Skip to main content
A run’s execution status and its eval verdict answer different questions. Succeeded means the session finished successfully; it does not necessarily mean it passed every criterion. Judging and analytics can finish after execution.

​
Read the status

Failed or timed-out runs receive a run-credit refund. A successfully executed session that fails an eval criterion does not count as an execution failure. See Billing and credits for how charges are calculated.

​
Stop a session

Stop a run when the agent is looping, working on the wrong thing, or no longer needed. This prevents further AI usage. In the web app, select Stop in the session header or the live session pane. The button changes to Stopping… while Gauge ends the agent and saves its work. The run then shows Canceled. From the CLI or API, use:
The API equivalent is POST /api/v1/runs/{runId}/cancel. Sending the request more than once is safe. After a run stops:
  • The partial session events and diff stay available to inspect.
  • The run is not judged, so it has no verdict.
  • The run is left out of pass rates and other results.
A run stopped while still queued costs 0 credits. For runs using Gauge’s provider keys, a run stopped while the agent is working is charged a base fee plus the AI usage measured until it stopped. Gauge releases the rest of the credits it reserved for the run. See Billing and credits.

​
Inspect a run

Replace <run-id> with the ID from the run list:
logs returns events available so far. A research task may make no repository changes, so an empty diff alone is not a failure. If a run has finished but its insight or verdict is pending, allow the analysis to complete. Runs without a finished judgment are left out of the pass rate. Check how many runs were judged, as well as the percentage.

​
When no runs are created

  1. Check the organization with gauge status.
  2. Inspect the saved configuration and cadence with gauge evals get <eval-id>. For preference prompts, use gauge preference get <prompt-id> and check that scenarios exist.
  3. Check gauge billing status for balance and workload eligibility.
  4. If a provider key is required, resolve that connection before trying again. See Configure runs.
Manual cadence does not create recurring runs. Launch the saved configuration with gauge evals run <eval-id>, or set a recurring cadence with gauge evals edit <eval-id> --cadence weekly. Preference prompts additionally need at least one repository/persona scenario. See the quickstart. If a command refers to shared scenarios, presets, or schedules, update the CLI and use the measurement’s own settings. Those old API routes return 403 with an upgrade message.

​
Retry deliberately

Read the original run’s status before launching a replacement, especially after a network timeout. A new launch can create additional billable work. Fix the cause first, then rerun the same task and criteria to make the comparison useful. For persistent failures, contact [email protected] with the organization, run/eval link, status, and visible error. Do not include provider keys or credentials.