mcpField
L
launch-run

launch-run

github.com/coval-aiTesting & quality361 installsUpdated 1d ago
Repository

About launch-run

A Coval skill for launching bounded evaluation runs. It checks the target workspace, cases, metrics, execution budget, and run status, and avoids blind retries or treating queued runs as complete.

Key features

  1. 01Verify organization and workspace
  2. 02Review cases and metric definitions
  3. 03Plan a bounded run request
  4. 04Launch runs with Coval CLI
  5. 05Check run ID and observed status
  6. 06Avoid unverified retries

Use cases

  1. 01Launch a single test for an untested connection
  2. 02Run approved cases within a session budget
  3. 03Return a run ID for launch-only requests

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/coval-ai/coval-external-skills --skill launch-run

After installing, type/launch-runin Claude Code to confirm it loaded.

Details

From the project README

Launch a Coval run

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:

  1. 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.
  2. 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.
  3. 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.
  4. Execute coval --agent runs launch --input-json @run.json. Check ok and exit status. Save the returned run ID. If the response is ambiguous, inspect matching runs before retrying; never blindly retry a launch.
  5. 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.

Comments

Markdown supported