linghungegeg/Linghun

AGI-oriented, hallucination-resistant AI coding runtime grounded in evidence, tools, memory, agents, and verification.

What it solves

Linghun is an AI-powered programming terminal designed for real-world software engineering rather than simple Q&A. It addresses common AI coding failures such as "hallucinations" where models claim to have fixed a bug without reading the code or verifying the result, context window overflow caused by large logs, and the friction of working in complex environments like Windows with Chinese paths and PowerShell.

How it works

Instead of relying solely on prompt engineering, Linghun implements a system-level "engineering exoskeleton" that wraps the LLM. It integrates the model into a controlled runtime that manages:

  • Evidence-First Loop: Forces the model to bind conclusions to actual file reads, tool outputs, and Git states.
  • Verification Gates: Distinguishes between partial, mock, and full smoke tests to prevent the model from claiming a task is "done" without proof.
  • Architectural Guardrails: Uses an "AntiCodeBlob" system to prevent the creation of massive, unmaintainable files and functions.
  • Managed Environment: Handles Windows-specific process guarding, path compatibility, and Git stable points for easy rollbacks.
  • Workflow Matrix: Breaks complex goals into phases, roles, and evidence requirements, supporting multi-agent exploration and background jobs.

Who it’s for

  • Individual Developers: Those who want to use natural language to drive real projects without learning complex tool syntaxes.
  • Professional Engineers: Developers needing a reliable AI partner that respects architectural boundaries, manages Git states, and provides verifiable delivery.
  • Enterprise Teams: Organizations requiring strict permission boundaries, local execution privacy, and the ability to connect internal tools via a standardized Capability Runtime.

Highlights

  • Anti-Hallucination System: Constraints the model to provide evidence for its claims, reducing "confident but wrong" answers.
  • Windows-First Design: Full support for PowerShell, Chinese paths, and native process tracking to ensure stability on Windows.
  • Git Integration: Automatic creation of stable points and managed worktrees to make AI-driven changes reversible.
  • Capability Runtime: An extensible "universal plug" allowing external software to expose capabilities to the AI via a manifest-based connector.
  • Multi-Model Routing: Ability to assign different models to specific roles (planning, execution, review, summary) based on the task.
  • Context Optimization: High cache hit rates (92-96%) achieved through indexing and noise control to reduce token costs.