feat: add start_conversation and continue_conversation tools

[infinitimeless]

Problem

When asking Claude to hold a multi-turn conversation with a local model, Claude defaulted to chat_completion (stateless) rather than create_response. This meant:

• The system prompt (persona, tone, instructions) had to be re-sent on every turn
• Claude manually managed conversation history instead of the bridge handling it
• Responses had no persistent identity or memory across turns

Solution

Two new tools that make multi-turn conversations first-class:

start_conversation(system_prompt, first_message, ...)

Starts a stateful session by sending a system prompt + opening message to /v1/responses. Returns:

• response_id — chain this into every subsequent turn
• message — the model's first response
• model — the model that was used

continue_conversation(response_id, message, ...)

Sends the next message in the session by passing previous_response_id automatically. The system prompt is preserved by LM Studio throughout — never needs to be re-sent.

Typical Workflow

1. start_conversation(
     system_prompt="You are a friend at a bar. Keep it casual and fun.",
     first_message="Hey! How's it going?"
   )
   → { response_id: "resp_abc123", message: "Hey! Not bad, just unwinding..." }

2. continue_conversation(
     response_id="resp_abc123",
     message="Same. Work's been insane this week."
   )
   → { response_id: "resp_def456", message: "Ugh, tell me about it..." }

3. continue_conversation(
     response_id="resp_def456",
     message="If you could go anywhere tomorrow, where would you go?"
   )
   → { response_id: "resp_ghi789", message: "Honestly? Northern Portugal..." }

Other Changes

• Default max_tokens raised from 1024 → 2048 in chat_completion and text_completion to prevent responses being cut off mid-sentence

Requirements

• LM Studio v0.3.29+ (for /v1/responses endpoint support)