collab-board

Strict turn-based collaboration between two AI models is enforced by alternating PRIMARY and SECONDARY turns through gated PLAN and IMPLEMENTATION phases, where each model must back claims with evidence and challenge the other to catch confidently-wrong answers. The session is managed through a split interlinked Markdown tree under .collab-board/.

Adversarial AI peer review is a skeptical collaboration process where a PRIMARY and SECONDARY model take strict alternating turns, challenging each other's claims with evidence. It prevents blind spots by gating the PLAN phase so both models must agree and resolve all open points before any implementation begins.

Yes, adapter-based delegation supports using different AI models like Claude and Codex for turn-based planning and implementation. The PRIMARY model drives the task while the SECONDARY model's turns are delegated through adapters, keeping per-turn context bounded regardless of session length.

To start a collaboration session, use the collab-board skill to scaffold a new session with a Node.js CLI, which creates a split interlinked Markdown session tree under .collab-board/. You then drive the strict turn-based planning and implementation loop until both phases agree and the work is complete.

The best way to keep AI context small during long sessions is enforcing strict alternating turns where each model reads only the files it needs within a split Markdown session tree. This bounded per-turn context approach ensures the context size remains manageable regardless of total session length.

Single AI model implementation produces confidently-wrong answers because it lacks skeptical peer review to challenge assumptions and catch blind spots. Introducing a strict turn-based collaboration with a SECONDARY model enforces adversarial review, requiring evidence-backed claims before advancing through gated phases.