Provider-neutral contract
Set CHAT_MODEL_URL, CHAT_MODEL_KEY and CHAT_MODEL_NAME for chat. Equivalent CODE_ variables configure code text. The worker sends a POST with model, capability, instructions, context, task, acceptance criteria and a maximum output length. The gateway should return HTTP 200 with a nonempty text field.
Important boundaries
Set CHAT_MODEL_FORMAT=responses for direct OpenAI Responses, bedrock for AWS Converse, or gateway for the text contract above. Equivalent CODE_ settings allow Claude Sonnet coding alongside OpenAI coordination. Bedrock needs a model/inference-profile ID, AWS region and AWS SDK credentials; it does not use MODEL_URL/KEY. Configure BEDROCK_THINKING for the selected Claude version and explicitly map unsupported effort names. Other vendor formats need a trusted gateway. Credentials stay on the server. Configure chat even for code Pods because management, review and final steps use it.
Failure and usage
Invalid output, non-200 status or the 45-second timeout blocks the Pod. MAX_MODEL_CALLS_PER_POD bounds attempts, but the Pod budget is only context. Token and currency billing controls require a provider-specific adapter.
Other modalities
Image, video, audio and vision slots are reserved. Text embeddings support company knowledge retrieval with NVIDIA or Azure. They need typed artifacts, storage, polling/cancellation, usage accounting and validation before they can run. The full adapter contract is in web/docs/PROVIDERS.md in the source repository.