Start a Coval evaluation

Help the customer answer a product question, not accumulate evaluation resources. Find their situation, explain the next useful step in one sentence, and follow the matching skill. Reuse supplied context; ask only for missing information.

Establish the starting point

Identify the agent's job, the decision the customer wants to make, and what evidence they already have: an endpoint, a test set, completed conversations, an observed failure, or reviewer labels. A first-time Coval user may already have production recordings; don't send them through synthetic onboarding.

For a connected account, check coval --version, coval --agent agent doctor, and the relevant resource's context/--help. Confirm the intended organization, API environment and workspace with the customer or their existing configuration. Doctor checks authentication/connectivity; it does not prove the organization identity. Never print a key or inspect other organizations to find a match.

Start with a small, scoped inventory. --agent responses wrap payloads in data; check ok and the process exit code. A list can be one page, not the whole account. No credentials yet? Explain setup using the CLI installation guide, and continue designing locally if the customer wants that.

Route by evidence

These are installed skill names, not shell commands. Locate the selected skill in the agent's installed skill catalog. If absent, explain the missing skill and install it only in the user's intended scope, or perform the bounded task using current public documentation. Do not invent an executable /skill command or assume sibling directories survive a single-skill installation.

Keep the next step small

• Reuse existing resources. Don't create a second agent, rewrite a production metric, attach defaults, schedule runs, or enable notifications merely to start.
• For a new voice connection, propose one case, one persona, one iteration, concurrency one. A request to plan or audit does not authorize a paid run. Honor an already authorized execution budget; don't ask again inside it.
• Before execution, state the selected cases, metric IDs, maximum simulations, call-duration limit, and stopping condition. Count base + every mutation and every persona. Reruns and metric tests consume budget too.
• When the customer already knows the failure, don't require a large discovery study before adding a useful regression test. Label assumptions as hypotheses.
• Finished execution is not proof of agent quality. Missing outputs, failed metrics, simulator mistakes and unvalidated judges must remain visible.

Do not call a starter run statistically conclusive or release-certified. End with the evidence obtained and the single next action it supports.