Failed or stuck runs

Ask:

Investigate run RUN_ID. Show its prompt and lifecycle diagnostics, and link to its browser trace for output. Identify the failure and whether it changed anything externally. Do not create a duplicate run; recommend a safe recovery.

Call get_run({"runId":"RUN_ID"}) then get_run_diagnostics({"runId":"RUN_ID"}). Expected evidence includes state, execution, trace, artifact, harnessLog and activity. Check launch, prompt delivery, output capture, claim release and cleanup rather than treating a missing trace as proof the agent never ran.

Evidence	Recovery
Queued	Inspect active runs/concurrency with `get_job`; wait for a slot.
Launch or permission failure	Browser-test the execution connection; check repository/model access and tags. Use `list_exe_connections` then `diagnose_exe_integration({"connectionId":"CONNECTION_ID"})` if appropriate. This probe uses a disposable VM.
Wrong/empty prompt values	Read actual trigger slug/reflection and persisted context; safely replace the full job configuration.
Failed with useful output	Fix the reported cause, inspect the browser trace and check external changes, then request a new manual run with a new idempotency key.
Starting/running/blocked but stuck	Request `stop_run`, poll, and inspect cleanup. `kill_run` currently has identical semantics. Escalate persistently stuck runs to the hosted operator with run ID and safe diagnostics.
Missing/partial trace	Inspect output and artifact state; ask the owner/operator about REST-only retained-artifact replay. See advanced operations.

Worked recovery configuration: use the original job ID and

invoke_job({"jobId":"JOB_ID","prompt":"Retry after correcting repository access","idempotencyKey":"investigation-RUN_ID-retry-1"})

only after fixing the cause and confirming the old run is terminal. Expected result is a new run ID. Verify its rendered prompt/state with get_run and successful output in the browser trace; a diagnostic probe passing does not itself prove the job succeeds.

Get started

Work with your agent

Automate jobs

Troubleshooting

Reference

Self-hosting

Failed or stuck runs