konjo-quality

# Konjo Quality Framework — Agent Reference ## Why This Exists AI-assisted code produces **1.7× more logical and correctness bugs** than traditional development (CodeRabbit 2026). AI agents change tests so broken code passes instead of fixing the code. AI self-review is architecturally circular — it checks code against itself, not against intent. This framework provides external ground truth via three independent walls that cannot be reasoned past. The Konjo Critic (Wall 3) uses `claude-opus-4-6` in a separate session. The builder has blind spots from the construction process; the critic comes in cold with a different capability profile to reduce correlated failures. --- ## The Three Walls | Wall | When | What | Blocks | |------|------|------|--------| | **Wall 1** | Pre-commit hook | Format, lint, unwrap scan, DRY (staged only), TODO scan | The commit | | **Wall 2** | CI / GitHub Actions | Coverage, mutation, complexity, size, docs, audit, review | The merge | | **Wall 3** | CI (PRs only) | Claude Opus adversarial review against 10 mandatory questions | The merge | --- ## Hard Quality Thresholds (all enforced by CI) | Metric | Hard Block | Target | Tool | |--------|-----------|--------|------| | Line coverage | ≥ 80% | ≥ 95% | cargo-llvm-cov / pytest-cov | | Mutation survival (changed files) | ≤ 10% | 0% | cargo-mutants / mutmut | | Cognitive complexity per function | ≤ 15 | ≤ 10 | clippy / radon | | Lint violations | 0 | 0 | clippy / ruff | | Dead code warnings | 0...