Changelog Generator (Rove)

Drafts release notes as Changesets — one .changeset/<name>.md file per change — in Rove's house style. scripts/release.sh (via changeset version) later consumes them into `packages/kobe/CHANGELOG.md` and the GitHub release body. See `docs/RELEASING.md` for the full flow.

When to use

• The user says "draft changelog", "write release notes", "add a changeset", "what changed since v0.X.Y", or similar.
• After landing a user-facing change that has no changeset yet (e.g. it was committed before this skill existed, or in a batch that skipped them).
• Before cutting a release tag — to backfill changesets for anything user-facing that slipped through, so the generated notes are complete.

Rove project conventions (load-bearing)

Every rule in this section overrides the generic guidance further down.

File format — a changeset, not a CHANGELOG edit

• Do not hand-edit packages/kobe/CHANGELOG.md or invent a ## [Unreleased] section — that file is generated.

• Each change is a new file .changeset/<two-words-random>.md (the changeset CLI names it; if writing by hand, any unique kebab name works). Shape:

  ---
  "@sma1lboy/rove": patch
  ---
  
  Single-line user-facing summary. This text lands verbatim under the next release and in the GitHub release body.

• The canonical frontmatter bump key is @sma1lboy/rove, with value patch | minor | major.

• Default to patch for every change, including features and pre-1.0 breaking changes. Use minor or major only when the user explicitly requests that bump in the current turn.

• The bump type is the category — Changesets groups output under ### Minor Changes / ### Patch Changes automatically. Don't write ### Added/### Fixed headings yourself.

• One changeset per coherent change. A batch that did three user-visible things → three changesets (or one with three bullets if they're one feature). Prefer the changeset CLI: bun run changeset (interactive) writes the file for you.

HARD RULE — no soft wraps

Every bullet, every paragraph in a changeset body must be on a single line. Do not wrap at column 70/80/whatever. The line can be 400 chars long; that's fine.

Why: GitHub renders release bodies with GFM's hard-break extension. Each newline inside a list item or paragraph becomes a <br> tag. Soft-wrapped text renders as a narrow column broken every ~70 chars on the live release page, which looks broken (KOB-13).

Voice

• Present tense, user-perspective. "Add X", "Fix Y", "Move Z" — not "Added X", not "I added X".
• Lead with what changed, not why. The why goes in a follow-up clause if it's non-obvious.
• Short bold lead-in for headlines (**The thing** — explanation...) is the established pattern.
• Reference internal anchors with backticks (`task.new`, `ctrl+,`, `packages/kobe/src/foo.ts`) rather than prose.

Filtering

Pull in: features, behaviour changes the user can see/feel, bug fixes affecting user-visible behaviour, distribution / packaging / install changes.

Skip: pure refactors, internal test additions (UNLESS a milestone), CLAUDE.md / docs / skills / memory / agent-config tweaks, dependency bumps with no behaviour delta, CI tweaks (unless a new gate the user cares about). A change that needs no release can still record that explicitly with bun run changeset -- --empty.

When in doubt, ask "would a Rove user reading this on github.com/Sma1lboy/rove/releases care?" If no → skip (or empty changeset).

How to draft

1. Find the cut point: the latest ## [<version>] heading in packages/kobe/CHANGELOG.md, or the last v* tag.
2. Run git log --no-merges <last-tag>..HEAD --pretty=format:'%h %s%n%b%n---' to get the commit set, and check .changeset/*.md for changes that already have one (don't duplicate).
3. Group the un-covered user-facing changes. Write one changeset file per coherent change, each with the right bump type and a single-line summary per the rules above.
4. Prefer bun run changeset so the CLI writes the file; only hand-author the .changeset/<name>.md if scripting a batch.
5. Surface the new changeset file(s) and ask the user to skim before committing them.

Example output (a changeset file)

.changeset/swift-pandas-cheer.md:

---
"@sma1lboy/rove": patch
---

**The Tasks pane fills its tmux pane and adapts to its width** — the task list now stretches to 100% of the pane as you drag the tmux split. On a narrow pane the secondary columns step aside so the task name stays readable: the branch label drops first, then the changes chip, and the title ellipsises only when it must.

Note: the summary is one long line. No newlines inside it. That's the only reliable way to make the GitHub release page render flowing text.

What to avoid

• ❌ Hand-editing packages/kobe/CHANGELOG.md or recreating a ## [Unreleased] section — it's generated from changesets now.
• ❌ Soft-wrapping a changeset summary at column 70 because "it looks nicer in the editor". Render-time soft-wrap exists for a reason.
• ❌ Changesets like "Refactor X to use Y pattern" — internal change, skip (or --empty).
• ❌ Auto-generating from git log without filtering. Most commits are noise.
• ❌ Touching the version number in package.json — changeset version (run by scripts/release.sh) owns that.