law/anti-patterns.md

Soft-warn patterns the airlock filter (Layer 2) surfaces during PR review. Unlike invariants, anti-patterns do not auto-reject; they raise a flag to the LLM judge (Layer 3) which decides whether the surrounding context justifies the pattern.

Each pattern has an id, a regex, a severity, and a description. The LLM judge sees the description and the surrounding diff context.

A-1 · MRO improver legacy imports

Legacy module reference deprecated · port forward to the current data path.

A-2 · cleanlab/fiftyone import

AD-toolchain experimental tools · production data path goes through zdata.

A-3 · ordinal-tag identifiers in source

Ordinal-style labels in code · use bullet ordering or feature names.

A-4 · hand-rolled JSON Web Token

Use the audited jose library for JWT.

A-5 · console.log in src/

Use the worker logger or emit() to trace · not stdout.

A-6 · trademarked centralised code host name

No reference to the trademarked host name in user-facing copy.

A-7 · brand drift

Primary tagline is "Agent Projects by Email". Adjacent brand words that imply a different positioning trigger this warn so the LLM judge can read context.

A-8 · suffix-versioned filenames

Replace cleanly · no suffix-versioned filenames.

A-9 · TypeScript any in app code

Use unknown and narrow at boundaries.

A-10 · hand-rolled HTTP server (Rust)

Use axum / actix / hyper · never roll the wire format.

Severity ladder

• reject — hard stop (used only when the pattern is itself a sign

of crypto / parser hand-roll). Currently no anti-patterns are at this level; that's what invariants are for.

• warn — soft signal · LLM judge decides.
• info — surface for human review at PR time, no impact on verdict.

— additions land via PR to this file; severity bumps require explicit review.