Emit early progress frames for streaming tool calls

[adambalogh]

Summary

Add support for emitting lightweight progress frames during streaming tool calls on buffered providers (OpenAI/Anthropic). This allows clients to display tool activity immediately (e.g., "writing file…") while the complete, argument-bearing tool call is still accumulated and flushed at the end of the stream for signing.

Problem

On buffered providers like OpenAI and Anthropic, tool calls are held until the stream ends so the signature can cover the complete arguments. This means clients receive no "tool starting" signal until generation finishes — a large write_file call can appear frozen for minutes while the file is being generated. Additionally, the lack of streaming bytes can cause read timeouts.

Solution

• Add _MockStreamChunk test helper to simulate LangChain AIMessageChunk objects yielded by model.stream(), with support for tool_call_chunks and usage_metadata
• Emit progress frames in chat_controller.py when tool call chunks arrive with a name but incomplete arguments:
  • Progress frame carries only the tool index, id, and name — never the accumulating arguments
  • Keeps bytes flowing to the client and prevents timeout
  • The authoritative, complete tool call (with full arguments) is still flushed once after the stream loop, ahead of the signed final frame
• Add comprehensive test (test_streaming_tool_call_emits_early_progress_frame) that:
  • Simulates realistic streaming chunks (name in first fragment, arguments trickling in)
  • Verifies an early progress frame announces the tool name without arguments
  • Confirms the complete arguments arrive in a later frame
  • Ensures progress frames arrive before the signed final frame

Implementation Details

The progress frame logic is gated by checking whether the tool call chunk has a name field but the accumulated tool call is incomplete. This ensures:

• Early frames are emitted only when the tool name becomes known
• Arguments are never leaked in progress frames (only in the final signed frame)
• The mechanism is transparent to non-buffered providers (e.g., Gemini) that already stream complete tool calls

https://claude.ai/code/session_01DKT5v2Ep7RU5Law5xh2L4i