feat(dash): real chat surface against primary slot

[thinmintdev]

What was dummy / what was wired (/diagnose summary)

The v3 dashboard's main chat surface (/, root route) showed a fully scripted demo conversation:

• ChatActive in ui/src/dash/dashboard.jsx hardcoded ~250 lines of JSX — a fake "refactor my slot manager" exchange, a fake read_file tool block, a fake generate_image tool block, and a fake image attachment row.
• The composer's onSend was () => setDraft(\"\") — typing and clicking send just cleared the input. Nothing was POSTed.
• The Tweaks-panel composerState (idle / sending / streaming / swap / offline / no-tools) was real and useful, but only ever driven by the Tweaks selector — never by an actual request.

What was already wired and reused as-is:

• useSlots() (5s poll) gives the persona's live model_id.
• POST /v1/chat/completions is mounted in src/hal0/api/routes/v1.py (curated surface, not the proxy catch-all) and accepts standard OpenAI {model, messages, stream}.
• The .composer-banner state machine + .msg / .bubble CSS classes in dashboard.css.

The fix

• New ui/src/api/hooks/useChatCompletions.ts — streamChatCompletion and chatCompletion. Streaming uses fetch + a manual ReadableStream reader (not EventSource — POST not supported) to parse data: {…}\\n\\n frames and the terminal data: [DONE]. Handles Qwen3.5's reasoning_content split: surfaces both content and reasoning_content buffers via onDelta, with content preferred for the final answer (falls back to reasoning_content for very short prompts where the model never emits real content).
• New ui/src/dash/chat.jsx — extracted Composer, ChatActive, ChatEmpty, PersonaPicker from dashboard.jsx. Rewrote ChatActive around a real useChat hook that owns message state, single-flight send, and an AbortController so the streaming Stop button actually works. Errors render as a red .msg row instead of being swallowed. Enter sends; Shift+Enter inserts newline; empty drafts disable the send button. The persona slot's model_id || model is what we send as the OpenAI model field.
• dashboard.jsx shrinks by ~300 lines — now owns only SnapshotStrip / MemoryMap / HealthCard / ThroughputCard / DashboardView.
• main.tsx adds the ./dash/chat.jsx side-effect import after dashboard.jsx so the window-globals are installed before <App /> mounts.

Why a new file rather than editing in place

Sibling PRs fix/sidebar-live-counts and fix/dashboard-memmap-throughput-live are also touching dashboard.jsx. The throughput agent's work-in-progress on the LXC was already adding useHardware() to MemoryMap (verified by stashing + diffing). Extracting the chat surface to its own file eliminates the per-hunk-git-add dance that the memory note feedback_multi_agent_one_file warns about. dashboard.jsx here only loses the chat-surface code; the sibling's MemoryMap / ThroughputCard hunks land cleanly on top with no overlap.

Endpoints called

When	Method	Path	Notes
Stream send	POST	/v1/chat/completions	Accept: text/event-stream, stream: true. Body: {model: <persona.model_id>, messages: [...], stream: true}

No backend changes.

Verification commands actually run

On hal0-dev (this worktree):

cd /tmp/hal0-chat-live/ui
npm run typecheck       # clean
rm -rf node_modules/.vite dist && npm run build    # 114 modules, 803ms
npx playwright test dashboard-v3 --reporter=line   # 2/2 pass (existing smoke)

On hal0 LXC (/opt/hal0), checked out this branch, rm -rf node_modules/.vite dist && npm run build, systemctl restart hal0-api. Then:

# Pre-load primary so the first chat doesn't wait for a cold load
ssh hal0 'curl -sX POST http://127.0.0.1:13305/v1/load \\
  -H content-type:application/json \\
  -d {\"model_name\":\"Qwen3.5-0.8B-GGUF\"}'
# → {\"status\":\"success\",\"model_name\":\"Qwen3.5-0.8B-GGUF\",...}

# Live Playwright spec: opens http://10.0.1.142:8080/, types
# \"Reply with exactly: hello world\", clicks send, asserts on a real
# assistant bubble + no dummy strings + send-disabled-when-empty.
HAL0_E2E_LIVE=1 HAL0_E2E_BASE_URL=http://10.0.1.142:8080 \\
  npx playwright test _chat-live-verify --reporter=line
# → 3 passed (24.2s)

The real assistant bubble captured by the live spec:

=== Final assistant bubble ===
\"hello world\"
=== /Final ===

i.e. the prompt "Reply with exactly: hello world" round-tripped through the dashboard → /v1/chat/completions → Lemonade → Qwen3.5-0.8B-GGUF and the model replied with exactly hello world as instructed. The throwaway _chat-live-verify.spec.ts was removed before commit — it depends on the LXC being reachable with a live primary slot and isn't gated for CI.

After the LXC verification, the LXC working tree was restored to its pre-test state (sibling-throughput WIP on dashboard.jsx, sibling-lemonade WIP on idle.py) and rebuilt.

Deferred (out of scope per brief)

• Streaming reasoning split (think/answer surface — there's a separate scaffold).
• Tool-calling / function-calling visualisation (the prototype's toolblock UI is gone for now).
• Multimodal attachments / voice input (the icons stay as "coming soon" toasts).
• Persistence across page reload (no localStorage; conversation lives only in the page session).
• Model picker in the composer (persona pick still routes via the existing PersonaPicker).
• Backend changes — the proxy works, no new routes needed.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com