oscar-gilg

Interactive onboarding for zombuul. Walks through prerequisites, API keys, .env creation, and verifies everything works. Run this after installing the plugin.

End-to-end smoke test of zombuul. Pre-flight, then 3 parallel agents (API, pod list, full pod lifecycle), then a serial e2e step that runs /zombuul:run-experiment on a tiny canned spec to exercise the orchestration + PR flow. Use before publishing a release. Takes ~10–15 minutes and costs roughly $0.10 in RunPod GPU time.

Run an experiment from a spec or description. If given a path to an existing spec, runs it directly. If given a natural language description, synthesizes a spec first, then runs it. Argument $ARGUMENTS — `<spec_path_or_description> [--remote] [--no-pr]`. Pass `--remote` to execute autonomously on a GPU pod (survives local disconnects); otherwise runs local-first. Pass `--no-pr` to skip the draft-PR-on-the-experiment-branch flow.

Finalize an experiment after the main work is done — sync pod results back, run analysis, spawn report review, commit and push. Argument $ARGUMENTS — `<spec_path> [--pod <pod_name>]`. Invoked automatically by `/zombuul:babysit --on-complete` when a long-running job finishes, or directly by you when an experiment doesn't use a babysitter (no pod, no cron).

Spin up a RunPod GPU pod. Argument $ARGUMENTS — optional pod name, optionally followed by `--remote` to install Claude Code + zombuul plugin on the pod (needed when the pod itself will run an agent, not just receive SSH commands).

Pause a RunPod pod: stop GPU billing; keep /workspace only (container disk is wiped on resume).

Provision a RunPod pod: wait for setup, configure SSH, sync .env and data. Argument $ARGUMENTS — JSON with keys: pod_id, pod_name, ip, port, and optionally spec_path and data_dirs.

Resume a paused RunPod pod.

Audit numeric claims by re-deriving each value from data and comparing to what's recorded. Argument $ARGUMENTS — optional scope (claim name, section, producer, or --manual/--changed/--all). Follows each claim's derivation, flags discrepancies with the standard `nature` taxonomy.

Log numeric claims added to a document into the claim registry. Argument $ARGUMENTS — optional. With no args, runs retrospectively over recent turns + git diff. With a prompt, runs the task prospectively and logs as it goes. With a file path, scopes retrospective logging to that file. Starts with `--commit` to also refresh macros + audit surface at the end.