To ensure we are using correct events, using the realtime types from the
openai sdk. Added conversion tests as well.




---
[//]: # (BEGIN SAPLING FOOTER)
* #1135
* __->__ #1134

The UI demo is nice, but this let you print more logs to the console (at
the cost of being ugly)

---
[//]: # (BEGIN SAPLING FOOTER)
* #1141
* __->__ #1139

Without this, the audio playback contends for CPU with the event queue
and mic input, which is bad

MCP recently added `structuredContent` as an additional field in tool
results. This lets us use that field when sending the response to the
model.

#1101)

### Problem
1. **Incorrect parameter reference**: Suggests passing
`output_schema_strict=False` to `Agent()`, but this parameter doesn't
exist
2. **Vague guidance**: The phrase "make the output type strict" doesn't
explain HOW to make types strict

### Solution
Updated the error message to provide accurate, actionable guidance:

**Before:**
```
"Either make the output type strict, or pass output_schema_strict=False to your Agent()"
```

**After:**
```
"Either use a dataclass, Pydantic model, or TypedDict, or wrap your type with AgentOutputSchema(your_type, strict_json_schema=False)"
```

---------

Co-authored-by: Kazuhiro Sera <seratch@openai.com>

## Summary

This PR implements real-time streaming of function call arguments as
requested in #834. Previously, function call arguments were only emitted
after the entire function call was complete, causing poor user
experience for large parameter generation.

## Changes

- **Enhanced `ChatCmplStreamHandler`**: Added real-time streaming of
function call arguments during generation
- **New streaming logic**: Function call arguments now stream
incrementally as they are generated, similar to text content
- **Backward compatibility**: Maintains existing behavior for completed
function calls
- **Comprehensive testing**: Added tests for both OpenAI and LiteLLM
models
- **Example implementation**: Created demonstration code showing the new
streaming capability

Closes #834

### Summary

Fixed `MCPUtil.invoke_mcp_tool()` to return `"[]"` instead of logging an
error when MCP tools return empty content. Empty results (`None, []`)
are valid outputs and should be treated consistently with
function tools. The `None -> "[]"` and `[] -> "[]"` behavior comes from
the MCP Python SDK's[ FastMCP implementation](
https://github.com/modelcontextprotocol/python-sdk/blob/6566c08446ad37b66f1457532c7f5a14f243ae10/src/mcp/server/fastmcp/utilities/func_metadata.py#L500).


  ### Test plan

- Added comprehensive test coverage in
`test_mcp_fastmcp_behavior_verification()` that verifies all FastMCP
edge cases
  - Enhanced FakeMCPServer to support custom content testing scenarios
  - Ran full test suite - all tests pass
  - Verified the fix handles `None, [], {}, [{}], [[]]` cases correctly

  ### Issue number

  Closes #1035

  ### Checks

  - [x] I've added new tests (if relevant)
  - [ ] I've added/updated the relevant documentation
  - [x] I've run `make lint` and `make format`
  - [x] I've made sure tests pass

Co-authored-by: thein <thein@MacBook-Pro.local>

Breakage from #1142

Add a web demo that uses a fastapi backend. Also remove the CLI UI demo.

Documentation (both written and code)