Support multiple LLM providers (Claude, Gemini, local models)

[Aditya8840]

Problem

The agent loop in agent.py is tightly coupled to OpenAI's Python SDK. Every LLM interaction — client init, chat completion, tool-call parsing, message serialization — uses OpenAI-specific APIs directly. This makes it impossible to use Claude, Gemini, Ollama, or any other provider without rewriting the core loop.

Hardcoded OpenAI touchpoints:

Location	What
agent.py:4	from openai import OpenAI
agent.py:112	client = OpenAI() — relies on OPENAI_API_KEY env var
agent.py:138-143	client.chat.completions.create(model=..., tools=..., tool_choice="required")
agent.py:145-146	response.choices[0].message / .model_dump(exclude_none=True)
agent.py:152-154	tool_call.function.name / tool_call.function.arguments — OpenAI tool_calls structure
main.py:20-22	--model defaults to gpt-4o via DROIDPILOT_MODEL env var

Expected behavior

Users should be able to choose their LLM provider via CLI flag or env var:

droidpilot "Open Settings" --provider anthropic --model claude-sonnet-4-20250514
droidpilot "Open Settings" --provider ollama --model llama3

Suggested approach

Introduce a provider abstraction that normalizes:

• Client initialization (API key handling per provider)
• Chat completion with tool use
• Tool-call response parsing (each provider returns tool calls differently)
• Message history serialization

The tool definitions in actions.py already use a provider-agnostic JSON schema format, so those can be reused across providers with minimal adaptation.

Scope

• agent.py — extract LLM interaction into a provider interface
• main.py — add --provider CLI arg
• New: provider implementations (openai, anthropic, ollama at minimum)