agentleFS
TheodoreADTheodoreAD/agent-skills/skills/release-herald/SKILL.md

release-herald

Skill0 starsChanged 49 days agoMITView on GitHub

Use when writing the release notes for your own repository's release — drafting the GitHub release for a tag from every pull request and commit since the previous release, working out what users must change (breaking, API and configuration changes, deprecations), listing what the release publishes (packages, images, charts, on any registry), or creating that GitHub release as a draft through gh, on github.com or GitHub Enterprise Server. Handles gitflow (squash-merged feature PRs, release and hotfix PRs into main) and trunk repos, and says when PR descriptions or commit messages are too thin to write from. Not for judging someone else's package or its release cadence, and never cuts, bumps, tags or pushes anything.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Resolve: Run repo and check the host, the repository and the flow it reports.

    line 68
  2. Distinctive:

    Trunk: every commit between the two tags.

    line 108
  3. Distinctive:

    Read the surface diffs. A renamed key, a changed default or a removed parameter is a change the user must act on, whether or not anyone …

    line 136

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

A release users can act on

Release notes are for the people who use what a repository ships. They lead with what those people must change, then what is new for them, then everything else. This skill collects the whole change set between two releases through gh, without touching git. It grades how much the PR descriptions and commit messages actually say, and hands you a digest to write from. Writing is the part left to you.

Two rules hold throughout. Draft from what the pull requests and commits say, and never invent a change, a migration step or a version. Never tag, bump, push or publish on your own initiative: the default output is markdown, and creating the release is a separate step the user asks for.

What this skill reads, runs and writes

  • Reads: through gh, the repository's tags, commits, compare ranges, pull requests (title, body, labels, files, commits, linked issues), file diffs for the files that count as user-facing surface, .github/workflows/*, Actions variables, runs, jobs and deployments, and its file tree. Locally, its own config, cache and drafts.
  • Runs: gh repo view, gh api (REST and GraphQL, always with --hostname), and gh release view|create|edit (always with -R HOST/OWNER/REPO). Never git. With no --repo, gh itself reads the current directory's remote, once, to name the repository.
  • Writes:
    • setup --accept writes the per-repository config at ~/.config/release-herald/repos/<host>/<owner>/<repo>.toml.
    • collect writes the collected JSON and the digest to ~/.cache/release-herald/….
    • setup writes its proposal to ~/.local/state/release-herald/…, where drafts also go.
    • publish creates a GitHub release: a draft unless --live; --update replaces an existing release's notes. Nothing else changes on the host.
    • Nothing is ever written into the repository's working tree.
  • Network: the repository's own host only, through gh. Package registries are never contacted; where an artifact lives is read from the workflows and confirmed by the user.

The paths are the Linux layout on every OS, and $XDG_CONFIG_HOME, $XDG_CACHE_HOME and $XDG_STATE_HOME override them. A repository's config is keyed by its URL on its host, never by a working-tree path, so a worktree, a second clone, or a session outside any checkout all find the same file.

Run the script

scripts/herald.py holds every deterministic step. S is this skill's directory.

python3 $S/scripts/herald.py repo [--repo HOST/OWNER/REPO]       # host, flow, config path
python3 $S/scripts/herald.py setup                               # first time per repo: a proposal
python3 $S/scripts/herald.py setup --accept <proposal.toml>      # after the user confirms it
python3 $S/scripts/herald.py collect --tag v1.2.0                # change set, quality, digest
python3 $S/scripts/herald.py check <draft.md> --tag v1.2.0 [--dropped 12,15]
python3 $S/scripts/herald.py publish --tag v1.2.0 --notes-file <draft.md>   # a draft release
python3 $S/scripts/herald.py explain                             # how quality is graded

--repo takes HOST/OWNER/REPO, OWNER/REPO or any URL; without it, the current directory's repository. Pass it whenever the user names a repository or a host. A corporate GHES repository needs nothing else, because the host comes from the repository's URL and no host is written into the script. Every subcommand takes --json.

The procedure

  1. Resolve. Run repo and check the host, the repository and the flow it reports. If any of them is wrong, stop and ask.
  2. First time for this repository, or collect says the workflows changed: set up. setup proposes three things:

    • the flow — gitflow when a develop branch exists
    • the surface — which files count as API or configuration
    • the artifacts — read from the release workflows: a job's environment.url, registry logins, pypi-publish, uv/twine/npm publish, JFrog jf rt upload/jf docker push, docker push, helm push, aws s3 cp, release actions, task-runner publish steps — then checked against the previous release's runs and deployments

    Show the proposal to the user and have them confirm or correct it. A destination held in a secret, or published by a task runner such as inv dist.publish, cannot be read — ask for it. Then setup --accept. Do not spend effort reading registries to verify a destination; the user's answer is the source.

  3. Collect. collect --tag <tag> picks the previous release by version: a final release compares against the previous final (release candidates in between are skipped), a pre-release against the previous tag of any kind. --previous overrides. Read its summary.
  4. Report the signal before drafting, every time. Tell the user the quality counts — items, PR descriptions and commit messages, each graded empty / thin / adequate / rich. On a low-signal warning (half or more of the kept items thin or empty), say plainly that there is little to go on and list the weakest items with their links. Let the user choose: enrich those PRs first, or proceed, with those entries written from title and diff only.
  5. Draft from the digest, into the path collect printed (draft to:), following the template below.
  6. Check. check <draft> --tag <tag> --dropped <items you left out on purpose>. Fix every error; read every warning.
  7. Show the user the notes together with:
    • the range and flow used
    • what was excluded, and why
    • which entries are thin
    • any deprecation without a removal version
    • any artifact without a confirmed destination
  8. Only when asked, create the release: publish --tag <tag> --notes-file <draft> makes a draft. --live publishes, and needs the user's words saying so. --update replaces the notes of an existing release. publish refuses when the tag is not on the remote — it never creates a tag — and when a release already exists without --update. It re-runs check first.

How the change set is found

Trunk: every commit between the two tags. Each is mapped to its merged pull request, and commits with no PR are kept as direct commits.

Gitflow:

  • A release is a merged pull request into the production branch. It is identified by its base, so a release branch named by hand, rel-1.2 say, is found as surely as release/1.2. That PR is a container. Its changes are the PRs behind its commits, and its own description becomes release-level context for the overview.
  • Two comparisons are tried: from the previous tag, and from the previous release PR's branch tip. The one with fewer commits is used. When release PRs are squash-merged into main, the previous tag reaches back to the start of history and the previous release's tip does not. For a hotfix it is the other way round. For merge-committed releases the two agree.
  • Back-merges (sync/*), version bumps and merge commits are excluded and listed.

Squash-merged PRs contribute their description and their squash commit message. Merge-committed PRs contribute their description and their individual commit messages. Read both. In commits mode (--source commits, or automatic when most commits have no PR), every commit is its own entry.

Per item, the digest marks signals, which are evidence, not verdicts:

  • BREAKING? — a label, a !: title, or "BREAKING CHANGE" / "action required" in the text
  • surface — a changed file matching the repository's surface globs, with that file's diff included
  • deprecation, security (a CVE or GHSA id)
  • docs — the files a "See …" link should point at

Read the surface diffs. A renamed key, a changed default or a removed parameter is a change the user must act on, whether or not anyone labelled it.

The notes

<Overview: what users get, in their terms. Patch: one line at most. Minor: one to three sentences.
Major: one short paragraph.> <Always: "No action is required to upgrade." or "N changes require
action — see below.">

## ⚠ Action required

### Breaking changes

**<What changed>** ([#N](URL/pull/N)) Previously, <old>. Now, <new>. This affects <who>. To migrate,
<step or snippet>. <Opt-out, or none.>

### Configuration changes

- `old.key` → `new.key`; `x` now defaults to `Y` (was `Z`). ([#N](URL/pull/N))

### Deprecations

- `<thing>` is deprecated and will be removed in <version>. Use `<replacement>` instead.
  ([#N](URL/pull/N))

### Security

- <CVE/GHSA id>: <one line>. ([#N](URL/pull/N))

## Highlights

### <Feature>

<One to three sentences.> <Usage a reader can copy, e.g. "Pass `max_retries=` to `submit()` to stop
retrying after N attempts."> See [docs/retries.md](URL/blob/TAG/docs/retries.md).

## All changes

### Features

- Added <what> to <where> ([#N](URL/pull/N), @login)
  - <nested detail only where a reader needs it>

### Fixes / Performance / Documentation / Dependencies

- …

## Artifacts

- <kind>: <name> <version> — <destination>

**Full Changelog**: [PREV...TAG](URL/compare/PREV...TAG)
  • Omit every empty section. A patch release has no Highlights.
  • Past tense, from the user's side. Rewrite PR titles into what changed for a user. Never paste a title or a commit log.
  • Compact links. The link text is #N and nothing more, followed by the author as a bare @login. A direct commit is linked by its short sha. Every link is built from the URL the digest names as its link base: never a bare #N, which links only inside a release body, and never a typed host.
  • Action-required items come first and are repeated in place: each breaking or configuration entry also appears under "All changes" with a leading Breaking:.
  • Name the exact key, flag, function, parameter, endpoint or environment variable, in backticks.
  • A deprecation names its removal version. If none is given, write "in a future release" and tell the user.
  • Highlights carry usage and a link to the documentation at the tag. That is where "how to use the new feature" lives: next to the feature, near the top, where a skimming reader still is.
  • Group "All changes" by type. Collapse dependency-bot PRs into one line, unless a bump changes something users see.

Why this shape and not another is in references/research.md: 21 projects' release notes, the published writing guides, and the gh and API behaviour every step above relies on.

Configuration

The proposal setup writes is the config, with comments. The keys are:

keywhat it does
urlthe repository it belongs to
flowgitflow or trunk; delete the key to detect on every run
production_branch, integration_branchthe branch names, when they are not the default branch and develop
sourceauto, prs or commits
surfaceglobs for files whose change is user-facing
[exclude]labels, title_prefixes, head_patterns, internal_paths, dependency_authors
[[artifacts]]kind, name, destination (with {version} / {tag}), workflow, job, evidence, confidence
workflow_digestwhat the artifact answer was derived from; a change makes collect say so

Edit the file by hand whenever that is quicker than re-running setup.

Without the skill

references/prompt.md is the whole procedure as one prompt: the gh and GraphQL calls, the gitflow rule, the quality rubric and the template. Anyone can paste it into an agent that has gh, without installing anything. Its GraphQL and thresholds are the script's own, kept identical by a test.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.