andpuqing/gflow/gflow-ops

gflow Ops

Use this skill for local gflow work: inspect the daemon, queue, jobs, reservations, logs, and safely mutate scheduler state.

When To Use

• The user asks about gflowd, gqueue, gjob, gcancel, ginfo, gstats, or the local scheduler state.
• The task needs job inspection, failure diagnosis, or GPU availability checks.
• The task needs job submission, update, hold, release, or cancellation.
• The task needs validation of the local gflow mcp serve MCP server.

Do not use this skill for remote cluster orchestration or multi-tenant service design. gflow is local-first.

Preferred Order

1. Check liveness first.

Use MCP get_health if available. CLI fallback: gflowd status.

1. Inspect before mutating.

Use get_info, list_jobs, get_job, get_stats, list_reservations.

1. Prefer MCP tools over ad hoc shell commands when the tool server is available.
2. For validation, use short no-GPU test jobs in the current repo directory.
3. Clean up test jobs when done.

MCP Tools

Prefer these tools when the local MCP server is configured:

• get_health
• get_info
• list_jobs
• get_job
• get_job_log
• get_stats
• list_reservations
• submit_jobs
• update_job
• hold_job
• release_job
• cancel_job

get_job_log.text is the agent-friendly program output view. Use full_text only when the filtered output is missing important context. Use submit_jobs for both one-off test jobs and small parameter sweeps. It attempts jobs sequentially and reports per-job success or failure.

CLI Fallbacks

Use CLI only when MCP is unavailable or when testing the CLI itself.

• Health: gflowd status
• Queue: gqueue, gqueue -a, gqueue -s Running
• Scheduler info: ginfo, gstats
• Submit: gbatch ...
• Inspect: gjob show <job_id>, gjob log <job_id>
• Mutate: gjob hold <job_id>, gjob release <job_id>, gjob update ..., gcancel <job_id>

Safety Rules

• Treat all write operations as potentially user-impacting.
• Do not cancel, hold, release, or update existing non-test jobs without explicit user intent.
• For experimentation, create dedicated test jobs with gpus: 0 unless the user explicitly wants GPU scheduling behavior.
• Before using cancel_job, confirm the target job ID and whether it is a test job.
• After creating test jobs, cancel them if they are still active.

Reliable Validation Pattern

When you need a queued job for hold_job, release_job, or update_job testing:

1. Submit a blocker job with gpus: 0 and a long sleep.
2. Submit a dependent job with depends_on_ids: [blocker_id].
3. Operate on the dependent job while it is waiting on the blocker.
4. Cancel both jobs after validation.

This avoids interfering with GPU workloads and reduces scheduler races.

Examples

User request: "Check whether gflowd is healthy and show me the latest failed jobs."

Preferred flow:

1. get_health
2. list_jobs with a failure filter and explicit limit
3. get_job or get_job_log for the most relevant failures

User request: "Test whether hold/release works."

Preferred flow:

1. Submit blocker test job
2. Submit dependent test job
3. hold_job
4. update_job if needed
5. release_job
6. cancel_job both test jobs

User request: "Create a small parameter sweep with MCP."

Preferred flow:

1. get_health
2. Build the full set of test jobs with gpus: 0 unless GPU behavior is part of the test
3. submit_jobs
4. list_jobs or get_job to confirm the submitted IDs and states
5. cancel_job the test jobs after validation if they are no longer needed