kobe versioning + changelog are managed with Changesets. The published package is @sma1lboy/kobe (packages/kobe); packages/branding is private and never published.
When you land a change that affects what the published package does (a feature, fix, or behaviour change), record it as a changeset:
bun run changesetThis prompts for the bump type (patch / minor / major) and a summary, then writes a .changeset/<random-name>.md file. Commit that file with your change. Because each change is its own file, two parallel branches never collide on CHANGELOG.md the way appending to a shared [Unreleased] section did.
- The summary is the user-facing changelog line — write it in product voice, present tense ("Add X", "Fix Y", or a short narrative). It lands verbatim under the next release.
- A pure tooling / docs / CI change that doesn't touch the published package needs no changeset. If you want to record "intentionally nothing to release", run
bun run changeset -- --empty. - Bump type:
patchfor fixes,minorfor features,majorfor breaking changes. kobe is pre-1.0, so breaking changes still gominorby convention unless we decide otherwise.
scripts/release.shThis consumes every pending .changeset/*.md:
changeset version— computes the next version from the pending bump types, rewritespackages/kobe/package.json, and prepends the collected notes toCHANGELOG.md(then deletes the consumed changesets).- Runs
bun install, thenbun install --frozen-lockfile, sobun.lockmatches the workspace package versions before the release commit is made. - Re-runs Biome
--writeon the touchedpackage.json/CHANGELOG.mdso the generated JSON formatting can't fail the lint gate (Changesets and the release script both reserializepackage.json, which used to re-expand the single-linefilesarray). - Commits
chore: release — X.Y.Z, tagsvX.Y.Z, and (after confirming) pushesmain+ the tag.
The push triggers .github/workflows/release.yml, which gates on typecheck + test + build, then npm publishes, extracts the new CHANGELOG.md section as the GitHub release body, and builds the standalone binaries.
The release workflow's gate does not run lint. Run
bun run lintlocally (or rely on the push-triggeredci.yml) — a lint regression won't block a publish on its own.
GitHub renders release bodies with GFM's hard-break extension: every single newline inside a list item or paragraph becomes a <br>, which makes the release page look like a narrow column broken every ~70 chars. Write each changeset bullet (and each paragraph) as one long line. Editors can soft-wrap at display time. KOB-13 has the rationale; the changelog-generator skill knows this rule.
A prerelease tag (v0.7.0-experimental.0) publishes to an npm dist-tag named after the prerelease identifier (experimental), so latest stays on the stable line while testers opt in with npm i @sma1lboy/kobe@experimental. Use Changesets' prerelease mode (changeset pre enter experimental … changeset pre exit) to generate those versions.