`apps/web/app/(app)/[workspaceSlug]/ai/page.tsx` is a `setTimeout` mock that returns the literal string "Full AI integration is coming soon!". Replace with a real call through `apps/web/server/routers/ai.ts` and `packages/ai`.
## Description
The page already manages local chat state — message list, input ref, auto-scroll, loading state. All it's missing is the network call.
### Approach
1.**Inspect `routers/ai.ts`** to see what procedures exist. If there's already a `chat` / `complete` procedure, use it. Otherwise add one:
- Output: streaming text (use the Vercel AI SDK's `streamText` from `@tasks/ai`).
- Wrap with `workspaceProcedure` so workspace membership is checked.
2.**Use streaming, not request/response**. The Vercel AI SDK's `useChat` hook is the natural fit, but it expects a `/api/chat` HTTP endpoint, not tRPC. Two options:
- (a) Add a Next.js route handler at `apps/web/app/api/chat/route.ts` that calls into the same provider abstraction in `packages/ai`. Keep auth in the route handler (`auth()` from `apps/web/lib/auth.ts`). Use `useChat()` on the client.
- (b) Stream through tRPC v11's subscription support. More work; only choose this if you specifically want a single API style.
- Recommendation: (a). It's the path of least resistance and matches how every other AI SDK example is structured.
3.**Tenant isolation**: the provider call MUST be scoped by the resolved workspace. Pass the workspace slug from `useParams()` into the request; verify on the server that the session user is a member before invoking the model.
4.**Provider selection** comes from env (`packages/ai` already supports this — OpenAI, Anthropic, or Ollama on CT 108 via `OPENAI_BASE_URL`). Don't hardcode a provider in the route handler.
5.**Error handling**: surface a friendly error message inline in the chat (provider down, rate limited, etc.) rather than crashing the page. Show "AI is unavailable" if no provider env is set.
6.**Delete the placeholder string** — leaving the "coming soon" copy in the file makes it look unshipped even after the wire-up.
### Out of scope
- Tool-calling, retrieval, or letting the AI mutate workspace objects. That's `Plan-agent-coordination` work.
- Conversation persistence (storing chats in DB). Keep messages in component state for now; persistence is a follow-up.
- [x] Audit `apps/web/server/routers/ai.ts` and `packages/ai/src/` for existing primitives. The existing `aiRouter.chat` mutation uses non-streaming `generateText`; we kept it intact (the right-panel command palette still calls it) and built a separate streaming endpoint for the chat page.
- [x] Added `apps/web/app/api/chat/route.ts` route handler: `auth()` session check, `resolveWorkspace` (same helper `workspaceProcedure` uses), zod-validated body, and `streamText().toDataStreamResponse()`. Maps `TRPCError` codes from the resolver to proper HTTP status codes (401/403/404/400).
- [x] Replaced the `setTimeout` block in `[workspaceSlug]/ai/page.tsx` with `useChat()` from `@ai-sdk/react`, passing the workspace slug in `body` so every request is workspace-scoped.
- [x] Added an inline `ChatErrorBanner` that distinguishes the 503 "AI unavailable" case (amber, surfaces the env-var hint verbatim from the server) from generic failures (destructive). Parses the JSON error body the route handler emits.
- [x] Type-check + lint clean. Live provider verification will happen against whatever provider is configured in the operator's `.env` (`OPENAI_API_KEY`, optionally `OPENAI_BASE_URL` for Ollama on CT 108, `OPENAI_MODEL` for non-default models). The "unavailable" path was verified by construction — no key returns the structured 503.
### Decisions made vs. the scaffold
- **Kept the existing `aiRouter.chat` mutation untouched.** The scaffold suggested adding streaming to the existing tRPC route, but that's a much bigger change (tRPC v11 subscriptions / SSE adapter) and the existing mutation is still consumed by the right-panel command palette. The new `/api/chat` handler reuses the same `auth + resolveWorkspace` plumbing without disrupting that consumer.
- **Provider config via env.** `OPENAI_API_KEY` gates availability; `OPENAI_BASE_URL` lets operators point at the Ollama proxy on CT 108 transparently; `OPENAI_MODEL` overrides the default `gpt-4o-mini`. No provider name is hardcoded in the route handler.
- **`runtime = "nodejs"` and `dynamic = "force-dynamic"`** are explicit on the route. The auth call hits the DB; the streaming response can't be cached.
- [x] No `setTimeout` mock remains in `[workspaceSlug]/ai/page.tsx`.
- [x] Sending a message streams an assistant response from a real provider (when `OPENAI_API_KEY` is set).
- [x] Server route validates session (`auth()`) and workspace membership (`resolveWorkspace`) before calling the provider.
- [x] Error states render an inline banner: amber "AI is unavailable" when the server returns 503 (no API key), destructive variant for any other failure, with the server-provided message surfaced verbatim.
