feat: render streaming reasoning content

[Shinyaigeek]

Summary

@ag-ui/core fully defines the ReasoningMessage type and the REASONING_MESSAGE_START / CONTENT / END event sequence, but neither @openuidev/react-headless nor @openuidev/react-ui consumed them, so reasoning/thinking output was silently dropped end-to-end (#414).

This wires reasoning through, colocated on the assistant turn (a reasoning field on AssistantMessage) rather than as a separate message — matching how the streaming layer already folds text and tool calls into a single assistant message, and how the format converters already treat reasoning.

What changed

react-headless

• types/message.ts — extend AssistantMessage with an optional reasoning field; swap the Message union's assistant branch so role === "assistant" narrowing exposes it.
• stream/processStreamedMessage.ts — handle REASONING_MESSAGE_CONTENT / CHUNK, accumulating deltas into currentMessage.reasoning. The reasoning-only messageId is intentionally ignored; a later TEXT_MESSAGE_START swaps in the real assistant id and the spread preserves the accumulated reasoning.
• types/messageFormat.ts — the default identityMessageFormat now strips the client-only reasoning field in toApi, so accumulated thinking is not echoed back to the backend on every subsequent turn.

react-ui

• New Reasoning/ component dir: ReasoningContent (markdown) and a shared ReasoningSection (the collapsible BehindTheScenes panel).
• Applied to all assistant render surfaces — Shell, CopilotShell, BottomTray (via ReasoningSection) and OpenUIChat (nested in its existing merged tool-call panel).
• Exported from the package root.

Design notes

• Colocation over a separate message — the live stream builds one assistant message per run; reasoning as a field fits that grain. See the discussion in the issue.
• Reasoning streams expanded ("Working…") and collapses once the answer text arrives (toolCallsComplete={!!message.content}).

Out of scope (follow-ups)

• Restoring reasoning from server history (loadThread / the format converters) — currently the live-stream path only.
• Interleaved reasoning↔answer ordering (single reasoning string concatenates blocks).
• Markdown re-parse cost for very long reasoning during streaming.

Testing

• processStreamedMessage: reasoning accumulation, preservation across the TEXT_MESSAGE_START id swap, and absence (no reasoning events).
• identityMessageFormat: strips reasoning outbound, leaves other messages untouched.
• pnpm test green (react-headless), typecheck + build green (both packages).

Closes #414

🤖 Generated with Claude Code