ripgrep's default filtering shortens agent code searches compared with grep -r

Hypothesis

In repositories that contain build artefacts, vendored dependencies or virtual environments, an agent whose search tool is ripgrep with its documented default filtering (files matching .gitignore, .ignore and .rgignore globs, hidden files, binary files and symlinks are skipped) locates the relevant source file with fewer search invocations and fewer output tokens than the same agent using grep -r with no --exclude-dir, because its first result set consists of source files rather than of copies in node_modules, .venv, dist or .git.

Prediction

On a fixed set of code-navigation tasks (find the definition of X, find all callers of Y, find where configuration key Z is read) run in checked-out repositories after a build, the ripgrep condition shows a lower median number of search calls before the first correct file is opened, fewer bytes of search output per task, and no lower task success. The advantage shrinks to nothing in clean checkouts without ignored directories, and reverses for tasks whose answer lies in an ignored file (generated code, lock files), where ripgrep needs -uu or --no-ignore and the agent must know to use them.

Proposed test

1. Select repositories with realistic ignored content and a task list with known answers; snapshot them so that both conditions see identical files.
2. Run the same agent and model with the search tool bound to rg with defaults in one condition and to grep -rn without exclusions in the other; log every call, its output size and the file finally opened.
3. Compare calls-to-first-correct-file, output bytes and success rate, reporting distributions rather than means alone. Add a third condition, grep -r --exclude-dir with a hand-written list, to separate the effect of filtering from the choice of tool.

Status

No result is claimed. The hypothesis attributes the effect to filtering, not to ripgrep's speed; a grep invocation with a good exclusion list may perform equally. Outcomes will also depend on how readily the model narrows its own searches.