Problem Description

FileTools.search_content today has several limitations that make it unsuitable for anything beyond trivial lookups:

• Substring-only matching — no regex. An agent can't search for r"def \w+", r"class [A-Z]", or r"TODO.*priority".
• Pure Python — walks the tree, opens each file, scans line by line. Slow on monorepos.
• Hardcoded file types — only files with an extension in TEXT_EXTENSIONS are searched. .proto, .tf, .nix, .go, and everything else is invisible.
• Snippet-based output — returns file path + size + a character snippet around the first match. Useful for a human reader, not for an agent that wants structured per-line hits it can reason about.

Claude Code, Cursor, and every modern coding assistant reach for ripgrep as the search primitive. Agno should too, with a pure-Python fallback for users who don't have rg installed so the tool remains zero-dependency.

Proposed Solution

Add grep to FileTools with the following shape:

def grep(
    self,
    pattern: str,
    path: Optional[str] = None,
    output_mode: Literal["files_with_matches", "content", "count"] = "files_with_matches",
    include: Optional[str] = None,
    ignore_case: bool = False,
    context: int = 0,
    limit: int = 250,
    multiline: bool = False,
) -> str

Key properties:

1. ripgrep when available, Python re fallback otherwise. shutil.which("rg") resolved once at __init__; the JSON result shape is identical across backends so callers never special-case.
2. Three output modes. files_with_matches (default, paths only), content (per-line hits with line numbers), count (per-file totals). Gives the agent an explicit verbosity knob.
3. include glob — filename filter like "*.py" or "**/*.tsx".
4. multiline — pattern can cross line boundaries when needed (ripgrep -U --multiline-dotall, Python re.DOTALL).
5. Hard cap at GREP_MAX_LIMIT=1000 — protects the agent's context from accidental blow-up; docstring instructs the agent to narrow the search when truncated: true rather than raise limit.
6. search_content stays untouched — grep is additive, not a replacement.

Parameter names (ignore_case, include, context, limit) are aligned with the existing CodingTools.grep for consistent vocabulary.

Alternatives Considered

• Extend search_content with use_regex=True flag plus output mode flags. Rejected: clutter of flags on a method whose name doesn't promise regex. Cleaner to add grep as a sibling.
• Hard ripgrep dependency. Rejected: adds a non-Python dependency to what is today a zero-dep toolkit. The fallback path keeps Agno installable without any binary.
• Use fff.nvim or similar. Rejected: no Python binding, no standalone CLI, requires Neovim.

Additional Context

Implementation and test coverage in #7642.

Would you like to work on this?

• Yes, I'd love to work on it! (PR open)