SkillToolResolver

Return the skill tools that one metadata.include_tools name stands for.

A resolver lets a skill list tools whose real names are only known at runtime, such as generated MCP tool names, and lets one name stand for a family of tools. It's called with the name and the agent's runtime, on every model call while a read or pin of a skill listing that name stays in context, and again before one of its tools runs. It may return the tools or a promise of them.

• Return the same tools for the same name within a thread, or the prompt cache breaks: a change moves bytes at a position that has already been sent.
• Keep it cheap, or cache, since it runs on every model call.
• One resolver serves every thread, so scope any cache by what the tools depend on, usually something in runtime.context.
• Names are resolved concurrently, so cache the pending promise, not only its result.
• Exceptions propagate out of the model call or tool call. To degrade gracefully, catch inside the resolver and return [].

A name that exactly matches a tool already in the request is claimed by that tool and never passed to the resolver.