ADR-003: Provider abstraction

ADR-003: Provider abstraction

Status

Accepted

Context

Chrome Built-in AI availability varies by device/channel. Ollama is critical for early technical users. Features must not hard-code window.ai or a single backend.

Decision

All inference goes through LocalAIProvider with runtime capability detection. UI always shows the selected Provider; switching is explicit.

Consequences

• Higher initial engineering cost
• Enables future WebGPU / OpenAI-compatible adapters without rewriting features
• Prevents silent cloud fallback