Build the infrastructure to route prompts to different LLM providers (Anthropic, OpenAI) within a single workflow. This is a prerequisite for MVP 3 but also generally useful.
Current state: the agent system likely targets a single model per workflow. This task adds the ability to call different models within the same orchestration flow.
Requirements:
- A way to specify which model to use for a given inference call (e.g. Opus for spec writing, GPT-5.3 for review, Sonnet for execution)
- Model selection can be driven by task metadata (complexity) or workflow step
- Unified response format regardless of provider (the orchestrator shouldn't care which model answered)
- Error handling: if a provider is down, graceful fallback or clear error
This is infrastructure that MVP 2 doesn't need (it uses one model) but MVP 3 and beyond require.
Acceptance criteria:
- Can make inference calls to at least 2 different providers (Anthropic + OpenAI) in the same workflow
- Model selection is configurable per workflow step
- Responses are normalized to a common format
- Provider failures are handled gracefully