[FEATURE] Support indexed code search tools for reduced token consumption

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code's Grep tool uses ripgrep for every code search. This works, but in a
typical session the agent makes 50-200 searches, and each one returns raw text lines
with no structural context. The agent then has to read full files to understand what
the results mean.

I measured this on a real codebase (ripgrep's source, 208 files): a 3-query
investigation task consumed 20,580 tokens -- 45% of that was file reads just to
understand grep results.

By replacing grep with an indexed search that returns function signatures + call
graph context (using trigram indexing + tree-sitter), the same task took 2,814
tokens. 87% reduction, measured.

The core issue: the Grep tool returns lines. Agents need structure -- function
bodies, who-calls-what, what-breaks-if-I-change-this. That understanding currently
requires multiple tool calls (grep + read + grep + read), which burns tokens and
adds latency.

Proposed Solution

Allow Claude Code to use indexed/structural search tools alongside or instead of
ripgrep. This could be:

1. A hook or config that lets users specify an alternative search command

(e.g., a tool that returns JSON with function signatures + call graph)

1. Native support for search results that include structural context

(enclosing function, callers, callees)

1. MCP server integration for code intelligence tools

I built an open-source prototype (Hypergrep) that demonstrates the approach:

• Trigram-indexed search (4.4ms warm vs ripgrep's 31ms)
• Returns full function bodies instead of lines (tree-sitter, 16 languages)
• Call graph queries (--callers, --impact) in microseconds
• Semantic compression: 3 output layers with token budget fitting
• 120 tests, MIT license

GitHub: https://github.com/marjoballabani/hypergrep
Benchmarks: https://github.com/marjoballabani/hypergrep/blob/main/BENCHMARKS.md

Not suggesting Claude Code adopt Hypergrep specifically -- just that the
architecture would benefit from supporting structured search results rather
than raw grep lines.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_