VexJoy Agent

Essays and writing behind this toolkit live at vexjoy.com.

AI agents skip steps.

"Looks correct" replaces running tests. "Trivial change" replaces verification. The agent confidently ships broken code because nothing structurally prevented it from skipping the work.

Harnesses have a second problem: given only a skill list, they do not route eagerly enough, or correctly enough. Good skills sit unused. So this toolkit connects the skills, agents, and workflows we want directly into the harness, automatically. You don't have to understand what is here. Say what you want in plain English and you get all the value we have put into it: the right specialist with the right methodology, behind gates that demand exit codes, not assertions.

44 domain agents, 139 workflow skills, 88 hooks, 129 scripts. Agents carry knowledge, skills enforce methodology, hooks block incomplete work, scripts handle determinism.

Works across Claude Code (/do), Codex ($do), Factory (/do), Reasonix (/do).

What It Looks Like

$ claude

> /do debug this Go test

  Routing: go-engineer + systematic-debugging
  Phase 1/4: Reproduce: running test, capturing failure...
  Phase 2/4: Hypothesize: 3 candidates from stack trace...
  Phase 3/4: Verify: isolated root cause in connection pool timeout
  Phase 4/4: Fix: patch applied, test passing, PR opened

  ✓ Delivered: PR #847, fix connection pool timeout in health check

The router reads intent, picks a Go agent paired with a debugging skill, and runs the full lifecycle. You typed one sentence. The system did the rest.

The Pipeline

  ROUTE        PLAN         EXECUTE      VERIFY       DELIVER      LEARN
 ┌──────┐    ┌──────┐    ┌──────┐    ┌──────┐    ┌──────┐    ┌──────┐
 │ /do  │───▶│ Task │───▶│Agent │───▶│Tests │───▶│  PR  │───▶│Record│
 │Router│    │ Plan │    │+Skill│    │Gates │    │Branch│    │Evolve│
 └──────┘    └──────┘    └──────┘    └──────┘    └──────┘    └──────┘

Anti-Rationalization

This is the single thing that separates it from "agent with a system prompt."

Hooks fire automatically. Gates block completion. Skills encode counter-arguments at every skip-worthy step. The agent verifies or it doesn't finish.

For what I do, the difference is enormous. If you're doing simple single-file edits, maybe less so.

Knowledge Work Is First-Class

The same routing serves knowledge work. The content engine researches, drafts in a calibrated voice, validates against 397 AI patterns, and repurposes finished pieces for each platform. /html turns any request into a single self-contained HTML file: report, slide deck, prototype, data viz, diagram. Non-engineers who try the toolkit consistently name the HTML artifacts as the thing they love. No code, no setup beyond the installer.

It Proves Its Own Changes

Changes to the toolkit itself ship with evidence. New skills get blind A/B tests against a no-skill baseline before merge. Routing and writing-standard decisions carry measured verdicts; PHILOSOPHY.md cites the numbers. Experiments that lost go into the negative-results registry, what-didnt-work.md, so no future session re-runs a known-dead path.

The automated nightly evolution loop (/evolve, writes to evolution-reports/) ran regularly through mid-May 2026. It is currently dormant; recent evidence has come from manual PRs instead.

Installation

git clone https://github.com/notque/vexjoy-agent.git ~/vexjoy-agent
cd ~/vexjoy-agent
./install.sh

Links into ~/.claude/ and mirrors into ~/.codex/, ~/.factory/, ~/.reasonix/ — each mirror only when that runtime is detected (its command on PATH or its home dir already exists). The installer asks symlink (live updates via git pull) or copy (stable snapshot).

Want only part of the toolkit? Run ./install.sh --configure to pick which skills, agents, and hooks install, or copy .local.example/profile.yaml to .local/profile.yaml and edit. No profile file = full install, unchanged behavior. Credit: @thomasvan. Details: .local.example/README.md.

Full setup: docs/start-here.md

Mirrors agents, skills, and supported hooks into ~/.codex/. The original six-hook allowlist was correct for Codex v0.114, when tool hooks only intercepted Bash. Current support requires Codex v0.144.1+ and classifies the 76 Claude hook registrations as 26 native, 36 adapter-backed, and 14 unsupported (62 supported). These are registration counts, not unique hook files. The installer also preserves explicit per-subagent model routing for GPT-5.6 Sol by setting the MultiAgent V2 compatibility keys documented in openai/codex#31814.

Codex now exposes apply_patch to tool hooks. VexJoy's adapter converts each patch operation into the Write/Edit payload expected by existing guards, but it cannot intercept writes performed through unified_exec, unmatched MCP tools, WebSearch, or other unsupported tool paths. PreCompact and Stop adapters also receive less telemetry than Claude Code: Codex does not provide Claude's conversation_history or session_data. This is expanded compatibility, not full Claude parity.

After install or any hook-definition change, run /hooks in Codex and review the new definitions before trusting them. Codex hash-trusts hook commands and skips changed, unreviewed definitions.

Gemini CLI support removed (deprecated upstream, transitioned to Antigravity CLI); Antigravity support pending CLI maturity. Per Google's transition announcement, Gemini CLI stops serving requests on 2026-06-18 for Google AI Pro / Ultra and free Gemini Code Assist for individuals. Gemini API integrations (image-gen backends, sprite pipeline, GEMINI_API_KEY) are unaffected and stay in the toolkit.

If a prior install mirrored into ~/.gemini/, remove the stale mirrors with:

rm -rf ~/.gemini/skills ~/.gemini/agents ~/.gemini/hooks ~/.gemini/scripts ~/.gemini/antigravity/plugins/vexjoy-agent

Mirrors agents (as "droids"), skills, and all hooks into ~/.factory/. Hook config merges into ~/.factory/settings.json with paths rewritten.

Mirrors skills, scripts, and the allowlisted hooks (scripts/reasonix-hooks-allowlist.txt) into ~/.reasonix/ (no agent or custom-command surface, so neither is installed; the /do router rides in as a skill). Reasonix fires only 4 events (PreToolUse, PostToolUse, UserPromptSubmit, Stop), so only hooks for those events are allowlisted. Hook config is written to the hooks key of ~/.reasonix/settings.json in Reasonix's native flat shape (one entry per hook, match regex over the tool name); the generator builds absolute python3 commands, so no path rewrite is applied. MCP/model/permissions in ~/.reasonix/config.json are user-owned and left untouched.

The toolkit supplies its own routing, domain knowledge, methodology, and enforcement. The default system prompt duplicates most of that.

claude --system-prompt "."

Strips built-in tool-use instructions. The toolkit's agents, skills, hooks, and CLAUDE.md provide equivalent coverage.

Four Layers

Full skill catalog: docs/skills.md.

┌─────────────────────────────────────────────────┐
│  SKILL.md                                       │
│  ┌─ Frontmatter ─────────────────────────────┐  │
│  │ triggers, pairs_with, success-criteria     │  │
│  └────────────────────────────────────────────┘  │
│  Reference Loading Table (conditional imports)   │
│  Phased Instructions (numbered, with gates)      │
│  Verification (evidence requirements)            │
└─────────────────────────────────────────────────┘

Built with the Toolkit

A game built entirely by Claude Code using these agents, skills, and pipelines:

Choose Your Path

I just want to use it Install, learn /do, done.

I do knowledge work Writing, research, data analysis, moderation, HTML artifacts. No code.

I'm a developer Architecture, extension points, adding agents and skills.

I'm an AI power user Routing tables, pipelines, hooks, learning DB.

I'm an AI agent Machine-dense inventory. Tables, paths, schemas.

I'm on LinkedIn 🚀 Thought leadership. Agree? 👇

Philosophy

• Zero-expertise operation. Say what you want. The system classifies, dispatches, enforces, delivers.
• LLMs orchestrate, programs execute. Deterministic work belongs to scripts. LLM judgment handles design decisions, diagnosis, review.
• Density. Every word carries instruction, rule, or decision. Cut everything else.
• Breadth over depth. Right context ensures correctness. Unfocused context adds cost.
• Structural enforcement. Exit codes enforce what instructions can't. Quality gates are automated, not advisory.
• Everything pipelines. Complex work decomposes into phases. Phases have gates. Gates prevent cascading failures.

Full design philosophy: PHILOSOPHY.md

Maintenance

Two report-only scripts surface upkeep work; both print a digest and never edit, delete, or block.

• python3 scripts/harvest-corrections.py clusters captured user corrections by routed domain and suggests one-line doc fixes. Run weekly by habit, or schedule it via /schedule.
• python3 scripts/stale-skill-scan.py --top 20 ranks stale skills/agents as pruning candidates. Run quarterly; see docs/deprecation-template.md.

Scheduled work follows the same boundary as everything else: judgment uses agents; repeatable plumbing uses scripts.

Contributing

See CONTRIBUTING.md.

License

MIT. See LICENSE.