openclawandroid OpenClaw PR Maintainer
Use for the requested OpenClaw issue/PR operation. Root AGENTS.md owns repair,
product, source-trust, and approval policy; this skill owns the GitHub workflow.
A pasted ref is context, not permission to publish or expand the task.
Select the scope
- Review, triage, or list: read-only. Explain the change, actionable findings, useful evidence, and material uncertainty. Do not edit, assign, label, comment, or close unless separately authorized.
- Fix only: investigate, repair locally, and validate. Publishing still requires ship/land or equivalent scoped authority.
- Land/ship or autonomous repair: finish the authorized scope through current source proof, review, exact-head CI, native merge, and remote verification. Preserve contributor credit and unrelated work. Routine steps need no repeated permission; security, schema, product, release, and other root gates remain.
- Queue/discovery: read triage. Batch live reads and shared discovery; do not turn one named PR into an unsolicited queue sweep.
- Author or regression attribution question: read author and provenance research. Account age and contribution-history queries are conditional, not a prerequisite for every PR.
The lead remains hands-on. Use bounded independent workers where they add useful evidence; serialize shared checkout/ref mutations and GitHub writes. Do not create user-owned app tasks to hide internal subtasks. Preserve a stalled worker's patches, claims, and proof before taking over or reassigning its work.
Inspect the actual item
Start with git status -sb; preserve unrelated changes. Use available local
gitcrawl data for related work when useful, then bare PATH gh with narrow
JSON fields for live decisions. Stale/missing archives fall through to gh;
do not broadly sync archives just to begin. PR source comes from gh pr view/diff and the checkout, not web search.
gh pr view <pr> --repo openclaw/openclaw \
--json number,title,state,body,author,assignees,baseRefName,headRefName,headRefOid,files,comments,reviews
gh issue view <issue> --repo openclaw/openclaw \
--json number,title,state,body,author,assignees,comments
Check assignment before deep work. Mention another assignee and treat a fresh assignment as active ownership unless the user directs otherwise. Assignment alone is not a block when stale or explicitly overridden. Claim an unassigned named item only under assignment or landing/autonomous-resolution authority, using the verified authenticated login. Never remove others' assignments without direction. Account/activity research belongs to its conditional route.
Establish the outcome
Read the affected owner, entry points, callers, relevant siblings, tests, docs, history, and current main as needed to support a conclusion. Inspect dependency contracts directly when they determine behavior; apply the root's personal Codex-source inspection requirement when relevant. Investigate contrary evidence before defending a verdict. Do not fill an arbitrary evidence matrix or invent a rejected alternative merely to satisfy a form.
- Already fixed: prove the original behavior is repaired on current main, with the canonical commit/PR and matching source or runnable proof. Under closure authority, comment that evidence and close; otherwise report it.
- Real defect: reproduce it, repair its owning invariant, and prove the repaired path and affected siblings. Prefer a coherent simplification over a symptom guard. Size and LOC are context, not correctness gates.
- Existing useful PR: improve the editable branch rather than replacing it unnecessarily. If it cannot safely be updated, create an authorized credited replacement before closing the source. Never squash a contributor-owned PR after replacing its ancestry; use a maintainer-owned replacement.
- Uncertain or product decision: state the concrete gap. Do not merge a speculative repair or automatically close work as out of scope.
Use $openclaw-testing to choose proportional proof and its source-trust/host
route. Trusted development proof defaults local; remote proof needs an environment
or isolation reason. Unavailable optional live proof may use meaningful boundary
proof with the limitation stated. Explicit live requests, external API contracts,
and changes whose risk requires authenticated execution keep their required proof.
Never describe mocks, skipped checks, or an older head as live evidence.
Review and publish
Before committing or landing nontrivial code, run $autoreview and resolve
verified actionable findings. Use it for a review-only request when that request
asks for independent autoreview; ordinary findings-only review does not need a
second reviewer by default. Do not add another mandatory CLI review pass.
Reopen review for substantive changes or unresolved concerns, not a patch-identical
rebase or a mechanical head change. Address real human/bot findings and explain
rejected ones; bot scores and Rank-up lists do not create separate obligations.
Use the current PR template. Lead with the plain-language problem and concrete
user impact; keep the explanation short and leave implementation inventories in
the diff or optional details. Keep important risks, migrations, required actions,
useful evidence, known gaps, and contributor credit visible and current. Do not
invent a user benefit for internal-only work. Explain material tradeoffs when they
matter; do not require universal LOC tables, provenance fields, or alternate-fix
essays. CHANGELOG.md is release-owned; user-facing release-note context stays
in the PR/commit. Omit agent transcripts unless explicitly requested.
Before public mutation, verify author/committer and authenticated writer identity. Use body files for shell-sensitive text. Post findings only under the authorized write scope, and update an existing relevant comment/body instead of duplicating it. Publication, media, exact-head CI, and native landing mechanics are in landing; read that reference before publishing or merging.
Verify the final merge/closure state and source rather than trusting a local summary. Report the problem, owner-level change, important proof and limitations, credit, and linked final state in concise prose. Record worthwhile follow-ups; do not manufacture another task after a bounded request is complete.