Diagnosing a Run

Begin with one explicit run ID. Do not ask Chat to infer the relevant “latest” run when multiple executions could fit.

Inspect the run

1. Open Executions and select the failed run.
2. Copy its run ID.
3. Review the terminal state, first failed block, stage order, row transitions, duration, bounded output, and approval evidence.
4. Ask Chat: Diagnose run <run-id>. Separate observed evidence from your inference.

Chat can request authoritative diagnostic evidence for that run. The bounded result correlates durable history and session events, identifies the failure anchor, and reports unavailable or partial evidence explicitly. Sensitive Script source and raw model prompts are not required for the normal diagnostic summary.

Read the answer correctly

• Run state, block state, durations, row counts, provenance, and persisted approval events are evidence.
• “The delimiter caused this failure” or “this expression is wrong” is MODEL_INFERENCE, even when well grounded.
• Current Canvas or Saved Flow configuration may differ from the historical run snapshot. Chat must identify which it is discussing.

Use Logs only after the run evidence points to an operational area. Logs without exact run linkage are not a substitute for the run record.