Inference as a Tool

When a work chain needs “inference” (reasoning, summarization, planning), treat it as a tool with a well-specified contract rather than an implicit side-effect.

• Invocation kinds: prefer runtime.invocation: inference (LLM/human) alongside runtime.invocation: manual for human-only steps. Use entrypoint: none when no script exists and the action is conversational.
• Skill shape: define inputs/outputs so tool calls can be validated and logged. For LLM-backed calls, accept messages + context and return text plus any structured fields the chain needs.
• Routing: callers supply credentials via headers (e.g., X-OpenAI-Key) or session context; skills should not embed provider keys.
• Audit: log tool calls and results as skill executions, even for inference-only steps, to keep chains observable.
• Fallbacks: allow manual/human fulfillment by setting invocation: manual and recording the resulting text in the same output schema.

This keeps inference interchangeable (LLM or human), observable, and composable with other skills in an agential semioverse.