agentleFS
xarleynxarleyn/dsh-plugins/.agents/skills/release-plugins/SKILL.md

release-plugins

Skill3 starsChanged 36 days agoMITView on GitHub

Ship a release wave of this monorepo — version-plan pre-flight, the

  • Deletes or force-pushes

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 3 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Plans: Every publishable project with commits its newest reachable release tag does not cover needs a committed plan:

    line 50
  2. Very specific:

    Run the set the release actually depends on.

    line 92
  3. Distinctive:

    Before trusting a green typecheck or verify, ask whether it could be green only because build output is sitting in the tree.

    line 112
  4. Very specific:

    Run the kit's manual playbooks — docs/manual-testing/smoke.md after every deploy, docs/manual-testing/wave.md for a wave (a row per changed …

    line 138
  5. Distinctive:

    Only then move the deployment's plugin list and repeat the smoke pass there.

    line 145
  6. Very specific:

    Read the publish step's log, not the registry's mood.

    line 173

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “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

Release the plugins

A release is derived from committed .nx/version-plans/*.md files, runs only through .github/workflows/release.yml, and publishes with npm Trusted Publishing (OIDC — never an npm token). docs/RELEASING.md is the canonical runbook; this skill is the operational path plus the traps that the workflow, the gates and RELEASING.md do not spell out on their own.

The one invariant to keep in your head: the order is publish → tag → push. The branch never advances to a state the registry does not already reflect, so a failure before the push leaves the branch and the version plans exactly as they were and is fixed by rerunning the workflow. Anything you do by hand that puts a version on npm outside that order breaks the invariant the rerun relies on.

The shape of a run

prepare resolves the release (plans applied, release commit created, packages selected, registry pre-flight, dependency-range check) → gates and one Package / <name> job per released package run side by side → publish uploads tarballs, publishes in dependency order, verifies each version installs, then creates one annotated wave tag release/<date>, pushes commit + tag, and creates one GitHub Release for the whole wave.

Inputs: dry_run (default true), first_release (false), publish_only (false), create_github_releases (true). The workflow releases whatever ref it was dispatched on (github.ref_name), so a branch can produce test versions without touching main.

0. Pre-flight

Is the ref pushable? The workflow's last step pushes the ref it ran on, and that push is the point of no return for everything in that history. Before dispatching from main or a branch, confirm with the maintainer that the ref is meant to be pushed. Machine-specific history constraints live outside the repo: read .private/README.md and follow whatever local pre-flight it points at before choosing the ref.

Plans. Every publishable project with commits its newest reachable release tag does not cover needs a committed plan:

pnpm release:check                 # reads COMMITS — uncommitted work is invisible
node scripts/verify-package-hygiene.mjs --version-plans-only

The second command is the same one the workflow runs first: it requires at least one plan Nx can parse. A plan that lost its opening --- fence is silently ignored by Nx — no bump, no changelog entry, not even counted — so an unparseable plan is worse than a missing one.

Predict the versions. The wave bumps only the projects that have a plan; their plans merge into the highest bump. release.version.updateDependents is never in nx.json, so no dependent package is dragged in — if a predicted version looks wrong, look for a plan you forgot, not for a dependent bump. (updateDependents was always at one point; versions resolved by an old run are not evidence about what this run will do.)

dsh-qa-surface plans need a matching curated entry in plugins/dsh-qa-surface/src/client/components/QaChangelog.tsx in the same change — the hygiene gate enforces it, and plugins/dsh-qa-surface/tests/client/components/qa-sidebar-changelog.test.tsx pins the details: QA_CHANGELOG[0] must be the next planned version, QA_VERSION must equal package.json.version, the entries from the current version down must equal the versions parsed out of CHANGELOG.md, and entries below the current version must not be added. A released section is frozen — write the new text for the planned version, never into a section that already shipped.

The manifest. pnpm plugins:manifest if any plugins/*/package.json changed; the hygiene gate fails on drift.

Leak sweep. The repository is public and the tarballs are permanent. Sweep before a release with the local scanner (.private/leak-scan/scripts/leak-scan.mjs --all; see .private/README.md). Plan and changelog prose is published twice — in the repo and in the wave's release notes — so it must read neutrally: describe what a user gets, never that something internal was removed, renamed or misconfigured.

1. Local verification

Run the set the release actually depends on. --skip-nx-cache is not a performance question: the Nx cache has restored green output for tasks that fail, so a cached run is not evidence.

NX_SKIP_NX_CACHE=true pnpm check      # lint, format, typecheck, test, build, check:files, verify, deps:check
pnpm test:release                     # workflow/script contract tests — via pnpm, never `node --test`
pnpm release:check                    # plan gate, reads commits
pnpm tarball:verify                   # every package; or tarball:verify:packages plugins/<dir>

pnpm check covers deps:check and verify:logging (inside verify) but does not cover tarball:verify or release:check — run those three separately. Per-package equivalent, which is what the release's own matrix runs:

pnpm nx run-many -t lint typecheck test build verify --projects=@yadsh/<name> --skip-nx-cache
pnpm tarball:verify:packages plugins/dsh-<name>   # takes a PATH, not a package name

Before trusting a green typecheck or verify, ask whether it could be green only because build output is sitting in the tree. A verifier that reads lib/ is honest only after a build, and a fresh checkout is emulated by rm -rf plugins/*/lib packages/*/lib followed by pnpm nx run-many -t typecheck --skip-nx-cache — in a clean checkout lib/ does not exist and shims the local tree never exercises are the ones that break.

Optionally rehearse the wave's dependency resolution before spending a run:

node scripts/workspace-packages.mjs --format=tsv > "$TMPDIR/rows.tsv"
node scripts/publish-release.mjs --check --tsv="$TMPDIR/rows.tsv"

1b. Stand acceptance

Green gates prove the code; they say nothing about a deployment. The failures that reach users are composition failures — a tool that exists but is out of the conversation's reach, an expert whose policy names tools the runtime refuses, a reviewer whose service is missing, a model pin that no longer matches — and every one of them passes lint, typecheck, test and verify.

So a wave that will be deployed is accepted on a stand, not in CI:

  1. Stage it on the test stand (the deployment kit's dev plugin list, or its release list) and restart the container.
  2. Run the kit's manual playbooks — docs/manual-testing/smoke.md after every deploy, docs/manual-testing/wave.md for a wave (a row per changed package: package → manual check → evidence), docs/manual-testing/signatures.md when something refuses. The kit's scripts/qa-smoke-evidence.mjs collects the evidence from the plugin logs and exits non-zero on a FAIL.
  3. Record the round in a protocol file: a round without evidence and without a written result did not happen, and the next incident starts from zero.
  4. Only then move the deployment's plugin list and repeat the smoke pass there. The two stands differ in pins and service set, so a difference between them is itself a finding.

2. Dispatch

gh workflow run release.yml --ref main -f dry_run=true  -f first_release=false \
  -f publish_only=false -f create_github_releases=true
# review resolved versions, dependent bumps and generated changelogs
gh workflow run release.yml --ref main -f dry_run=false -f first_release=false \
  -f publish_only=false -f create_github_releases=true
gh run watch <run-id> --exit-status

--ref is not optional in practice: the release runs on the ref you name, and "the branch I happened to be on" is not a decision. Set first_release=true only for the repository's first release. Set create_github_releases=false for a routine update that should still get the commit, changelog, wave tag, artifact and npm publication but no GitHub Release.

A dry run exercises less than it looks like: it resolves plans and versions and Select release packages then reports count=0, so gates, projects and publish are all skipped. Replay, the tarball gates and publication are only ever proven by a live run.

3. Confirm the wave

Read the publish step's log, not the registry's mood. npm's CDN answers stale: a version document (https://registry.npmjs.org/@yadsh%2F<name>/<version>) can 404 for minutes after a successful publish, and the package document (/latest) lags further. Trust + @yadsh/<name>@<version> lines in the publish log, then re-query after a pause. Against a wave, prefer the whole set:

node scripts/workspace-packages.mjs --format=tsv > "$TMPDIR/rows.tsv"
node scripts/publish-release.mjs --verify-install --tsv="$TMPDIR/rows.tsv"

Then check that the state a user would see matches the repository: the wave tag exists and points at the release commit, the GitHub Release carries every tarball of the wave, and pnpm release:check reports nothing unreleased. The release commit (chore(release): publish, authored by github-actions[bot]) carries the versions, the changelogs and the deletion of the consumed plans — after it, the released projects need no plan, and new work needs a new one.

A version's health is its provenance: npm view <name>@<version> --json should show dist.attestations for anything the workflow published. A version without it was published by hand.

The release commit is pushed by the bot, so any PR CI that runs on it sits in action_required (GitHub suppresses recursive runs) until someone approves it: gh api -X POST repos/<owner>/<repo>/actions/runs/<id>/approve.

4. When it fails

Nothing before the push has left the runner, so the default move is to fix the cause and rerun. The exception is "npm has it, the branch does not" — that one needs an explicit decision. Both, with the exact commands and preconditions, are in references/recovery.md.

Traps

The catalogue with symptoms, root causes and fixes is references/pitfalls.md. The ones that cost this repository a release:

  • A brand-new package cannot be published by the workflow. Trusted Publishing cannot create a package name; npm answers the first publish with a misleading 404. Bootstrap it once by hand, register the Trusted Publisher, rerun. prepare checks package existence before it fans the validation out and before anything is pushed, so the failure costs seconds.
  • 404 on a package that already exists means the workflow is not one of its publishers — same fix, different cause.
  • Publish order must be dependency order. plugins-before-packages once published dsh-plugin-kit last, leaving versions on npm that no consumer could install. publish-release.mjs owns the order now and holds back a package whose dependency did not publish; never reintroduce a shell loop.
  • git am refuses the release-commit patch with does not match index when a bin target is committed non-executable: pnpm install chmods it 0755 on POSIX runners and the patch carries a mode change. Commit the exec bit (git update-index --chmod=+x <file>) so the chmod is a no-op. Windows never reproduces this — a local -rwxr-xr-x proves nothing.
  • Nx removes consumed plans outside its own commit. The prepare job stages the tree and folds the deletion in, or the next run applies spent plans again.
  • dry_run and a cached local run both lie. See above.
  • A flaky retry hides a real defect. Nx marks a task flaky and passes on retry; a test that fails under load and passes isolated may still be broken. Fix it or report it, never write it off.
  • The tarball gate is the only gate that compares exports against the real build. A copied exports map passes typecheck, tests and pnpm check, and first fails at pack time — on a push.
  • Someone else's uncommitted work is usually in the tree. A red run may not be yours, and git add -A will take their files. Stage explicit paths and assert the staged count; git add with a nonexistent pathspec adds nothing, silently.
  • The release test finds its own package managers. It reads npm_execpath as a location rather than as node source — a standalone pnpm.exe is spawned directly, a .cjs/.js wrapper through node — and falls back to the install beside node or to PATH. So node --test scripts/release-workflow.test.mjs reproduces what the gate runs.

References

  • references/preflight.md — the check sequence, what each gate covers, and the fresh-checkout emulation.
  • references/pitfalls.md — incident catalogue: symptom, root cause, fix, and how to catch it before a release.
  • references/recovery.md — failure matrix, publish_only semantics, registry verification recipes.
  • Canonical: docs/RELEASING.md, docs/VERIFICATION.md, .agents/skills/create-plugin/references/release-and-gates.md, AGENTS.md.

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.