Review Severity

Two Tiers

• Every review finding is blocking or advisory
• The test is behavioral — does fixing the finding change what an agent or the pipeline does?
• Blocking: the fix changes behavior or closes a contract gap
• Advisory: the fix changes only presentation

Blocking — Gates the Merge

• Correctness and security defects
• Policy-contract violations: a carve-out's unmet preconditions, no-secrets, ci-safety gate-evasion, surface-sync that breaks publish
• A rule directive whose violation changes agent behavior
• A style finding whose fix changes meaning — an atomic-bullet split that alters what the bullet directs

Advisory — Never Gates

• Pure prose and style: connective or em-dash placement, a presentation-only atomic-bullet split
• CHANGELOG wording, naming taste, synonym preference
• Copilot findings are always advisory regardless of Copilot's review state — even a Copilot CHANGES_REQUESTED never gates the agent's flow
• Anything whose fix changes only presentation, not behavior

Gating Predicate

• Any blocking finding present → the reviewer posts CHANGES_REQUESTED and the merge gates
• Only advisory findings → the reviewer posts COMMENTED and the merge is allowed
• The policy reviewer's posted state already encodes this — the event is derived from per-finding severity (see the skills/onboard-repo/templates/post-review.sh header; consumers carry a copy at .github/codex-review/post-review.sh)
• The merge watcher gates on the policy reviewer's CHANGES_REQUESTED alone (see skills/release/watch-pr-reviews.sh header)
• Copilot never gates

Split Reading From Acting

• Read every finding in full first — severity never licenses skipping a body (see rules/reviewer-feedback-reading.md)
• Blocking → fix before merge
• Advisory → acknowledge
• Fold an advisory in only when a blocking round is already happening
• Otherwise defer the advisory to a follow-up PR or issue and reference it from the current PR
• Never burn a dedicated re-review round on a lone advisory