AGENT SKILL

pr-pipeline-tauceti

Use when a Lean or Mathlib-style change is intended for a TauCetiProject/TauCeti pull request, including local readiness, PR drafting, review iteration, or current-head review checks.

License 许可
GPL-3.0
最近更新
2026-11-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
pr-pipeline-tauceti
description
Use when a Lean or Mathlib-style change is intended for a TauCetiProject/TauCeti pull request, including local readiness, PR drafting, review iteration, or current-head review checks.

02 资源文件 2

03 SKILL.md 正文

<SUBAGENT-STOP> If dispatched for one bounded PR stage, execute that stage only and return its evidence. </SUBAGENT-STOP>

Input: checkout, diff/context, write scope. Read references/tauceti-contract.md; the live AGENTS.md overrides it and requests. Empty diff: roadmap-only; verify target/claim, return evidence/blockers, then stop read-only: no lean-prepare-pr, Lean/build, PR draft, branch, PR, or write. If formalization is needed, using-leanpowers completes first; REQUIRED SUB-SKILL: one lean-prepare-pr on exact snapshot; validate; rerun if stale. Use lean-apply-feedback.

  1. Resolve checkout/remote/base/head/branch/in-flight. Read AGENTS.md, README.md, COORDINATION.md, PR template, and TauCetiRoadmap. New math needs an exact merged-roadmap target and a live claim: an intention issue the requesting human filed, assigned to their login and unexpired (TauCetiRoadmap README §Coordinating work; body gets Closes TauCetiProject/TauCetiRoadmap#N), or our worker lease refs/tauceti-claims/author/<focus>/<id>. None, or any live claim by another owner: stop. Upstream both are cooperative; the stop is local policy. Maintenance names its area or uses Roadmap: none only for genuinely general, cross-cutting, infrastructure, or dependency work.
  2. Keep one topic; no stacked PRs, drafts included: a dependent opens only after its prerequisite lands on main. Branches prepared ahead on the human's fork may build on each other. Before opening, rebase onto current main (git rebase --onto <upstream>/main <prerequisite-head> <branch>), verify the diff against main is this one topic, and redo steps 3 and 5. Inspect paths; reuse Tau Ceti/Mathlib first. List each rule of AGENTS.md §The rules of the repo at the observed commit with its check; reject violations and accidental human-owned edits. Preserve intentional scripts/, .github/, or lakefile changes and stop for human review; only documented pin files may move forward. Code from another pin goes through lean-upgrade to TauCeti's current pin first.
  3. Validate the one lean-prepare-pr result: paths/diff, style/review/fidelity/verification receipts, checks, linter, lake build, and axiom/full-type inspection. Record commit, pins, failures, and stale evidence; green build is not semantic or roadmap approval.
  4. Draft the live PR body with one standalone Roadmap: line, result, receipts, scope, limits. Add <!--tauceti-target:v1 {"focus":"<area>","id":"<canonical-target-id>"}--> for roadmap work. The newest literal tauceti-scoreboard comment must contain tauceti-meta:v1.head_sha exactly equal to the actual PR head; otherwise unreviewed.
  5. Before the first push and every fix push: the commit to push is a new snapshot, so refresh step 3's receipts and build/lint/axiom checks on it, then run scripts/local-review.sh <checkout> <title-and-body-file> (minutes; run in background). It applies TauCetiReview's own engine and current-main rubrics to that commit and posts nothing; exit 0 means every rubric approved, 1 findings, 2 or 3 the engine could not run. Record head and TauCetiReview SHA. Fix findings via lean-apply-feedback, commit, rerun: at most three fix rounds after the first run. If any rubric is still red, or a finding can be shown wrong, stop: give the human the head SHA, red rubrics, scoreboards and options (more rounds, open as is, drop). Pushing a head with a red local rubric needs a fresh explicit instruction from the human naming that head; earlier or general authorization does not cover it. Upstream review is not automatic at present (TauCeti README §Review), so opening does not guarantee a verdict. If the engine cannot run, judge the diff once per rubric file (after _common.md) at that SHA, verdict plus evidence. Common misses to check: provenance only in the PR body rather than the module docstring (lean-prepare-pr); hypotheses stronger than the proof uses; inline show where a named have belongs; general lemmas in a specialised file. Then stop at readiness unless remote mutation is explicitly authorized now (using-taucetipowers: authorization). Record the observed head. New branch: git push --force-with-lease=<branch>: <remote> HEAD:<branch>. Existing branch: git push --force-with-lease=<headRefName>:<observed_oid> <remote> HEAD:<headRefName>. A stale lease means re-observe, never plain/force push. Run gh pr create only after body approval.
  6. After push, watch CI and the newest scoreboard for the exact head (tauceti-review <PR> is the read-only rerun; --post only when authorized). No scoreboard for the current head means unreviewed: run read-only tauceti-review <PR> and report it to the human. Done means CI green and every rubric green for the current head. Findings go through lean-apply-feedback: fix by pushing through step 5 again; reply in the rubric thread (when authorized) only to contest a finding that is wrong. Stop there: merging, closing and de-duplicating are the repo's CI and bors, never the agent; never --admin. A bors merge shows the PR as CLOSED, titled [Merged by Bors] - …, with no mergedAt; confirm the commit on main. Points raised after merge go in a follow-up PR.

Return paths/diff, claim evidence, receipts, rubric-gate result with TauCetiReview SHA, draft body, remote actions, review status, and blockers.

04 同仓库更多 Skills