editor

Feature audit ​

The AUD story closing every feature (ADR-0020). It checks the feature in depth and the whole project in breadth. Each check leaves evidence in the story's Findings table; a problem is fixed in the audit, or becomes a story.

Procedure: the /audit skill. Sections A to E are first reviewed by the independent auditor subagent (fresh context, read-only, ADR-0021); each of its findings is reproduced before being recorded.

A. Mathematics (feature, in depth) ​

  1. Re-derive every formula of the feature from its @see (derivation or reference), without looking at the code.
  2. Recompute every expected test value by hand; compare with the test.
  3. Look for missing cases: degenerate input (0, empty, aligned, coincident), sign, bounds, very large values.
  4. Mutation spot-check: break each function once (flip a sign, swap an operator, change a constant) and check that at least one test fails; restore. Record each mutation and the failing test.
  5. Properties: every function with a stateable invariant has test.prop properties (ADR-0022); add the missing ones.

B. Sources ​

  1. Feature: re-open every @see target online — still reachable, still saying what is cited; update the [verified] YYYY-MM-DD date in docs/references.md.
  2. Project: the weekly link check (links.yml) is green — run it on demand if needed; [unverified] rows: try again, or keep them uncited.
  3. No @see points to an [unverified] row (lint guarantees it; check no eslint-disable bypasses it).

C. Provenance (no copied code) ​

  1. Every function is written from its derivation or specification, not from third-party code; its TSDoc formula matches its body.
  2. No comment, identifier pattern or structure taken from a known library; when in doubt, search the distinctive line on GitHub code search.
  3. Licenses: GPL / LGPL code was only read (docs/research/ §4); no dependency added without its row in docs/tooling/dependencies.md.

D. Duplicates ​

  1. Exported API (npm run docs:api, or src/**): no two functions for one concept, no near-identical helpers.
  2. Derivations and references: no two documents for the same result; no URL under two IDs (tests/references.test.ts).
  3. Tests: no copy-pasted test hiding a missing case.

E. Consistency ​

  1. Code ↔ TSDoc ↔ derivation ↔ docs/domain/ ↔ ADRs say the same thing.
  2. Open questions of docs/domain/README.md: still open, or answered by the feature.
  3. AGENTS.md, CLAUDE.md, docs/conventions/, docs/tooling/ describe what the project really does.
  4. Backlog: statuses, feature plan, epic table up to date.

F. Quality gates ​

  1. npm run check:all green; coverage 100 %; npm run build green.
  2. No .skip / .only in tests; every eslint-disable has a justification still valid.
  3. The feature is demonstrable (playground or tests) for its VAL story.
  4. Agent configuration (AGENTS.md, CLAUDE.md, .claude/rules/, skills, agents) still matches the project; contradictions removed.

G. Next feature ​

List the sources and questions the next feature will need: they feed its research spike.

H. Light evolvability check ​

The six questions of the evolvability review in one line each (size, tools, fitness functions, patterns, agent configuration, process). What deserves action goes to the improvement log, treated at the epic retrospective (ADR-0023).

Findings table (in the AUD story) ​

CheckResultEvidenceAction
A4 — mutation of filletSetback (sign of τ)caughtfillet-setback.test.ts "direction" fails—

Regression and end-to-end tests are added when the applications exist (ADR-0020), not by the feature audits.