scrollIntoView walks up the DOM and scrolls every ancestor, including
the panel with overflow:hidden — this pushed the header and messages
area off-screen after long LLM responses. Using container.scrollTo
limits scrolling to only the messages div.
- Add comprehensive CSS for markdown elements in chat bubbles (headers,
lists, tables, blockquotes, links, code blocks, horizontal rules)
- Integrate KaTeX via marked-katex-extension for inline ($) and display
($$) math rendering with DOMPurify whitelist for MathML/SVG elements
- Add normalizeDisplayMath() pre-processor to handle LLM output where $$
delimiters appear inline rather than on their own lines
- Add dark theme KaTeX styles matching the SpiceBook color palette
- Fix asyncio.to_thread usage in chat SSE endpoint for streaming
- Mount FastMCP at /mcp with tools for notebook CRUD, simulation,
and cell execution. Includes status resource and circuit_assistant
prompt.
- Add SSE streaming chat endpoint at /api/chat/stream backed by
GPU LLM gateway (qwen3). Chat widget sends notebook context
(SPICE cells, markdown notes) so the assistant can reference the
user's circuit.
- React floating chat panel with zustand-persisted conversation
history, streaming token display, reasoning collapse, and
keyboard shortcuts.
- Refactor main.py from deprecated on_event("startup") to lifespan
context manager with combine_lifespans for MCP integration.
- Add notebook_id path traversal validation, decouple get_engine()
from HTTPException for MCP compatibility, fix HTTP client init
race condition with asyncio.Lock.
- Update Caddy labels for /mcp/* routing and SSE streaming on
backend reverse proxy.
