[sfc-gh-kganesan]

HUMAN:
I tested this skill end to end on Cortex Training against current Arctic Platform, on two paths: Harbor with Snowflake sandboxes, and Harbor with HARBOR_ENV=host. The direct code localization recipe was also run. Every command in it was run as written. Paths we didn't run are marked untested rather than left out.

The Snowflake sandbox path needs Snowflake-AI-Research/Arctic-Platform#404. Please merge this after it lands. The host and code localization paths already work on Arctic Platform main.

---

AGENT:

Why

Following cortex-training-rl against current Arctic Platform main fails before training starts:

• The skill runs only cortex-training login. The Arctic Platform launchers read the Cortex connection only from ARCTIC_CORTEX_*, so their preflight stops with "no Cortex connection".
• It lists openhands-sdk and openhands as HARBOR_AGENT values. The Harbor launcher now refuses agents that Harbor installs inside the environment, because task code there could read the rollout's API key.
• It says the default HARBOR_ENV is HostEnvironment. The default is now the Snowflake sandbox, and the Harbor launch sets no HARBOR_ENV_KWARGS, so the launcher's preflight stops before provisioning GPUs.

Summary

• Preflight exports ARCTIC_CORTEX_* from the login config JSON.
• The Harbor launch sets HARBOR_ENV_KWARGS (base_image: sandbox-base, which needs no image repository), and the HARBOR_AGENT / HARBOR_ENV / HARBOR_ENV_KWARGS descriptions match the launcher.
• Troubleshooting covers the two preflight refusals, a Snowflake run where every rollout errors because the registry rejects image pushes, and rollouts failing on the 100-sandbox quota or an expired session, with a cleanup snippet.
• Only commands and paths that were run end to end:
  • A run's Cortex job is found by its comment (cortex-training list | jq ...). Nothing writes the $CKPT_DIR/cortex_job_id the skill used to read.
  • cortex-training capacity (it has no --hardware flag).
  • The OPENHANDS_TOOLS example drops file_editor, which the Harbor recipe doesn't provide.
  • HARBOR_ENV lists snowflake and host as tested; the Snowflake kwargs include role and an egress_eai for GitHub, with login guidance for runs that log in again every 20 minutes.
  • The launcher's pre-GPU first-task check (Harbor RL on Snowflake sandboxes, end to end: base_image, errored rollouts, sandbox reliability, session renewal, and a pre-GPU task check Snowflake-AI-Research/Arctic-Platform#404) and the errors it reports are in the launch step and troubleshooting.
  • The Harbor reference numbers are the 25-step Snowflake run, and the reward-graph text matches what the graph script produces.
• From an OpenHands agent following the skill from a fresh home directory: step 2 clones SkyRL and exports SKYRL_HOME, rule 4 says how to check login configs without printing them, and step 4 says when a step is done and that a run ends by itself. Launch and monitoring cover evaluation during training (eval/all/avg_score).

Issue Number

None.

How to Test

Follow the skill on a fresh clone of Arctic Platform main:

1. Preflight, then export the four ARCTIC_CORTEX_* variables as shown.
2. Run the direct code-localization launch from step 3 with trainer.max_training_steps=3. With ARCTIC_CORTEX_* exported, the launcher gets past its credential check, provisions Cortex, and trains.
3. Run the Harbor launch from step 3 with its HARBOR_ENV_KWARGS and trainer.max_training_steps=3.

Results so far, from a fresh clone of Arctic Platform main:

• Direct code localization: completed 3 steps on Cortex (4 training + 4 sampling H200s), with OpenHands code localization: install [sft,cortex] on the driver, not [rl] Snowflake-AI-Research/Arctic-Platform#391's launcher fix. Average reward 0.078, 0.109, 0.229 over 64 rollouts per step, with 1 errored rollout in step 1 (an OpenHands SDK ConversationState race) and 0 after. The log ends with Training done!.
• Harbor on Snowflake sandboxes with this skill's HARBOR_ENV_KWARGS: completed 3 steps on Cortex with 64 concurrent sandboxes and no image push. Average reward 0.781, 0.294, 0.208 (maximum 3.0); 188 of 192 rollouts were graded, and 4 sandbox transport failures were masked as .error. The log ends with Training done!. This needs Harbor RL on Snowflake sandboxes, end to end: base_image, errored rollouts, sandbox reliability, session renewal, and a pre-GPU task check Snowflake-AI-Research/Arctic-Platform#404.
• Harbor on HARBOR_ENV=host: completed 3 steps, average reward 0.391, 0.172, 0.208, 0 errored rollouts.
• Harbor on Snowflake sandboxes, 25 steps (the README's SWE-smith example, Qwen3.5-4B, 64 rollouts per step, Harbor Snowflake sandbox: base_image replays task Dockerfiles without mirroring Snowflake-AI-Research/Arctic-Platform#392–Snowflake sandbox: renew the session before it expires, and stop leaking sandboxes Snowflake-AI-Research/Arctic-Platform#395): average reward 0.59 for steps 1–10, 1.81 for steps 11–20 and 2.39 for steps 21–25, with a best step of 2.54. 1,596 of 1,600 rollouts were graded, and the run took 1 h 35 min.
• Harbor on Snowflake sandboxes, 3 steps, from a fresh clone of Harbor RL on Snowflake sandboxes, end to end: base_image, errored rollouts, sandbox reliability, session renewal, and a pre-GPU task check Snowflake-AI-Research/Arctic-Platform#404 with this skill's kwargs: the first-task check passed in 20 s, then 192 of 192 rollouts were graded (average reward 0.94, 0.18, 0.19). The job ended CANCELLED and no sandboxes were left. Both monitoring.md snippets run against the 25-step log.
• An OpenHands agent following this skill (OpenHands CLI, Claude Sonnet 4.5, fresh home directory, Harbor on Snowflake sandboxes): it ran preflight, exported the data, launched in tmux, monitored, drew the reward graph, and stopped the run, whose job ended CANCELLED. Steps 1 and 2 averaged 0.156 and 0.223. Where it went off the skill, the skill now says what to do: it had to find SkyRL itself, printed a login config, stopped during step 3's update because it counted trajectory directories, and drew the graph with matplotlib.
• Evaluation during training (trainer.eval_before_train=true trainer.eval_interval=2, 4 steps): eval/all/avg_score was 0.29 before training, 0.66 after step 2, and 0.50 after step 4, on the 32 SWE-smith eval tasks.

Video/Screenshots

N/A (CLI).

Notes

The base_image option, the Snowflake sandbox reliability fixes and the pre-GPU check all ship in Snowflake-AI-Research/Arctic-Platform#404, which is merged. Training on Harbor tasks other than SWE-smith localization, and the skill's evaluation flags on such tasks, need Snowflake-AI-Research/Arctic-Platform#405. Merge this after it.

[sfc-gh-kganesan]

@neubig could you review this? It updates cortex-training-rl for the latest Arctic Platform Harbor recipe, which was released today:

1. Secure sandboxes by default. Each training rollout now runs in its own Snowflake sandbox. The skill sets this up, so model-generated commands never run on the user's machine.
2. No image push needed. Tasks run on Snowflake's ready-made sandbox image, so users can train on their Harbor tasks without building or uploading container images.
3. Problems caught before GPUs start. The launcher now tries one task first, and the skill explains its messages, so a setup mistake costs seconds instead of GPU time.
4. Simpler setup and cleanup. The skill now connects the user's Cortex credentials for them, checks GPU availability, and confirms the GPUs are released when the run ends.
5. New reference results. A 25-step run on Snowflake sandboxes reached a mean reward of 2.39 out of 3.0 over its last 5 steps.

Every command in the skill was run end to end on Cortex Training.

[openhands-release-bot]

🚀 Released in v0.39.0.