# FastHarness

**Repository Path**: daoos_admin/FastHarness

## Basic Information

- **Project Name**: FastHarness
- **Description**: No description available
- **Primary Language**: Rust
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-04-04
- **Last Updated**: 2026-04-06

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# FastHarness (Rust Edition) 🦀

FastHarness is a high-performance, robust, and highly extensible AI Agent framework completely written in Rust. It serves as a 100% exact parity port of the Python-based `OpenHarness` project, offering a lightning-fast asynchronous engine, robust type safety, and zero-warning compilation.

Built on top of `tokio`, `reqwest`, and `serde`, FastHarness orchestrates multi-agent collaboration, dynamic tool loading, persistent session memory, and complex prompt engineering through an interactive REPL interface.

---

## 🌟 Key Features

### 🚀 Core Engine
- **Fully Asynchronous**: Powered by `tokio` for non-blocking I/O, tool execution, and background task management.
- **Server-Sent Events (SSE)**: Native streaming support for LLM responses using `eventsource-stream`.
- **Dynamic Prompt Engineering**: Context-aware prompts that automatically inject OS, architecture, shell info, and parse `CLAUDE.md` / `MEMORY.md` workspace files.
- **Persistent Configuration**: Global `~/.openharness/settings.json` and project-level `.fastharness/settings.json` overrides.

### 🛠 Tool Ecosystem (41 Built-in Tools)
FastHarness includes a massive arsenal of tools out of the box:
- **Filesystem & OS**: Read, write, edit files, Glob pattern matching, Grep regex search, and raw Bash execution.
- **Web & Information**: Web searching and HTML-to-Markdown fetching (`scraper`, `html2md`).
- **Swarm & Orchestration**: Spawn local background agents (`agent`), manage teams (`team_create`), and control background tasks (`task_create`, `task_stop`, `task_output`).
- **Workflow & Memory**: Interactive Todo list management, memory persistence, and strict Plan Mode (`enter_plan_mode`).
- **Extensibility (MCP)**: Adapters for Model Context Protocol (MCP) servers to expose external tools to the LLM.

### 🔌 Extensibility & Middleware
- **Markdown Skills**: Dynamically load custom tools from Markdown files with YAML frontmatter located in `.fastharness/skills/`.
- **Lifecycle Hooks**: Global event interceptors for `session_start`, `session_end`, `pre_tool_use`, and `post_tool_use`.
- **Permission Governance**: Interactive `[y/N]` prompts for destructive tools (like bash execution or file modification) seamlessly integrated into the async stream.

### ⌨️ Command Registry (Slash Commands)
A powerful REPL driven by `rustyline` with 60+ planned slash commands. Currently implemented:
- `/help` - Show available commands.
- `/status` - Display active model, context path, and session status.
- `/clear` - Clear the terminal and conversation history.
- `/cost` - Real-time estimation of token usage and billing (powered by `tiktoken-rs`).
- `/resume` - Restore the last saved session for continuous workflow.
- `/vim`, `/voice`, `/keybindings` - Extensible stubs for advanced interaction modes.

---

## 📦 Installation & Setup

### Prerequisites
- Rust 1.70+ and Cargo
- An Anthropic API Key

### Build from Source
```bash
git clone <your-repo-url>/FastHarness
cd FastHarness
cargo build --release
```

### Environment Variables
Set your Anthropic API key before running:
```bash
export ANTHROPIC_API_KEY="sk-ant-api03-..."
```

---

## 🎮 Usage

Start the interactive REPL:
```bash
cargo run --release
```

Or pass a specific prompt directly via CLI:
```bash
cargo run --release -- --prompt "Write a Python script to scrape a website"
```

### CLI Arguments
```text
Usage: fast_harness [OPTIONS]

Options:
  -p, --prompt <PROMPT>              Execute a single prompt and exit
  -c, --cwd <CWD>                    Target working directory (default: current dir)
  -m, --model <MODEL>                Override the default LLM model
      --max-turns <MAX_TURNS>        Maximum number of autonomous agent turns
      --backend-only                 Run as a headless backend server (RPC over stdin/stdout)
      --permission-mode <MODE>       Set permission mode (full_auto, default, plan)
      --system-prompt <PROMPT>       Override the base system prompt
  -h, --help                         Print help
```

---

## 🏗 Architecture & Modules

- `src/engine/` - The core Query loop, SSE parsing, and message tracking.
- `src/tools/` - Implementation of the 41 built-in agent tools.
- `src/commands/` - The Slash Command Registry and command handlers.
- `src/permissions/` - Governance, denied tools lists, and interactive confirmation.
- `src/coordinator/` - Agent Swarm logic, Team registries, and background task spawning.
- `src/services/` - Background daemons including Session Storage, Token Estimation, and Cron.
- `src/prompt/` - Dynamic prompt construction and environment context injection.
- `src/skills/` & `src/hooks/` - Dynamic extensibility and lifecycle interceptors.
- `src/ui/` - Presentation layers including REPL, React Launcher (stub), and Textual App (stub).

---

## 🤝 Roadmap & Missing Features (Compared to Python OpenHarness)
While the structural architecture is 100% aligned, the following integrations are currently stubs and will be implemented in future PRs:
1. **True MCP IPC Communication**: Connecting `McpToolAdapter` to actual `stdio` subprocesses.
2. **LSP Integration**: Connecting `LspTool` to local Language Servers (e.g., `rust-analyzer`).
3. **Voice Mode**: Implementing real-time Speech-to-Text via `cpal` or external APIs.
4. **WebSocket Bridge**: Enabling the `ReactLauncher` to serve the web UI.

---

## 📄 License
This project is licensed under the MIT License.