https://github.com/bit-web24/autotest

An Agentic AI for testing random code
https://github.com/bit-web24/autotest

Last synced: 4 months ago
JSON representation

An Agentic AI for testing random code

• Host: GitHub
• URL: https://github.com/bit-web24/autotest
• Owner: bit-web24
• Created: 2025-09-30T19:50:41.000Z (9 months ago)
• Default Branch: master
• Last Pushed: 2026-02-11T17:07:17.000Z (5 months ago)
• Last Synced: 2026-02-12T01:40:49.757Z (5 months ago)
• Language: Python
• Homepage:
• Size: 371 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Autotest

An Agentic AI for testing random code and write code agentically.

## Architecture & Design

This project implements a sophisticated **Supervisor-Worker** agentic workflow using **LangGraph** and **Groq**. The system is designed to autonomously analyze, plan and code software tasks.

### Core Components

The architecture consists of a central **Supervisor** that orchestrates specialized sub-agents:

1.  **Supervisor Agent** (`agent.supervisor`)

    *   **Role**: The project manager and orchestrator.

    *   **Responsibility**: Decomposes complex user requests into sub-tasks, delegates work to the correct sub-agent (`planner` or `coder`), monitors progress, and handles state handoffs.

    *   **Implementation**: Built using `langgraph_supervisor`.

    *   **Model**: Defaults to `llama-3.1-8b-instant` via Groq.

    *   **Memory**: Uses `AsyncSqliteSaver` (SQLite) for persistent state management and checkpointing.

2.  **Planner Agent** (`agent.planner`)

    *   **Role**: The strategic thinker.

    *   **Responsibility**: Creates detailed implementation plans, breaks down tasks, and manages the "todo" list.

    *   **Tooling**: Access to the filesystem via MCP (Model Context Protocol).

    *   **Type**: ReAct Agent.

3.  **Coder Agent** (`agent.coder`)

    *   **Role**: The executor and engineer.

    *   **Responsibility**: Writes code, executes commands, fixes errors, and implements the plans generated by the Planner.

    *   **Tooling**:

        *   **Filesystem MCP**: For reading/writing files and running commands (`npx @modelcontextprotocol/server-filesystem`).

        *   **Context7 MCP**: For query based context retrieval (`npx @upstash/context7-mcp`).

    *   **Type**: ReAct Agent with a specialized system prompt for code analysis and enhancement.

### Data Flow

```mermaid

graph TD

    User([User Request]) --> Supervisor

    

    subgraph "Core Orchestration (LangGraph)"

        Supervisor[Supervisor Agent]

        State[(Shared Graph State)]

        Persistence[(SQLite Persistence)]

        Summarizer[Summarizer / Reflector]

        

        Supervisor <--> State

        State <--> Persistence

        State <--> Summarizer

    end

    subgraph "Worker Agents (ReAct)"

        Planner[Planner Agent]

        Coder[Coder Agent]

    end

    subgraph "MCP Infrastructure"

        FS_MCP[Filesystem MCP Server]

        Context_MCP[Context7 MCP Server]

    end

    Supervisor -- "Delegates Task" --> Planner

    Supervisor -- "Delegates Task" --> Coder

    

    Planner -- "Read/Write Files" --> FS_MCP

    Coder -- "Read/Write/Execute" --> FS_MCP

    Coder -- "Semantic Search" --> Context_MCP

    

    Planner -- "Status/Plan" --> Supervisor

    Coder -- "Results/Code" --> Supervisor

    

    Supervisor --> User

```

### Key Features

*   **Model Context Protocol (MCP)**: The agents use MCP to standardize tool access (e.g., filesystem, external context).

*   **State Management**: Uses `SummaryState` to summarize past interactions and `AsyncSqliteSaver` to persist multiple global states for different chat sessions via unique thread IDs.

*   **Memory Handoffs**: Sub-agents manage task-specific memory by reading from and writing to the shared graph state, ensuring the Supervisor can transition between planning and coding without loss of context.

*   **Reflection**: The agents (especially Coder) are designed with a "reflection" loop to analyze their own work and correct errors before finishing.

*   **Pre-Model Hooks**: A `summarization_node` runs before model calls to optimize context window usage.

### Directory Structure

*   `agent/src/agent/supervisor`: Main entry point and orchestration logic.

*   `agent/src/agent/planner`: Planner agent definition and tools.

*   `agent/src/agent/coder`: Coder agent definition and tools.

*   `agent/src/agent/tools`: Shared tool definitions (legacy/utility).

*   `agent/src/agent/models.py`: LLM configurations (Groq, OpenAI, etc.).

*   `agent/src/agent/settings.py`: Configuration and environment variables.