What problem does it solve? Running disciplined, multi-session research inside a repository is hard: claims get promoted without evidence, negative results are lost, novelty is asserted without prior-art checks, and agent roles approve their own work. This Skill scaffolds a portable research operating system into any repo — a director, 17 specialist roles, persistent ledgers, workflows, and strict review gates — so research progress is tracked, auditable, and resumable across sessions. ## Core Features & Use Cases - Conservative installer: scripts/install_research_os.py scaffolds the research tree (claims ledger, experiment registry, literature notes, workflows, Makefile gates) without overwriting existing project files, with --dry-run preview and digest-based ownership tracking. - Specialist agent roster: 17 narrow roles (literature scout, theorist, experimentalist, adversarial reviewer, replication reviewer, synthesis writer, and more) with enforced role boundaries so no agent approves its own work. - Evidence discipline: an evidence ladder from idea to formally proved, mandatory invariants (no improvement claim without a baseline, no novelty claim without novelty review, negative results retained), and make review-gates / make validate-claims checks before synthesis. - Use Case: Point it at an existing ML codebase to run a brownfield loop: map the repo, reproduce the baseline, run the cheapest decisive experiment, pass replication and adversarial review, and record the promoted claim with its evidence in research/claims/claims.yaml. ## Quick Start Ask the agent to install the research operating system into your target repository and run one research loop, for example: "Use codex-research-os to install the research OS into /path/to/my-repo and run the first brownfield loop."
