1. Premise

Two problems with the current arrangement (a CLI eval tool plus plain-text search):

Input mangling. Code reaches a CLI tool through shell quoting, so it passes through two layers of escaping that must be hand-authored — fish quoting rules plus the tool's own argument parsing. Backslashes, $, and \uXXXX escapes get eaten. A tool call whose argument is a single JSON string removes the shell layer entirely; the remaining layer is serialized by the harness, not hand-written.

Unstructured output, and too much of it. A CLI returns one interleaved, ANSI-colorized stream: value, stdout, and exception all mixed, silently truncated. Meanwhile plain-text search cannot distinguish a definition from a usage, cannot resolve an alias, and cannot tell two same-named vars apart — which in a repo with one query.cljc and one impl.clj per module is most of the work.

The deeper cost is context. Chatty tooling — fetch a list, then N calls to enrich each element — is what actually fills a context window. A single call that reduces server-side and returns only the answer is worth an order of magnitude more than a faster one.