Modern software development has entered a fragmented era of AI assistance. A developer might start their morning in Cursor, shift to a Claude-powered environment for complex architectural reasoning, and then move to a different agent for rapid prototyping. The friction is not in the intelligence of these models, but in the repetition of context. Every time a developer switches tools, they find themselves duplicating the same project specifications, coding standards, and architectural constraints across multiple configuration files. This redundancy creates a synchronization nightmare where a change in the project's naming convention must be manually updated in three different hidden dot-files just to keep the AI agents aligned.

The Architecture of VHK v2.9.0

To solve this context drift, VHK (CLI-based AI Coding Harness) has released version 2.9.0. The core philosophy of the tool is to treat project rules as a single source of truth rather than tool-specific configurations. At the center of this system is a single `RULES.md` file. VHK monitors this file and automatically synchronizes its contents across eight different AI tool formats, including `.cursorrules` and `CLAUDE.md`. This ensures that regardless of which agent is currently active in the IDE, the model is operating under the exact same set of constraints and project memories.

From a technical standpoint, VHK is designed for modern JavaScript environments, requiring Node 22 or higher. The tool is distributed via npm and can be deployed globally using the following command:

npm i -g @byh3071/vhk

Beyond simple file synchronization, VHK integrates deeply with the Model Context Protocol (MCP), incorporating 35 distinct tools to extend the capabilities of the connected agents. The project is released under the MIT license and is available on GitHub, allowing the community to extend the synchronization logic as new AI coding agents enter the market.

Moving From LLM Trust to Mechanical Verification

While rule synchronization solves the input problem, VHK v2.9.0 addresses a more critical failure point in AI coding: the hallucination of completion. Most AI agents claim a task is finished based on the LLM's internal probability of success, often ignoring the fact that the code fails to compile or the tests are crashing in the background. VHK introduces a receipt function that shifts the definition of done from a linguistic claim to a mechanical fact. Instead of asking the LLM if the task is complete, VHK monitors the exit codes of critical system commands such as `tsc`, `test`, and `build`.

If the TypeScript compiler returns a non-zero exit code, the task is not complete, regardless of how confident the AI agent sounds. This creates a hard verification loop where the agent is forced to iterate until the machine confirms success. This transition from trust-based to verification-based workflows removes the need for developers to manually run build commands just to check if the AI actually solved the problem.

Furthermore, VHK implements a self-evolution mechanism through a memory and pattern system. As developers interact with the agent, the tool captures lessons learned during specific sessions and stores them in `memory/pattern`. These patterns are then proposed as candidates for the main `RULES.md` file. This transforms the project documentation from a static file into a living organism that evolves based on actual coding outcomes. The addition of Korean natural language routing further streamlines this process, allowing developers to manage these complex orchestrations using intuitive commands.

This shift toward a centralized, verified harness suggests a future where the specific AI model becomes a pluggable commodity, while the project's context and verification logic remain the permanent, sovereign property of the developer.