Add agent_chat tool for agentic workflows with tool-calling

[lambertmt]

Summary

This PR adds support for multi-turn agent conversations where the local LLM can request tool calls that the orchestrating system (Claude) executes.

Features

• agent_chat tool - Start or continue agent conversations with tool-calling support
• Stateful conversation management - Conversations persist with unique IDs for multi-turn workflows
• Few-shot prompt format - Reliable JSON tool call output from instruction-following models
• Multi-strategy JSON parsing - Handles code blocks, inline JSON, and permissive matching
• Tool result continuation - Feed tool results back to continue the conversation
• Auto-cleanup - Conversations expire after 30 minutes
• list_conversations debug tool - Monitor active agent sessions

Use Case

This enables a hybrid architecture where Claude (or other orchestrators) can:

1. Delegate analysis/reasoning tasks to local LLMs
2. Receive structured tool call requests from the local model
3. Execute tools using the orchestrator's capabilities
4. Feed results back to continue the agent loop
5. Get final answers at zero API cost for the delegated reasoning

Example

// Start agent conversation
agent_chat({
  task: "Find Python files and count lines",
  tools: [{ name: "run_command", description: "Run shell command", parameters: {...} }],
  context: "Working in /home/user/project"
})

// Returns: { type: "tool_call", tool_call: { name: "run_command", arguments: {...} }, conversation_id: "..." }

// Continue with tool result
agent_chat({
  conversation_id: "...",
  tool_result: { tool_name: "run_command", result: "file1.py\nfile2.py" }
})

// Returns: { type: "final_answer", content: "Found 2 Python files..." }

Testing

Tested with gpt-oss-120b model via llama-server. The few-shot prompt format produces reliable JSON tool calls.

Test Plan

• TypeScript compiles without errors
• Tool call JSON parsing handles multiple formats
• Conversation state persists across calls
• Direct llama-server testing confirms tool call output
• Full integration test with Claude Code as orchestrator

---

🤖 Generated with Claude Code