Troubleshoot tasks and connections
Use execution evidence to distinguish waiting, running, interrupted and review states.
The task has not started
Confirm model or Agent availability, assignment, budget and grants. Assignment does not mean an executor has claimed the task; look for an actual run record.
Heartbeats without progress
A heartbeat confirms a live connection, not productive progress. Check stage changes, work notes and files. Record the cause of a stall or interruption before an owner decides to retry.
A file cannot be read
Check the granted directory, file existence, current account access and OS permissions. Isolate the issue with a simple test file instead of broadly expanding access.
The result is incorrect
Check inputs, revisions, date ranges and acceptance criteria. Link the issue to evidence and specific changes, then create a new iteration. Keep human review for consequential conclusions.
Prepare a useful issue report
Include time and timezone, installed version, company or project identifier, task and run identifiers, reproduction steps, expected and observed behavior, and steps already tried. Redact screenshots and logs; exclude credentials, personal records and full business datasets from public issues.
Time / timezone:
Client and service versions:
Task / run / request ID:
Steps to reproduce:
Expected / observed:
Affected users and operations:
Redacted evidence:
Workaround already tried:When to pause and resume
For duplicate writes, unexpected sending, unauthorized reads or unusual spend, pause the affected executor or connection, retain records and notify the owner. Before resuming, verify access, unresolved operation results and one small sample. Avoid bulk replay of all failures. FDE response times follow the service agreement.