Use quick-eval when installed; it covers planning, bounded execution and the
result audit. Do not bypass its session budget when invoked from that workflow.
If this skill is installed alone, follow the same minimum launch contract:
- Confirm the intended organization, workspace and API environment. Read the agent, persona, exact cases and selected metric definitions. Check command help and errors. Authentication is not proof of tenant identity.
- Prepare a structured launch request with explicit
options.test_case_ids, iteration count, concurrency, metric IDs and a verified per-call duration limit. Count cases × iterations × (1 + mutation count), per persona, including any prior runs in the session. Concurrency does not reduce total cost. - Show the concrete plan and bounded stopping condition. A request to plan or inspect is not paid-run authority; honor already granted execution authority and ask only when scope/budget must expand. Default to proposing one call for an untested connection, not the entire test set.
- Execute
coval --agent runs launch --input-json @run.json. Checkokand exit status. Save the returned run ID. If the response is ambiguous, inspect matching runs before retrying; never blindly retry a launch. - Return the run ID and actual observed status. Queued is not completed. Follow
the user's requested scope: for a completed evaluation, poll and audit results
as in
quick-eval; for launch-only, hand back the run without claiming quality.
Use the current runs schema. Do not use
/eval/*, invent command flags, add schedules, change agent defaults, or rerun
unattended. A local plan is not a server-enforced spending cap.