cut-release -- assess, sanitize, bump, lint, open PR

This skill drives the release-cut workflow end-to-end on the current worktree. It STOPS at "PR open"; tagging is the human-gated trigger that fires the release workflow.

When to use

Trigger this skill on any of these intents:

• "cut a release"
• "ship v0.x" / "ship the release"
• "release prep" / "prepare the release"
• "bump and open the PR"
• "open the release PR"
• "what kind of release do we need now"
• "we have enough merged for a release"
• Any phrasing that ends in opening a release PR.

Do NOT trigger on:

• "tag v0.16.0" -- tagging is the post-merge step the operator runs.
• "publish a blog post" / "draft release notes for the website".
• "what changed since v0.15.0" -- enumeration only, no release.
• "regenerate uv.lock" / raw version-string edits.
• "rollback to v0.15.0" -- release reversal is a different skill.

Companion primitives

The skill DEPENDS on these existing primitives in the same source tree. Do not duplicate their content; reference them.

• .apm/skills/apm-strategy/SKILL.md -- versioning judgement and breaking-change lens. Auto-loads via its own trigger on CHANGELOG.md and release-pipeline edits. Treat it as the authoritative lens for "is this BREAKING" and "does this positioning change warrant a migration note".
• .apm/instructions/linting.instructions.md -- canonical CI-mirror lint chain. Reference; do not re-derive ruff / pylint command lists in this skill.
• .github/instructions/changelog.instructions.md -- Keep-a- Changelog format contract. Reference; do not redefine the format.

Output charset rule

Per .github/instructions/encoding.instructions.md, source files in this bundle (SKILL.md, assets, scripts) MUST stay in printable ASCII. The PR body and changelog entries this skill produces are also written to repo files (CHANGELOG.md, PR body), so they MUST stay ASCII too. Use -- for em dashes, [!] / [+] / [x] for status markers, and so on.

Procedure

Run these phases in order. Reload plan.md (the session memento) at the start of each phase and after every tool return.

Phase 0 -- ground state

1. Read the current branch with git rev-parse --abbrev-ref HEAD. The skill assumes you are on a release-prep branch (or are willing to create one). If you are on main, STOP and ask the operator to put you on a branch.
2. Resolve the last release tag with git describe --tags --abbrev=0.
3. Persist the goal to plan.md:
   • Range: <last-tag>..HEAD.
   • Acceptance: release PR is open, lint mirror green, no version bumped beyond what the cycle warrants.

Phase 1 -- enumerate merged PRs

Run scripts/list-changes-since-tag.sh (no arguments). It emits JSON on stdout: one object per merged PR with fields pr, title, labels, author, paths_summary, is_user_facing_guess.

The script's heuristic for is_user_facing_guess is intentionally conservative -- it flags as INTERNAL only when the diff is fully contained in .apm/, .github/instructions/, tests/, docs/, or matches chore(repo). Treat the field as a HINT, not a verdict; the LLM makes the final call per assets/entry-sanitizer.md.

DO NOT rely on training-recalled PR titles. Truth #5: pretraining is frozen. Every PR title in the changelog must come from the script's stdout.

Phase 2 -- pick the bump (B10 CHECKPOINT 1)

Load assets/semver-rubric.md. Walk every PR from Phase 1 against the rubric. The rubric outputs PATCH, MINOR, or ESCALATE-TO-MAJOR.

Show the operator:

• The chosen bump and the next version number.
• Per-PR rationale (which signal fired).
• The "why patch/minor" one-liner that will go into the PR body.

If the rubric outputs ESCALATE-TO-MAJOR (>= 1.0.0), STOP and surface the escalation per assets/semver-rubric.md "Major bump" section. Do not bump to 1.0.0 without explicit operator confirmation.

Wait for the operator's confirm before continuing.

…