| name | zig-deep-correctness-review |
| description | Deep behavioral correctness review (pass 3): concurrency, algorithms, SIMD parity, pipeline determinism, resource lifetime, outliers, and test-coverage gaps. Use when the user asks for deep correctness, concurrency review, determinism audit, race hunting, or /zig-deep-correctness-review. Goes beyond idiom/surface reviews. |
Zig Deep-Correctness Review
Pass 3. Do not re-report idiom/naming/stdlib/surface ReleaseFast issues
already covered by best-practices or idiom-lint. Focus on behavioral correctness.
Agents: zig-review-specialist.
Themes: .grok/skills/zig-workflows/references/module-presets.md →
deep-correctness-review.
Phase 1 — Per-theme review (parallel)
One zig-review-specialist per theme unit (background: true,
capability_mode: read-only). Use the preset focus text as the unit brief.
For each finding give concrete failure: inputs/state → wrong behavior / crash /
leak / nondeterminism. Cite real file:line. Prefer few high-confidence findings
over speculation. Clean units with zero findings are valid.
Kinds: bug, race, determinism, numerical, leak, test-gap, anomaly,
contract. Severity: high|medium|low. Also: title, detail, fix_direction,
durable, durable_mechanism, durable_rationale.
Phase 2 — Adversarial verify (high bar)
For each finding, spawn a verifier that defaults to skepticism:
- Trace control/data flow; try to refute
- Races: concurrent access + write, no happens-before?
- Determinism/numerical: does divergence actually manifest?
- Test-gap: truly untested and load-bearing? (search tests)
CONFIRMED only with a concrete failing case or confirmed coverage gap.
Keep CONFIRMED/PLAUSIBLE with is_real.
Phase 3 — Synthesize
Split cleanly:
- top_bugs — real defects ranked by severity + fix direction
- test_gaps — highest-value untested invariants + narrow scenario + blast radius
- durable_items — only net-new lint/agent/doc items (no duplicates of existing guidance)
- summary
If nothing survives verification, report that the reviewed surface is clean — that
is a valuable result.
Done criteria
- All themes reviewed and verified
- Ranked bugs / test gaps / durable items delivered as Markdown
- No code edits unless the user asks
