Skip to main content

Debug using artifacts

Stub

This How-to is a stub. The artifact model exists (candidate sets, decisions, evaluations), but a fully polished “artifact debugger” UX is planned.

Goal​

You will debug a run using the canonical artifact chain: CandidateSet → SelectionDecision / binding decision → EvaluationResult → RecoveryAction.

When to use this​

  • A run failed and you need to understand why.
  • You want to audit what decisions the system made.

Prerequisites​

  • Durable artifacts are enabled and stored
  • You can query artifacts by run_id / node_run_id

Steps​

  1. Start from the failed node_run_id (or the root).
  2. Retrieve the candidate set(s) that were generated.
  3. Retrieve the binding decision(s) and the chosen NodeTypeRef.
  4. Retrieve the execution output artifacts.
  5. Retrieve evaluation and recovery artifacts (if present).

Verify​

  • You can narrate a run as a sequence of durable decisions and outputs.
  • The evidence chain is complete (no “it just happened” gaps).

Troubleshooting​

  • Missing links between artifacts → ensure stable IDs (subtask_id, candidate_set_id) are carried through.
  • Artifacts too verbose → summarize for UI, keep full data in storage.
  • Sensitive data in artifacts → enforce redaction and re-run.

Cleanup / Rollback​

  • None.

Next steps​