release

Add a release (version bump + changelog) to the current branch before its PR is created or updated. Every PR must include a release.

This skill is non-interactive. Decide the version and changelog yourself; do not ask the user to confirm the classification, version, or wording. Only stop and report if a script returns an error.

Steps

1. Gather release info

Run the info script to collect merged PRs and calculate the version:

bash .claude/skills/release/scripts/release-info.sh 2>&1

Capture the JSON output (errors may arrive on stderr, hence 2>&1). If the error is already_released, the branch already carries its release commit (each branch gets exactly one; later commits such as review fixes ship under it). There is nothing to do, so skip the remaining steps and report that. This is not a failure. stale_release means another PR already shipped this branch's version; stop and tell the operator to bump the version manually by updating all four version fields (.claude-plugin/plugin.json, the codagent entry in .claude-plugin/marketplace.json, .codex-plugin/plugin.json, .cursor-plugin/plugin.json) and renaming the branch's ## <version> heading in CHANGELOG.md to match, then rerun this skill. Updating only some of these can make the next run report already_released and leave the rest stale. For any other "error", stop and report the message to the user.

2. Decide the classification and version

The script's classification comes from conventional commit prefixes, which is only a heuristic. Review it and adjust on your own judgment:

• Internal changes (tooling, CI, repo automation, skills used only by maintainers such as this one) are patch-level even if prefixed with feat:.
• User-facing new capabilities in shipped skills are minor.
• Breaking changes to shipped skills' behavior or interfaces are major.

Recompute new_version from current_version if you change any category (major → X+1.0.0, minor → X.Y+1.0, otherwise X.Y.Z+1). Briefly state the chosen version and classification in your output, then continue without waiting.

3. Write changelog descriptions

For each PR in the JSON output, write a one-sentence description that is slightly more informative than the raw PR title. Strip the conventional commit prefix from the title (feat:, fix:, chore:, etc.) and expand it into a sentence that gives enough context to understand the change without clicking through.

If an entry has "is_branch_commits": true and "number": 0, it represents unmerged commits on the current branch. Read the commit messages (git log origin/main..HEAD --oneline --no-merges) and write a description summarizing those changes. Omit the PR link for these entries since they have no PR yet — just use a bullet with the description.

4. Format the changelog section

Build the changelog section for the new version. Only include sections that have entries. Sort entries by PR number ascending within each section.

Format:

## <new_version>

### Major Changes

- [#<number>](https://github.com/Codagent-AI/agent-skills/pull/<number>) <description>

### Minor Changes

- [#<number>](https://github.com/Codagent-AI/agent-skills/pull/<number>) <description>

### Patch Changes

- [#<number>](https://github.com/Codagent-AI/agent-skills/pull/<number>) <description>
- <description without link> (for branch-commit entries with no PR)

5. Write the changelog section to a temp file

Write the formatted changelog section (from step 4) to a temporary file:

TMPFILE=$(mktemp)
cat <<'CHANGELOG_EOF' > "$TMPFILE"
<paste the formatted changelog section here>
CHANGELOG_EOF
echo "$TMPFILE"

6. Commit the release

Run the release script with the new version and temp file path:

bash .claude/skills/release/scripts/create-release-pr.sh <new_version> <tmpfile_path>

On a feature branch, this commits the release onto the branch and pushes it; if the branch has no PR yet, the calling workflow (e.g. codagent:push-pr or codagent:finalize-pr) creates it next. On main, it creates a release/v<new_version> branch and opens a release PR.

7. Report

Print the version and the PR URL returned by the script (or note that no PR exists yet).

Clean up the temp file:

rm -f "$TMPFILE"