GitHub Repository

Claude Code skill for adversarial code review - spawns a GPT‑5.6 Sol reviewer in a herdr split pane to attack your diff, then interrogates its findings before relaying verdicts.

1 stars

Adversarial code review setup with herdr, Claude and GPT-5.6-sol

Name: Adversarial code review setup with herdr, Claude and GPT-5.6-sol
Availability: InStock
Author: overflowy

by overflowy·Jul 21, 2026·2 points·1 comment

Visit Project View on HN

AI Analysis

●●SolidBig BrainNiche Gem

Spawns a sandboxed GPT-5.6 agent in a split pane to attack your diffs.

Strengths

•Herdr multiplexer enables real-time observation of the adversarial agent's thought process.
•Kernel-level sandboxing via agent-safehouse prevents the reviewer from accessing SSH keys.
•Interrogation loop forces Claude to validate findings against actual code before reporting.

Weaknesses

•Setup requires four distinct tools (herdr, CLIProxyAPI, safehouse, skill) before first use.
•Dependence on CLIProxyAPI for non-Anthropic models adds a fragile middleware layer.