A Claude Code subagent that finds code without eating your context

Today's file is a subagent for Claude Code. You ask where something lives in code the session has never seen, the search runs in a separate agent, and what comes back is addresses rather than file contents.

Why a separate agent

Everything the main session reads stays in its context. To find where a function is defined and who calls it, it may open a dozen files, and by the time the real edit starts a good part of the context is taken up by text it no longer needs. A subagent reads in its own context, and only its answer comes back, in lines like these:

app/Services/Billing.php:142 - charge() is defined here
app/Http/Controllers/CheckoutController.php:58 - the only caller

Plus two or three lines on how the thing is built. Its prompt calls retelling a file instead of giving addresses a failed job.

What to change

Ours runs on opus with effort: high, because a wrong address costs more than the search. Put sonnet or haiku in the model line for a cheaper search, or inherit to run it on whatever the main session runs on.

It has no Edit or Write tool, so it cannot change a file through them. It does have Bash, for grep and find, which makes read only a rule in its prompt rather than a lock. If you want the lock, take Bash out of the tools line.

Install

One file goes into ~/.claude/agents/, and a new session picks it up. The word PROACTIVELY at the start of its description nudges Claude to call it on its own before an edit. The command and a one-line check are in the note below.