notque/vexjoy-agent

What it solves

VexJoy Agent removes the need for users to manually manage a catalog of specialized AI agents and tools. It allows users to make plain-English requests and automatically routes them to the correct domain knowledge, methodology, and enforcement scripts, ensuring that complex tasks are completed with evidence-based verification rather than just LLM confidence.

How it works

The system operates through a four-layer pipeline:

1. Agents: Provide domain-specific knowledge, such as failure mode catalogs and idiom tables.
2. Skills: Implement phased methodologies with strict exit criteria that prevent steps from being skipped.
3. Hooks: Trigger on lifecycle events to block incomplete work without consuming LLM tokens.
4. Scripts: Handle deterministic tasks like linting and test running to remove LLM judgment from repeatable work.

It integrates with various AI CLIs (Claude Code, Codex, Factory, and Reasonix) via a /do router that pairs the request with the appropriate agent and skill.

Who it’s for

• Developers looking for automated architecture and debugging workflows.
• Knowledge workers performing research, drafting, and data analysis.
• AI power users who want to manage routing tables and telemetry.

Highlights

• Anti-Rationalization: Uses automated gates and exit codes to demand evidence (e.g., test output) instead of allowing the LLM to claim a task is "trivial" or "correct."
• Multi-CLI Support: Compatible with Claude Code, Codex, Factory, and Reasonix.
• Content Engine: Capable of researching and drafting content across 397 writing patterns and producing HTML artifacts like slide decks and diagrams.
• Deterministic Execution: Separates LLM orchestration from program execution to ensure reliability.