Volc

Volc ML Platform

Workflow

1. Determine whether the user wants an explanation, command examples, or live execution. Do not execute commands when the user only asks for instructions or a command template.
2. Confirm the installed CLI surface before relying on command syntax:
   • volc version
   • volc configure --help
   • volc ml_devinstance --help
   • volc ml_task --help
3. Read references/cli-commands.md for exact flags and reusable templates.
4. Read references/ml-platform-docs.md when official documentation, current product behavior, or a doc citation is needed.
5. Inspect current state with list/get/status/log commands before proposing a state-changing action.
6. Keep unresolved values as explicit placeholders. Never invent instance IDs, task IDs, queue names, image IDs, regions, endpoints, or credentials.
7. Distinguish clearly between commands that were executed and commands that are only suggested.

Remote Development Workflow

For requests such as "在远程开发机起任务":

1. Locate and inspect the development instance with volc ml_devinstance list and volc ml_devinstance get --id <devinstance-id>.
2. If it is stopped, show or run volc ml_devinstance start --id <devinstance-id> only after the user confirms the instance and cost impact.
3. Use the SSH command or endpoint provided by the platform. Do not guess the host, port, or username.
4. From the remote workspace, verify the code path, entrypoint, image, and resource queue before submitting with either:
   • volc ml_task submit for custom jobs.
   • volc ml_task sbatch for Slurm scripts.
5. Monitor with volc ml_task list, volc ml_task get, volc ml_task logs -f, volc ml_task instance list, and volc ml_task top.

Task Submission

• Prefer a reproducible config generated by volc ml_task template --path <file> --example true, then customize it before submission.
• Validate the entrypoint, user-code path, image, resource queue, framework, node count, and GPU requirements before submitting.
• Show the final command and unresolved placeholders when required values are missing; do not submit a partially inferred job.
• Use volc ml_task submit for custom jobs and volc ml_task sbatch for Slurm scripts. Verify each subcommand with --help because available flags can vary by CLI version.

Guardrails

• Confirm before state-changing, destructive, or cost-impacting commands, including starting or stopping development instances, rebooting, cancelling or deleting resources, and submitting long-running jobs.
• Do not print, request, store, or commit AK/SK credentials. Use an existing profile or placeholders such as $VOLC_AK, $VOLC_SK, $VOLC_REGION, and $VOLC_PROFILE.
• Prefer --resource_queue_name over deprecated --resource_group_id; use --resource_queue_id only when a queue name is unavailable.
• Remember --user_code_path/--cp uploads files to TOS; .gitignore under that path affects uploaded code.
• For symlinks, choose explicitly: --copy-links copies referents, while --links recreates symlinks.
• Limit log output to what is needed for diagnosis and avoid reproducing credentials or sensitive environment values.