harness-synthesizer

Synthesizes executable validation harnesses for AI agent tool calls.

1|Updated May 21, 2026
One-click install
npx skills add https://github.com/hiddink-ai/hiddink-harness --skill harness-synthesizer
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: harness-synthesizer
Source: https://github.com/hiddink-ai/hiddink-harness/tree/main/templates/skills/harness-synthesizer
Command: npx skills add https://github.com/hiddink-ai/hiddink-harness --skill harness-synthesizer

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill addresses the lack of structured constraint enforcement in AI agent tool calls, preventing execution errors and unauthorized actions by generating code-level verifiers.

Core Features & Use Cases

  • Multi-Mode Validation: Supports verifier (post-hoc), filter (pre-execution), and policy (action ranking) modes to suit different safety requirements.
  • 2-Stage Isolation: Implements Base64 encoding and subprocess-based execution to prevent shell injection and memory contamination.
  • Use Case: A developer can generate a hard-enforcement filter for a git-manager agent to block destructive commands like force pushes or hard resets on protected branches.

Quick Start

Use the harness-synthesizer skill to generate an advisory verifier harness for the golang-expert agent.

Frequently Asked Questions about harness-synthesizer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I prevent AI agents from executing unauthorized or destructive tool calls?

To prevent unauthorized AI agent tool calls, you can generate executable validation harnesses that enforce constraints. These harnesses act as code-level verifiers, blocking invalid actions and preventing execution errors in agent workflows.

What is the best way to add safety guardrails for AI agent workflows?

The best way to add safety guardrails is implementing multi-mode validation across verifier, filter, and policy modes. This approach provides advisory or hard-enforced constraints, ensuring structured and safe tool interaction for AI agents.

How can I block destructive commands like force pushes on protected branches in automated workflows?

You can block destructive commands by generating a hard-enforcement filter harness. For example, a filter can be synthesized for a git-manager agent to strictly prevent force pushes or hard resets on protected branches.

How does Base64 encoding and subprocess execution secure AI agent tool interactions?

Base64 encoding and subprocess execution secure AI agent tool interactions through a 2-Stage Isolation pattern. This mechanism prevents shell injection and memory contamination, ensuring structured and safe tool execution.

What is the difference between verifier, filter, and policy modes for AI agent validation?

Verifier, filter, and policy modes differ by execution timing and strictness: verifier provides post-hoc validation, filter applies pre-execution blocking, and policy handles action ranking to suit different safety requirements.