Leonxlnx/unlazy

Anti-laziness skill for AI agents. Core: the Depth Tree method, which splits a task N layers deep and gives every leaf the full time budget of the whole task, so effort multiplies with depth. Grounded in 2025-2026 research on model laziness, underthinking and premature completion.

What it solves

It addresses "laziness" and partial compliance in AI agents during substantial software engineering tasks. It prevents agents from claiming work is complete when it is not by requiring runnable, evidence-backed gates that must be verified before a task is considered finished.

How it works

The system uses a "gate contract" (a Markdown ledger) where each requirement is paired with a specific shell command (CHECK:) and an expected output (EXPECT:). A checker script executes these commands and verifies that the actual output matches the expectation. To prevent unauthorized execution, commands must be approved by the user. The project also supports scoped pipelines for parallel work, where agents can claim ownership of specific file paths to avoid write conflicts, and a "Stop hook" for tools like Claude Code that blocks the session until all gates are met.

Who it’s for

Developers using AI agents (such as Claude Code or Codex CLI) for complex, long-horizon coding tasks who need a rigorous way to ensure the agent actually completes every part of the requested refactor or feature.

Highlights

  • Runnable Gates: Replaces vague AI promises with executable shell checks and strict output matching.
  • Evidence-Based Verification: Generates SHA-256 digests of checks and outputs to detect structural drift in the requirements.
  • Parallel Coordination: Provides a scoping and leasing system to manage multiple agents working on disjoint parts of a repository.
  • Integration Hooks: Includes a specialized hook for Claude Code to prevent the agent from finishing a session prematurely.

Written about in

Related

  • Project
    notque/vexjoy-agent

    A toolkit that connects plain-English requests to specialized AI agents and workflows, using a four-layer system of agents, skills, hooks, and scripts to enforce evidence-based task completion.

  • Project
    AMAP-ML/LongHorizon-Harness

    A loop engineering framework that turns existing AI agents into long-running computer-use systems by wrapping them in a structured plan-act-verify cycle across GUI and CLI environments.

  • Project
    appautomaton/latex-arxiv-SKILL

    An agentic harness for writing verified AI/ML review papers in LaTeX, featuring a gated, issue-driven workflow that prevents citation hallucinations.

  • Project
    microsoft/agent-governance-toolkit

    A governance framework for autonomous AI agents that provides deterministic policy enforcement, identity management, and audit logging to prevent unauthorized actions.