get_run, then use stop_run({"runId":"RUN_ID"}) for a starting, running or blocked run. The response requests stopping; poll until terminal and inspect diagnostics for cleanup.
Current behavior: kill_run calls the same stop operation as stop_run. It does not force a queued run to terminate or release its slot immediately. Both return operation_failed when the run is not active. Do not treat kill as stronger recovery; contact the hosted operator if a run remains stuck.
For a timed-out manual submission, repeat invoke_job with the same idempotencyKey. The response contains the original invocationId, runId, and duplicate: true. Retain the key you submitted: current get_run does not return invocation.idempotency_key. Never send its internal invocation.claim_key or a manual: prefix back as the key. If the key was lost, inspect existing runs before requesting another.
An intentional new run after failure needs a new key. First verify that the prior run is terminal, inspect whether it made external changes, and fix the cause. A retry may repeat those external effects: invocation deduplication is not transactional rollback of the agent’s work. For queue_full, wait for capacity and reuse the key for the intended request. See failed runs.