Update SKILL

[alex000kim]

Updates to SKILL based on several autoresearch runs from https://blog.skypilot.co/scaling-autoresearch/

[gemini-code-assist]

Code Review

This pull request updates the SkyPilot skill documentation by adding advanced command options, explaining workdir synchronization behavior, and providing examples for parallel experiment submission. It also refines the recommended agent feedback loop and lists common mistakes. The review feedback suggests making the Python script for job ID extraction more robust against empty results and using explicit job IDs in log commands to avoid ambiguity when multiple jobs are present.

[gemini-code-assist]

The Python one-liner to get the latest job ID is clever, but it's not robust against cases where no jobs are running on the cluster. If sky queue returns no jobs, json.load(sys.stdin)['exp-vm-01'] will likely be an empty list, causing max() to raise a ValueError and break the script. It's better to handle this edge case gracefully.

Suggested change

	job_id=$(sky queue exp-vm-01 -o json \
	| python3 -c "import sys,json; jobs=json.load(sys.stdin)['exp-vm-01']; print(max(j['job_id'] for j in jobs))")
	job_id=$(sky queue exp-vm-01 -o json \
	| python3 -c "import sys, json; jobs = json.load(sys.stdin).get('exp-vm-01', []); print(max(j['job_id'] for j in jobs) if jobs else '')")

[gemini-code-assist]

For robustness, especially in programmatic use by an agent, it's better to explicitly use a job ID with sky logs. The current command sky logs mycluster --status might be ambiguous if multiple jobs are on the cluster, even if it defaults to the latest one. This would also make the example consistent with the recommendation in the callout on line 408.

Consider mentioning how to obtain the job ID (e.g., from sky launch output or sky queue) and using it here and in the following steps. For example:

4. **Wait for completion**: Get the job ID (e.g., from sky launchoutput), then runsky logs mycluster <JOB_ID> --status...

[romilbhardwaj]

Thanks @alex000kim!

[romilbhardwaj]

instead of sky logs mycluster <JOB_ID> --status, would just streaming logs with sky logs mycluster be better for an agent? So if an application causes the job stall (e.g., timeout + retry loops), the agent can look at the output to infer next steps?

[alex000kim]

good point, updated!