Resolution Flow

The chat request, end to end

The chat stream posts directly from the browser to ai-assist (:8080), bypassing the Next.js proxy. This is an intentional asymmetry: all other data calls proxy through the studio server, which injects server-derived identity; the chat POST carries identity/ACL/project key in the client body.

The lifecycle, step by step

1. Destructure { messages, context, sessionId, provider }. ACL defaults to least privilege — reads true, writes false.
2. Rate limit per userEmail (in-process sliding window).
3. Load or create session in MongoDB (non-fatal) and inject Redis working memory.
4. Two-block prompt — a large static system prompt (cache-friendly) plus a dynamic block (user, role, ACL, page context, working memory).
5. Build tools — commerce, ticket and render-only UI tools, sharing one per-request approval gate.
6. streamText loop, stopping after a bounded number of steps or when an approval card is emitted.
7. onFinish — persist the working-memory state and the user + assistant turns.

The approval gate

Writes are gated with two layers of defense:

• Same-step: emitting an approval card sets the gate to pending; write tools bail out.
• Cross-step: the loop ends after an approval card, so the actual write only runs on the next user turn — after the rep clicks Approve.