[AAASM-2886] 📝 (skill): Restructure release-tag-cut to Agent Skills spec (EXAMPLES + REFERENCE + bundled script)#1044
Conversation
Extracts the version-literal sed sequence + Cargo.lock regeneration from the release-tag-cut SKILL.md (steps 2-4) into a single reusable shell script. Behavior: - Refuses no-op invocations (current == target). - Enumerates every Cargo.toml declaring the current literal via git grep. - Surfaces the file list to stdout BEFORE mutating. - sed -i.bak replaces in place + cleans up .bak siblings. - Runs cargo update --workspace to regenerate Cargo.lock. - Exits non-zero with a clear message on any sed/cargo failure. Refs AAASM-2886.
…ut SKILL.md Gives the LLM concrete trigger criteria (alpha-series cut on green master with clean tree + draft notes) and disqualifying scenarios (SDK-only, GA, hotfix, pre-conditions not met) so the right skill is picked at the right time. Refs AAASM-2886.
Documents the invocation syntax (/release-tag-cut <X>), the required context (main checkout + `remote` remote name), and how the <X> placeholder binds to the operator-supplied target version everywhere in the executable plan. Refs AAASM-2886.
Concrete walk-through of cutting v0.0.1-alpha.10 from a v0.0.1-alpha.9 baseline: actual commands + expected output for steps 1-6, including the 16-file / 43-literal footprint from AAASM-2849 and the final gh run watch confirming release.yml is queued. Refs AAASM-2886.
…KILL.md Documents the two verification commands (git ls-remote --tags + gh run list) the operator runs to confirm the skill succeeded, with expected output for each, plus the recommended /release-validate-channels follow-up once release.yml finishes. Refs AAASM-2886.
…g-cut SKILL.md Lists the five release.yml jobs the operator MUST NOT replicate by hand: GitHub Release creation, cargo publish per workspace crate, Homebrew tap PR, downstream SDK repository_dispatch fanout, and (post-AAASM-2883) the node-sdk + python-sdk FFI source-pin bump PRs. Prevents duplicate publishes and audit-log divergence. Refs AAASM-2886.
Slims down inline bash in steps 2 (enumerate + sed) and 4 (cargo update) by delegating to scripts/release-tag-cut.sh. Step 3 keeps only the atomic Cargo.toml-only commit. Step 4 keeps only the separate Cargo.lock commit plus the cargo generate-lockfile fallback note. Refs AAASM-2886.
Claude Code review — AAASM-2886CI3/3 green: Scope vs. AAASM-2886
Quality observations
VerdictReady for approval and merge. Strongest v2 PR (most ACs, most CI coverage, helper script extracted cleanly). Companion to AAASM-2888 (homebrew-tap-merge enhancement) which extracts the verify-tap-sha256 helper. — Claude Code |
Claude Code review — AAASM-2886 (spec-compliance restructure)1. CI state3/3 green incl. 2. Anthropic Agent Skills spec compliance
The 3. Scope vs AAASM-2886 (re-scoped)Matches: EXAMPLES.md + REFERENCE.md split, bundled script, lean SKILL.md, description rewrite. VerdictReady for approval and merge. Strongest CI coverage of the five (dev-verify exercises the bundled script). Clean lossless restructure. — Claude Code |
Description
Restructures the
release-tag-cutskill to the Anthropic Agent Skills spec(progressive disclosure). The skill previously crammed the full executable
plan, a verbose worked example, and all post-conditions into a single 376-line
SKILL.md. This PR splits that content across the spec's three disclosurelevels and bundles the helper script inside the skill directory.
Restructure:
descriptiontostate both what the skill does and when to use it ("Use when an
operator is ready to cut a new alpha-series agent-assembly release…"), the
spec's discovery signal. Third-person, well under 1024 chars.
SKILL.mdfrom 376 → 142lines (spec target < 500, recommended < 200). It now carries only a lean
overview, When/When-NOT-to-use, How-to-use, a concise 6-step plan as short
bullets, brief pre/post-conditions, the do-NOT-manually-run callout, the
"does not do" list, and a
## Detailed referencessection linking one leveldeep.
EXAMPLES.md— the full alpha-10 worked example, verbatim, with a TOC.REFERENCE.md— the detailed per-step executable-plan prose (commands,edge-cases, no-op guard rationale, full pre/post-conditions, auto-handled
rationale), with a
## ContentsTOC at the top (it is > 100 lines).scripts/release-tag-cut.sh—git mv'd from repo-rootscripts/intothe skill dir so the script is bundled with the skill (forward-slash,
one-level-deep reference, stays executable
755).No content was lost — everything was relocated. All links from
SKILL.mdareone-level-deep relative links (
[EXAMPLES.md](EXAMPLES.md),[REFERENCE.md](REFERENCE.md),scripts/release-tag-cut.sh).Type of Change
Breaking Changes
Does this PR introduce any breaking changes to public APIs or behaviour?
The skill's runtime behaviour is unchanged — only its file layout and
discovery metadata. The repo-root
scripts/release-tag-cut.shpath is gone;the bundled copy lives at
.claude/skills/release-tag-cut/scripts/release-tag-cut.sh.Related Issues
Testing
Describe the testing performed for this PR:
Manual verification:
wc -l SKILL.md→ 142 (< 500 spec limit, < 200 recommended).name+ improveddescriptionpresent (head -5).REFERENCE.mdandEXAMPLES.mdexist; both carry a## ContentsTOC.scripts/release-tag-cut.shexists inside the skill dir and is executable(
755); repo-rootscripts/release-tag-cut.shremoved.SKILL.mdlinks are one-level-deep relative links.Checklist
cargo fmt,cargo clippy)