feat: Langfuse observability, chat forking, Prompt Lab, and split-screen comparison

[TerminallyLazy]

Summary

• Langfuse observability integration — traces every agent interaction (LLM generations, tool calls, iterations) with a built-in trace viewer for drill-down analysis
• Chat forking — fork any conversation from any message point, creating an independent branch with full history up to that point
• Prompt Lab — analyze and optimize prompts using LLM-powered refiner/judge agents, accessible from trace viewer generation observations
• Split-screen comparison — side-by-side view of original vs forked chat with shared messages dimmed and divergent messages highlighted

Details

Langfuse Observability

• langfuse_helper.py manages client lifecycle (init/flush/shutdown) with lazy initialization
• 10 agent lifecycle extensions cover the full trace: init, trace start, iteration, generation start/end, tool span start/end, utility calls, flush
• Observability settings tab with connection test endpoint
• Trace viewer modal with span/generation drill-down, token counts, cost, and timing

Chat Forking

• fork_context() deep-copies via serialize/JSON-round-trip/deserialize with optional truncation at any log position
• Fork buttons on message action bars and sidebar, fork indicator on forked chats
• Auto-naming with collision check ("Chat (fork)", "Chat (fork 2)", etc.)

Prompt Lab

• /prompt_refine generates 1-3 improved prompt variants via utility model
• /prompt_judge scores variants on clarity, efficiency, fidelity, and safety (1-5 scale, rejection threshold < 3)
• Three-panel modal: original prompt (read-only) | editor with suggest/test | results with scoring
• Entry point from trace viewer on GENERATION observations

Split-Screen Comparison

• /chat_logs read-only endpoint fetches logs from any context without switching active context
• Fork point auto-detection from fork_info metadata with content-matching fallback
• Shared messages rendered at 45% opacity, divergent messages highlighted with accent border
• Compare button in sidebar for forked chats, also accessible from Prompt Lab

Files changed

40 files, +5,093 lines (26 new files, 14 modified)