harness.Run now writes one ai_runs row per run that reaches the provider, covering the repair turn: skill, version, provider, model, temperature, status (ok, repaired, bad_output, provider_error), latency, summed token counts when reported, error text and the per-turn request and response bodies. Every stored string is redacted first: the decrypted API key, Authorization and API-key header values, Google Chat webhook URLs and URLs with key-like query or fragment parameters or user info. JSON is decoded and its strings redacted, so escapes cannot hide a credential. A row that cannot be written fails the run. Purge deletes rows created before a cutoff (RunRetention is 30 days), clearing suggestion_batches and week_plans references in the same transaction. Migration 0006 rebuilds ai_runs with the new status CHECK and a created_at index; a pre-harness running or error row stops it. Maestro task 53e382d8-1db3-4555-a31d-e44cf55cc98d. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>