---
argument-hint: "[packages...] [version] [--beta] [--dry-run]"
compatibility:
  Requires Bun, Git, and uv. Staged npm approval requires the shared Chromium browser with the 1Password extension.
disable-model-invocation: true
effort: high
model: sonnet
name: release-bumper
skill-dependencies:
  - chromium-browser
  - cli-gh
description: "Cut a release: bump versions, write changelogs, commit, tag, and approve staged npm publishes."
---

# Release Bumper

If a slash or dollar invocation already added these instructions to the conversation, follow them directly. In that
case, do not invoke this skill again through a skill tool.

Release one package or several packages with version bumps, changelog entries, commits, and tags.

A non-dry-run invocation authorizes the repository-local version edits, changelogs, commits, and annotated tags defined
by this workflow. Do not add a generic confirmation gate after the agent derives the release plan. Ask only when an
unresolved package selection or dependency-range policy changes the release set or approach. GitHub release creation
remains a separate external write governed below.

A request to push, ship, or publish the release authorizes the tag pushes in step 12. It also authorizes approval of the
staged npm versions that those pushes create, as step 14 defines. Do not ask the user to approve them by hand.

## Arguments

- `packages`: optional package names or directories. Omit in a single-package repository.
- `version`: optional explicit semver. Valid only for one user-selected package.
- `--beta`: create or advance a `-beta.X` prerelease.
- `--dry-run`: preview without modifying files, committing, or tagging.

## Helper Interface

Resolve `<skill-dir>` from this `SKILL.md`. Keep helper stdout as JSON and diagnostics on stderr.

```sh
bun run "<skill-dir>/scripts/plan-release.ts" \
  [--cwd <repo>] [--beta] [--dry-run] [--version <semver>] \
  [--package <name-or-dir>]...
```

The read-only discovery output has `schemaVersion: 2`. It reports package identity, complete per-package `changedFiles`,
workspace edges and declared ranges, previous-tag facts, selected targets, and worktree state. `changeHints` are
filename-based, explicitly non-authoritative navigation hints. Never use them to decide release relevance or changelog
inclusion.

After the agent decides every stable patch/minor/major version, write discovery JSON to a temporary file and run:

```sh
uv run "<skill-dir>/scripts/finalize-release-plan.py" \
  --discovery <discovery.json> \
  [--version <package>=<semver>]...
```

The finalizer performs beta and explicit-version transitions, stable prerelease promotion, npm-range satisfaction,
simple dependency-range suggestions, and dependency ordering. It reports complex ranges, peer ranges, dependency cycles,
and stable versions not supplied by the agent as unresolved decisions. When an unsatisfied edge adds a dependent, choose
that package's release version and rerun with another `--version` assignment. The finalizer never chooses a regular
release magnitude or dependency policy.

For every stable changelog written, validate its deterministic structure:

```sh
uv run "<skill-dir>/scripts/validate-changelog.py" \
  --file <CHANGELOG.md> --version <semver> --date <YYYY-MM-DD> [--tag <tag>]
```

This checks the expected release and date, heading/category order, allowed categories, list structure, and release-link
tag. It does not judge importance, wording, or semantic category.

## Workflow

1. Run discovery with the user arguments mapped directly. Exit `2` means the target is not a releasable Git/package
   repository. Exit `64` means invalid input. Stop on either.
2. Require `workingTree.clean`. Do not absorb unrelated work.
3. Resolve unknown or ambiguous package selection. An explicit user version remains single-package only. If a target's
   `previousTags` version is below its manifest version, that version shipped untagged and the reported base is stale:
   create the missing annotated tag on its release commit (for example `docs: release <version>`) using the observed tag
   convention, include it in the push command, and rerun discovery. Ask only when no release commit is identifiable.
4. Inspect each target's complete `changedFiles` and the net diff from its previous tag. Decide whether the surviving
   changes warrant a release. Runtime environments, refactors, documentation, tests, and tooling can all be relevant in
   context. Filenames never decide this.
5. For every relevant stable target without an explicit version, choose patch, minor, or major from the consumer-facing
   change. For beta releases, let the finalizer compute the mechanical transition.
6. Run the finalizer. Review unsatisfied workspace edges. Accept its suggestion only for a simple dependency range when
   that policy fits. Choose peer and complex range policy explicitly. Add dependents and their agent-chosen release
   versions, then rerun until the package set and dependency order are resolved.
7. For a dry run, report the ordered package/version plan, range edits, changelog/tag/commit actions, and agent-decided
   skips. Stop before writes.
8. For a stable release, read `references/common-changelog.md` and write consumer-facing entries from the bounded net
   diff. The agent owns entry selection, wording, importance, and category. Beta releases do not update changelogs.
9. Update manifests and accepted dependency ranges. Validate every stable changelog with the helper, in a single shell
   invocation that loops over the changelogs and reports each result.
10. Format once using the repository's narrowest established command.
11. Commit and tag dependencies before dependents. Use one commit and one annotated tag per package:
    - single-package commit: `docs: release <version>`.
    - monorepo commit: `docs: release <package> <version>`.
    - single-package tag: follow observed `v<version>` or bare-semver facts.
    - monorepo tag: follow observed package tag facts, defaulting to `<package-dir>@<version>`.
12. Push only when the user's request asks to push, ship, or publish the release. In that case, push the release
    commits, then push only the tags created by this execution, with at most three tags per command. Otherwise, do not
    push, and after success recommend an exact `git push origin <tag>...` command containing only the tags created by
    this execution. Never use `--tags`.
13. Before the final report, inspect `.github/workflows/` once for both the criteria in this step and the criteria in
    step 14. For this step, look for an active workflow that creates or publishes GitHub releases from pushed tags. A
    filename such as `release.yml` is a hint, not proof. Use `$cli-gh` to run `gh release list` read-only in the same
    turn as that inspection and check whether the repository has an established history of maintained GitHub releases.
    If it does, offer to create a GitHub release for each new tag, pending the user's approval, according to these
    rules:
    - One to three tags and applicable release CI exists: do not offer manual release creation. In that case, the tag
      push should trigger CI.
    - No applicable release CI exists: offer to create one release per new tag with `gh release create` through
      `$cli-gh`.
    - More than three tags will be pushed together: offer to create one release per tag with `gh release create` through
      `$cli-gh` even when release CI exists, because GitHub does not create tag push events above that threshold. See
      [GitHub's push-event limits](https://docs.github.com/en/actions/reference/workflows-and-actions/events-that-trigger-workflows#push).

    Never create a GitHub release without the user's approval. If release history cannot be verified, report it as
    unknown and do not offer the write.

14. Before the final report, also use the single `.github/workflows/` inspection from step 13 to look for an active
    workflow that runs on pushed tags, sets `id-token: write`, and runs `npm stage publish` (staged) or `npm publish`
    (direct). A filename is a hint, not proof. If such a workflow exists:
    - Report that pushing the recommended tags triggers npm publication.
    - Recommend push commands with at most three tags each, because GitHub creates no tag push events above that
      threshold. See the push-event limits link in step 13.
    - For staged publishing, each version stays unpublished until someone approves it with 2FA. When this execution
      pushed the tags, read `references/npm-staged-approval.md` and approve each staged version through
      `$chromium-browser`. Do not stop to ask the user to approve it.
    - When the tags are not pushed, report that approval is pending. If the user pushes the tags later in the same
      session, run the approval workflow then.
    - Never run `npm publish`, `npm stage approve`, or `npm stage reject`.

## Safety and Completion

Helper failures mean malformed input, violated invariants, or failed validation. An unresolved agent decision is data in
the JSON, not a helper failure. Discovery and dry-run are read-only. Do not write changelogs before the final stable
package set is known, and do not infer a tag convention when discovery reports observed facts.

Dry-run completion requires a discovery-backed, agent-reviewed action preview with zero writes. Release completion
requires validated manifests and stable changelogs, formatting, one commit and annotated tag per package in dependency
order, and a report of created commits/tags, agent-decided skips, the exact tag-push command or commands, any applicable
GitHub release proposal, and any CI npm publication notice. When this execution pushed tags to a staged-publish
workflow, completion also requires a `npm view` result for each approved version, or the exact human gate that blocks
it.

Use `### ⛔ Release stopped — working tree is not clean`, `### ⚠️ Release decision required`,
`### 🔎 Release preview — no files, commits, or tags written`, or `### 🏁 Release complete` as applicable. Keep helper
JSON, versions, hashes, tags, commands, and changelog text exact and undecorated.
