VexJoy Agent

What it is – VexJoy Agent is a toolkit that lets you issue plain‑English commands to large‑language‑model (LLM) agents (Claude, Codex, Factory, Reasonix, etc.). A router (/do or /d) classifies the request, selects an appropriate agent (domain knowledge) and skill (method), then runs a multi‑phase pipeline (plan, execute, verify, deliver) while enforcing hooks that require evidence before proceeding.

Key concepts

• Agents (43) – curated knowledge packs (e.g., Go‑engineer, testing‑automation‑engineer) that know the domain vocabulary and failure‑mode catalogs.
• Skills (62) – reusable procedural methods that guide the agent through a task (e.g., systematic‑debugging, testing‑preferred‑patterns).
• Hooks (79) – lifecycle checks that block progress unless required evidence (test output, proof of change) is provided.
• Scripts (156) – automation plumbing for validation, scheduling, and orchestration.
• Jev router (/d) – uses TypeSafe’s Jev model to preserve intent: it restates the desired outcome, judges alignment, and records a receipt in learning.db before dispatching the agent.

How it works

/do  →  Router  →  Task Plan  →  Agent+Skill  →  Tests (gates)  →  PR  →  Record

The router picks an agent‑skill pair, runs the task through four phases (reproduce, hypothesize, verify, fix), and finally opens a pull‑request with the fix.

Installation

git clone https://github.com/notque/vexjoy-agent.git ~/vexjoy-agent
cd ~/vexjoy-agent
./install.sh   # installs into ~/.claude, ~/.codex, ~/.factory, ~/.hermes, ~/.reasonix

Options include dry‑run, target selection, and uninstall/rollback. The installer is idempotent and can be configured via a local profile.yaml.

Typical usage

$ claude
> /do debug this Go test

The system routes the request to the go‑engineer agent with the systematic‑debugging skill, runs the four verification phases, patches the code, and opens a PR.

Supported runtimes

• Claude Code (/do)
• Codex ($do)
• Factory (/do)
• Reasonix (/do)
• Hermes (installed but not detailed in the README)

Extensibility

• Add new agents, skills, or hooks by following the SKILL.md front‑matter schema.
• Use the provided scripts (agent-scheduler.py, crontab-manager.py, etc.) to automate recurring judgments or maintenance.
• Optional Jev Auto‑Compact plugin can prune context after 60 % token usage, with evidence logged via jev-compact-evidence.py.

Use cases

• Automated debugging and PR generation for codebases.
• Knowledge‑work pipelines: research, drafting, HTML report generation (/html).
• Structured, evidence‑driven workflows where every step must be verified before moving on.

Design philosophy (summarized)

• Outcome‑first: describe the result, let the router pick the tools.
• Deterministic code for computable parts; models only for judgment/generation.
• Dense, rule‑rich prompts; minimal unnecessary context.
• All work is pipelined with enforced gates to prevent silent failures.

License – MIT.

---

All details are taken directly from the repository’s README; no additional features are inferred.