Streaming and Partial-Result UX

PROJECT CONTEXT:
My agent goes silent for 30-90 seconds and users assume it is broken. Terminal agents like opencode won users by streaming tokens, showing tool activity live, and making long work feel alive and interruptible.

GOAL:
Add end-to-end streaming: token streaming from the model, live tool-activity events, and graceful partial results on interruption.

REQUIREMENTS:
- Token-level streaming of model text to the client (SSE or websocket)
- Tool lifecycle events streamed: started, result summary, failed
- A structured event protocol with types, not raw text splicing
- Interruption support: cancel mid-run and receive a coherent partial result
- Graceful rendering of interleaved text and tool activity

CONSTRAINTS:
- The event protocol must be versioned and typed - clients should not parse prose
- Interruption must leave state consistent (no half-applied edits)

DEFINITION OF DONE:
- First token reaches the client in under 2 seconds on a normal run
- Tool calls appear as live activity lines with running/success/failure states
- Cancel mid-generation yields a partial transcript plus a clean run state
- A slow 60-second task shows continuous signs of life throughout

COMMON FAILURES TO AVOID:
- Buffering the entire response and calling a spinner 'streaming'
- Interleaving raw tool JSON into the text stream and breaking the client
- Cancellation that kills the process but leaves files half-edited
- No heartbeat during long tool runs so users still think it froze