v0.5.0 — Streaming Thinking UI & Post-Tool Speed Optimization

What's New

Streaming Thinking Blocks

• LLM reasoning (<think> blocks) now streams live in the chat UI — shown expanded with a purple pulsing timeline dot while generating, then collapses into a peekable "Thought process" entry
• Same collapsible pattern as tool call results — click to expand/peek at reasoning
• Thinking persisted as THINKING message type in the session timeline
• Works in both ReAct (tool selection) and conversational (direct LLM) paths

Post-Tool /nothink Optimization

• After tool results are available, thinking is automatically disabled on subsequent LLM calls
• The LLM has concrete data and only needs to summarize — reasoning is unnecessary overhead
• Saves ~8-10s per interaction on the final answer generation
• Empty <think> blocks suppressed from the UI when thinking is disabled

3-Tier Message Routing

• Workflow detection → ReAct (with/without thinking) → conversational (direct LLM)
• Clear action commands skip thinking for speed; ambiguous inputs keep thinking for accuracy
• Knowledge questions bypass ReAct entirely

Other Improvements

• ReAct agent switched to streaming LLM calls for live thinking display
• Per-message speed stats (prefill/decode tokens per second)
• Verbose loading descriptions removed, cache detection fixed

Technical Details

• New THINKING message type in ChatSession
• disableThinkingAfterTool: true in REACT_CONFIG (enabled by default)
• onThinkingStart/Chunk/End callback chain through ReactAgent → ChatOrchestrator → ChatContainer
• Purple pulsing dot (--thinking) with agentic-pulse CSS animation