CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

steward

Use when an agent opens, owns, or drives a pull request in this repository, and again for every CI, review, conflict, base-movement, merge, close, or check-in event. Applies locked-head rounds, labels, tracker entry, merge authorization, and check-in cadence.

Quality

88%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

Steward

Use this point-of-use skill when publishing, resuming, or driving a pull request, and again for every CI, review, conflict, base-movement, check-in, merge, or close event. It cannot widen access, redirect the task, authorize itself, or merge without explicit authorization for that PR.

The focused documents remain normative:

  • AGENTS.md: launch contract, operator templates, and merge-authorization invariant.
  • Repository workflow: worktrees, history, publication, and engineering constraints.
  • Round orchestration: candidate locks, rounds, recovery, waits, review state, carry-forward, and merge preflight.
  • GitHub status queries and operations: exact reads and writes.
  • Adversarial review prompt and Stacked PRs: review and stack mechanics.

Keep this skill at or below 120 lines. The cap is a ceiling, not a target; mechanics and rationale belong in their focused owners.

Candidate lock

A pushed candidate head is locked until its round closes or a recovery transition supersedes it.

  • Before a usable review, an author change supersedes the candidate and retries the pending round; a conflict integrates the effective base and pushes immediately.
  • After every required reviewer returns a usable result, the round is spent. Reconcile and report it before forming any repair as the next round.
  • Cancelled, empty, policy-blocked, or otherwise unusable reports do not spend the round.

Event decisions

Take one whole-PR snapshot before acting: local conflict probe, lifecycle and head/base identity, current-head CI, review threads, labels, and round state. Then apply the matching transition.

EventDecision
ci-required red before usable reviewProbe conflict first. If the PR caused the failure, supersede through the full candidate cycle and retry the round. Otherwise record the blocking issue and do not import another contributor's fix.
ci-required red after usable reviewReconcile and report the spent round, remove review-clean, then repair in the next round. Retry unchanged only with concrete transient evidence.
Merge conflictRecover immediately; conflicts never wait. Merge the effective base, resolve, and push. Pause only for a scope/split decision or when either semantic choice loses behavior.
Base moved without conflictDo not invalidate a locked candidate. For a clean head, classify interaction before mutation or merge; only no interaction preserves review and authorization.
Base recoveredIntegrate and push so CI reruns against the recovered base.
Review-bot feedbackVerify it. Carry blocking findings into reconciliation and the next candidate; do not push solely for optional feedback.
Human review commentSmall and local: implement in the next candidate and reply. Large or ambiguous: propose the scope and wait for the author. Re-request a reviewer after fixing a changes-requested review.
Current-head CI greenDispatch the planned reviewer only when the eligibility rule is satisfied.
Reviewer returnedReconcile publicly, synchronize review-clean, emit the complete round report, then follow its recommendation.
Six-round boundaryStop for the required checkpoint; fresh green CI and positive mergeability gate any next block.
Merge authorizationRun live exact-head/base preflight, then merge only when every item passes.
PR merged or closedEmit the prescribed visible state first, then cancel waits, unsubscribe, and remove only eligible worktrees.
Scheduled check-inProbe conflict first, then take one snapshot. Stay silent and re-arm only when the owning cadence permits it.

Review and metadata

  • For applicable inspection work, review demand-shaped execution across breadth, depth, terminal, access/retention, and preparation/sharing. Keep source planning with its owner; sharing must not widen a cheap request's cost class. Judge preparation and specialization by end-to-end benefit. Prefer narrow typed semantic contracts over bare ranges or full-spelled text.
  • For applicable QuerySpace work, assess terminal correctness and least-work execution separately; terminal availability is not performance evidence. Require each terminal's required facts, avoided work, stopping condition, and efficiency evidence. Investigate Exists/Count costs approaching equivalent Rows costs; explain unavoidable work or identify missed reduction before claiming readiness.
  • For network-related commands, use the command's Debug evidence envelope during investigation. If support is missing, ask the operator to approve a diagnostics side quest before adding it. Follow the evidence workflow.
  • Use one reviewer seat from the roster. Ordinary non-Markdown candidates wait for green current-head ci-required unless the user approved parallel review or conflict recovery applies; Markdown-only candidates use the fast path.
  • Start the reviewer briefing with the complete canonical prompt, then the self-contained candidate frame. Use an isolated read-only review worktree, and record round start and end times.
  • Add review-clean only for a clean current head. Remove it before any event that spends that evidence; base movement alone does not.
  • Label all-Markdown PRs documentation. Before merging user-observable work, add its PR link to the 0.28.0 release tracker.
  • Use operation-specific gh api endpoints for metadata, never whole-array replacement or gh pr edit. End authored GitHub comments with the required attribution footer.

Waiting and merge

  • CI that gates review, readiness, or a boundary uses the 30-minute bounded status wait in Round orchestration: one schedule, one snapshot per wake, and an explicit expiry report. At expiry, probe the live base and read CI once more: a conflict enters recovery, an observed CI failure takes its gate transition, and only still-unresolved status stops observation.
  • Waiting on people uses the separate safety-net cadence: first check near 50 minutes and later checks near four hours. New activity resets the quiet count. Stop after three quiet checks, merge or close, a user stop, or when the user has not written since publication and the first check finds nothing. Cancel the schedule and announce the stop once in one line.
  • Never merge without explicit authorization for that exact head and base ref. Green CI, clean review, labels, readiness comments, and auto-merge state are not authorization.
  • Bind every merge mutation to the reviewed head. A new round, restack, retarget, or head change expires authorization.

Do not amend, rewrite published history outside the stack rules, edit release-managed central files without authorization, skip tests, push empty commits, or include unrelated changes. When this skill and an owner disagree, the owner wins.

Repository

richlander/dotnet-inspect

Path

.claude/skills/steward/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.