Intent Replay

A customer withdraws a booking change while an agent's proposed update is waiting to execute. A transcript or successful tool message cannot establish which instruction reached storage. Intent Replay helps QA engineers and agent developers connect the accepted request, model proposal, authorization decision, action journal and independent local booking read, then preserve the investigation as a repeatable regression. The demo uses one real recorded Meta proposal for a fictional appointment on 25 November 2026 in Europe/London, initially 14:00 with 15:00 requested. After an accepted withdrawal, a deliberately faulty worker commits 15:00; a guarded worker suppresses the update and keeps 14:00. A matched no-withdrawal control uses the same proposal; both workers save 15:00. Engineers can inspect and export evidence, reopen saved results, or rerun the frozen input without another model call. A documented Windows command prints expected versus observed results, saves full reports, and returns zero only when required expectations match. An expected execution FAIL can therefore be a passing regression check. The public source runs locally without API keys using the retained captures. Fresh-clone verification passed 57 Python tests, 58 mocked UI checks and both corrected saved regressions; an unexpected injected outcome correctly returned exit 1. The prototype covers the included local SQLite booking backend. The worker defect is controlled fault injection, not a demonstrated model failure. Factual accuracy remains unassessed; execution PASS does not imply overall task success. Arbitrary external-agent compatibility, voice quality and production readiness are not validated. The video is 2:46 with stock synthetic narration and recorded model output; generation itself was not filmed.