Search tools

Parent: Tools and workspace.

grep searches file contents with a regex. glob lists files whose paths match a pattern. Both run in-process and do not need rg, fd, or RTK. Prefer these tools over shell search for workspace inspection. grep content mode mints chainable [path#TAG] snapshots only when the selected edit tool is hashline. Match text is search preview only.

• Patterns: grep takes a Rust/regex pattern. glob takes a path glob; a pattern with no / (for example *.rs) matches nested paths as **/*.rs.
• Path: grep path may be a directory or a single file. A file path searches that file only, including when a directory walk would have skipped it as ignored or hidden.
• Defaults: directory walks honor .gitignore, skip hidden files, and never follow symlinks. Pass include_hidden when you need dotfiles.
• Order: results come back in walk order, sorted by name within each directory, so repeat runs agree and a capped result is the first N paths shown rather than an arbitrary sample.
• Caps: results are bounded (default 200). grep also caps matches per file and trims long lines. Every capped, timed-out, or cancelled search says so in its summary, including when it found nothing.
• Output: grep groups matches by file. Content mode shows N | text match previews. When the selected edit tool is hashline, each file is prefixed with a chainable [path#TAG] header. Copy the tag and line numbers into PUT and CUT operations. Do not copy preview bodies because they may be truncated. Use read_file when you need exact line text. Set output_mode to files_with_matches or count when you only need paths or tallies; default is content.
• Permissions: both request read access only. Workspace-scoped searches, plus the user's global AGENTS.md, skill trees, and agent definitions, work in every permission mode, including plan. Other searches outside the workspace are denied in plan and ask first in auto, allow_edits, and supervised.