Troubleshooting
Start with docker compose ps -a, run details and the relevant service logs.
| Symptom | Check |
|---|---|
| Bot does not answer | Channel membership, mention, include_ids, reconcile_from, token |
| Workflow fails to load | validate, YAML, string revision, overlapping routes |
unknown_profile | Exact profile name, profile activity and account activity in Settings |
validation_error | error.details for the invalid field path |
capacity_exhausted | Active work, pending sandbox cleanup and concurrency limit |
sandbox_provision_timeout | AgentBox availability and worker logs. The initial provisioning deadline expires separately from the run timeout |
sandbox_cleanup_pending / sandbox_provision_conflict | Check sandbox.error, worker logs and resources in AgentBox. An empty search does not prove absence. The execution slot stays occupied until deletion is confirmed. The sandbox state remains unavailable afterwards |
| Model authentication error | Account authorization, API/worker access to its S3 store, and model access |
| File preparation error | Mattermost access from AgentBox, Python and file size |
| Answer arrives without file | Outbox instructions, after_run errors and file limits |
| Web reports no access | ORPHEUS_BROWSER_AUTH, SSO and public origin |
| Updates arrive late | Reverse-proxy SSE buffering and connection state |
| SAML callback returns 401 | Signatures, ACS, time, AuthnStatement and metadata |
| Old model or tools used | Saved session configuration and workflow revision |
Collect diagnostic context
Record the user action, time, session_id, run_id, error code and phase. Check whether a fresh test task reproduces the issue. Remove secrets and unnecessary user data before sharing logs.
Do not start by deleting the database or recreating every session. A publication failure does not mean the agent must repeat its work. First check whether the result already exists in the destination system.