overflowy/herdr-adversarial-review

This is a Claude Code skill that adds a second, independent AI reviewer to check your code changes before you ship them. Instead of asking the same assistant that wrote the code to also review it, this skill spins up a separate model, GPT-5.6 Sol, running in its own live terminal pane, hands it your diff or plan along with a description of what you were trying to achieve, and waits for it to finish reviewing. Once the second reviewer produces its findings, Claude does not simply repeat them back to you. It interrogates each finding against the actual code, pressing the reviewer on anything that seems disputable, and then reports every finding as Confirmed, Disputed, or Unverified. The author notes that the adversarial reviewer is instructed to look for problems and will sometimes overstate or invent issues, so getting a verdict that nothing holds up is treated as a perfectly valid outcome. Setting this up requires several separate tools working together. Herdr is a terminal pane manager that creates the reviewer's pane and tracks whether it is idle, working, or blocked. CLIProxyAPI is a local proxy that lets the standard Claude command line tool talk to a non-Anthropic model by presenting an Anthropic compatible interface backed by an OpenAI login. Agent-safehouse provides sandboxing on macOS so the reviewer, which runs with reduced permission checks, cannot touch anything outside the project folder. Finally, a small shell function called safecodex ties these pieces together with the right environment variables so that typing one command launches the sandboxed second opinion reviewer. After installing all four pieces, the skill itself can be added through Claude Code's plugin system or by copying its files into your skills folder by hand. From there, you trigger a review either with a slash command or just by asking Claude in plain language to poke holes in your diff or get a second opinion before you ship. It is released under the MIT license.