菜单
Issue-first, PR-gated delivery — no direct main pushes, independent review required, deterministic merge gates. Use whenever a task implements a GitHub issue and ships through a PR (assigned by an orchestrator, by project rules, or self-invoked by the implementing agent).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | github-driven-workflow |
| description | Issue-first, PR-gated delivery — no direct main pushes, independent review required, deterministic merge gates. Use whenever a task implements a GitHub issue and ships through a PR (assigned by an orchestrator, by project rules, or self-invoked by the implementing agent). |
Enforce a fail-closed GitHub delivery workflow: every change traces to an issue, lands on a branch, ships through a PR, and merges only when all gates pass.
Whenever a task implements a change from a GitHub issue and delivers it through a PR — invoked by an orchestrator, by project instructions, or by the implementing agent itself. It controls the full lifecycle from issue intake through merge.
Do not invoke for a single sub-step (e.g. "open a PR", "check CI") unless the full workflow context is already established.
Before writing code:
main or equivalent).If no issue is identified, stop and request or create one.
Inspect the issue for scope and acceptance criteria. If either is missing or ambiguous, update the issue or request updates before writing code.
main.issue-<id>-<slug> from the default branch and switch to it.Implement the change on the issue-<id>-<slug> branch.
Run repo-appropriate validation. Record commands and output.
The PR must include:
Closes #<issue> in the body.- [ ]) only for known remaining work; every unchecked box blocks merge.Independent review is required in principle. A qualifying review is review evidence produced by an actor other than the implementation author, durably visible on the PR.
Run the bundled acquisition script. Resolve this path from the github-driven-workflow skill root, not from the target repository root:
scripts/acquire-review.py <OWNER>/<REPO> <PR_NUMBER> [kind]
In this source repository the same helper lives at skills/github-driven-workflow/scripts/acquire-review.py; in an installed skill it remains scripts/acquire-review.py relative to the installed skill directory.
Treat any nonzero exit as "review not acquired" and proceed to authorized bypass per below. Project-level customization of acquisition logic is documented in the skill's README.md.
The bundled default is reviewer-neutral: it picks copilot or codex uniformly at random (or honors an explicit [kind] third argument) and dispatches a single asynchronous review request, printing route: <kind> (dispatched) on success. The bundled script does not prefer any specific automatic reviewer; projects that want a different selection policy supply one via the REVIEW_ACQUIRE_SCRIPT override (see the skill's README.md). (dispatched) means only that an async request was sent — the §8 evidence gate is not yet satisfied. Callers must not equate route: <name> alone with merge-readiness — only the §8 gate determines that. Wait briefly and re-check §8; if evidence does not accrue within a reasonable wait, dispatch a different kind or proceed to authorized bypass per below. Override implementations may also emit (evidence) when their route posts a durable artifact at dispatch time.
Acceptable evidence on the PR:
@codex review (independent regardless of who posted the request) — typically a formal Review event, occasionally a top-level comment that the agent recognizes as a review.Reviewed-by: <reviewing-entity-id> distinct from the implementer. Independence is judged by the recorded identity, not by the GitHub poster.Self-reviews, local notes, unlinked claims, and generic comments do not qualify. Pick the lowest-friction route available; do not exhaust slow async routes when a faster durable route is already available. Asynchronous routes (Copilot, @codex) require waiting; if no response appears within a reasonable wait, switch routes rather than block indefinitely.
When no review route is viable, record the bypass on the PR with a comment citing the authorization:
gh pr comment <N> --body 'Bypass: independent review waived. Authorization: <provenance>. Reason: <reason>.'
Accepted provenance:
owner=$(gh repo view <owner>/<repo> --json owner --jq .owner.login)
test "<commenter-login>" = "$owner"
owner is the org login matching no human account), the comment must come from an account the org owner has explicitly delegated, citing that delegation; verification compares the commenter against the delegated login. Generic admin permission alone is not sufficient.Record the cited provenance (and verified <commenter-login> on the owner path) alongside the bypass evidence in §8.
Run all checks before merging.
PR state
gh pr view <N> --json state,isDraft \
--jq '{open: (.state == "OPEN"), notDraft: (.isDraft == false)}'
Both must be true.
CI checks
gh pr view <N> --json statusCheckRollup \
--jq '.statusCheckRollup | map({name, state})'
Empty array ⇒ pass. Any non-SUCCESS state, or command error ⇒ stop.
Labels
gh pr view <N> --json labels \
--jq '[.labels[].name] | any(. == "blocked" or . == "do-not-merge" or . == "needs-decision")'
Must return false.
Unresolved review threads
gh api graphql -f query='
{
repository(owner: "<owner>", name: "<repo>") {
pullRequest(number: <N>) {
reviewThreads(first: 100) {
nodes { isResolved }
pageInfo { hasNextPage endCursor }
}
}
}
}'
Count nodes where isResolved is false. Must be zero. Paginate with after: "<endCursor>" while hasNextPage is true. Query error ⇒ stop.
Unchecked task boxes
gh pr view <N> --json body --jq '.body | test("- \\[ \\]|\\* \\[ \\]")'
Must return false.
Independent review evidence or authorized bypass
The gate passes when any of the three clauses below is satisfied. They mirror the evidence types §7 accepts; do not filter by state or author.login (independence is enforced at evidence-recording time per §7).
Clause 1 — formal Review event. Any formal GitHub PR Review event satisfies this clause regardless of the reviewer's identity (humans, Copilot, Codex, or other reviewer accounts that produce formal reviews). Reviewer agents that respond as a normal comment rather than a formal Review event are matched by Clause 2 (Reviewed-by: marker) or, when owner-authored with a review verb, by Clause 3.
gh pr view <N> --json reviews --jq '.reviews | length >= 1'
Returning true satisfies the gate.
Clause 2 — Reviewed-by: comment artifact. Codex CLI artifact comments and other reviewer agents recorded per §7.
gh api repos/<owner>/<repo>/issues/<N>/comments \
--jq '[.[] | select(.body | test("(?m)^Reviewed-by:\\s*\\S"))] | length >= 1'
Returning true satisfies the gate. Honor system per §7: the <entity-id> after Reviewed-by: must be non-empty and distinct from the PR author identity, but its validity is not machine-verified.
Clause 3 — owner comment-as-review or authorized bypass. The owner-as-review path requires both an owner-login match and an explicit review verb in the same comment.
owner=$(gh repo view <owner>/<repo> --json owner --jq .owner.login)
gh api repos/<owner>/<repo>/issues/<N>/comments \
--jq --arg owner "$owner" '[.[] | select(
(.user.login == $owner and (.body | test("(?i)\\b(Approved|Reviewed|LGTM|Changes requested)\\b")))
or (.body | test("^Bypass: independent review waived\\. Authorization: .+\\. Reason: .+\\."))
)] | length >= 1'
Returning true satisfies the gate. For org-owned repos, substitute the delegated account per §7 in the owner comparison. An authorized bypass recorded per §7 is matched by the second sub-clause and needs no separate query.
Do not gate on reviewDecision. It aggregates branch protection rules; without an approving-review rule, it stays empty regardless of review count.
Do not add gates beyond this set. mergeStateStatus, reviews[].author.login filtering, and any check not listed in §7's evidence types are anti-patterns in autonomous environments.
Auto Mode note: Child agents may attempt to add gates "for safety" (commonly
reviewDecision == APPROVED, author-login filtering, or restricting to clause 1 only). The three-clause set above is authoritative and reflects §7's full evidence vocabulary; additions or restrictions are self-fabrications and should be rejected.
Cite the evidence (review count, comment URL, or bypass comment URL plus cited provenance) in the merge note.
Merge only when every gate passes. If any gate fails, fix, revalidate, or leave the PR open with a comment stating the exact blocking condition.
Stop before implementation or merge when any required state cannot be verified.
Stop conditions:
main, or PR is missing or draft.Closes #<issue>.SUCCESS, pending, or command errored.When stopped, state the exact blocking condition and the action needed to unblock.
Procedural guidance only. Does not configure GitHub branch protection, CI workflows, or repository permissions.
Audit the structural and logical integrity of a mathematics paper — quantifier scope, domain-of-definition guards, proof/computation honesty, theorem hierarchy, contribution-list discipline (mapping every claimed contribution to its formal result, prior baseline, concrete gain, and independent significance), stale claims, notation accuracy at introduction, standing assumptions, and undefined relation symbols or coined terms in summary prose — without touching prose style or rhetorical prose inflation (structural inflation of contribution lists and result hierarchy is in scope).
Audit a LaTeX mathematics document for notation drift — symbols used under different names across sections, orphaned macros, undeclared aliases (including cross-language pairs where a term defined in one language is later used in another), symbols or relation signs used with no definition site anywhere, symbol reuse across scopes, and first-use-after-gap without back-reference — without evaluating mathematical correctness or prose style.
Review or rewrite mathematical prose so an edit, paraphrase, terminology migration, or explanatory description preserves the exact mathematical meaning fixed by its semantic source — referent identity, operation versus result, domain/codomain and representation level, quantifier and equality mode, role and provenance, notation role, and classified-before-replacement terminology migration with concept-cluster synchronization — without auditing claim-versus-proof strength (use math-claim-integrity), notation bookkeeping (use math-notation-consistency), Japanese style or translation events (use wabun-math-style), or audience prerequisites and ordering (use exposition-flow).
Detect and correct Japanese-language anti-patterns in mathematical writing — epistemic hedges weakening proved claims, incorrect verb tense, passive/active confusion in proofs, ambiguous particles, unjustified 明らか, decorative connectives, redundant meta-discourse, double negation, stacked の, generic/general conflation, borrowed non-mathematical vocabulary, vague category names for explicit formulas, full-width semicolons or untranslated English, canonical-terminology drift, and translation-level concept conflation — without evaluating mathematical correctness or structural hierarchy.
Turn repository-grounded source material into a self-contained, audience-admissible document or presentation by modelling the intended audience, extracting information units with explicit discourse roles, and enforcing dependency-ordered exposition in which every load-bearing statement follows the prerequisites needed to interpret it — without factual verification or citation checking (use sot-integrity), prose polishing (use deslop-prose), mathematical-correctness review (use the mathematical-writing skills), or typography and slide layout.
Workspace-level Git cleanup — discover repositories under the current directory and prune merged branches, merged linked worktrees, stale remote-tracking refs, and stale worktree metadata.