Release Labtasker

Treat the Client and Server as independently installed distributions that always ship with one shared version. Separate release preparation from externally visible publication.

Establish scope and authority

Determine the requested target version and whether the user asked only to prepare, or explicitly authorized tagging, pushing, creating a GitHub release, or publishing to PyPI. Do not infer authorization for later stages.

Inspect the current branch, tags, remotes, git status, and all pending diffs. Do not include unrelated worktree changes in release commits or artifacts. Never reuse the v1 worktree's release scripts or workflow for v2.

Prepare the version

Use the deterministic helper from the repository root:

uv run python .agents/skills/release/scripts/set_version.py VERSION
uv lock
uv run python .agents/skills/release/scripts/set_version.py --check VERSION

The helper updates the workspace version, both distribution metadata files, both public __version__ values, and the Claude Code plugin/marketplace version. It maps a PEP 440 prerelease such as 2.1.0rc1 to the corresponding plugin SemVer 2.1.0-rc.1. uv lock owns lockfile updates. Review the resulting diff and update user documentation only when the release changes documented behavior. Do not create a version-only changelog file when the repository has no maintained changelog.

Validate release readiness

Run the complete ordinary gate from AGENTS.md. Build artifacts only from the exact reviewed source tree. Then smoke-test the release wheels in clean virtual environments, confirming that:

• labtasker-client imports labtasker and provides its CLI without Server dependencies;
• labtasker-server imports and provides its CLI without the Client package;
• the labtasker convenience metapackage installs matching Client and Server distributions and both CLIs;
• installed metadata and __version__ equal the target version; and
• all three source distributions and all three wheels are present.

For a fully release-gated Linux release, also require the distributed integration suite for the exact commit, either locally with PyTorch and Accelerate installed or through the repository's distributed CI. Report that gate as pending rather than pretending it passed when it cannot be verified.

Write consistent release notes

Summarize user-visible changes since the immediately previous published release. Do not describe internal refactors, tests or CI maintenance as product changes, and do not promise compatibility that the specification does not provide. Use short imperative-present bullets such as Add, Change and Fix, name the affected Client, Server, CLI or Worker when that distinction matters, and combine closely related commits into one outcome. End every bullet with a period and order bullets by user impact, with compatibility and behavior changes before additions and fixes. Use vVERSION as the GitHub Release title.

Use this shape for every GitHub Release. The compatibility notice is conditional, and the comparison line is omitted only for the repository's first release:

> [!IMPORTANT]
> This release changes Client-Server compatibility. Upgrade the Server to
> vVERSION first, then upgrade all Clients and Workers to vVERSION. Earlier
> Clients cannot [...exact consequence...].

## Changes

- Add ...
- Change ...
- Fix ...

**Full Changelog**: [vPREVIOUS...vVERSION](https://github.com/luocfprime/labtasker/compare/vPREVIOUS...vVERSION)

Put a GitHub [!IMPORTANT] admonition before ## Changes when a release changes Client-Server interoperability, supported version combinations, API or persisted database compatibility, or requires an upgrade or migration action. State the exact consequence and action instead of writing only “compatibility changes.” Cover, when applicable:

• whether upgrading both Client and Server is required or only recommended;
• which mixed-version combinations remain supported;
• the required upgrade order;
• whether database migration is automatic or requires a manual step; and
• what fails or becomes unavailable if versions are mixed.

Use required only when mixed versions are unsupported or unsafe. If the current and previous v2 Clients remain supported by the new Server, say that explicitly and use recommended for a same-version upgrade. Do not add an admonition when there is no compatibility or upgrade information worth calling out; repeated empty warnings make real warnings less visible.

For a compatible release that merely benefits from matching versions, use this form rather than implying that a coordinated upgrade is mandatory:

> [!IMPORTANT]
> This release includes Server changes. Upgrading the Client and Server together
> is recommended. Existing v2 Clients remain compatible.

For an incompatible release, use required, give the supported versions and upgrade order, and state any migration behavior in the same notice. Do not use vague notices such as “Server changes” without telling the reader what to do.

Keep the Full Changelog comparison as the final line even when GitHub generated it automatically. Compare the immediately previous published release tag with the new tag, including across a major-version boundary, and use the visible vPREVIOUS...vVERSION label. Do not replace the curated summary with the commit comparison and do not add a comparison link when no previous release exists. Keep the single ## Changes heading for ordinary releases rather than switching between Highlights, What's Changed, Bug fixes and generated commit groups.

Publish only with explicit authorization

Immediately before any tag, push, GitHub Release, or PyPI operation, recheck the exact commit, clean-worktree state, version alignment, artifact contents, and authorization. Use an annotated vVERSION tag unless the repository establishes a different convention.

The reviewed .github/workflows/release.yml workflow publishes all three distributions when a GitHub Release is published. It requires a vVERSION tag, the ordinary and real distributed gates for the exact commit, clean-wheel smoke tests, and protected pypi, pypi-client, and pypi-server environments. PyPI must register that workflow with environment pypi-client for labtasker-client, pypi-server for labtasker-server, and pypi for labtasker. Do not add a manual publishing trigger, improvise credentials, or copy the v1 single-package workflow.