Feature request : add command for grepai

Summary

Add rtk grepai to filter grepai semantic code search output. Despite being AI-native, grepai's default output is verbose
for LLM consumption — it includes full code bodies, separator lines, and metadata that waste tokens. Its built-in --toon flag paradoxically makes output 2-3x
larger.

Motivation

grepai is increasingly used in Claude Code workflows for semantic search (grepai search) and call graph tracing (grepai trace callers/callees). These are
high-volume commands that return large outputs with embedded code blocks that the LLM already has access to via other tools.

Current pain:

• grepai search "auth logic" (10 results) → ~716 tokens of code bodies, separators, whitespace
• grepai trace callers login (27 callers) → ~319 tokens with repeated file entries and code snippets

Proposed Filters

filter_search() — 90%+ savings

Strip code bodies, keep one line per result:

Before (716 tokens)

Found 10 results for: "authentication logic"
────────────────────── Result 1 (score: 0.68) ──────────────────────
File: backend/src/some/path/api/auth.py:223-228

def authenticate_user(credentials):
"""Authenticate user with the given credentials."""
validated = validate_credentials(credentials)
...
After (48 tokens)

grepai: "authentication logic" — 10 results

0.68  auth.py:223-228              (backend/src/.../api/)
0.38  SKILL.md:330-394 (+51 lines) (.claude/skills/auth-flow/)

filter_trace() — 85%+ savings

Group callers/callees by file, deduplicate:

Before (319 tokens — 27 separate caller entries with code)

After (30 tokens)

login (function) @ auth.py:117 — 27 callers in 8 files:

authRepository.test.ts  18x (lines: 61,96,126,...)
test_exercise_library.py  2x (lines: 120,145)
conftest.py:598  1x

Passthrough

Other subcommands (init, watch, status, graph) pass through unfiltered but still tracked.

Hook Rewrite

Only rewrite search and trace (the filtered subcommands):

grepai search "query"        → rtk grepai search "query"
grepai trace callers login   → rtk grepai trace callers login
grepai init                  → (no rewrite, stays raw)

Expected Savings

┌────────────────────────────┬──────────────┬───────────────┬──────────────────┐
│         Subcommand         │ Input tokens │ Output tokens │     Savings      │
├────────────────────────────┼──────────────┼───────────────┼──────────────────┤
│ search (10 results)        │ ~716         │ ~48           │ 93%              │
├────────────────────────────┼──────────────┼───────────────┼──────────────────┤
│ trace callers (27 callers) │ ~319         │ ~30           │ 91%              │
├────────────────────────────┼──────────────┼───────────────┼──────────────────┤
│ trace graph                │ ~52          │ ~52           │ 0% (passthrough) │
├────────────────────────────┼──────────────┼───────────────┼──────────────────┤
│ status                     │ ~26          │ ~26           │ 0% (passthrough) │
└────────────────────────────┴──────────────┴───────────────┴──────────────────┘

Start by locating the existing CLI command entry points and hook rewrite logic for rtk grepai. Trace how grepai search and trace callers/callees output is handled, then compare behavior with the requested filtered examples. Done means search and trace are compact and deduplicated while init, watch, status, and graph pass through unchanged.