Repository navigation
Add cut-release skill covering the Homebrew tap step - #323
vpetersson-bot wants to merge 3 commits into
Conversation
The Homebrew tap lives in a separate repo (Screenly/homebrew-screenly-cli) with no automation linking it to our release workflow, so when it is skipped nothing fails — `brew install screenly-cli` just keeps serving the previous version. It had drifted a full release behind (v26.8.0 while we shipped v26.9.0). Add a `cut-release` skill that documents the whole release path and treats the tap bump as a required release step rather than a follow-up, plus a standalone drift check. Also record the details that are easy to get wrong: the non-zero-padded CalVer month, the four files to bump and which spell the version with a `v`, and the mcpb/manifest.json placeholder that must not be touched. Expand the README release section to point at the skill and spell out the formula fields. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: multica-agent <github@multica.ai>
Follow-up to the tap change: the formula no longer carries an explicit `version` field at all, so the runbook's "bump both fields" instruction was wrong in both directions. Homebrew derives the version from the tag and strips the leading `v`, which leaves `tag:` as the only line a release touches and makes the version underivable from anything but the tag. The previous text also told the reader to keep the `v` prefix as an established convention. That was backwards: Homebrew tokenises a leading `v` as a StringToken sorting below any NumericToken, so `v`-prefixed versions compare as older than every bare one, and the FormulaAudit/Version cop rejects the prefix outright. Record that, and that the switch is one-way. Also correct the distribution-channel notes. nixpkgs is a separate package from our in-repo flake.nix, and while it self-updates via nix-update-script the bot has gone quiet: nixpkgs is still on 1.2.1 while we ship 26.9.0, so `nix-shell -p screenly-cli` serves a pre-CalVer build. Add a spot-check for it, and note that flake.nix needs no action because it reads the version from Cargo.toml. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: multica-agent <github@multica.ai>
102c18c to
442da3a
Compare
|
|
||
| ```bash | ||
| # what the tap currently pins vs. what we just shipped | ||
| gh release view --repo Screenly/cli --json tagName -q .tagName |
There was a problem hiding this comment.
Agreed. The step now runs gh release view vYY.M.MICRO, so an unpublished release fails instead of resolving to the previous one. Fixed in 6eb5e1a.
| ```bash | ||
| gh release view --repo Screenly/cli --json tagName -q .tagName | ||
| gh api repos/Screenly/homebrew-screenly-cli/contents/Formula/screenly-cli.rb \ | ||
| --jq .content | base64 -d | grep 'tag:' |
There was a problem hiding this comment.
This block is gone from the README; it now links to the skill, which uses base64 --decode. Fixed in 6eb5e1a.
salmanfarisvp
left a comment
There was a problem hiding this comment.
@vpetersson-bot thanks, this runbook is useful. I checked its claims against the repo and nixpkgs, and they hold: the mcpb/manifest.json placeholder plus the jq injection in release.yml, flake.nix reading Cargo.toml through fromTOML, the target list, attestation and Docker tags, the four version-carrying files and their v prefixes, and nixpkgs still on 1.2.1. The nixpkgs finding alone makes this worth merging.
A few things to fix before it goes in:
- The same content is in two places. The README's new step 4 repeats most of the skill (the no-
version-field explanation, the sha256 note, the verify commands). The copies will drift apart. Could the README keep a short step 4 (bumptag:in the tap'sFormula/screenly-cli.rb, then verify) and link to the skill for the details? - The PR description is out of date. It says the tap is "a full release behind." That was fixed by homebrew-screenly-cli#18, and the tap now matches
v26.9.0. Please update it so reviewers aren't misled. - SKILL.md, step 3: "CI is not run-owned — kick it off and check back; don't block on it" reads as agent-internal jargon. Suggest "The workflow takes a while; check back on it rather than waiting."
- SKILL.md, step 0: "the release after
26.8.2is26.9.0, not26.8.3" only holds when the month has changed. Suggest "a new month resets MICRO to0(e.g. after26.8.2, the first September release is26.9.0)."
Nothing here blocks the content itself. Happy to approve once these are addressed.
Point the README's Homebrew step at the skill instead of duplicating it, view the release by its tag, use the portable base64 --decode, and reword the MICRO reset and the CI wait. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Copilot encountered an error: Your billing is not configured or you have Copilot licenses from multiple standalone organizations or enterprises. To use premium requests, select a billing entity via the GitHub site, under Settings > Copilot > Features.
|
@salmanfarisvp thanks for checking the claims against the repo and nixpkgs. All four points are addressed in 6eb5e1a:
I also took Copilot's points: the tap step now looks up the release by its tag, and the decode uses |


What does this PR do?
Adds a
cut-releaseagent skill and makes syncing the Homebrew tap an explicit, required release step.The tap (Screenly/homebrew-screenly-cli) had drifted a full release behind (pinned to
v26.8.0afterv26.9.0shipped). That is now fixed by homebrew-screenly-cli#18, and the tap matchesv26.9.0; this PR is about keeping it from happening again. The reason it drifts is structural: it is a separate repo, nothing inrelease.ymlpoints at it, and skipping it fails silently —brew install screenly-clijust keeps serving the previous version. The README had it as a one-line trailing bullet.Changes:
.claude/skills/cut-release/SKILL.md— full release runbook, following the same format as thecut-release/anthias-releaseskills inScreenly/Anthias. Covers the tap bump as step 4 of the release (not a follow-up), plus standalone drift checks for both Homebrew and nixpkgs.README.md— step 4 is now a short required step (bumptag:in the tap's formula, then verify) that links to the skill for the details, so the two don't drift apart.The skill also records the details that are easy to get wrong:
26.8.0, not26.08.0) —Cargo.tomlis parsed as strict SemVer.vprefix:Cargo.tomlis bare,action.ymlandDockerfilearev-prefixed,Cargo.lockis regenerated viacargo build.mcpb/manifest.jsonmust not be bumped — its0.0.0is a deliberate placeholder the release workflow overwrites from the tag.flake.nixmust not be bumped either — it readsversionfromCargo.tomlviabuiltins.fromTOML, so it follows automatically.The
vprefix, and what else it touchesThe companion tap PR (homebrew-screenly-cli#18) drops the
vfrom the formula version by deleting theversionfield rather than rewriting it — Homebrew derives the version from the tag and strips thevon its own, which leavestag:as the only line a release touches. The runbook here is updated to match; an earlier revision of this branch said to keep thev, which was backwards.Worth knowing, since it explains why: Homebrew tokenises a leading
vas aStringTokenthat sorts below anyNumericToken, sov-prefixed versions compare as older than every bare one, andFormulaAudit/Versionrejects the prefix outright.This also surfaced a second, larger drift. nixpkgs (
nix-shell -p screenly-cli, advertised in our README) is a different package from our in-repo flake, and it is still on 1.2.1 — four releases back, predating the CalVer switch entirely. Itsnix-update-scriptbot last filed a PR in July 2026 (1.2.0 -> 1.2.1) and has filed nothing for1.2.2,26.8.0or26.9.0. That is not fixed here — it is someone else's repo — but the skill now includes a spot-check so it gets noticed at release time.GitHub issue or Phabricator ticket number?
SIRI-98
How has this been tested?
Docs-only — no code paths touched.
Version facts were verified against the repo rather than written from memory: the four version-carrying files were found by grepping for the current version, the
mcpb/manifest.jsonplaceholder behaviour was read out ofrelease.yml, andflake.nixwas read to confirm it derives fromCargo.toml.The Homebrew claims were verified by running the real
Versionclass from Homebrew 7.0.1 rather than asserting them: token breakdown,26.9.0 > v26.8.0(so the transition upgrades cleanly), andVersion.detecton the git URL + tag returning"26.9.0"(so theversionfield is safely removable). Theghdrift-check commands in the skill were run as written.docs.ymlis unaffected — itsdocs-help-mdjob diffs generated CLI help againstdocs/CommandLineHelp.mdand does not look at README prose. It passed on the previous push.Checklist before merging
🤖 Generated with Claude Code