[dmorrissey-aai]

Adds gateway.decisions.create() to LLMGateway and AsyncLLMGateway. It calls the gateway's POST /v1/decisions, which answers predicate, choice, and score questions about an input.

• Request: questions are plain dicts typed by LLMGatewayDecisionQuestionParam. Optional settings are named in the signature so IDEs show them, and any other kwarg is forwarded unchanged. The gateway rejects unknown fields itself.
• Response: LLMGatewayDecisionResponse / LLMGatewayDecisionAnswer. It has its own usage model, because decisions doesn't return prompt_tokens/completion_tokens. probabilities[].value is typed Any: on pydantic v1, a Union[str, int] would coerce one of the two.
• Tests and docs: unit tests for sync and async; examples in CLAUDE.md and README.md.

Testing

• Unit suite passes on pydantic 2 and on pydantic 1.10 (gateway tests).
• Live run against production with gpt-6-luna:
  • sync and async calls, with both a string input and a list of input items
  • all three question types parsed, with choice values as str and level indexes as int
  • a 400 validation error and an unknown kwarg each surface as LLMGatewayError with .errors

🤖 Generated with Claude Code