adversarial-panel

Runs isolated multi-model Codex panels with blind answers, cross-critique, and evidence-weighted synthesis.

1|Updated Jul 16, 2026
One-click install
npx skills add https://github.com/sota411/codex-config --skill adversarial-panel-sota411
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: adversarial-panel
Source: https://github.com/sota411/codex-config/tree/main/user-skills/adversarial-panel
Command: npx skills add https://github.com/sota411/codex-config --skill adversarial-panel-sota411

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve? High-stakes decisions and contested technical claims suffer when a single model answers alone, since one model can hallucinate, share its own blind spots, or anchor on weak assumptions. This Skill orchestrates 2-3 independent Codex panelists that answer blindly, attack each other's claims, and defend or concede under critique, so conclusions rest on verified evidence rather than confident prose. ## Core Features & Use Cases - Isolated three-round debate: Round 1 blind answers, Round 2 cross-critique with mandatory falsification attempts, and Round 3 final positions with explicit concede/defend responses to every critique. - Sandboxed panelist execution: Each panelist runs in a bubblewrap mount/PID namespace with a read-only workspace, per-panelist private directories, denied access to auth files and /proc, and filtered environment variables. - Schema-validated outputs with fail-fast degradation: JSON schemas enforce claims with confidence and falsification conditions; invalid outputs trigger one retry, and runs abort when fewer than two valid panelists remain. - Use Case: Before approving an irreversible database migration, run the panel so a first-principles analyst, an outside-view estimator, and a falsifier independently assess the plan, attack each other's reasoning, and surface the strongest surviving evidence. ## Quick Start Ask the agent to use the adversarial-panel skill to adversarially verify your important decision with multiple independent Codex models.

Frequently Asked Questions about adversarial-panel

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run a multi-model adversarial debate with Codex?▼

Write a self-contained brief to a temporary file, then run scripts/run_panel.py with --brief, --workspace, and two or three --panelist entries in role=model:effort form. The runner executes blind answers, cross-critique, and final positions, returning a run_dir and record_path as JSON.

When should I use an adversarial panel instead of a single model answer?▼

Use it when a wrong answer would cause real harm to cost, safety, design, or major decisions, and when claims are contested or verifiable by execution, recalculation, or primary sources. For local code review, the review skill remains the primary workflow.

What are the system requirements for the adversarial panel runner?▼

The runner requires Linux with bubblewrap (bwrap) installed, since each panelist runs in its own mount and PID namespace. If bubblewrap is missing, the runner fails fast instead of launching child Codex processes without isolation.

How does the panel prevent panelists from seeing each other's answers?▼

In Round 1 each panelist receives only the shared brief and its own methodology, with no conversation history or peer answers. Sandbox mounts expose only a read-only /workspace and a private /panel directory, denying access to other panelists' artifacts.

What happens when a panelist returns invalid or empty output?▼

Invalid outputs are retried once under identical conditions. If a three-panelist run still has two valid panelists, it degrades and discloses the dropout; fewer than two valid panelists aborts the run as a failure.

Why does the panel forbid voting and averaging of conclusions?▼

Voting and confidence averaging hide disagreement without testing it. The skill instead ranks evidence by strength, prioritizing reproducible execution and primary sources over independent agreement, and preserves minority opinions backed by verifiable counterexamples.