ai-sdk skill can change dependencies, provider and model outside requested scope

The ai-sdk skill can expand a narrow task into dependency, provider or model changes that were not requested. I confirmed these instructions in skills/ai-sdk/SKILL.md at commit 358400ace0bc12928f0d402184dae60973a3e95e.

Examples:

• Line 317 instructs installing ai when bundled documentation is missing, even when the current task only needs documentation lookup.
• Line 332 defaults to Vercel AI Gateway unless the user names another provider; it does not explicitly preserve an already configured provider.
• Line 333 selects the highest available model version rather than the configured model or a compatible version within the task's cost and behavior constraints.

A concrete trigger is a small fix in an existing AI SDK application with a pinned model and direct provider. The expected behavior is to preserve that architecture, look up documentation matching its installed version, and change dependencies or models only when needed for the requested fix.

Suggested change: prefer installed-version docs, use official remote docs when absent, and preserve the incumbent provider/model unless migration is requested or a demonstrated incompatibility requires a decision. Discovery ordering and model version numbers should not implicitly authorize a migration.

This is an instruction-content finding, not a claim that every model will follow the problematic instructions. Related but separate: #158 discusses overbroad skill injection.

Start in skills/ai-sdk/SKILL.md, especially lines 317 and 332-333, and compare the instructions with the issue's requested scope. Done means the guidance prefers installed-version or official remote docs and preserves the configured provider and model unless migration or incompatibility is explicitly required.