Agent Testing troubleshooting

Start with the HTTP status and stable error.code, then inspect the affected test, target, or invocation. Do not automate recovery from raw message text, and redact customer history, session cookies, API keys, and Tool Secrets from diagnostics.

Save or library errors#

validation_failed means the typed definition, folder move, selection, or request field is invalid. Check that Next Reply and Tool Invocation history ends with a user entry, Simulation has a scenario and condition, the definition is under 256 KiB, and a folder is not being moved into its descendant.

conflict_test_revision means someone saved a newer immutable revision. Reload, compare current_revision_id, reapply the intended edit, and submit once. not_found can also mean the resource is private or belongs to another workspace; do not use it to infer that a hidden resource exists.

Invocation does not start or finish#

workspace_not_activated means an Operade operator must activate the workspace and set a positive daily allowance. HTTP quota_exceeded during create or resubmit means preflight stopped before persistence. A worker-level quota stop appears as failure_code: quota_exceeded on a run.

If work stays pending, verify the test worker and Redis queue, then inspect lease age and heartbeat evidence. Do not manually finalize rows. Durable recovery will retry safe mocked work and interrupt work that may have caused a real Tool side effect.

Bucketing and recovery errors#

invocation_not_ready means either at least one run is not terminal or the invocation used only one repeat. Wait for unfinished runs; for a single-repeat invocation, run again with more than one repeat because failure bucketing does not apply. bucketing_not_retryable means there is no failed bucketing attempt eligible for retry, or another worker still owns the bucketing lease. Neither error changes raw run results.

Use resubmit when you need a new invocation. The default preserves the old snapshot; explicitly choose current Main Draft or Live to test new behavior. See Past runs and results, Durable execution and recovery, and the global error index.