Code Quality Standard

Applies to: Human- and AI-authored production code
Version: 1.1 | Updated: 2026-07-16

---

Purpose

AI speed must not lower the bar. Code from agents is held to the same standard as human code — often higher on tests and plan conformance.

Minimum bar

Correctness

• Matches acceptance criteria and approved plan
• Error paths handled; no silent swallows of security-relevant failures
• Null/empty/boundary cases explicit where the type system doesn’t already force them

Design

• Functions do one job; names describe behavior
• No unnecessary duplication of existing helpers
• Comments explain why, not narrate what
• No dead or commented-out code committed

Safety

• No secrets in source
• Parameterized queries / safe HTML / safe shell
• AuthZ on new endpoints and tool wrappers
• Untrusted model output not executed as code or SQL

Tests

• Unit tests for new logic
• Integration tests for new boundaries (API, DB, queue)
• Tests assert behavior, not incidental implementation details
• Coverage not used as a vanity metric to excuse missing critical paths

Types and tooling

• No any / untyped public APIs without an explicit waiver
• Lint and typecheck clean for touched packages
• Formatting matches repo defaults

AI-specific rules

1. Plan conformance — unexpected files require a plan revision, not a shrug
2. No drive-by refactors — out-of-scope cleanup belongs in its own change
3. Cite sources — when copying patterns from the codebase, match local idioms
4. Prove it — agent claims of “tests pass” must match CI, not chat transcript

Related

• Coding checklist
• Coder agent
• Definition of Done