Launch agentic RL on Iris
Read .agents/ops/coreweave.md, the selected configuration, and the current
cloud.iris.iris_backend interface before constructing a command. Resolve cluster, runtime profile,
credentials, capacity, retry policy, and artifact destinations at execution time.
Workflow
- Start from a clean committed revision and classify the harness from the resolved configuration.
- Validate model, topology, rollout concurrency, sandbox provider, verifier, checkpoint policy,
artifact paths, and campaign registration/publication policy.
- Run the complete launch with
--dry-run. Inspect the generated job identity, runtime commit and profile,
resource roles, command, secret references, retries, and durable destinations.
- Confirm external prerequisites using non-secret evidence: provider access, snapshot headroom,
task data, model access, and target-cluster capacity.
- Submit the reviewed command and preserve its resolved configuration with the experiment record.
- Observe through initialization into completed trials, verifier rewards, advancing training steps,
and a checkpoint when due. Use
rl-job-health-deep-dive when state alone cannot prove progress.
Safety
- Never expose credentials in commands, logs, or artifacts.
- Never patch live pods or remote checkouts.
- Do not cancel or mutate a running job without authority.
- A running controller state without advancing trials or training is not proof of health.
- On failure, preserve the first causal evidence and recommend the smallest reproducible correction.
1---2name: rl-agentic-launch-iris3description: Validate, submit, and observe agentic MarinSkyRL training on Iris when the configuration uses Harbor, Daytona, terminal-bench, or another sandboxed agent harness.4---56# Launch agentic RL on Iris78Read `.agents/ops/coreweave.md`, the selected configuration, and the current9`cloud.iris.iris_backend` interface before constructing a command. Resolve cluster, runtime profile,10credentials, capacity, retry policy, and artifact destinations at execution time.1112## Workflow13141. Start from a clean committed revision and classify the harness from the resolved configuration.152. Validate model, topology, rollout concurrency, sandbox provider, verifier, checkpoint policy,16 artifact paths, and campaign registration/publication policy.173. Run the complete launch with `--dry-run`. Inspect the generated job identity, runtime commit and profile,18 resource roles, command, secret references, retries, and durable destinations.194. Confirm external prerequisites using non-secret evidence: provider access, snapshot headroom,20 task data, model access, and target-cluster capacity.215. Submit the reviewed command and preserve its resolved configuration with the experiment record.226. Observe through initialization into completed trials, verifier rewards, advancing training steps,23 and a checkpoint when due. Use `rl-job-health-deep-dive` when state alone cannot prove progress.2425## Safety2627- Never expose credentials in commands, logs, or artifacts.28- Never patch live pods or remote checkouts.29- Do not cancel or mutate a running job without authority.30- A running controller state without advancing trials or training is not proof of health.31- On failure, preserve the first causal evidence and recommend the smallest reproducible correction.
marin-community