muse-skills

Review-loop skills for Muse sessions, built from a real incident: two round-1
reviewers approved with 0 blocking findings while an independent review found
13 real issues. These skills encode the protocol that caught everything in
rounds 2–3.

Skill	What it does
skills/adversarial-review	Pre-verdict adversarial probe pass over a green-gated diff: a runnable catalogue of probes (imports, worktrees, guard evasions, vendoring, builds, help text, docs-as-written) plus a refutation pass before APPROVE.
skills/fix-verification	Post-fix closeout: one traceability row per finding, diff-to-row mapping, per-row repro re-runs on the final head (incl. after rebases), status-field consistency, and a Revision rounds: N/3 counter.

Background (evidence, not required reading): the miss analysis, skills review,
and cross-review that produced these live in the originating work log, and each
skill names the concrete misses its steps catch.

Use in a Muse session

Skills install from a local path. Clone once, install each skill (user scope
shown; omit --scope user for project scope):

git clone https://github.com/seb-patron/muse-skills.git
muse skills install muse-skills/skills/adversarial-review --scope user
muse skills install muse-skills/skills/fix-verification --scope user

Then invoke by name (adversarial-review, fix-verification) or let the
session load them when a review/fix round starts.

Structural validation

The lint validates tracked skill roots, YAML metadata, the 14 probe identities,
and selected instruction fragments. It uses a pinned PyYAML dependency; CI runs
the lint and its regression tests on every PR. From the repository root:

python3 -m venv .venv
.venv/bin/python -m pip install -r evals/requirements.txt
.venv/bin/python evals/check_skills.py
.venv/bin/python -m unittest discover -s evals -p 'test_*.py' -v

On Windows, use .venv\Scripts\python.exe instead of .venv/bin/python.
This checks document structure and text presence; it does not execute Muse or
establish that Muse loads or follows the skills correctly. See
evals/README.md for the contract and its limitations.

Behavioral evaluation

The Promptfoo harness under evals/behavioral runs
frozen revie