OpenRouter-compatible streaming can crash on empty tool-call arguments #1150

Summary

When using OpenRouter-compatible chat providers, the streaming tool-call parser in src/lib/models/providers/openai/openaiLLM.ts can throw on empty or partial function.arguments chunks.

This leaves the UI stuck on Brainstorming... because the research request dies before the client receives a clean end/error state.

Symptoms

• UI stays on Brainstorming... for a long time
• Server logs show an error similar to:

Error:  is empty

• Research never transitions into Research Progress or a proper error state

Root Cause

Some OpenAI-compatible providers stream tool-call arguments incrementally and may emit an empty or incomplete arguments chunk before the final JSON is complete. The current implementation parses each chunk immediately and throws.

Proposed Fix

• Treat empty arguments chunks as {}
• Gracefully ignore parse failures for partial chunks until the JSON stream is complete
• Continue streaming instead of crashing the request

Additional Context

The generic Brainstorming... state gives little feedback when research has not started yet, so a small UI improvement to show elapsed time and likely wait reasons can help debugging.