release

name

release

description

Cut a zen-linear release end to end. Resolves the version from the commit range, checks the docs that are about to represent a shipped version, curates docs/release-notes/vX.Y.Z.md into copy for the person downloading, runs the gate, hands over the tag command, then verifies what actually published. Manual invocation only, never auto-run: use it when Drew asks to cut a release, not because main looks ready.

Release (zen-linear)

.github/workflows/release.yml already handles the mechanical half: it runs the tests, cross-compiles five targets, writes the checksums, and cuts the release from the notes in the tagged commit. This skill does not reimplement any of it.

What it does not do is judgement. Which commits a stranger cares about, what the version should be, whether a doc about to become the documentation for a shipped version is still true, and how the notes read. That is this skill.

Say almost nothing. Apply these rules silently. Never restate a rule or justify a step you are about to take. When a rule stops you, name what is blocked in one line and what you need to continue.

Drew pushes the tag. Phases 1 to 5 stop at handing over the command. A published tag is permanent and cannot be renumbered. Phase 6 resumes once he reports it pushed: a release is not done until the artifacts are verified.

1. Scope the release

• git fetch --tags, then confirm main is checked out, clean, and even with origin. Anything else stops here. Local main silently carrying commits that never reached origin is a real state this has been in, and tagging it cuts a release from a tree nobody reviewed.
• git log <last-tag>..HEAD --no-merges for the range. History is one squash commit per PR, so the commit titles are the unit.
• Sort every commit into user-facing or internal, and show the split. Internal work is most of a typical range: CI, agent config, refactors, test-only changes. It does not appear in the notes.
• A commit is user-facing if someone who only runs the tool would notice, or if it changes what they download, install, or read.

Before the first release there is no last tag. The range is the whole history, and most of it is the fork's. Scope that one off what the tool does today rather than off the log.

Then propose the bump. docs/CONTRIBUTING.md under "Version numbers" governs which, and get Drew's confirmation before going further.

2. Check the docs the release makes public

Derive the set rather than reading all six. git diff --name-only <last-tag>..HEAD gives the surfaces the range touched, and "Docs that describe the code" in docs/CONTRIBUTING.md maps each one to the documents that describe it. Read those.

This matters more here than at merge time. A doc that is wrong at this moment becomes the documentation for a shipped version, and the release page links into it.

Then work the other way, from phase 1's user-facing list. Every entry on it has to be findable in a document that says what the code now does. That is the check the map cannot make: the map catches a doc whose subject moved, and this catches a change that arrived with no doc at all. A change nothing describes is either a gap to fill now or work no user sees, and phase 1 already decided it was user-facing, so it is usually the first.

docs/install.md earns the most attention of any of them. It names the platforms the release carries, the requirements, the auth paths, where the binary lands, and what --version reports. All of those are claims a release makes true or false.

docs/keys.md is the one most likely to have drifted, because a key changes in a PR that had no reason to think about it. ? renders the same keys the code holds, so the app is the reference, not the last version of the file.

Fix what the range made wrong before writing the notes, and report which documents were read and which changed. Phase 5 commits them.

3. Curate the notes

Write docs/release-notes/vX.Y.Z.md, for the person downloading rather than the person who wrote the patch. Drew's voice: terse, considerate, stoic, no strong adverbs, no em-dashes.

Structure:

• No H1. gh titles the release from the tag, so one in the notes repeats it.
• One line saying what this release is.
• ## New / ## Fixed / ## Changed, only the ones that apply. Bold lead-in naming the thing, then plain sentences about what changed. A first release has no before to differ from, so it says ## What it does instead of all three.
• ## Install.
• ## Known gaps, where there are any worth naming.

Check ## Install against what docs/install.md currently says rather than copying the previous file forward. It goes stale the moment a platform or an install path changes, and it is the section a reader acts on.

There is no card block and no appcast. Nothing reads these but a person on the release page.

Run the mechanical checks before showing anything. They are cheap and catch what rereading your own copy does not:

grep -c '—' docs/release-notes/vX.Y.Z.md    # must be 0
grep -nEi 'seamless|powerful|beautiful|just works|simply|easily|quickly' docs/release-notes/vX.Y.Z.md

Then show Drew the notes and get approval. Copy is the deliverable here.

4. Gate

make all, fully green: gofmt, go.mod tidiness, golangci-lint, the race suite, and the build. Run it directly, never through a pipe that swallows the exit code.

5. Hand over

Commit the notes and phase 2's doc fixes to main. Doc-only changes go straight to main here, and the pre-push hook rejects pushes to it, so the commit is yours and the push is his.

Then confirm the tree is clean again before handing anything over. Phase 1's check does not run twice, and a doc fix left uncommitted means the release links to the stale page phase 2 existed to correct.

Then give Drew both commands, in this order, with the version named explicitly:

git push origin main
git tag -a vX.Y.Z -m "zen-linear vX.Y.Z" && git push origin vX.Y.Z

main first, and it is not a style preference. The release job checks out the tag and reads docs/release-notes/${GITHUB_REF_NAME}.md out of it. A tag cut before the notes commit is on main fails the run, and the tag is already published by then.

The hook guards refs/heads/main only, so the tag pushes without --no-verify. Never suggest --no-verify for either.

Stop here. Do not run them.

If the tag goes first anyway, the run fails on the missing notes and re-running it fails identically, the tag being fixed to a commit that does not carry them. Do not move a published tag. Push the notes to main, then cut the release against the tag by hand, from the archives the failed run left as artifacts:

gh release create vX.Y.Z dist/* --notes-file docs/release-notes/vX.Y.Z.md --verify-tag

6. Verify what published

Resume once Drew reports the tag pushed, and check what landed rather than trusting the workflow's exit code.

• gh release view vX.Y.Z --json assets shows six assets: four tarballs, the windows zip, and checksums.txt. A missing archive is a matrix leg that failed after the others published.
• Download one archive and match it against the published checksum. The checksums are generated in a different job from the builds.
• Run the installer against the live release, with INSTALL_DIR at a scratch path. Nothing before a release exists can exercise its download path, so this is the only time it is ever tested.
• The installed binary reports the version rather than dev. A dev here means the ldflags stamp broke, and the binary is otherwise fine, which is why nothing else catches it.
• Run it against a real Linear workspace, not just --version. Load a list, open an issue, save a field.

No Linear status work. Shipped tickets are already Done from their merges.

Hand back

## Released vX.Y.Z

**Release** <link>

**Shipped**
- <what a user gets>

**Verification**
- Docs: <n> read, <n> updated
- Assets: <n> of 6
- Checksum: <matched or not>
- Installer: <result against the live release>
- Version: <what the installed binary reports>
- Live run: <what was exercised against a real workspace>

**Known gaps**
- <what shipped unfinished, and where it is tracked>

The run is not complete while an asset is missing, a checksum is unmatched, the installer is untested against the live release, the binary reports dev, or a user-facing change in the notes has no document describing it.