skills

Review diffs, commit ranges, fixup chains, or PRs across languages and report findings ordered by severity. Enforces primitive-obsession and smart-constructor blockers, property-test and boundary-test requirements, complexity thresholds, commit structure against the canonical form, and per-language rules (Rust, TypeScript, Lean, SQL, bash, Ruby, Markdown, regex). Also the home for changing review rules or codifying review feedback.

Adversarial review of formal/Lean and other high-stakes artifacts, run as an ORACLE HUNT rather than a lens checklist: build a ranked target table of every public observation and the hidden predicate it computes, attack the top targets with two-run distinguishers and compiled witnesses, and classify each result by enforcement rank (kernel theorem / export drill / runtime bridge / operator policy / evaluator judgment / honest doc) and threat scope. Presuppose defects and find-and-prove them — no claim without a witness; for what you can't break, prove it sound. Use to review your own formal work before delegating, and as the prompt spine for adversarial subagent reviews. Grounded in the formal-methods literature, hardened on a machine-checked Lean model.

Pre-flight decision procedures run BEFORE authoring a slice/rung of a formal (Lean or similar) system: identity design, carrier/representation choice, primitive forcing, increment scoping, seam/interface extraction, effect channels, recon discipline, and proof-engineering forward hygiene. Each slice kind has a keyed question set with the failure mode each question pre-empts — answering them before building prevents the defects the review-time sibling (find-and-prove) would otherwise catch after the fact. Use when planning the next increment of a formally-proven system, choosing between design candidates, or deciding whether a design round is needed at all.

Break work into the smallest atomic steps. Verify the foundation first (step 0), then order steps so partial progress always leaves the system valid or better. Owns the canonical atomic commit form: the closed semantic action-verb set, transformation-priority ordering, and commit-message body rules.

Minimize representable states so invalid states are impossible. Parse, don't validate: replace primitive obsession with domain types, smart constructors, and boundary parsing; bound every range and cardinality; tighten test matchers; normalize so each fact has one determinant; ratchet tooling thresholds so the state space only narrows. Applies to types, signatures, tests, commits, documentation, and tooling configuration.

Notation and inference rules for state-space minimization: S/C/I_repr/I_reach state sets, boundary morphisms, strictness and dominance rules, the encoding order, proof obligations, reception semantics for text rewrites, and the skill-gap rules. Two uses: the output layer when the deliverable is formal, and a reasoning mode when it is not — working in the calculus activates formal-methods reasoning and surfaces constraints that prose analysis misses. Semantics live in the state-space-minimization skill's reference modules.