Agent Claude Code
Post-run output auditor for experiments. Dispatched by harvest-and-report command. Reads the Red Team Brief's validation criteria and a sample of raw outputs. Checks for data quality issues, degenerate outputs, and reward hacking. Returns CLEAN/ANOMALIESFOUND with specific examples.
Agent Claude Code
Adversarial pre-flight reviewer for experiments. Dispatched by experiment-preflight command. Reads the Red Team Brief, experiment.yaml, code, and dry-run output. Finds reasons the run will waste compute. Returns PASS/FAIL with specific findings. Must NOT receive design conversation history — only files. No sunk cost…
Command Claude Code
Create or update a benchmark/dataset/task reference file. Use when adding a new benchmark, correcting evaluation details, or enriching an existing entry with new information.
Command Claude Code
Sync experiment data and artifacts to the live Research Dashboard. Run after any artifact is produced, experiment state changes, or notes are updated.
Command Claude Code
Pre-flight review: red-team brief, adversarial review, canary job proposal. Run before any compute.
Command Claude Code
Full RACA experiment tutorial — walk through designing, red-teaming, running, and reviewing an experiment end-to-end. Takes 20-30 min with a cluster.
Command Claude Code
Find the best compute option for a job across all configured clusters. Reads .raca/clusters.yaml, checks queue status, estimates cost, and presents a ranked comparison.
Command Claude Code
Post-run harvest: download results, validate, upload to HF, sync dashboard, alert user. Run when artifacts are produced — don't wait for job completion.
Command Claude Code
RACA onboarding — welcome, connect clusters, set up HuggingFace, launch dashboard.
Command Claude Code
Update RACA to the latest version. Pulls new skills, rules, tools, and commands without touching your experiments, notes, or API keys.
Skill Claude CodeCodex
Maintains knowledge about the RACA visualization website. Answers questions about what visualizations are currently supported, how to check if artifacts can be displayed, and how to add new visualizer tabs. Also handles code edits to the frontend/backend when a new visualization type is needed. Run this skill when the…
Skill Claude CodeCodex
Activated when users discuss experiments — designing, running, reviewing, or managing them. Handles folder creation, dashboard sync, state tracking, and lifecycle enforcement. Users may enter at any stage. Meet them where they are, but ensure red-teaming and validation happen before compute runs.
Skill Claude CodeCodex
Core job execution skill. Takes an experiment design or canary specification and handles the full job lifecycle: write sbatch script, submit, monitor, handle artifacts, resume on failure, and report results. Run this skill when the user says "run this job", "submit the experiment", "launch the canary", or "run on ".
Skill Claude CodeCodex
Walk the user through connecting a new compute cluster (SLURM, RunPod, or local GPU) to RACA. Writes config to .raca/clusters.yaml and verifies connectivity. Run this skill when the user says "add a cluster", "set up a cluster", "connect to HPC", or "configure RunPod".
Skill Claude CodeCodex
Walk the user through setting up RunPod as a compute provider for RACA. Covers account creation, API key, GPU selection, pricing, pod vs serverless, known issues, and writing the RunPod entry in .raca/clusters.yaml. Run this skill when the user says "set up RunPod", "add RunPod", "configure RunPod", or "I want to use…