Can a prompt replace Gorgias?

Build a help desk for one online store, where the order context is always on screen.

Stack: your choice, Postgres, Docker Compose. Mail in through a signed inbound webhook, mail out through a transactional provider on a domain you own with SPF, DKIM and DMARC. Commerce data through a private app on your own store — Shopify's Admin API is the reference here, and the README should be explicit that this is a custom app for one store, not a public app for the app store.

The ticket view is three columns: conversation in the middle, customer context on the right, actions at the bottom.

Context panel, resolved from the sender's email address: lifetime order count and value, the last five orders with status and fulfilment tracking, and any open returns. Cache it with a short TTL, show the fetch time, and degrade to "context unavailable" rather than blocking the reply when the API is down — an agent must always be able to answer.

Actions, and this is where care is needed. Each is a form with a confirmation step, and every execution writes an immutable audit row with the actor, the payload sent and the response received, rendered in the ticket thread:
- Resend the fulfilment tracking email.
- Refund an order, fully or partially, with a reason. Cap the amount at the order total in the code, not just in the UI.
- Cancel an unfulfilled order.
Refunds move money; treat every one as requiring an explicit second confirmation showing the exact amount.

Macros: canned replies with {{orderNumber}}, {{trackingUrl}}, {{customerFirstName}} interpolated from the resolved context, so the common answer is one keystroke.

A live-chat widget: a small embeddable script, a websocket to the server, and messages that land in the same ticket queue as email. Persist the visitor id in local storage so a returning visitor keeps their thread. Keep the widget under 20KB and make it work without cookies.

Routing: tickets that mention refund, return or chargeback words get a tag automatically; everything else lands unassigned in one queue.

Write tests for the refund amount cap, for the audit row being written before the ticket is updated, and for the context panel degrading gracefully on an API timeout.

Do not build social messaging channels, voice, or an autonomous AI agent.