The review doctrine
- +Review Contract: every PR gets a computed contract — change class (production, migration, one-off script, test, config, docs, generated, revert) from deterministic signals like draft flag, labels, branch prefix, paths, title, and size. The LLM fills intent only when metadata is silent. Visible on every review.
- +Depth follows the contract: one-off scripts get a single balanced reviewer (correctness + data safety) instead of the full specialist squad; docs and generated changes skip the second pass. The security/migration floor never relaxes.
- +Review Laws: one severity rubric everywhere, silence is a valid review, no praise filler, style is the linter's job. Every finding must carry a concrete failure scenario, file:line evidence, and a suggested fix.
- +Judge scoring now runs on every review, every plan — class-aware thresholds, hard cap of 10 inline comments, near-threshold findings fold into a collapsed Minor notes section. No minimum-comment behavior.
- +Team-feedback memory: dismissals become semantic memories with the reason and change kind; repeated dismissed patterns auto-suppress (security never suppressed). One-off-script dismissals don't silence production reviews.
- +Re-reviews resolve Argus's own fixed comments (“Resolved by <sha>”) and only post what's new.
- +Glass Box footer on every review: contract, what was checked, suppressed count, duration. Gauge tracks address rate — whether comments actually led to code changes, human-fix weighted, per category per change class.
- +Oversized PRs get an honest reduced-confidence note and a split recommendation instead of a fake-thorough review.