Proposal: post-review sanity check when all specialists return NO FINDINGS

[Ziadstr]

Problem

When the review specialist army dispatches 3-5 agents on a 500+ line diff, there is no validation if all specialists return NO FINDINGS. The system trusts their output unconditionally.

In practice, agents can:

• Rationalize away real issues ("this pattern looks intentional")
• Hit context limits and give up silently
• Miss issues due to prompt phrasing or model variance

A 500-line diff with zero findings from 5 specialists is suspicious. Not impossible, but worth a second look.

Proposed solution: lightweight post-review sanity check

After the specialist merge step, before Fix-First:

1. If diff > 200 lines AND all specialists return NO FINDINGS: flag as suspicious
2. Re-dispatch one specialist (testing or security, most likely to find something) with an explicit prompt: "A previous review of this diff found zero issues. Review again with extra scrutiny. Focus on what might have been missed."
3. If the re-run also returns NO FINDINGS: accept it (genuine clean diff)
4. If it finds issues: include them in findings with a note "found on second pass"

This adds one extra agent call in the rare case where a large diff gets a clean sweep, zero cost in the common case.

Why this matters

From experience building a similar review system: we added an "anti-rationalization" hook that fires PostToolUse and catches when agents produce suspiciously clean results on objectively complex code. It's caught real misses where the agent skipped auth checks, missed N+1 queries, or rationalized away a race condition.

The cost is minimal (one conditional re-dispatch), the signal is high, and it prevents false confidence in the review output.

Alternative: confidence floor

Instead of re-dispatching, require at least one finding on diffs > 200 lines. If all specialists return NO FINDINGS, output a warning in the review report: "Large diff reviewed with zero findings. Manual review recommended." This is cheaper but less actionable.

Not submitting a PR for this since it touches the review pipeline architecture and should be a design decision. Happy to implement either approach if there's interest.