nemoclaw-maintainer-cut-release-tag

작성자: nvidia

origin/main에서 사전 태그 날짜가 있는 변경 로그 항목을 확인한 후 결정적 NemoClaw semver 릴리스 태그를 생성하고, 릴리스 하우스키핑을 처리하며, 초안을 작성합니다…

npx skills add https://github.com/nvidia/nemoclaw --skill nemoclaw-maintainer-cut-release-tag

Cut Release Tag

Cut one signed annotated semver tag from a generated plan. Use the release scripts for tag writes and nemoclaw-maintainer-e2e for maintainer-requested workflow dispatches. Do not improvise raw tag, push, version-bump, or other release-state GitHub writes.

Treat these as separate states:

  • Tag can be cut: the release entry and required image checks pass. The maintainer chooses to proceed with the displayed documentation coverage and general E2E state. The release brief records both decisions and contains no unresolved prompts.
  • Tag cut: the remote signed tag exists and peels to the planned candidate.
  • Post-tag follow-through: after reporting the tag as cut, continue the same task. Monitor latest, release labels, public documentation, and release images. Draft the Announcement and report lkg state.

Hard Rules

  • Use the exact requested version. Generate the plan with --version vX.Y.Z; never infer a bump.
  • Tag only the candidate captured in the plan.
  • By default, plan origin/main without an exception. For urgent QA qualification, a maintainer may select an exact historical ancestor with --candidate <full-sha> --exception <reason>.
  • Require the release entry for a current-main plan. A historical plan records its explicit release-entry exception in the signed release brief.
  • Treat documentation coverage as maintainer context, not a tag gate. Show the exact coverage point, later commits and PRs, review and check state, changed paths, and open managed docs PRs.
  • Record the maintainer's documentation decision in the signed release brief.
  • Require applicable GHCR base and managed-image publication evidence.
  • Treat E2E as maintainer context, not a tag gate. Show the newest full E2E result and let the maintainer run focused tests, run the full suite, or proceed with the displayed status.
  • Record every displayed or requested E2E result and the decision in the release brief, the signed Markdown release record. Record a plain-language exception reason when the status is exceptional or a requested run remains unresolved.
  • Pass the exact final release brief to release:cut with --message-file. The file becomes the signed tag annotation; do not maintain another exception record.
  • Ask the maintainer to paste the plan's full confirmation phrase before cutting.
  • Push only the planned semver tag. Never push or move latest or lkg here.
  • Report the tag as cut immediately after remote readback. This report is a progress checkpoint, not the final response.
  • Continue the same task through post-tag follow-through. Do not make a post-tag result a tag gate.
  • Ask before a workflow rerun. Never create a GitHub Discussion.
  • Never move, delete, or replace an existing remote semver tag unless the maintainer starts a protected-tag remediation.
  • Follow the release-train policy and the shared Git and GitHub Access Hard Stop.

Workflow

Copy this checklist and update it as you work:

Release tag:
- [ ] 1. Generate the exact-version plan and release-brief template
- [ ] 2. Verify required candidate evidence
- [ ] 3. Show E2E context and record the maintainer's decision
- [ ] 4. Finish and review the Markdown release brief
- [ ] 5. Confirm, cut, and read back the signed tag
- [ ] 6. Follow tag-triggered work and draft the Announcement

1. Generate the Plan and Brief Template

Run the default current-main plan:

npm run release:plan -- --version vX.Y.Z

For an accepted urgent QA qualification, run:

npm run release:plan -- --version vX.Y.Z \
  --candidate <full-lowercase-40-sha> \
  --exception "<plain-language reason>"

Do not pass --exception without a historical candidate. Do not select current origin/main with an exception.

The script writes ../nemoclaw-release-vX.Y.Z/plan.json. Show the maintainer:

  • previous tag and peeled commit;
  • requested version;
  • candidate commit and headline;
  • exact range from the previous commit through the candidate; and
  • the full confirmation phrase derived from the version and candidate.

After later reads of remote state, keep this candidate when all of these remain true:

  • the candidate is still an ancestor of origin/main;
  • the previous release tag still peels to the commit recorded in the plan;
  • the candidate's release entry remains valid, or the historical plan retains its explicit release-entry exception; and
  • the candidate's own required evidence remains valid.

New commits on main do not invalidate that plan. A managed documentation PR or branch for a later commit does not invalidate it either. Regenerate the plan only when the release range, version, or candidate must change.

The plan is immutable once written. The helper creates the brief only when it is absent; fill that same brief in place. A same-version candidate change starts a new release workspace, such as ../nemoclaw-release-vX.Y.Z-replan-2/. Never overwrite a plan or pair a new plan with an old brief. Pass that workspace explicitly:

npm run release:plan -- --version vX.Y.Z \
  --output ../nemoclaw-release-vX.Y.Z-replan-2/plan.json

Use the returned plan directory for the brief and cutter commands that follow.

Create the brief template now, before collecting evidence, so each result can be recorded as it is read:

node --experimental-strip-types --no-warnings \
  .agents/skills/nemoclaw-maintainer-day/scripts/handoff-summary.ts \
  --plan ../nemoclaw-release-vX.Y.Z/plan.json \
  --output ../nemoclaw-release-vX.Y.Z/release-brief.md

The helper refuses to overwrite an existing brief. Fill this generated file in place; do not rerun the helper after evidence has been added.

2. Verify Required Candidate Evidence

Read and follow Candidate Evidence. It owns the executable reads for the release entry, documentation coverage, and applicable base-image verifier.

Show the complete documentation coverage evidence. Offer the maintainer the three choices defined there. If the maintainer requests documentation work or stops, do not continue to E2E or tag confirmation. If the maintainer proceeds, record the exact decision line in the release brief.

Do not offer the general E2E proceed option until the release entry and image checks pass and the maintainer chooses to proceed with the displayed documentation coverage. Record the returned paths, URLs, PR state, commit ranges, review state, check state, and image identities in the release brief.

3. Present General E2E and Ask for a Decision

Use nemoclaw-maintainer-e2e to find the newest completed or active full E2E run. Show these details instead of reducing the run to one passing/failing label:

  • candidate SHA and full-run SHA;
  • status and conclusion;
  • workflow attempt, created, started, and last-updated timestamps, plus age at inspection;
  • workflow URL and Release qualification job URL; and
  • any failed, cancelled, skipped, queued, or still-running results.

Offer exactly these choices:

  1. Run focused tests selected by the maintainer.
  2. Run the full suite.
  3. Proceed with the status as shown.

After a requested run, show its same fields and add every result to the release brief. A requested run remains unresolved while it is queued or running, or after it ends without success. It becomes resolved only when a successful result covers the same requested scope.

Use Exceptions: None only when the maintainer accepts a successful full result for the candidate and no requested run remains unresolved. Otherwise, ask for and record one concise plain-language reason. The reason must say what differs or remains unresolved and why the maintainer is proceeding. Selecting “Proceed with the status as shown” is the decision, not the reason. Stop and ask the maintainer why before continuing when a reason is required. This exception applies only to E2E. It never replaces the current-main release entry, a historical plan's release-entry exception, the documentation coverage decision, or required image evidence.

4. Finish and Review the Release Brief

Replace every TODO_RELEASE_BRIEF prompt in that Markdown file with:

  • the complete canonical release entry and its repository path for a current-main plan, or the plan-bound historical release-entry exception;
  • the latest included cumulative docs PR, coverage commit, later commits and PRs, changed-path result, review and check state, open managed docs PRs, and maintainer decision;
  • exact-candidate E2E workflow, attempt, and successful base-image-publication job URL;
  • the newest full E2E result and every focused or full rerun result, including SHA, time, age, status, conclusion, and URLs;
  • the maintainer's E2E choice; and
  • a final Exceptions: None line or Exceptions: <plain-language reason> line.

Keep the helper's compact range counts and QA focus. Do not duplicate every commit or risky file; the canonical release entry and compare range already preserve that detail.

Do not put secrets in the brief. Show the complete rendered file to the maintainer. Because this exact public Markdown becomes the signed tag annotation, make any correction in the file before asking for confirmation.

5. Confirm, Cut, and Report the Tag

Ask the maintainer to paste the plan's exact phrase:

CONFIRM RELEASE vX.Y.Z <full-candidate-sha>

After receiving that exact phrase, run the cutter immediately. The cutter reads the immutable plan, validates the signed release brief's documentation decision, and checks remote tag state before the push.

Run:

npm run release:cut -- \
  --plan ../nemoclaw-release-vX.Y.Z/plan.json \
  --message-file ../nemoclaw-release-vX.Y.Z/release-brief.md \
  --confirm "CONFIRM RELEASE vX.Y.Z <full-candidate-sha>"

Require the script's remote readback to show that the signed annotated tag exists and peels to the planned candidate. Report the tag, candidate, plan path, brief path, and readback. Then continue the same task.

6. Follow Tag-Triggered Work and Draft the Announcement

Start these operations together:

  1. Load nemoclaw-maintainer-release-notes. Draft release-note-draft.md from the plan's immutable range. Open the completed draft in the requested editor. Never create the Discussion.
  2. Find the tag-push runs for these workflow files:
    • .github/workflows/release-latest-tag.yaml
    • .github/workflows/docs-publish-public.yaml
    • .github/workflows/base-image.yaml
  3. Bind each run by workflow path, event=push, release tag, and planned candidate. Retain its run ID and attempt. Monitor the three runs concurrently until they reach terminal results.

Classify the effects that each workflow owns:

  • For Release / Latest Tag, verify that latest identifies the release tag. Verify label carry-forward and released-label deletion.
  • For Docs / Publish Public, require the publish job to succeed.
  • For Images / Base Images, require Publish complete managed images to succeed. Report Pi candidate failures separately; they do not determine production promotion.

A failed post-tag workflow does not change tag success. Report the failing job and recovery path. Ask before a rerun. Bind and monitor the new attempt. The managed-image workflow supports failed-job reruns that reuse successful producer artifacts from the same run.

After image classification, read the peeled lkg commit. This skill never moves lkg. If production promotion succeeded and lkg differs, show the current and proposed releases and ask for separate maintainer authorization. If the maintainer moves lkg, monitor .github/workflows/release-lkg-brev-image.yaml and its returned downstream production-image run.

Send the final response only after:

  • all three automatic workflows are terminal and their effects are classified;
  • the Announcement draft exists; and
  • lkg already identifies the release, is ineligible because production promotion failed, awaits an explicit maintainer decision, or its authorized downstream production run is classified.

Keep the semver tag immutable.

Recovery

  • Missing release entry in a current-main plan: finish the candidate's documentation work and generate a new plan for the resulting commit. A historical plan uses only its plan-bound explicit exception.
  • Documentation coverage shows a gap, failed checks, unapproved changes, unsupported paths, or an open managed docs PR: show that state. Let the maintainer proceed, create or update a docs PR, or stop. If documentation work changes the candidate, generate a new plan.
  • Required GHCR evidence fails: repair and rerun only the affected image work. Do not replace it with the general E2E proceed decision.
  • General E2E is old, incomplete, failed, or from another SHA: show it and offer focused, full, or proceed. Record the decision and reason in the brief.
  • Candidate is no longer on origin/main, the previous release changed, or the version is no longer available: stop and generate a new plan.
  • Signing or access fails: report the exact error and follow the shared hard-stop guidance. Do not improvise tag commands.
  • A post-tag workflow fails: report that state and its rerun path separately. Do not move the already-published semver tag.

nvidia의 다른 스킬

compileiq-debug
nvidia
무언가 잘못되었을 때 사용: Search()가 멈추거나, 모든 평가가 INVALID_SCORE를 반환하거나, 점수가 개선되지 않거나, 모든 설정이 동일한 숫자를 반환하거나, ptxas 오류 등이 발생할 때
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
compileiq-validate-result
nvidia
검색이 완료된 후, 속도 향상을 청구하거나 ACF를 발송하기 전에 사용합니다. dump_results CSV를 로드하고, 상위 K개 후보(단일 목표)를 추출합니다…
changelog-audit
nvidia
릴리스 전에 Warp CHANGELOG.md를 감사합니다: 누락된 항목 복구, 사용자 영향별 정렬, 항목 언어 다듬기, 줄 바꿈, (릴리스 브랜치 모드) 비교 업데이트…
maintain-dynamic-plugins
nvidia
NeMo Relay 동적 플러그인 로더, 매니페스트, Rust 네이티브 SDK, gRPC 워커 프로토콜, Python 워커 SDK, 문서, 테스트 및 릴리스 워크플로 커버리지를 유지 관리합니다.
dgx-diagnose
nvidia
일반적인 DGX Station GB300 문제 진단 — CUDA 충돌, 잘못된 GPU 타겟팅, vLLM/SGLang 컨테이너 버그, MIG 상태 문제, NVLink/Fabric Manager 오류,…