Show HN: Adversarial Code Review paired agents, zero noise,validated findings

github.com · rainmod · 15 hours ago · view on HN · tool
quality 3/10 · low quality
0 net
AI Summary

A code review system using two adversarial LLM agents—one finding vulnerabilities, one challenging findings with counter-evidence—to overcome the problem of LLMs agreeing with themselves and to produce zero-noise validated security findings.

Author here. We built this because one AI reviewing code and checking its own work just agrees with itself — the research confirms LLMs can't reliably self-correct on hard reasoning. The fix: split it into two agents with opposed goals. A reviewer agent finds problems. A dev agent tries to disprove each finding with specific counter-evidence from the codebase.

Three verdicts: VALID, INVALID, AMBIGUOUS. Only what survives reaches your team.

Agents are auto-generated per service — point the skill at your repo, it scans your stack and produces the full set. ~30 min of human tuning per service to add tribal knowledge. Repo includes the shared preamble (quality constitution), reviewer/dev templates, and the auto-generation skill. Happy to answer questions.