ctxl-cli 0.3.0

Context Engineering CLI — Reduce AI agent token waste with compressed codebase skeletons, task-focused instructions, and session checkpoints. Zero AI, pure parsing.

ctxl — Context Engineering CLI for AI Agents

Reduce token waste. Prevent hallucination. Zero AI used.

ctxl (pronounced "contextual") is a developer CLI tool that helps you manage the context window of AI coding agents (GitHub Copilot, Cursor, Claude, etc.) by generating compressed codebase skeletons, task-focused instructions, and session checkpoints — all using deterministic parsing, not AI.

The Problem

AI coding agents (Copilot, Cursor, etc.) read your entire codebase to build context. A 3,000-line PySpark file burns ~750 tokens every time the agent references it. Over a session, your context window fills with noise, leading to:

• 🔴 Hallucination — the model starts making things up
• 🔴 Token waste — you pay for irrelevant context
• 🔴 Lost focus — the agent forgets your actual task

The Solution

ctxl gives you three commands that prepare your environment before the AI agent reads it:

ctxl map          # Generate a compressed codebase skeleton (~95% token reduction)
ctxl init         # Generate task-focused Copilot instructions
ctxl checkpoint   # Save session state for safe /clear workflows

Zero AI models. Zero API calls. Zero tokens burned by this tool.

Installation

pip install ctxl-cli

Quick Start

ctxl map — Codebase Skeleton

Generate a compressed structural map of your codebase with line numbers:

ctxl map                      # Map current directory
ctxl map ./src                # Map a specific directory
ctxl map -e .py               # Only Python files
ctxl map -o codebase.md       # Save to file
ctxl map --clipboard          # Copy to clipboard for pasting into AI chat

Before (raw file, ~750 tokens):

class DataPipeline:
    def __init__(self, spark, config):
        self.spark = spark
        self.config = config
        self.source_path = config.get("source_path", "/data/raw")
        # ... 60 more lines of implementation
    
    def clean_data(self, df, drop_nulls=True):
        string_cols = [f.name for f in df.schema.fields ...]
        # ... 20 more lines

After (ctxl map output, ~50 tokens):

L22: class DataPipeline:
    L25: def __init__(self, spark: SparkSession, config: Dict)
    L33: def load_data(self, table_name: str, filters: Optional[Dict] = None) -> DataFrame
    L42: def clean_data(self, df: DataFrame, drop_nulls: bool = True) -> DataFrame
    L52: def transform(self, df: DataFrame, rules: List[Dict]) -> DataFrame
    L66: def validate(self, df: DataFrame) -> bool
    L75: def save(self, df: DataFrame, partition_cols: List[str] = None) -> str

Line numbers (L22:, L42:) let the AI agent navigate directly to the right location.

ctxl init — Copilot Instructions

Generate a .github/copilot-instructions.md file that GitHub Copilot reads natively:

ctxl init "Fix the data pipeline ETL bug"
ctxl init "Add authentication" -f auth.py -f models.py
ctxl init "Refactor tests" --no-map

ctxl checkpoint — Session State

Save your progress before running /clear in Copilot Chat:

ctxl checkpoint save \
    -t "Fix ETL pipeline" \
    --done "Found the bug in clean_data()" \
    --state "Pipeline runs but output has wrong column order" \
    --next "Fix column ordering in transform()" \
    --file "data_pipeline.py"

ctxl checkpoint list          # List all checkpoints
ctxl checkpoint show          # Show latest checkpoint

Supported Languages

ctxl map uses Tree-sitter for parsing and supports:

Language	Extensions
Python	.py
JavaScript	.js, .jsx
TypeScript	.ts, .tsx
Java	.java

More languages can be added easily via Tree-sitter grammars.

How It Works

Your Codebase (10,000+ tokens)
        │
        ▼
   Tree-sitter Parser (deterministic, local, free)
        │
        ▼
   AST → Extract signatures + line numbers
        │
        ▼
   Compressed Skeleton (~500 tokens)
        │
        ▼
   AI Agent reads skeleton instead of raw code
        │
        ▼
   90-95% fewer tokens burned 🎉

License

MIT