agentleFS
zajcazajca/pohunek/.agents/skills/pr-handoff/SKILL.md

pr-handoff

Skill9 starsChanged 2 days agoMITView on GitHub

Turn a finished, gate-green milestone into a stack of small sequential PRs (one per slice branch, unsigned commits, bottom-up), each described from the milestone's audit evidence, then update the milestone's issue. Use when the user asks for a PR for a branch, or after a milestone-build harness run reports all DoD items met with the gates green.

  • Commits and pushes

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 3 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 4 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    Before replacing a published PR head, record its active pull_request run IDs (gh run list --branch <slice-branch> --event pull_request …

    line 78
  2. Distinctive:

    Run steps 1–5 for each slice, bottom-up, observing the capacity rule above.

    line 94
  3. Very specific:

    Stage explicitly: Add files by path — never git add -A.

    line 98
  4. Very specific:

    Commit — unsigned: Use git commit --no-gpg-sign.

    line 101
  5. Distinctive:

    Review gate: No label or other trigger starts the review.

    line 125
  6. Very specific:

    Do not merge here — merging is the merge-advance skill or the user's PR flow on GitHub, bottom-up.

    line 149
  7. Distinctive:

    Do not release, tag, or delete branches.

    line 155

“Common” is measured: kept by at least 6 in 10 of the 4 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

pr-handoff — commit, push, open a PR

Bridges the gap between "milestone implemented, gates green" (the milestone and milestone-review skills stop there) and a published pull request. Authorization semantics: a milestone-build harness run whose final audit reports every DoD item met with all gates green pre-authorizes this handoff — publishing is then part of the autonomous loop, no further ask. This intentionally supersedes the default "commit/push only when the user asks" for that one path (see .agents/instructions/workflow.md, "Accepted harness trade-offs"); for any other branch, wait for the user to request it. This preauthorization is for publishing only — it does not authorize merging the PR or anything beyond it.

Preconditions

  • The full CI-mirror gate set passes on the branch. For a harness milestone that is the gate set in .lh-harness/workflows/milestone-build.md (already run by the milestone's audit); for a plain branch, run that same set by hand. Follow the gates skill together with the authoritative .agents/instructions/verification.md commands, including Hermes compatibility, SDK workspace, and real-daemon checks. A partial run never authorizes this handoff.
  • You are on a milestone slice branch, not on main itself. The bottom slice is based on main; every other slice is based on the slice below it (see the milestone skill and pullRequests in .github/agent-workflow.json).
  • The milestone's GitHub issue is known (explicit URL/number from the user or task; otherwise resolve per the github-workflow skill — unique unambiguous match only, otherwise ask).
  • The milestone's verification evidence is available: gate results, DoD verdicts with path:line proof, test counts. From a harness run, take them from the final auditor report and the final response episode.

One PR per slice — the stack

Publish the milestone as a stack of small sequential PRs, one per slice, bottom-up. For each slice, from the bottom:

  • Its gate set passes on that slice branch alone; a later slice never excuses a red lower one.
  • --base is main for the bottom slice and the previous slice's branch for every other; --head is the slice's own branch. The diff each PR shows is therefore only its own slice.
  • The description names the stack position and neighbours (Stack 2/3: #<prev> → **this** → #<next>, filling in #<next> once it exists) and the DoD items this slice satisfies.
  • Non-final PRs reference the issue with Refs #N; only the final PR uses Closes #N, so the issue cannot close before the whole stack lands.
  • A review fix to a lower slice is committed on that slice's branch; restack the branches above it with git rebase --update-refs from the top branch, re-run the applicable gates on every rebased slice (its inputs changed with its base), and keep those higher branches local. Push the changed lower branch first, wait for its current head's CI and review, then publish higher branches one at a time, bottom-up, using --force-with-lease for rewritten remote heads. Do not batch-push the restacked branches.

GitHub Actions capacity

The repository has limited concurrent Actions capacity. Coordinate with other agents publishing independent PRs: check active PR runs before a head push and avoid starting another expensive run while capacity is saturated. Never cancel another agent's run just to make room. For a stack, commit and verify every slice locally first, then push and open only the bottom ready PR. Publish the next slice after the previous PR's current-head CI and review finish; a local rebase of a higher slice does not need a push yet. Record unpublished slices as pending in issue tracking, then add each PR link when that slice is published.

Before replacing a published PR head, record its active pull_request run IDs (gh run list --branch <slice-branch> --event pull_request --json databaseId,headBranch,status,url) and confirm each belongs to that PR. If a run has not completed, cancel it with gh run cancel <run-id> before the new push or immediately afterward. Read its status again with gh run view <run-id> --json status,conclusion; a cancel request alone is not confirmation. Confirm it completed with cancelled (or had completed before cancellation). The CI workflow's same-ref cancel-in-progress is a fallback, not a reason to leave an obsolete run consuming capacity. Watch only the new head after the push.

A milestone that is one small concern is a stack of one: a single PR on main with Closes #N.

Steps

Run steps 1–5 for each slice, bottom-up, observing the capacity rule above; repeat step 6 after each publication to record its PR link and the remaining unpublished slices as pending.

  1. Stage explicitly. Add files by path — never git add -A. Exclude transient files the user does not want committed (historically idea.md, harness run state).
  2. Commit — unsigned. Use git commit --no-gpg-sign; commits in this repo are NEVER signed, and --no-gpg-sign keeps that true even on machines whose global git config forces a signer. Message: concise, imperative, English. Never add a Co-Authored-By trailer or any "generated with" footer.
  3. Push. git push -u origin <slice-branch> only when that slice is next to receive CI and review. Keep the other slices local until their turn.
  4. Build the PR description from evidence, not memory. English body, before/after structure:
    • Summary — what changed and why, one paragraph; link the tracked issue with Closes #N on the final PR of the stack, Refs #N on the others; state the stack position and the DoD items this slice covers.
    • Before / After — what was missing, what is observable now.
    • Verification — the gate results, per-item DoD verdicts with path:line evidence, test counts, and any verification step worth repeating by hand. This section mirrors the final audit report; do not invent or embellish results.
  5. Open and verify. gh in a non-interactive agent shell never prompts: save the assembled title and body to a scratch file and pass them explicitly — gh pr create --base <main-or-previous-slice-branch> --head <slice-branch> --title "<title>" --body-file <file> (remove the scratch file afterwards). Confirm the URL, then check the initial CI status (gh pr checks <n>). Report the PR number, URL, and CI state.

    Review gate. No label or other trigger starts the review. The maintainer's manager (the coordinator or operator) starts the independent review session for each pushed head (a fresh Opus high effort session that did not author the change; with no manager named, deliver-issue Phase 5 step 1 says who starts it), which posts a PR conversation comment authored by zajca that carries <!-- opus-pr-review:v1:<repo>#<N>@<40-char head sha>:high=<h>:medium=<m>:low=<l> -->. Report the PR number and head sha to the manager. The gate holds only for a marker whose author is zajca and whose sha is the PR's current head, with high=0 and medium=0; LOW-only passes and each LOW finding becomes a follow-up issue. A marker for an earlier head, or from any other author, does not count. deliver-issue Phase 5 has the watch command and the merge criteria.

  6. Update the milestone tracking. Via the github-workflow skill, comment the ordered list of stack PR links on the milestone issue with the standard handoff content (branch/worktree, HEAD revision, scope covered, exact checks run with real exit results including any skipped/failed ones, remaining work/blockers, and worker run IDs from the harness run), and keep the project status In Progress — a PR opening is never Done. Done happens when the work has verifiably landed on the remote default branch (via merge-advance or the PR being merged); verify every write from the API response.

Constraints

  • Do not merge here — merging is the merge-advance skill or the user's PR flow on GitHub, bottom-up. The repository does not delete head branches on merge, so after a lower PR merges retarget the next one explicitly (gh pr edit <n> --base main); if the lower PR was squash- or rebase-merged, rebase the next slice onto origin/main first (git rebase --onto origin/main <old-lower-branch> <slice-branch>).
  • Do not release, tag, or delete branches.
  • Issue/project writes follow the github-workflow skill's safe-persistence rules; a blocked sync is reported as blocked, never claimed successful.
  • PR text only references secrets-free evidence; never quote environment values, tokens, or logs that could contain them.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.