feat(ai): SSE chat streaming + exhaustive today-listing prompt rules

[thomasluizon]

Refs thomasluizon/orbit-ui-mobile#159. Paired frontend PR: thomasluizon/orbit-ui-mobile#160 (merge THIS one first — the SSE endpoint must be live before the clients consume it).

SSE chat streaming

New POST /api/chat/stream next to the unchanged buffered POST /api/chat (kept — it's public surface for API-key/agent consumers). Event protocol (data: frames, camelCase, enums as strings, nulls omitted):

event	payload	meaning
started	—	headers flushed, feeds client watchdog
round	iteration	tool round running (keepalive during silent tool work)
delta	text	answer fragment
reset	—	round emitted text but ended in tool calls — client clears the draft
final	full ChatResponse	authoritative result (actions, correlationId, relatedSurfaces)
error	status, error, code?	same shapes the buffered endpoint returns (PAY_GATE→403+code, Result failure→400+code, validation→400, unhandled→500)

• AiIntentService switches to CompleteChatStreamingAsync when a sink is provided: content deltas flush immediately (First content token after Nms logged for the ~2s TTFT acceptance check), tool-call deltas accumulate via an index-keyed builder (args joined across chunks, empty args normalized to {}), and the AssistantChatMessage is rebuilt so the tool loop's conversation context is identical to the buffered path. Null sink → the exact pre-existing code path.
• A round isn't known to be final until it ends, so every round streams; tool rounds simply emit no content. The rare content-before-tool-calls case emits reset.
• Pre-stream failures (validation, auth, 429 rate limit) stay plain HTTP status codes — SSE only starts after validation. Paygate check, quota counting (background increment), correlationId, persistence: all untouched.
• 15s per-attempt NetworkTimeout + retry policy from Harden calendar auto-sync event dedupe #156 still apply; for unbuffered responses the timeout acts per-read (server-side idle guard), and retries only cover the pre-headers leg — mid-stream drops surface as an error event.

Today-listing completeness (user-reported regression, prompt-only)

"liste meus hábitos de hoje" was dropping due-today habits (entire Before Bed group skipped despite being fully labeled [TODAY] in the index). The index data was already correct — done-today habits drop out naturally via DueDate advancement (desired), and remaining habits carry TODAY/OVERDUE. The failure was pure model listing decay, encouraged by rule 10's "keep responses concise".

Fix: the index instruction now demands enumerating EVERY labeled entry and verifying the list against the heading's exact counts (N due today, M overdue — computed with the same predicates as the labels, so they're a ready-made checksum); rule 12 forbids omitting/sampling/summarizing listing answers; rule 10 scopes conciseness to prose, never lists. No data changes, no extra LLM round — the single-round speed from #156 is preserved, and done-today habits stay excluded.

Tests (+16; suite now 3,143)

• AiIntentServiceStreamingTests — real ChatClient pipeline over scripted OpenAI SSE bytes: text-round deltas, tool-args split across chunks, content-then-tools reset, mid-stream drop → failure, buffered path untouched.
• ChatStreamEventTests — serialization contract locks (camelCase, string enums, null omission) matching the shared Zod union.
• Handler: round per iteration + Domain→Application sink bridging; null-sink passthrough.
• Prompt sections: exhaustiveness lines locked (gated on UserToday so the null-today contract holds); GlobalRules additions locked.
• AgentCatalog guard test forced the ProcessChatStream capability mapping (included).

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate failed

Failed conditions
1 Security Hotspot
65.0% Coverage on New Code (required ≥ 80%)

See analysis details on SonarQube Cloud

[claude]

SSE streaming lands cleanly. The implementation is solid across all the areas that matter:

Authorization & rate-limiting are both class-level attributes on ChatController, so the new /stream endpoint inherits them without any extra decoration needed. Pre-stream validation (message length, image, history, clientContext) correctly returns plain HTTP errors before headers are committed; post-commit failures route through SSE error events — exactly the right split.

Streaming pipeline in AiIntentService is well-structured: buffered and streaming paths share the CompletedRound abstraction, tool-call args are accumulated correctly across chunks via StreamingToolCallBuilder, and the content-before-tool-calls reset case is handled and tested. The AssistantChatMessage reconstruction ensures the tool-loop conversation context is identical on both paths.

Tests are thorough: real ChatClient pipeline over scripted SSE bytes covers text deltas, split tool-args, content-then-reset, and mid-stream drop; serialization contract tests lock the camelCase/string-enum/null-omit wire format the shared Zod schema depends on; handler tests verify round keepalives and Domain→Application sink bridging.

Prompt fix for today-listing is minimal and targeted: the exhaustive-enumeration instruction is gated on UserToday.HasValue (correct null contract), and rule 10/12 changes are tested.