Release Semver

Semantic Versioning Release Process

Intro

Releases are routine, not heroic. Decide the version bump from the nature of the changes, draft a changelog, update version files in one commit, tag with vX.Y.Z, and publish. The goal is a release anyone on the team can perform from the same checklist.

Overview

Determine the version bump

Semantic versioning has three numbers: MAJOR.MINOR.PATCH.

• Patch (0.1.0 → 0.1.1): bug fixes only, no API changes.
• Minor (0.1.0 → 0.2.0): new features, backward-compatible.
• Major (0.1.0 → 1.0.0): breaking changes.

Pre-1.0 projects bend the rules: minor bumps are allowed to break compatibility, and patch bumps cover both features and fixes. Once you ship 1.0, the rules become strict.

Pre-release checklist

• All tests pass on the main branch.
• No uncommitted changes in the working tree.
• Dependencies reviewed and up to date where appropriate.
• CHANGELOG or release notes drafted.
• A root LICENSE file exists and matches the intended project license. If the project ships release archives, the archive build must include that LICENSE at the top level. For processkit, scripts/build-release-tarball.sh copies root LICENSE into processkit-vX.Y.Z/.

Draft the changelog

Group entries under Added, Changed, Fixed, and Removed. Reference issue and PR numbers. Note breaking changes prominently — call them out in their own section if a major bump is involved.

The release flow (single turn, bulletproof)

A release is one continuous sequence — you do not stop halfway. /pk-release vX.Y.Z runs every step below in order, in the same turn, and is not considered done until the final verification passes. See DEC-20260422_1348-SnowyWolf.

1. Bump version in every relevant file (aibox.lock, Cargo.toml, pyproject.toml, package.json, lockfiles).
2. Finalize CHANGELOG — rename the [vX.Y.Z-candidate] section to [vX.Y.Z] — YYYY-MM-DD; ensure Added/Changed/Fixed/Removed grouping; call out breaking changes if a major bump.
3. Regenerate provenance / transitive artifacts — for processkit, scripts/stamp-provenance.sh vX.Y.Z.
4. Commit the bump with message chore(release): bump to vX.Y.Z.
5. Tag the commit: git tag -a vX.Y.Z -m "<tag summary>".
6. Push branch then tag: git push origin main && git push origin vX.Y.Z. A tag push alone does not create a GitHub Release — it is a git ref, not a distribution-channel artifact.
7. Create the GitHub Release with notes extracted from the CHANGELOG and attach the release archive plus checksum when the project produces binary/archive assets. For processkit, build first with scripts/build-release-tarball.sh vX.Y.Z; the tarball must contain top-level LICENSE and CHANGELOG.md. The tag commit must also contain root LICENSE, so GitHub's generated source archives carry the license. The single canonical command:

   gh release create vX.Y.Z \
     dist/processkit-vX.Y.Z.tar.gz \
     dist/processkit-vX.Y.Z.tar.gz.sha256 \
     --repo <org>/<repo> \
     --title "vX.Y.Z — <one-line summary>" \
     --notes-file <(awk -v v=X.Y.Z '
       $0 ~ "^## \\[v" v "\\]"   { f=1; next }
       f && /^## \[/              { f=0 }
       f
     ' CHANGELOG.md) \
     --latest
   

8. Verify the Release is live. This is the release's completion gate. For archive-based releases, verify the Release has the expected tarball/checksum assets and that extracting the tarball shows a top-level LICENSE; also verify the tag itself contains root LICENSE:

   gh release view vX.Y.Z --repo <org>/<repo>
   tar -tzf dist/processkit-vX.Y.Z.tar.gz | grep '/LICENSE$'
   git show vX.Y.Z:LICENSE >/dev/null
   

   If this exits non-zero, the release is incomplete — return to step 7. Do not report the release as done until step 8 succeeds.
9. Other channels (crates.io, PyPI, npm, container registry, homebrew tap, etc.) — run after step 8 so the GitHub Release remains the canonical artifact pointer.

Recovery pattern — if a tag was pushed in a prior session without the Release, run gh release create and gh release view for that tag directly; /pk-publish exists as an alias for this recovery scenario.

pk-doctor release_integrity check (detection layer) — walks every local v* git tag, probes GitHub for a matching Release, and WARNs on any tag without one. Run /pk-doctor --category=release_integrity after a release to confirm; run /pk-doctor routinely to catch any historical gap. Requires gh CLI + auth; emits INFO when gh is unavailable.

Example

User says: "Let's release — we fixed two bugs and added a feature." The agent recommends a minor bump (because of the new feature), drafts changelog entries grouped Added/Fixed, updates version files, creates the bump commit, and tags it.

This skill also provides the /pk-release and /pk-publish commands for direct invocation — see commands/pk-release.md and commands/pk-publish.md.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• Bundling a breaking change into a patch or minor release. Consumers rely on semver guarantees to control when they take breaking changes. A breaking change in a patch release violates the contract and causes silent runtime failures in consumer code.
• Skipping the changelog because "git log has it". Git log is not a changelog. It lacks context, grouping, and the human judgment that makes a changelog useful. Every release needs a curated changelog with Added/Changed/Fixed/Removed sections.
• Bumping all version files in separate commits. A brief window where Cargo.toml says 1.2.0 but pyproject.toml still says 1.1.0 is a broken state. Update all version files in a single atomic commit.
• Tagging before CI passes. A tag should point to a commit that has been fully built and tested. Tagging a commit that later fails CI means the published artifact may not match what was verified.
• Publishing from a developer machine without a clean checkout. Uncommitted local changes — WIP files, debug flags, console.log statements — may sneak into the artifact. Publish from CI with a clean checkout or from a fresh clone.
• Re-tagging a published version to fix a mistake. Package registries (crates.io, PyPI, npm) prohibit or block re-publishing to an existing version. Cut a new patch version (e.g. 1.2.1) to fix a mistake in 1.2.0; never overwrite a tag.
• Pre-1.0 projects that promise stability they cannot keep. If the API is still moving, stay below 1.0 and say so. Cutting 1.0.0 too early creates a compatibility obligation that slows future improvements.
• Treating git push --tags as "the release is out." Pushing a tag creates a git ref; it does not create a GitHub Release, a crates.io release, a PyPI release, or any other distribution-channel artifact. Downstream consumers (package managers, gh api, the Releases page, auto-sync tools like aibox) read from the channel, not the tag. Until gh release create (or the equivalent channel publish command) has run, the release is not visible. The flow above bakes this in: the release is not complete until step 8 (gh release view) succeeds. See DEC-20260422_1348-SnowyWolf.
• Stopping after /pk-release prepare-like work. In an older split where /pk-release prepared and /pk-publish published, it was easy to do only half. The current flow is single-turn: /pk-release runs every step through GitHub-Release verification. /pk-publish remains only as a recovery alias for historical tags without Releases.

Full reference

When pre-1.0 ends

Cut 1.0 when the public API is something you are willing to support without breaking. Until then, keep the version below 1.0 honestly — there is no shame in 0.17.0, and it signals to users that the surface may still move.

Version bumps for special changes

Anti-patterns

• Bundling breaking changes into a patch release
• Skipping the changelog "because git log has it"
• Tagging a commit that was not built and tested
• Publishing from a developer machine without a clean checkout
• Re-tagging a published version to fix a mistake (cut a new patch instead)
• Pre-1.0 projects that promise stability they cannot keep