Make assistant support flows survive real customers and real developers

[codeyogi911]

Why

A live customer conversation on the browser assistant surfaced four compounding failures around one delayed-delivery question:

1. "My delivery is delayed." never reached the deterministic order intake — the trigger required the literal word "order", so the model improvised, gluing its own question onto the scripted contact-card line.
2. A dropped WebSocket erased the flow — contact details, the pending order lookup, and the conversation history all lived on the connection, and mobile browsers drop sockets constantly. After reconnecting, the customer's order number hit the off-topic guardrail dead end.
3. A bare order-number reply was not recognized outside the scripted continuation sentinel.
4. Store order-name formats containing "/" (e.g. fiscal-year formats such as #2026-27/7903) could not be parsed at all, so lookups were impossible for every order in such a store.

There was also no way to test any of this before production, because Turnstile's strict claim verification rejects Cloudflare's documented testing keys — and the eval:voice gate turned out to be unrunnable on a fresh clone (workspace-local wrangler path that npm hoists away).

What

• Durable session state: contact, verification, pending escalation, and pending order flow move from per-connection state into Durable Object session storage. A dropped socket keeps history and state for a 10-minute grace period and resumes mid-flow on reconnect; a scheduled cleanup wipes everything if nobody returns.
• Broader deterministic intake: shipment nouns (delivery, package, parcel, shipment) near tracking signals trigger the scripted order flow, not just "order".
• Bare order numbers always enter the order flow: with contact on file they look up directly; without, they request the contact card once. "/" is accepted in order numbers end to end.
• Local Turnstile test mode: on the local surface only, a successful siteverify under Cloudflare's documented testing secrets is accepted without claim checks. Non-local surfaces keep strict verification, so a production misconfiguration still fails closed.
• Working eval harness + prompt hardening: eval:voice resolves wrangler through the package; order-result guidance is strengthened (ticket offer is now non-optional). Evals pass 12/13; the remaining case (an unfamiliar-product how-to that must search before answering) is model-flaky and pre-existing — a deterministic forced-search fix was tried and reverted because it conflicts with the no-tools-on-off-topic bar. Tracked as follow-up.
• Developer-machine hygiene: the vitest pool pins bindings a local .dev.vars would leak into tests, and the public-readiness scan judges what git can publish (git ls-files) instead of failing on properly ignored local state.

Validation

• npm run check green: 219 tests (including new reconnect-resume, bare-number, trigger-phrase, and testing-secret coverage), all typechecks, generated-file check, and the public scan — with a .dev.vars present.
• npm run eval:voice: 12/13 with the known-flaky case above.
• Live WebSocket replay of the original customer scenario against wrangler dev: deterministic card on the delivery phrasing, contact survives a simulated connection drop, and the bare order number after reconnect performs a real lookup with a truthful provider-failure reply instead of the off-topic refusal.

🤖 Generated with Claude Code