Handle string and dict-shaped responses from OpenAI-compatible backends so non-standard providers no longer crash on missing choices fields. Add regression tests to keep SDK, dict, and plain-text parsing paths aligned.
- Remove litellm dependency entirely (supply chain risk mitigation) - Add AnthropicProvider (native SDK) and OpenAICompatProvider (unified) - Merge CustomProvider into OpenAICompatProvider, delete custom_provider.py - Add ProviderSpec.backend field for declarative provider routing - Remove _resolve_model, find_gateway, find_by_model (dead heuristics) - Pass resolved spec directly into provider — zero internal lookups - Stub out litellm-dependent model database (cli/models.py) - Add anthropic>=0.45.0 to dependencies, remove litellm - 593 tests passed, net -1034 lines
