---

name: tldr-code description: Token-efficient code analysis via 5-layer stack (AST, Call Graph, CFG, DFG, PDG). 95% token savings. allowed-tools: [Bash] keywords: [debug, refactor, understand, complexity, "call graph", "data flow", "what calls", "how complex", search, explore, analyze, dead code, architecture, imports]

TLDR-Code: Complete Reference

Token-efficient code analysis. 95% savings vs raw file reads.

Quick Reference

---

The 5-Layer Stack

CopyLayer 1: AST         ~500 tokens   Function signatures, imports
Layer 2: Call Graph  +440 tokens   What calls what (cross-file)
Layer 3: CFG         +110 tokens   Complexity, branches, loops
Layer 4: DFG         +130 tokens   Variable definitions/uses
Layer 5: PDG         +150 tokens   Dependencies, slicing
───────────────────────────────────────────────────────────────
Total:              ~1,200 tokens  vs 23,000 raw = 95% savings

---

CLI Commands

Navigation

Copy# File tree
tldr tree [path]
tldr tree src/ --ext .py .ts        # Filter extensions
tldr tree . --show-hidden           # Include hidden files

# Code structure (codemaps)
tldr structure [path] --lang python
tldr structure src/ --max 100       # Max files to analyze

Search

Copy# Text search
tldr search <pattern> [path]
tldr search "def process" src/
tldr search "class.*Error" . --ext .py
tldr search "TODO" . -C 3           # 3 lines context
tldr search "func" . --max 50       # Limit results

# Semantic search (natural language)
tldr semantic search "authentication flow"
tldr semantic search "error handling" --k 10
tldr semantic search "database queries" --expand  # Include call graph

File Analysis

Copy# Full file info
tldr extract <file>
tldr extract src/api.py
tldr extract src/api.py --class UserService      # Filter to class
tldr extract src/api.py --function process       # Filter to function
tldr extract src/api.py --method UserService.get # Filter to method

# Relevant context (follows call graph)
tldr context <entry> --project <path>
tldr context main --project src/ --depth 3
tldr context UserService.create --project . --lang typescript

Flow Analysis

Copy# Control flow graph (complexity)
tldr cfg <file> <function>
tldr cfg src/processor.py process_data
# Returns: cyclomatic complexity, blocks, branches, loops

# Data flow graph (variable tracking)
tldr dfg <file> <function>
tldr dfg src/processor.py process_data
# Returns: where variables are defined, read, modified

# Program slice (what affects line X)
tldr slice <file> <function> <line>
tldr slice src/processor.py process_data 42
tldr slice src/processor.py process_data 42 --direction forward
tldr slice src/processor.py process_data 42 --var result

Codebase Analysis

Copy# Build cross-file call graph
tldr calls [path]
tldr calls src/ --lang python

# Reverse call graph (who calls this function?)
tldr impact <func> [path]
tldr impact process_data src/ --depth 5
tldr impact authenticate . --file auth  # Filter by file

# Find dead/unreachable code
tldr dead [path]
tldr dead src/ --entry main cli test_  # Specify entry points
tldr dead . --lang typescript

# Detect architectural layers
tldr arch [path]
tldr arch src/ --lang python
# Returns: entry layer, middle layer, leaf layer, circular deps

Import Analysis

Copy# Parse imports from file
tldr imports <file>
tldr imports src/api.py
tldr imports src/api.ts --lang typescript

# Reverse import lookup (who imports this module?)
tldr importers <module> [path]
tldr importers datetime src/
tldr importers UserService . --lang typescript

Quality & Testing

Copy# Type check + lint
tldr diagnostics <file|path>
tldr diagnostics src/api.py
tldr diagnostics . --project              # Whole project
tldr diagnostics src/ --no-lint           # Type check only
tldr diagnostics src/ --format text       # Human-readable

# Find affected tests
tldr change-impact [files...]
tldr change-impact                        # Auto-detect (session/git)
tldr change-impact src/api.py             # Explicit files
tldr change-impact --session              # Session-modified files
tldr change-impact --git                  # Git diff files
tldr change-impact --git --git-base main  # Diff against branch
tldr change-impact --run                  # Actually run affected tests

Caching

Copy# Pre-build call graph cache
tldr warm <path>
tldr warm src/ --lang python
tldr warm . --background                  # Build in background

# Build semantic index (one-time)
tldr semantic index [path]
tldr semantic index . --lang python
tldr semantic index . --model all-MiniLM-L6-v2  # Smaller model (80MB)

---

Daemon (Faster Queries)

The daemon holds indexes in memory for instant repeated queries.

Daemon Commands

Copy# Start daemon (backgrounds automatically)
tldr daemon start
tldr daemon start --project /path/to/project

# Check status
tldr daemon status

# Stop daemon
tldr daemon stop

# Send raw command
tldr daemon query ping
tldr daemon query status

# Notify file change (for hooks)
tldr daemon notify <file>
tldr daemon notify src/api.py

Daemon Features

Daemon Socket Protocol

Send JSON to socket, receive JSON response:

Copy// Request
{"cmd": "search", "pattern": "process", "max_results": 10}

// Response
{"status": "ok", "results": [...]}

All 22 daemon commands:

Copyping, status, shutdown, search, extract, impact, dead, arch,
cfg, dfg, slice, calls, warm, semantic, tree, structure,
context, imports, importers, notify, diagnostics, change_impact

---

Semantic Search (P6)

Natural language code search using embeddings.

Setup

Copy# Build index (downloads model on first run)
tldr semantic index .

# Default model: bge-large-en-v1.5 (1.3GB, best quality)
# Smaller model: all-MiniLM-L6-v2 (80MB, faster)
tldr semantic index . --model all-MiniLM-L6-v2

Search

Copytldr semantic search "authentication flow"
tldr semantic search "error handling patterns" --k 10
tldr semantic search "database connection" --expand  # Follow call graph

Configuration

In .claude/settings.json:

Copy{
  "semantic_search": {
    "enabled": true,
    "auto_reindex_threshold": 20,
    "model": "bge-large-en-v1.5"
  }
}

---

Languages Supported

---

Ignore Patterns

TLDR respects .tldrignore (gitignore syntax):

Copy# .tldrignore
.venv/
__pycache__/
node_modules/
*.min.js
dist/

First run creates .tldrignore with sensible defaults. Use --no-ignore to bypass.

---

When to Use TLDR vs Other Tools

---

Python API

Copyfrom tldr.api import (
    # L1: AST
    extract_file, extract_functions, get_imports,
    # L2: Call Graph
    build_project_call_graph, get_intra_file_calls,
    # L3: CFG
    get_cfg_context,
    # L4: DFG
    get_dfg_context,
    # L5: PDG
    get_slice, get_pdg_context,
    # Unified
    get_relevant_context,
    # Analysis
    analyze_dead_code, analyze_architecture, analyze_impact,
)

# Example: Get context for LLM
ctx = get_relevant_context("src/", "main", depth=2, language="python")
print(ctx.to_llm_string())

---

Bug Fixing Workflow (Navigation + Read)

Key insight: TLDR navigates, then you read. Don't try to fix bugs from summaries alone.

The Pattern

Copy# 1. NAVIGATE: Find which files matter
tldr imports file.py              # What does buggy file depend on?
tldr impact func_name .           # Who calls the buggy function?
tldr calls .                      # Cross-file edges (follow 2-hop for models)

# 2. READ: Get actual code for critical files (2-4 files, not all 50)
# Use Read tool or tldr search -C for code with context
tldr search "def buggy_func" . -C 20

Why This Works

For cross-file bugs (e.g., wrong field name, type mismatch), you need to see:

• The file with the bug (handler accessing task.user_id)
• The file with the contract (model defining owner_id)

TLDR finds which files matter. Then you read them.

Getting More Context

If TLDR output isn't enough:

• tldr search "pattern" . -C 20 - Get actual code with 20 lines context
• tldr imports file.py - See what a file depends on
• Read the file directly if you need the full implementation

---

Token Savings Evidence

CopyRaw file read:    23,314 tokens
TLDR all layers:   1,189 tokens
─────────────────────────────────
Savings:              95%

The insight: Call graph navigates to relevant code, then layers give structured summaries. You don't read irrelevant code.