用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/PaulRBerg/agent-skills --skill release-bumper命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| argument-hint | [packages...] [version] [--beta] [--dry-run] |
| compatibility | Requires Bun, Git, and uv. |
| disable-model-invocation | true |
| effort | high |
| model | sonnet |
| name | release-bumper |
| skill-dependencies | ["cli-gh"] |
| user-invocable | true |
| description | Cut a release: bump versions, write changelogs, commit, tag. |
If these instructions are already present in the conversation from a slash or dollar invocation, follow them directly; do not invoke this skill again through a skill tool.
Release one package or several packages with version bumps, changelog entries, commits, and tags. Supports single-package repositories, workspace monorepos, stable releases, beta releases, and dry runs.
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.
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.Resolve <skill-dir> from this SKILL.md. Keep helper stdout as JSON and diagnostics on stderr.
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:
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 assignment. The finalizer never chooses a regular release magnitude or dependency policy.
--versionFor every stable changelog written, validate its deterministic structure:
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.
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.
Require workingTree.clean. Do not absorb unrelated work.
Resolve unknown or ambiguous package selection. An explicit user version remains single-package only.
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.
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.
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.
For a dry run, report the ordered package/version plan, range edits, changelog/tag/commit actions, and agent-decided skips. Stop before writes.
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.
Update manifests and accepted dependency ranges. Validate every stable changelog with the helper.
Format once using the repository's narrowest established command.
Commit and tag dependencies before dependents. Use one commit and one annotated tag per package:
docs: release <version>;docs: release <package> <version>;v<version> or bare-semver facts;<package-dir>@<version>.Do not push. After success, recommend an exact git push origin <tag>... command containing only the tags created
by this execution; do not use --tags.
Before the final report, inspect .github/workflows/ 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 read-only to 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:
$cli-gh.$cli-gh even when release CI
exists, because the multi-tag push is not expected to trigger that automation reliably.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.
Helper failures mean malformed input, violated invariants, or failed validation; an agent decision remaining unresolved 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, and any applicable GitHub release proposal.
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.