refactor: introduce LLM provider adapter ABC with Portkey gateway support

[crivetimihai]

Summary

• Introduces an Abstract Base Class (LLMProviderAdapter) for LLM provider HTTP proxy adapters, consolidating provider-specific wire-format logic that was previously scattered across ~6 if/elif chains in llm_proxy_service.py
• Adds concrete adapter implementations for OpenAI, Azure OpenAI, Anthropic, Ollama (native + OpenAI-compat), and Portkey
• Adds a registry/factory (get_adapter, get_portkey_adapter) that maps LLMProviderType to adapter instances
• Adds Portkey external LLM gateway runtime support (original PR scope)
• Reduces llm_proxy_service.py from ~987 lines to ~456 lines by extracting provider-specific logic into self-contained adapter classes

Architecture

                      ┌─────────────────────┐
                      │  LLMProviderAdapter  │  (ABC)
                      │─────────────────────│
                      │  build_request()     │
                      │  transform_response()│
                      │  transform_stream()  │
                      │  uses_sse_framing    │
                      └──────────┬──────────┘
           ┌──────────┬─────────┼─────────┬──────────┐
           │          │         │         │          │
      OpenAI    Anthropic    Azure    Ollama    Portkey
      Adapter    Adapter    Adapter   Adapter   Adapter

Adapter classes

Adapter	Extends	Provider types
OpenAIAdapter	ABC	openai, openai_compatible, cohere, mistral, groq, together
AzureOpenAIAdapter	OpenAI	azure_openai
AnthropicAdapter	ABC	anthropic
OllamaNativeAdapter	ABC	ollama (native /api/chat)
OllamaOpenAICompatAdapter	OpenAI	ollama (/v1 endpoint)
PortkeyAdapter	OpenAI	portkey (+ gateway routing)

Key design decisions

• Stateless singletons: Adapters are reused across requests (no per-request state)
• Injected get_api_key: Adapters receive API key decryption as a callable, keeping them decoupled from LLMProxyService internals
• Gateway routing stays separate: should_route_provider_via_llm_gateway() remains in llm_provider_service.py — it's an orthogonal routing decision, not a wire-format concern
• LangChain path untouched: mcp_client_chat_service.py is a different concern (creating SDK objects, not HTTP requests)

New files

mcpgateway/services/llm_adapters/
    __init__.py, base.py, registry.py,
    openai_adapter.py, azure_openai_adapter.py,
    anthropic_adapter.py, ollama_adapter.py, portkey_adapter.py

tests/unit/mcpgateway/services/llm_adapters/
    test_openai_adapter.py, test_azure_openai_adapter.py,
    test_anthropic_adapter.py, test_ollama_adapter.py,
    test_portkey_adapter.py, test_registry.py

Modified files

• mcpgateway/services/llm_proxy_service.py — replaced if/elif chains with adapter dispatch, removed extracted private methods
• tests/unit/mcpgateway/services/test_llm_proxy_service.py — updated tests to call adapters directly instead of deleted private methods

Test plan

• 66 new adapter unit tests pass (tests/unit/mcpgateway/services/llm_adapters/)
• 76 existing proxy service tests pass (behavioral no-ops verified at each phase)
• Full unit suite passes
• Formatting clean (make autoflake isort black)

🤖 Generated with Claude Code