Search Hierarchy

Search Tool Hierarchy

Use the most token-efficient search tool for each query type.

Decision Tree

Query Type?
├── STRUCTURAL (code patterns)
│   → AST-grep (~50 tokens output)
│   Examples: "def foo", "class Bar", "import X", "@decorator"
│
├── SEMANTIC (conceptual questions)
│   → LEANN (~100 tokens if path-only)
│   Examples: "how does auth work", "find error handling patterns"
│
├── LITERAL (exact identifiers)
│   → Grep (variable output)
│   Examples: "TemporalMemory", "check_evocation", regex patterns
│
└── FULL CONTEXT (need complete understanding)
    → Read (1500+ tokens)
    Last resort after finding the right file

Token Efficiency Comparison

| Tool | Output Size | Best For | |------|-------------|----------| | AST-grep | ~50 tokens | Function/class definitions, imports, decorators | | LEANN | ~100 tokens | Conceptual questions, architecture, patterns | | Grep | ~200-2000 | Exact identifiers, regex, file paths | | Read | ~1500+ | Full understanding after finding the file |

Hook Enforcement

The grep-to-leann.sh hook automatically:

1. Detects query type (structural/semantic/literal)
2. Blocks and suggests AST-grep for structural queries
3. Blocks and suggests LEANN for semantic queries
4. Allows literal patterns through to Grep

DO

• Start with AST-grep for code structure questions
• Use LEANN for "how does X work" questions
• Use Grep only for exact identifier matches
• Read files only after finding them via search

DON'T

• Use Grep for conceptual questions (returns nothing)
• Read files before knowing which ones are relevant
• Use Read when AST-grep would give file:line
• Ignore hook suggestions

Examples

# STRUCTURAL → AST-grep
ast-grep --pattern "async def $FUNC($$$):" --lang python

# SEMANTIC → LEANN
leann search opc-dev "how does authentication work" --top-k 3

# LITERAL → Grep
Grep pattern="check_evocation" path=opc/scripts

# FULL CONTEXT → Read (after finding file)
Read file_path=opc/scripts/z3_erotetic.py

Optimal Flow

1. AST-grep: "Find async functions" → 3 file:line matches
2. Read: Top match only → Full understanding
3. Skip: 4 irrelevant files → 6000 tokens saved

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: parcadei
• Source: parcadei/Continuous-Claude-v3
• License: MIT

Install and usage instructions live in the source repository linked above.