researchlisted

# Research Confirm external behavior against the current source instead of recalling it. The enemy is the version-specific detail recalled with confidence, wrong in the way that compiles and fails at runtime. The overcorrection is researching a call the repository already makes elsewhere: working code outranks a fresh fetch. ## When to use - A code decision depends on how a pinned external version behaves, and a wrong answer would still compile or type-check. - Not for a question the repository's own code answers, a version-only bump, or a concept with no version dependency. ## The loop 1. **Find the installed version** in the lockfile or manifest first, because the question is how this pinned version behaves, not how the library works. 2. **Fetch that version's first-party docs**, changelog or migration guide, never a blog post; without an exact match use the highest documented version not newer than the install, else the lowest newer one, and state both versions. 3. **Fetch only the section that answers the question** and quote at most ten lines, because a page read once is re-read every later turn. 4. **Cite every claim that shaped a decision** with its URL; an uncited claim is a recollection and is flagged as one. 5. **State what could not be confirmed** instead of filling the gap: "could not confirm X; proceeding on Y" is a valid outcome. ## Where this runs - Delegate the read to a `general-purpose` delegate on `sonnet` from `researcher-prompt.md` with the pinned