← ClaudeAtlas

releaselisted

Take a change from converged review to a proven release: two independent whole-surface reviews, the version decision, the gate run, tag, push, CI, assets, and a proof against the released binary. Reads the workspace gates, `.github/workflows/ci.yml`, and `gh run`; writes the version bump, the tag, and the release.
junyeong-ai/harnex · ★ 1 · AI & Automation · score 74
Install: claude install-skill junyeong-ai/harnex
# release ## 1 — Converge the review before deciding anything else Two reviewers over the whole changed surface, dispatched in parallel and **given no access to each other's findings** — a shared list turns the second reviewer into a confirmer. One reads with no authoring context; the other is a separate engine, so a blind spot has to be shared by both to survive. Then, for every claim either returns: - **Reproduce it against the built binary before fixing it.** A reviewer's finding is a hypothesis. A fix landed on an unreproduced claim is a change with no defect behind it. - **Mutation-test every guard the change adds or touches**, to the discipline `making-changes.md § Verification` states. A guard that passes its own mutation is watching nothing, and the suite still reports green. For Rust, the set is computed rather than enumerated: diff the release range into a file and run `cargo mutants --in-diff <file> --test-workspace=false --copy-vcs=true` — the git directory is copied because integration tests read it, and without it the unmutated baseline fails before any mutant runs. A missed mutant is an open claim until a test kills it; one another package's test would kill is confirmed by re-running it with `--test-workspace=true`. Shell and template guards have no such tool and keep the manual form. Do not start §3 while a claim is open. An unresolved finding after a tag is a finding that ships, and a review dispatched and not yet read is an open