Execution and Validation Flow

End-to-end flow

1. load task definition and create run directory
2. write input.md and run-progress.json (running)
3. execute by kind:
   • script: execute script body directly
   • agent: single-round by default; executor + user-simulator multi-round flow only when review: true
4. validate result (current rule: output must be at least 1 char)
5. write output.md / result.md / run.json / dialogue.* / error.md

Manual vs Scheduled Trigger

• manual trigger: Console UI calls /api/dashboard/tasks/run
• manual trigger is currently async acceptance: the API returns accepted=true, a guidance message, and executionId first, then the run continues in background
• scheduled trigger: task scheduler starts runs from cron or one-shot time rules
• scheduler has same-taskId serialization; manual trigger currently does not have the same anti-reentry protection

Why Multiple Run Records Appear

• every execution creates a new timestamp directory
• Console UI lists runs by scanning these timestamp directories
• so one accepted run request becomes one run record; two accepted requests become two records

Artifacts and Responsibilities

• input.md: execution input snapshot
• run-progress.json: in-progress status and phase updates
• output.md: final task output body
• result.md: human-readable summary
• dialogue.md / dialogue.json: multi-round agent revision trail
• run.json: final metadata and status
• error.md: failure summary when execution fails

Result fields

• status: final status (success/failure)
• executionStatus: execution-stage status
• resultStatus: validation status (valid/invalid/not_checked)
• executionId: unique execution id

Mermaid