# AnchorCli

An AI-powered coding assistant built as a .NET 10.0 console application, featuring the **Hashline** technique for safe, precise file editing.

## What is Hashline?

AnchorCli's unique approach to file editing. Every line returned by file tools is prefixed with a content-derived hash anchor:

```
1:a3|  function hello() {
2:f1|    return "world";
3:0e|  }
```

When editing, you reference these `line:hash` anchors instead of reproducing old content. Before any mutation, both the line number **and** hash are validated — stale anchors are rejected immediately.

This eliminates:
- Whitespace/indentation reproduction errors
- Silent edits to the wrong line in large files
- Entire-file rewrites just to change one line

## Features

- **Interactive REPL**: Chat with an AI model to edit files, manage directories, and execute commands
- **Comprehensive Toolset**: 12+ tools for file operations, editing, directory management, and command execution
- **AOT-Ready**: Native AOT compilation for ~12 MB binaries with no .NET runtime dependency
- **Rich CLI**: Beautiful terminal output using Spectre.Console with tables, rules, and colored text
- **Streaming Responses**: Real-time AI response streaming in the terminal
- **OpenAI-Compatible**: Works with any OpenAI-compatible API (OpenAI, Ollama, Cerebras, Groq, OpenRouter, etc.)

## Requirements

- .NET 10 SDK
- `clang` (for native AOT publish on Linux)

## Quick Start

```bash
export ANCHOR_API_KEY=your_key_here
export ANCHOR_MODEL=qwen3.5-27b       # optional, default: gpt-4o
export ANCHOR_ENDPOINT=https://api.openai.com/v1  # optional

dotnet run --project AnchorCli
```

## Native AOT Build

```bash
dotnet publish AnchorCli -r linux-x64 -c Release
./AnchorCli/bin/Release/net10.0/linux-x64/publish/anchor
```

The resulting binary is ~12 MB, has no .NET runtime dependency, and starts instantly.

## Environment Variables

| Variable | Default | Description |
|---|---|---|
| `ANCHOR_API_KEY` | *(required)* | API key for the LLM provider |
| `ANCHOR_ENDPOINT` | `https://api.openai.com/v1` | Any OpenAI-compatible endpoint |
| `ANCHOR_MODEL` | `gpt-4o` | Model name |
| `ANCHOR_MAX_TOKENS` | `4096` | Max response tokens |

## Available Tools

**File Operations:**
- `read_file` - Read a file (or a window) with Hashline-tagged lines
- `grep_file` - Search a file by regex — results are pre-tagged for immediate editing
- `find_files` - Search for files matching glob patterns
- `get_file_info` - Get detailed file information (size, permissions, etc.)

**Edit Operations:**
- `replace_lines` - Replace a range identified by `line:hash` anchors
- `insert_after` - Insert lines after an anchor
- `delete_range` - Delete a range between two anchors
- `create_file` - Create a new file with optional initial content
- `delete_file` - Delete a file permanently
- `rename_file` - Rename or move a file
- `copy_file` - Copy a file to a new location
- `append_to_file` - Append lines to the end of a file

**Directory Operations:**
- `list_dir` - List directory contents
- `create_dir` - Create a new directory (with parent directories)
- `rename_dir` - Rename or move a directory
- `delete_dir` - Delete a directory and all its contents

**Command Execution:**
- `execute_command` - Run shell commands (with user approval)

## Project Structure

```
AnchorCli/
├── Program.cs              # Entry point + REPL loop + AI client setup
├── Config/AppConfig.cs     # Environment variable configuration
├── Hashline/
│   ├── HashlineEncoder.cs  # Adler-8 + position-seed hashing
│   └── HashlineValidator.cs # Anchor resolution + validation
├── Tools/
│   ├── FileTools.cs        # read_file, grep_file, find_files, get_file_info
│   ├── EditTools.cs        # replace_lines, insert_after, delete_range, create/delete/rename/copy/append
│   ├── DirTools.cs         # list_dir, create_dir, rename_dir, delete_dir
│   └── CommandTool.cs      # execute_command
└── Json/AppJsonContext.cs  # Source-generated JSON context (AOT)
```

## How It Works

1. **Setup**: The AI client is configured with your API credentials and model preferences
2. **REPL Loop**: You interact with the AI through a conversational interface
3. **Tool Calling**: The AI can call any of the available tools to read/edit files, manage directories, or execute commands
4. **Hashline Validation**: All file edits are validated using the Hashline technique to ensure precision
5. **Safe Execution**: Commands require explicit user approval before running

## Supported Models

AnchorCli works with any OpenAI-compatible API endpoint, including:
- OpenAI (gpt-4o, gpt-4.1, etc.)
- Ollama (local models)
- Cerebras
- Groq
- OpenRouter (qwen3.5-27b, etc.)
- Any custom OpenAI-compatible server

## License

MIT License - See LICENSE file for details.