mentask 0.20.0

MentAsk: Universal AI Coding Agent with multi-provider support via models.dev

mentask

Autonomous AI Coding Agent for the Terminal

---

v0.20.0: THE SPICE MUST FLOW | Level 4 Autonomy: Self-Evolving Tooling Architecture

---

mentask is a professional, autonomous coding agent designed for complex software engineering. Powered by an advanced asynchronous reasoning loop and a multi-layer orchestration engine, it doesn't just edit code — it evolves its own capabilities to match the specific needs of your codebase.

No GUI. No cloud sync. No bloat. Just a high-performance terminal agent with hardened security and autonomous tool-forging capabilities.

---

What's New in v0.20.0: Level 4 Autonomy

The "The Spice Must Flow" update introduces a paradigm shift in AI agent capabilities: Self-Evolving Tooling. mentask is no longer limited by its pre-programmed toolset.

1. 3-Layer Plugin Architecture

A specialized hierarchy for total extensibility without compromising core stability:

• Layer 1: Core Tools (Native, Immutable) – The foundational "instincts" (File I/O, Shell, Security).
• Layer 2: Community Plugins (MCP) – Modular integrations with third-party services via the Model Context Protocol.
• Layer 3: Autonomous Plugins (Evolved) – Project-specific tools created and injected by the agent on-the-fly to solve repetitive tasks with native efficiency.

2. Autonomous "Forge" Capability (forge_plugin)

The agent can now architect, validate (via AST), and hot-reload its own Python modules. If a task requires repetitive specialized logic (e.g., massive audio demixing, complex CSV restructuring), mentask will forge a native tool to handle it, saving context tokens and increasing execution speed by orders of magnitude.

3. Persistent Hot-Reloading

New tools are saved to .mentask/plugins/ and immediately available in the agent's schema without restarting the session. These tools persist across sessions and remain isolated from the core application source code.

---

How it works

mentask operates via a Thinking -> Action -> Observation cycle managed by the AgentOrchestrator:

1. Environmental Awareness: Performs a recursive Project Blueprint scan to build a proactive system instruction.
2. Cognitive Loop: Processes intent using advanced multi-model providers (Gemini, DeepSeek, OpenAI).
3. Tool Evolution: If current tools are insufficient, the agent invokes the Forge Engine to expand its own capabilities.
4. Security Centinel: Every action passes through a TrustManager and Safety Layer (Path Traversal protection, MASS_DELETION guards).
5. Atomic Execution: File modifications use a temporary-write + rename strategy with automatic backups.

---

Features

Advanced Agentic Engine

Feature	Description
Self-Forging Tools	Agent creates and hot-reloads its own Python plugins (forge_plugin).
Autonomous Delegation	Spawns specialized sub-agents (Explorer, Verifier) for parallel research.
LSP Integration	Real-time syntax verification and self-correction via Ruff LSP.
Multimodal Intelligence	Native analysis of images, audio, and video demos.
Context Optimization	Proactive "Context Snapping" (summarization) to manage long-turn sessions.
MCP Support	Connect to any external MCP server for database, cloud, or API tools.

---

Installation

Prerequisites

• Python 3.10+
• A Google API Key (or OpenAI-compatible key for other models).

From Source

git clone https://github.com/julesklord/mentask
cd mentask.py
pip install -e ".[dev]"

---

Safety & Security

mentask implements a Hardened Trust Model:

• Path Isolation: The agent is restricted to whitelisted directories (TrustManager).
• Risk Analysis: Commands are categorized (SAFE, NOTICE, WARNING, DANGEROUS).
• Critical Asset Protection: Native protection for .git, .env, and lockfiles.
• Atomic Writes: Zero-risk file editing with automatic .bkp snapshots.

---

Architecture

flowchart TD
    CLI(["mentask CLI"]) --> Orchestrator(AgentOrchestrator)
    Orchestrator --> PluginLoader(PluginLoader)
    PluginLoader -. scan .-> UserPlugins[(.mentask/plugins/)]
    
    subgraph Tool_Hierarchy [3-Layer Toolset]
        Core[Core Tools]
        MCP[Community MCP]
        Dynamic[Evolved Plugins]
    end

    Orchestrator <--> Tool_Hierarchy
    Orchestrator <--> LLM[Gemini / DeepSeek / models.dev]

---

Contributing

Licensed under the MIT License. Built with precision for the modern engineer.

Created by julesklord.