166same name
- mendixlabs/mxcli★ 139same name
- antiv/mate★ 111same name
- michelbr84/GarraRUST★ 19same name
- jmagly/kintsugi-usb★ 18same name
- jmagly/carbonyl-agent★ 16same name
- itsOmidKarami/kraft★ 13same name
9 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- site/llms.txtllms.txt
- flake-triageSkill
- render-evidenceSkill
The file
We split this skill into the rules it gives an agent
7 rules. 0 are common, 5 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 18 edited versions we read keep these.
5
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 50
Commit and push as soon as the cheapest check that proves the change correct passes.
- Distinctive: line 61
Red on
maintoo: Check the base branch before root-causing. - Distinctive: line 90
If bot findings stop converging — each fix draws a new or reshaped one — stop pushing for them and say once what is still flagged.
- Distinctive: line 94
After pushing for a changes-requested review, re-request that reviewer.
- Distinctive: line 95
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 …
- Very specific: line 102
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 …
- Very specific: line 117
Keep a send_later check-in armed (about an hour out) while the PR is red, conflicted, or otherwise not mergeable.
“Common” is measured: kept by at least 6 in 10 of the 18 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
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:
./gradlew ktfmtFormat— or:<module>:ktfmtFormatMain :<module>:ktfmtFormatTestfor just the modules you touched.ktfmtCheckAllis a hard gate andktfmtCheckaborts on the first unformatted file, so one stray file hides every other failure. TypeScript:npm --prefix cli/serve-web run format. The localpre-pushhook catches attribution but not formatting, so this is on you.In a sandbox, pass
-Dorg.gradle.java.home=/root/.cache/coo-ee/jdk-gl/17and--no-daemon: without the JVM home the daemon starts on the wrong JVM and the build dies with a context mismatch, and--no-daemonavoids 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.- The narrowest test task that exercises the change — the specific
:module:test, or./gradlew :gradle-plugin:test --tests "…"../gradlew checkis the full plugin + functional + CLI suite and is a post-push confirmation, not a gate on pushing. - A new or changed module must apply
composeai.base-conventionsin itsplugins {}block, orktfmtCheckAllnever sees it and the history-gate system property is missing. See Important constraints. - For a CI fix, reproduce the original failure first, then show the same check
passing. A render task in particular goes
UP-TO-DATEoff a stale.error.jsonsidecar — 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
maintoo. 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 oncemaincarries 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 thedesign-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-triageskill before either fixing or rubber-stamping it. - A CMP Desktop render failing with
UnsatisfiedLinkErroronlibGL.so.1and 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 attributionruns on every PR — no path filter, same detector as thecommit-msgandpre-pushhooks — 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 anddesign-parityand 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 tookuser.emailfrom 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-evidenceskill for the capture recipe. Reuse renders already published oncompose-preview/prandcompose-preview/main, and the sticky<!-- preview-diff -->comment, before re-rendering anything. - Keep a
send_latercheck-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.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

