Investigate a codebase

Ask an agent a question about a codebase and you get an answer in the same confident register whether it read four files or none. The prose does not carry the difference, so the reader cannot tell a measured claim from a plausible one, and the failure only surfaces later — in the migration that dropped a column something still read, in the deletion that broke boot, in the "nobody calls this" that was a search with a typo in it.

This skill produces an answer whose evidence can be checked. It scores how much the question is worth before spending anything, splits the search so that the parts can genuinely disagree, refuses to call a zero-hit search "absent" without a control that fired, settles contradictions in a table instead of by tone, and closes with what it did not look at.

It reads. It writes nothing to the repository unless a run record is explicitly asked for.

What separates an answer from a guess

Five mechanics, each of which exists because its absence produced a wrong answer somebody acted on: