The model references functions, response fields, or options that aren't in the version you're on: `response.data.items` when it's `response.body.results`, a renamed hook, an option removed two majors ago. It's silent until that path runs. Two defenses: turn on a type checker (`tsc --noEmit`, mypy) so hallucinated APIs fail at build time, and feed the agent the real, current docs through an MCP like Context7 instead of trusting its training. When something looks off, paste the actual API reference and have it rewrite against that.
I'm on [library] version [X]. Before I trust this code, verify that every method, field, and option it uses actually exists in that version, not a hallucinated or renamed API. If you're not confident your training is current, say so and ask me to paste the docs. Then run the type checker and fix anything that doesn't resolve.