Yorrick's Agent Plugins

A collection of agent plugins by Yorrick Jansen (Claude Code, Codex, and opencode).

Installation

Every plugin here works with Claude Code, Codex, and opencode.

Claude Code

claude plugin marketplace add yorrick/agent-skills
claude plugin install <plugin-name>@yorrick

Codex

codex plugin marketplace add yorrick/agent-skills
codex plugin add <plugin-name>@yorrick

opencode

{
  "plugin": ["yorrick-agent-skills@git+https://github.com/yorrick/agent-skills.git"]
}

See .opencode/INSTALL.md for the update story and caveats.

Restart the CLI afterwards.

Updating

Both harnesses auto-update — differently, and Codex's behaviour is undocumented.

	Claude Code	Codex
Auto-update	Per-marketplace toggle — off by default for third-party, on for Anthropic's own	Always on while plugins are enabled; no dedicated opt-out
Fires	Once per session, at a random point within 10 min of start	At app-server startup (TUI, codex exec, desktop app)
Updates the installed plugin?	Yes, where enabled — version pins respected	Yes — reinstalls whenever the remote revision moved
Takes effect	After /reload-plugins or restart	Immediately; caches invalidated live

Enable it on Claude Code with /plugin → Marketplaces → yorrick → Enable auto-update.

Codex's check_for_update_on_startup config flag does not disable this — it only governs Codex's own self-update prompt (which the desktop app ignores anyway: codex#18543, closed as not planned). The plugin auto-upgrade has no off switch short of [features] plugins = false.

To update by hand:

# Claude Code — two steps; refreshing the marketplace alone does not upgrade
claude plugin marketplace update yorrick
claude plugin update <plugin-name>@yorrick

# Codex — one step, despite the help text saying "snapshots"
codex plugin marketplace upgrade

Verify against the cache rather than plugin list, which only reflects config:

ls ~/.claude/plugins/cache/yorrick/<plugin-name>/
ls ~/.codex/plugins/cache/yorrick/<plugin-name>/

Codex auto-update is a supply-chain path. Anything pushed to main here reaches every installed machine at next Codex startup, with no review step and no notification. Protect this branch accordingly.

opencode installs the repository as a git plugin spec and may pin the resolved commit in a lockfile, so a restart alone may not fetch a newer revision. If changes do not appear, remove the cached package (this also clears branch-pinned installs) and restart:

rm -rf ~/.cache/opencode/packages/yorrick-agent-skills*

Removing

claude plugin uninstall <plugin-name>@yorrick
codex plugin remove <plugin-name>@yorrick
# opencode: remove the plugin line from opencode.json, then clear the package cache above

Removing the marketplace is a bigger hammer — in Claude Code it also uninstalls every plugin that came from it. On Codex, removing a plugin while leaving the marketplace registered may see it reinstalled by the startup auto-upgrade; check the cache path above after a restart.

Renamed from claude-code-plugins

This repo was yorrick/claude-code-plugins. Existing registrations keep working through GitHub's redirect — refresh with the update commands above. Do not remove and re-add the Claude marketplace: removing its last registration also uninstalls its plugins.

claude plugin marketplace add yorrick/claude-code-plugins

Plugins

dev-loop

Automated development loop that composes existing Claude Code commands into a full lifecycle: brainstorm, plan, implement, create PR, then iteratively review (simplify + code review + security review) until clean.

Commands:

• /dev-loop — Full lifecycle: interactive brainstorming and planning, then automated implementation and review loop
• /review-loop — Review loop only on an existing PR (skip brainstorm + implementation)

Prerequisites:

• superpowers plugin (brainstorming, writing-plans, executing-plans skills)
• code-review plugin (/code-review:code-review)
• /simplify and /security-review (built-in)
• gh CLI (for creating PRs)

claude plugin install dev-loop@yorrick

self-improve-skill

Automatically reflects on skills used during sessions and proposes improvements.

claude plugin install self-improve-skill@yorrick

pr-review

Multi-provider interactive PR code review using Claude Code, Gemini CLI, and Codex CLI.

claude plugin install pr-review@yorrick

agent-session-monitor

Meta-agent skill for monitoring and steering Claude Code, Codex, and other terminal coding agents through tmux panes. The skill content is portable Agent Skills format; local Claude/Codex discovery should symlink to agent-session-monitor/skills/agent-session-monitor.

claude plugin install agent-session-monitor@yorrick

task-status

Shows a compact, terminal-safe visual summary of the current task: what is done, what is happening now, what comes next, what was deferred, and what is blocked. It estimates each active item's relative effort and displays one weighted overall completion percentage, excluding explicitly deferred work. When conversation evidence establishes a real branch or join, it also adds a compact ASCII dependency diagram. It is strictly read-only and uses only the current conversation as evidence.

Invoke it with /task-status:task-status in Claude Code or $task-status in Codex.

claude plugin install task-status@yorrick
codex plugin add task-status@yorrick

flowchart

Creates clear process diagrams and interactive walkthroughs for Claude Code and Codex. It covers source-backed flowchart design, format selection, plain labels, and keyboard checks on the actual inline or exported surface.

claude plugin install flowchart@yorrick
codex plugin add flowchart@yorrick

visual-design-review

Turns a system design into a concise visual review of consequential choices, then carries the user's decisions into the specification. It reuses available diagram skills, including flowchart, and rendering capabilities such as Codex's visualize skill without requiring a Codex-specific host in Claude Code.

claude plugin install visual-design-review@yorrick
codex plugin add visual-design-review@yorrick

jev-router

Sends each job where it costs least, using Jev, TypeSafe's decision model, on OpenRouter.

In Codex and opencode (0.2.0), every message you type goes to Jev with two typed questions: what is the smallest model that can do this job well, and does the message only make sense inside the conversation? When Jev gives its pick at least a 60% probability, and the message is not such a follow-up, the job goes to a helper on that size's model and thinking level, which signs off with Done by <model> at <level> thinking. Codex cannot switch the main model per message, so there the main model hands the job to a helper and relays its result; opencode can, so there the message itself moves onto the helper's model.

In Claude Code (0.3), Jev instead predicts how long the job is (how many model calls), and the router prices doing it in the session, which re-reads the whole conversation on every call, against briefing a fresh subagent. It starts in shadow mode (decide and log, tell the session nothing) until its fork check passes; /jev mode shadow|capture|live switches modes. See jev-router/README.md.

It only routes inside the harness you are in, leaves headless runs such as claude -p reviews alone unless you set JEV_ROUTER=on, and never blocks a message: if Jev is slow (over 2 s) or anything fails, the message goes through as if the router were not there.

It is off until you turn it on with /jev on ($jev on in Codex), and /jev status shows what it decided and what Jev has cost. While it is on, each message (up to its first 4,000 characters) goes to OpenRouter and TypeSafe (the company that makes Jev), and in Claude Code so do the last 1,500 characters of the assistant's previous reply, in every mode, shadow included. Keep it off for private work.

claude plugin install jev-router@yorrick
codex plugin add jev-router@yorrick

supabase-security

Access-control rules for Supabase projects exposed directly to a browser via PostgREST: the three enforcement layers (GRANT, RLS, triggers) and the order they run in, the traps that cause privilege escalation, and how anon keys, JWTs and service_role differ.

Ships a read-only audit script that bundles Splinter — Supabase's own SQL linter, the engine behind the dashboard's Security Advisor — vendored unmodified, and adds four rules it does not have (policies with no FOR/TO, missing RESTRICTIVE tenant isolation, the delete-and-reinsert column bypass, and TRUNCATE, which no policy applies to). See supabase-security/README.md for the credit and licence position.

claude plugin install supabase-security@yorrick

Local Development

For testing or development, load a plugin directly:

claude --plugin-dir ./dev-loop
claude --plugin-dir ./self-improve-skill
claude --plugin-dir ./agent-session-monitor
claude --plugin-dir ./task-status
claude --plugin-dir ./visual-design-review

opencode resolves a filesystem path as a plugin spec, so a checkout can be loaded directly:

{
  "plugin": ["/path/to/agent-skills"]
}