[BUG]: FREISA-GPT raises TypeError after every successful tool call #111

Description of the bug

ChatSession.process_user_request() raises TypeError whenever a tool call succeeds — i.e. on the main FREISA-GPT path.

LLMClient.get_response() is annotated -> str and returns a JSON string (llm_client.py:86):

return json.dumps(data["choices"][0]["message"])

But the caller indexes it as if it were a dict, at mcp_client.py:404:

final_response = await self.llm_client.get_response(messages)
...
return final_response["content"]   # <-- TypeError: string indices must be integers

Steps To Reproduce

This branch is taken whenever process_llm_response() actually executes a tool (so result != llm_response). Minimal reproduction of the type confusion:

import json

def get_response() -> str:                      # mirrors LLMClient.get_response
    return json.dumps({"role": "assistant", "content": "Woof!"})

final_response = get_response()
print(final_response["content"])                # mcp_client.py:404
# TypeError: string indices must be integers, not 'str'

Confirmed statically by mypy:

src/mcp_client_pupper/mcp_client.py:404: error: Invalid index type "str" for "str";
    expected type "SupportsIndex | slice[...]"  [index]

Additional Information

The exception is not caught anywhere on the way up: process_user_request() is awaited unguarded in PuppyVoiceAssistant._process_heard_text(), which is itself driven from asyncio.run(...) inside the sounddevice audio callback (puppy_voice_assistant.py:145). So the failure surfaces inside the realtime audio thread rather than as a clean error.

Possible fixes:

• Have get_response() return the parsed dict (and update the annotation and the other three call sites), or
• json.loads(final_response)["content"] at the call site.

The first is cleaner: llm_response is already json.loads-ed again inside process_llm_response(), so the round-trip through a string serves no purpose.

mypy also flags three Item "None" of "LLMClient | None" has no attribute "get_response" errors in the same file — ChatSession.llm_client is typed Optional[LLMClient] but never narrowed.