GitHub Repository

Two Claude Code skills that run a coding task through a multi-agent harness — plan → N parallel implementations → adversarial verification → judge. pantheon (Claude-only), pantheon-x (GPT-5.5 cross-model verify). MIT.

1 starsJavaScript

Code and reviews that only count if they survive an attack

Name: Code and reviews that only count if they survive an attack
Availability: InStock
Author: lolu1032

by lolu1032·Jun 17, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainShip It

Adversarial reviewers try to break builds before self-written tests rubber-stamp them.

Strengths

•Cross-model verification with GPT-5.5 shrinks single-model blind spots
•Deterministic pipeline with test-defined correctness before any code
•Honest README admits this is scaffolding around models, not new tech

Weaknesses

•Requires paid Claude Code plan plus OpenAI Codex plugin for full features
•Best-of-N and LLM-as-judge patterns already exist in agent frameworks

Post Description

Pantheon is a technique for generating each subagent, generating codes with different solutions, deciphering them and writing codes that ultimately survive. For Pantheon-X, GPT 5.5 is broken, GPT is broken and surviving is the last skill to write code written by Claude. For Pantheon-gap, it's a review skill that keeps checking if the actual review is true and determines whether it's a real error or not. In the case of Pantheon-gap-x, GPT 5.5 is a skill that performs better by cross-checking and determining whether the error is real or not. I'd appreciate it if you could actually use it and give me feedback through the problem