Simulation
Simulation works without model credentials. It assembles a labeled configuration brief from the saved company answers and selected evidence. It is useful for validating the plan and controls. It is not presented as AI-generated business work or proof of business impact.
Live text mode
Live mode uses the selected server-side provider for chat/code: AI_PROVIDER=nvidia, azure, bedrock, or legacy. NVIDIA uses NVIDIA_API_KEY with Lightning chat/code models by default; Azure uses its resource endpoint, API key and deployment names; Bedrock uses AWS credentials, region and model/inference-profile IDs. Explicit provider selection ignores legacy credentials, and no provider fallback occurs. Restart web and worker after changes. NVIDIA free endpoints are for prototyping and subject to provider limits. Vision/media execution remains unimplemented; embeddings support knowledge retrieval. Legacy mode calls the existing provider-neutral gateway or Responses adapter. Chat is also needed for management, quality review and final delivery, even if a code capability is requested. The worker sends instructions, company context, selected evidence, upstream outputs, task and acceptance criteria to the gateway.
When a provider is unavailable
The Pod blocks and reports the missing or failed provider. The application does not silently replace a failed live call with simulation. A blocked Pod can be retried after the cause is addressed.
Reserved capabilities
Image, video, audio and vision have configuration slots but no execution adapter. NVIDIA/Azure text embeddings use the separate knowledge retrieval adapter. Supplying a key alone does not activate them. See model gateways.