An assistant can sound confident even when it cannot prove whether a payment happened or whether a launch path was checked. Release Steward turns those questions into an inspectable workflow.

The assistant collects a synthetic shop name, stable order reference, provider-outcome fixture and optional mock evidence through a two-step intake form. The web simulator streams real local MCP tool events as it inspects replay safety and supplied launch evidence, saves a review decision with a stable operation ID, and reads the receipt back. Repeating the same review reuses the receipt. Changing the contract under that same operation ID is rejected.

In the unknown-payment fixture, the assistant stops for outcome reconciliation instead of treating a timeout as permission to retry. In the launch-review fixture, it says mock evidence is declared, not independently verified. No real payment or external action happens.

This is a simulated Alexa+ experience, not a connection to Alexa. Intent routing is scripted, not LLM-based. The MCP calls, server-side evaluation, receipt reuse and read-back are working code. This runs locally with single-process receipt storage. It is not a production integration.

What's new vs existing

The Replay Safety Kit and Launch Evidence engines are pre-existing components. New October 13 work: MCP wrapper, 3 typed tools, client orchestration, persistent local review receipts, simulator UI, two-step intake, streamed tool events and real end-to-end MCP tests. The demo shows the new features; reused engines are disclosed.

Built with

Node.js, Express, Zod and the open-source MCP TypeScript SDK. No AWS services, credits or paid inference used. No AWS mini-challenge claim.

Tests

Real MCP discovery/invocation, unknown-result stop, missing-evidence report, receipt reuse/read-back, conflicting-contract rejection and invalid scenario rejection pass. Intake validation, streamed tool sequence and receipt reuse are tested. Phone UI and the continuous form-filling demo are visually checked.

Built With

Share this project:

Updates

Submission history