Skip to content

ci(release): publish a signed packslip with each release - #4301

Merged
max-sixty merged 3 commits into
max-sixty:mainfrom
jdx:packslip
Sep 27, 2026
Merged

max-sixty merged 3 commits into
max-sixty:mainfrom
jdx:packslip

Conversation

@jdx

@jdx jdx commented Sep 27, 2026 •

Copy link
Copy Markdown
Contributor

This adds one step to the host job in release.yaml. The step writes a signed packslip manifest, packslip.sigstore.json, and it ships as a release asset alongside the archives.

A packslip lists each archive's digest, its platform, and where wt and git-wt sit inside it, plus the completions and agent skills that go with the release.

What this gets you and your users

  • Better security. Installers such as mise and aqua find worktrunk's downloads through a registry entry someone else wrote, and at best compare them to a checksum file hosted on the same release. With a packslip, an installer first checks that the manifest was signed by max-sixty/worktrunk's own release workflow (keyless Sigstore signing through GitHub OIDC), then checks each archive's digest and size against it before unpacking anything. mise also remembers the signer, so a release that turns up signed by a different workflow or key is not installed until the user explicitly decides to trust it. There's no signing key for you to store or rotate.
  • Direct control over how worktrunk is installed. The manifest comes from your release workflow, so you decide which archives exist for which platforms, where the executable sits in each, and what ships with it. Installers no longer depend on third-party registry entries that infer all of that from asset names and drift when a name or platform changes.
  • Completions that just work. Users get bash, zsh, fish and PowerShell completions from wt config shell completions for the exact version they have installed, with nothing to add to their shell rc files.
  • Your agent skills, matched to the version. The worktrunk and wt-switch-create skills ship with each release, pinned to the release commit, so an agent working with a given worktrunk release reads the skill written for that release, not whatever is on the default branch. mise skills ls shows them, and mise skills sync links them into a project's agent directory.

Once a release ships a packslip, I'll switch mise's registry entry for worktrunk to use it, so users get all of this from a plain mise use worktrunk with nothing to change on their end. Other installers can check the same file with packslip verify.

What it declares

  • Artifacts: the five worktrunk-*.tar.xz / worktrunk-*.zip archives. It leaves out checksums, installers, source.tar.gz, and wt.rb.
  • bin: wt git-wt
  • Resources:
    completion/bash,zsh,fish,powershell/wt=exec:wt config shell completions {shell}
    completion/bash,zsh,fish,powershell/git-wt=exec:git-wt config shell completions {shell}
    skill/worktrunk=repo:skills/worktrunk
    skill/wt-switch-create=repo:skills/wt-switch-create
    
    The completions use the same command Homebrew's formula runs. Nushell is left out because wt config shell completions nu prints the full shell integration, not only completions.

Release flow

Releases are immutable, and v0.79.0 is marked immutable, so the bundle can't be uploaded after gh release create. The step therefore runs just before that command, on the final archives: after the signed Windows zip has replaced the unsigned one. It writes the bundle with upload: false, and a following step copies it into artifacts/ only if the packslip step succeeded, so the existing artifacts/* glob attaches it to the release. The packslip step has continue-on-error: true: a packslip or Sigstore outage ships the release without a packslip rather than holding back crates.io, winget, Homebrew and AUR, and a bundle that failed its own verification never ships.

The host job gains id-token: write for signing and attestations: write: the action attests build provenance for the five archives and links it from the manifest. host already runs only for tag pushes, so PR runs don't reach this step. The action is pinned by SHA to v1.4.0 (Dependabot's github-actions entry keeps it current), and the step has timeout-minutes: 10 so a stalled Sigstore call can't hold the release until the job's default timeout.

Verification

  • Dry run against v0.79.0: I downloaded the five archives and ran packslip create with the same artifacts, bins, and resources, signing with a throwaway local key. packslip verify then checked all 5 of 5 artifacts and the 4 resources.
    • Each artifact got the right platform: darwin arm64/x64, linux-musl arm64/x64, and windows x64.
    • Each got the right paths: worktrunk-<target>/wt and worktrunk-<target>/git-wt, or wt.exe and git-wt.exe in the zip.
  • skills/worktrunk/SKILL.md and skills/wt-switch-create/SKILL.md exist at v0.79.0's commit.
  • I ran the v0.79.0 wt and git-wt with config shell completions for bash, zsh, fish, and powershell. Each printed a script that passes bash -n, zsh -n, fish --no-execute, or the PowerShell parser.
  • actionlint reports no new findings; the 7 shellcheck infos it prints are already on main.

For maintainers

Nothing beyond merging. The first packslip appears with the next release.

I maintain packslip and mise, and I'm happy to adjust anything here. That includes dropping the git-wt completions if you'd rather only register wt.

AI-assisted — Tool: Claude Code; model: anthropic/claude-opus-5-5; version: unavailable.

🤖 Generated with Claude Code

Sign a manifest of the release archives with the workflow's OIDC
identity and ship it with the GitHub release, so installers can verify
downloads and pick up version-matched completions and skills.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

@worktrunk-bot worktrunk-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Every downstream publish job (publish-cargo, publish-winget, publish-homebrew, publish-aur) has needs: host. Because of that, a failure anywhere in this step stops gh release create and skips crates.io, winget, Homebrew and AUR for the tag. Several things can fail it: the action's gh release download / gh attestation verify of the packslip binary, the provenance attestation, a Sigstore/Rekor outage while signing, or its own packslip verify. So an outage in the newest distribution channel would hold back the release on all the established ones. The inline suggestions make the step non-blocking and copy the bundle into artifacts/ only when the step succeeded. That matters because out: artifacts combined with continue-on-error would still ship a bundle that failed its own verify step. The cost is that a release can ship without a packslip, and nothing flags it except the step's annotation. Whether that beats blocking the release is a maintainer call.

Separately, for @max-sixty: this adds a third-party composite action to the job that assembles and publishes the release. It runs with github.token (contents: write) and id-token: write, and it has read/write access to artifacts/ right before those files are attached to the release. jdx/packslip was created this month, and @v1 is a moving tag that has advanced across several releases since then. The other third-party actions in this file are also tag-pinned, so this follows existing practice. Still, this is the one step placed between the signed Windows zip and gh release create. The author offered a SHA pin, and Dependabot's github-actions entry keeps SHA pins updated. Whether to take on the dependency, and how to pin it, is your decision, so I'm leaving this as a comment rather than an approval.

Comment thread .github/workflows/release.yaml Resolved
Comment thread .github/workflows/release.yaml Resolved
Run the packslip step with continue-on-error and stage the bundle into
artifacts/ only when it succeeded, so a Sigstore or packslip outage
ships the release without a packslip instead of holding back every
publish job, and a bundle that failed its own verify never ships.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@max-sixty

Copy link
Copy Markdown
Owner

@worktrunk-bot any downsides we should weigh? @jdx has the track record for us to adopt his early ideas

@worktrunk-bot

Copy link
Copy Markdown
Collaborator

No blocker from me. With the step non-blocking now, the release-path risk I raised in my review is handled. What's left to weigh:

  • Supply-chain trust in jdx/packslip@v1. This is the main one. The step runs third-party code inside host with contents: write and id-token: write, right before gh release create attaches artifacts/*. continue-on-error covers the action failing. It doesn't cover the action misbehaving, such as altering files in artifacts/. The repo is about three weeks old, and v1 is a moving tag. A SHA pin (which jdx offered, and which Dependabot keeps current) turns this into "trust what we reviewed" and costs nothing.
  • A hang still blocks the release. continue-on-error only applies when the step exits. Nothing bounds the step's runtime, so a stalled Sigstore/Rekor call would run until the job's 6h default timeout. Then host fails and every publish job is skipped. Adding timeout-minutes: 10 to the step closes that gap.
  • The signing identity becomes part of the contract. As the PR body describes, mise remembers the signer and holds back a release signed by a different workflow until the user re-trusts it. Renaming release.yaml, moving signing to another workflow, or transferring the repo would then prompt every mise user. That's fine, but it's worth knowing before any later workflow reshuffle.
  • wt config shell completions becomes an external contract. It's a hidden subcommand (src/cli/config.rs), documented as being "for package managers", and Homebrew already calls it. Each manifest records the command for its own version, so renaming it later only needs the resources: line updated in the same release. Installers run it at install time, not in CI, so a broken command wouldn't fail the release. It would show up as broken completions for users.
  • Release job maintenance. release.yaml is dist-generated with allow-dirty = ["ci"], so this becomes one more hand-maintained block to carry forward when dist regenerates. The existing custom publish jobs already work that way.

Upside, for balance: worktrunk has no provenance attestations today, and this adds them for the five archives as a side effect. It also gives mise users version-matched completions and skills without registry drift.

If you want to go ahead, I'd take the SHA pin and the step timeout. Both are small edits to this PR.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jdx

jdx commented Sep 27, 2026

Copy link
Copy Markdown
Contributor Author

Thanks! I took both of the bot's suggestions in 218a2cb: the action is now pinned by SHA to v1.4.0 (Dependabot will keep it current), and the step has timeout-minutes: 10, so a stalled Sigstore/Rekor call fails the step (which continue-on-error then absorbs) instead of hanging host.

On the signing-identity point, one update: a repository rename or transfer doesn't have to disrupt mise users. The signing certificate carries GitHub's numeric repository and owner IDs. mise keeps installing across a rename with the same repository ID, refuses a re-created repo at the old name, and asks for confirmation only on a real owner change (packslip 1.4.0 plus jdx/mise#13702, in review). Renaming release.yaml or moving signing to another workflow would still ask users to re-trust, so that part stands.

For wt config shell completions: each release's manifest records the command for that version, so renaming it later only means updating the resources: line in the same release.

AI-assisted — Tool: Claude Code; model: anthropic/claude-opus-5-5; version: unavailable.

@max-sixty
max-sixty enabled auto-merge (squash) September 27, 2026 15:07
@max-sixty
max-sixty merged commit f2996ba into max-sixty:main Sep 27, 2026
38 checks passed
jdx added a commit to jdx/mise that referenced this pull request Sep 28, 2026
Timoni and Worktrunk now publish signed packslips with their releases.
`mise use timoni@0.35.0` and `mise use worktrunk@0.80.0` can install
from the release manifests, verifying the workflow signature and
artifact digest while also receiving the completions and version-matched
skills declared by each project.

The `packslip:` backend takes priority starting with Timoni 0.35.0 and
Worktrunk 0.80.0. Earlier versions remain installable through their
existing Aqua (and, where configured, asdf or Cargo) backends. `mise
ls-remote` lists versions from the preferred backend, so older releases
can be listed with `mise ls-remote aqua:stefanprodan/timoni` or `mise
ls-remote aqua:max-sixty/worktrunk`. The registry declares the `timoni`,
`wt`, and `git-wt` executable names so lazy install shims can select the
right tool.

Validated both signed releases with `mise ls-remote packslip:…` and an
explicit `mise exec packslip:…@version`: `timoni --version` returned
`timoni version 0.35.0`, and `wt --version` returned `wt v0.80.0`. The
installed releases contain the declared executables. `mise run lint-fix`
passed.

Tracks #13712. Upstream release PRs: stefanprodan/timoni#659 and
max-sixty/worktrunk#4301.

*AI-assisted — Tool: Codex; model: OpenAI/unavailable; version:
unavailable.*


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **New Features**
* Added Packslip registry support for Timoni version 0.35.0 or later and
Worktrunk version 0.80.0 or later.
* Registered `timoni` as a Timoni binary, and `git-wt` and `wt` as
Worktrunk binaries.
* Existing Aqua support for Timoni and Aqua and Cargo support for
Worktrunk remain available.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants