🤖 Code Task Agent for VS Code

100% Free & Open-Source, autonomous AI coding agent for Visual Studio Code.
Built on the Hermes Agent Standard and OpenCode Architecture: Featuring Plan Mode vs Build Mode, Context Compaction Engine, Model Context Protocol (MCP), Image/Vision support, SQLite FTS3 Long-Term Memory, and BYOK (Bring Your Own Key). No subscriptions required.

🎁 100% Free & Open Source (BYOK)

CodeTask AI Agent is completely 100% free and open-source forever.

No Subscriptions or Paywalls: You have full control over your models and usage.
BYOK (Bring Your Own Key): Connect directly to OpenAI, Anthropic Claude, Google Gemini, DeepSeek, OpenRouter, or run 100% locally for free with Ollama.
Privacy & Security: Your API keys are stored encrypted locally inside VS Code's native SecretStorage. Code is never sent to third-party telemetry servers.

✨ Features Overview

🔑 BYOK (Bring Your Own Key): Connect to any LLM provider — OpenAI (gpt-4o), Anthropic (claude-3-5-sonnet), Google Gemini, DeepSeek, Ollama (Local), OpenRouter, or any custom OpenAI-compatible endpoint.
🛡️ Plan Mode vs ⚡ Build Mode:
- Plan Mode (Read-Only Analysis): Safe codebase research using read-only tools (read_file, list_dir, grep_search) to generate structured implementation_plan.md artifacts without mutating workspace code.
- Build Mode (Full Execution): Full read/write/terminal execution with auto-verification (tsc --noEmit, vitest, lint).
🗜️ Context Compaction Engine (compaction.ts): OpenCode-inspired automatic context pruning and turn summarization to prevent token overflow HTTP 400 errors during long sessions.
🔌 Model Context Protocol (MCP) Client: Full @modelcontextprotocol/sdk integration. Automatically discovers tools and resources from .mcp.json.
🧠 Long-Term Memory & Sessions: Session persistence powered by SQLite FTS3 (via sql.js WASM) and file-based context files (AGENTS.md, SOUL.md, CONTEXT.md, MEMORY.md).
🤖 Sub-Agent Delegation: Spawn specialized background worker subagents (researcher, coder, tester) to execute sub-tasks concurrently.
🖼️ Vision / Multimodal Support: Attach images directly in chat for vision-capable models (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5/3.6). Automatically strips image attachments when switching to text-only models (DeepSeek, etc.).
⚡ High Resilience Stream Engine: 25s per-chunk inactivity timeout, 3-attempt exponential backoff retry, automatic JSON argument repair, and auto-continuation loops.

🚀 Quick Start

Installation

Clone the Repository:

git clone https://github.com/code-task-agent/code-task-agent.git
cd code-task-agent

Install & Build:
```
npm install
npm run build
```
Launch in VS Code:
- Open the folder in VS Code (code .).
- Press F5 to open the Extension Host window.
- Click the Code Task Agent icon on the Activity Bar.

Package VSIX Extension

To build an offline .vsix installer:

npm run package

Install the generated code-task-agent-0.1.0.vsix via VS Code Extensions $\rightarrow$ Install from VSIX...

⚙️ Configuration (BYOK)

Click the Code Task Agent sidebar icon (or press Ctrl+Shift+P $\rightarrow$ Code Task Agent: Setup Provider & API Key).
Navigate to Settings:
- Click ＋ Add Provider Profile.
- Select your provider (OpenAI, Anthropic, OpenRouter, Google, DeepSeek, Ollama, Custom).
- Enter your model name (e.g. claude-3-5-sonnet-20241022, gpt-4o, deepseek-chat) and API Key.
API Keys are encrypted securely inside VS Code's native SecretStorage.

🏗️ Architecture

graph TD
    A[VS Code Extension Host] <-->|Webview postMessage| B[React 18 Webview UI]
    A --> C[AgentController]
    C --> D[AgentLoop]
    D --> E[Provider Layer]
    D --> F[Tool Registry]
    D --> G[Compaction Engine]
    D --> H[Modes System]
    
    E -->|OpenAI / Anthropic / Ollama / DeepSeek| I[LLM API Endpoints]
    F --> J[Built-in Tools: read/write/edit, grep, terminal]
    F --> K[MCP Manager stdio]
    F --> L[Worker Subagents]
    C --> M[Memory Store sql.js / FTS3]

📜 Standard Alignment

Hermes Agent Standard: Context files (AGENTS.md, SOUL.md, CONTEXT.md), progressive skill disclosure (SKILL.md), provider-agnostic stable system prompts.
OpenCode Harness Architecture: Event-driven ReAct loop, context compaction engine, Plan Mode vs Build Mode tool filtering, git snapshot safety nets.

📄 License

Distributed under the MIT License. See LICENSE for details.