feat(chat): stream AI replies over SSE (#159)

[thomasluizon]

Closes #159. Paired backend PR: thomasluizon/orbit-api#199 (merge that one first — the SSE endpoint must be live before these clients consume it).

Also carries the fix for the reported "liste meus hábitos de hoje is missing habits" bug — that fix is backend prompt rules in the paired PR (no client change needed); details there.

What streams now

Send a message → tokens render into the AI bubble as they arrive, on web and mobile identically. Typing dots show until the first token, then the bubble grows; tool-calling turns keep the dots during tool work (the backend emits round keepalives) and stream only the final answer.

• shared owns the protocol: chatStreamEventSchema (mirrors the backend record — contract-aligner verified 0 drift), an incremental SSE parser (createChatSseParser) + stream consumer (consumeChatSseStream), new chat-store actions appendToMessageContent/updateMessage, and CHAT_STREAM_IDLE_TIMEOUT_MS = 60s replacing the whole-response CHAT_SEND_TIMEOUT_MS.
• web: new dedicated BFF route app/api/chat/stream/route.ts that pipes the upstream body straight through (the catch-all proxy buffers, which would defeat streaming) with the same cookie-auth + single 401-refresh as the catch-all. The composer fetches it directly; the sendChatMessage Server Action is deleted (Server Actions can't stream).
• mobile: new lib/chat-stream.ts on expo/fetch (RN's built-in fetch has no readable response body) with SecureStore bearer + single refresh rotation. TextDecoder comes from Expo's winter runtime — no polyfill needed.

Failure semantics (the #156 affordances survive)

• The 60s bound is now an idle watchdog — reset on every received chunk, with round events covering long tool work. Firing aborts → 408 → the existing timeout copy + Retry button.
• SSE error events and pre-stream HTTP failures feed the same classifySendFailure flow: timeout/generic arm Retry, limit/upgrade don't, premium denials still route to /upgrade.
• The final event ALWAYS replaces the streamed draft content (covers the backend's StripJsonWrapper edge) and attaches actions/correlationId/relatedSurfaces; a reset clears the draft when a round turns out to be a tool round.
• Quota counting, optimistic aiMessagesUsed bump, query invalidations: unchanged.

Tests

• shared: parser (fragmentation, CRLF-split frames, comments, malformed-JSON resilience, forward-compatible unknown events) + consumer (terminal outcomes, incomplete-stream) + store actions. Suite: 875.
• web composer suite rewritten against mocked fetch + real ReadableStream SSE bytes: incremental render, reset-no-duplicate, idle-watchdog timeout + retry, limit no-retry, retry-no-duplicate-user-bubble, pre-stream classification. Suite: 1,460.
• mobile composer suite mirrors web on the openChatStream seam + offline gate + quota + pending-ops. Suite: 415.
• parity-checker: all PAIRED. Lint + type-check clean across workspaces.

Manual verification checklist (auth-walled, needs your eyes)

1. Web + mobile: ask a read-only question → text streams in; Render logs show First content token after Nms with N ≤ ~2000 after the final round.
2. "liste meus hábitos de hoje" → EVERY due/overdue habit listed (Before Bed group included), habits already done today absent; single AI API responded round in Render logs.
3. Tool turn ("create a habit to stretch daily") → answer streams after the tool work, action chips render, habit exists.
4. Kill connectivity mid-stream on mobile → banner + Retry within ≤60s; Retry resends without duplicating your bubble.
5. Free account at the message limit → limit copy, no Retry button; premium-gated reply still routes to /upgrade.
6. Support request via chat still carries [trace: …] in the email.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
orbit-ui-mobile-web	Ignored		Jun 10, 2026 6:22pm

[claude]

Solid end-to-end streaming implementation. The shared SSE parser and consumer are well-factored: CRLF normalization happens at concatenation time (so cross-chunk splits are handled correctly), unknown event types are silently dropped for forward compat, and the incomplete outcome cleanly covers the half-open case. Both platform hooks are structurally identical — same idle watchdog, same 401 refresh path, same final-always-wins semantics — which satisfies the cross-platform parity rule. The streaming BFF route correctly pipes the upstream body without buffering, mirrors the catch-all's auth pattern, and the x-accel-buffering: no header is the right touch for Nginx deployments. CHAT_SEND_TIMEOUT_MS and sendChatMessage are fully removed with zero remaining callers. Test coverage is thorough: fragmented frames, CRLF splits, comments/unknown events, watchdog fire, limit-no-retry, retry-no-duplicate-user-bubble, and pre-stream HTTP classification are all exercised.

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
81.2% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud