Agent Production Case Normalizer

Scenario

Use when a team wants to turn messy production incidents, user feedback, or agent traces into reusable evaluation cases.

Input and output

Input is a redacted array of trace records with case_id, user_input, observed_output, constraints, and optional expected_evidence. Output is a normalized case list, a review queue for missing fields, and a stable case_set_hash.

Execution and recovery

Run the README command. Malformed records return a visible error. Records with missing constraints or evidence are kept in review_queue rather than silently becoming passing tests.

Privacy boundary and acceptance

Remove private text, customer identifiers, credentials, cookies, local paths, and raw documents. The sample produces one normalized case and a non-empty hash. This is case preparation only; it does not call a model or agent.