- TypeScript 68%
- HTML 17.7%
- Shell 8.6%
- Python 5.7%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PRcFbeesDv5yn4JkAhLbtS |
||
| .agents/plugins | ||
| .claude | ||
| .claude-plugin | ||
| development-skills | ||
| labs | ||
| personal | ||
| plugins | ||
| scripts | ||
| .gitignore | ||
| ci-band-prototype.html | ||
| CLAUDE.md | ||
| README.md | ||
Personal Claude Code Plugin Marketplace
labs turns Claude Code into a small delivery team. Give it a task and it plans the work, builds it, has fresh agents review and test it, and opens the merge request. development-skills gives those agents shared rules for writing Go and SolidJS.
What labs does
/labs:compose <task>: ship a task as a merge request. Labs sizes the task and picks one of three tiers. An opus planner writes acceptance criteria and haiku implementers build against them in parallel. Fresh agents review the diff and grade the tests, a haiku builder runs the suite and lint, and labs opens the MR. Small tasks are planned and built in one thread; large ones give each parallel subtask its own git worktree.
/labs:review-army: the review you run before something important ships. Up to eleven reviewers read the same diff, each for one thing: spec, correctness, security, blast radius, concurrency, tests, simplicity, performance, operability, standards and tech debt. It runs your suite, mutates your guards to prove the tests catch them, and sends a verifier to disprove every blocking finding before it reaches you.
/labs:triangulate <bug>: a root cause you can trust. Three agents investigate from unrelated angles, blind to each other, and agree on one cause backed by a failing test.
Why not just ask Claude?
- The reviewer isn't the author. Every review is a fresh agent that sees the plan and the diff, so it has no stake in the shortcuts the implementer took.
- Tests come from the plan. On medium and large tasks the test-writer derives each assertion from the acceptance criteria, then breaks the code to prove every test can fail.
- Findings get weighed, then fixed. Each review finding is triaged FIX or DECLINE with evidence. Declines are logged, so the next reviewer doesn't raise them again.
- Each model does the job it's good at. Opus plans, haiku builds and writes tests against the plan, sonnet reviews and fixes, haiku runs the suite and reports only what failed.
Getting started
You need Claude Code installed and signed in. Then, inside a Claude Code session:
/plugin marketplace add https://git.bobparsons.dev/deadstyle/claude-marketplace.git
/plugin install labs@personal-marketplace
/plugin install development-skills@personal-marketplace
Restart Claude Code so the plugins load, then run /plugin to confirm both show as installed. The repo is public, so no account is needed.
Which to install:
labs: anyone. It works in any language: hand it a task with/labs:compose, review a branch with/labs:review-army. Start with its README.development-skills: if you write Go (Fiber, GORM) or SolidJS (TanStack Query). Thecoding-standardsandcommentsskills apply to any language. See its README.personal: skip it. It is wired to my homelab (AUR, Forgejo, Infisical).
To pick up new versions later:
/plugin marketplace update personal-marketplace
Plugins
| Plugin | What it ships |
|---|---|
labs |
Dev-workflow skills and hooks: the composer delivery tiers (compose → composer-inline / -lite / -full), review-army, triangulate, plan-it, triage-and-dispatch, renovate, retro / post-mortem, branch cleanup, a composer progress pane and herdr tab naming |
development-skills |
Layered skills for Go (Fiber, GORM) and SolidJS (TanStack Query) projects: coding-standards, comments and copy across languages, plus go and solid, which load their Fiber, GORM and TanStack Query guides on demand, with version supplements. A repo's own conventions skill sits on top and wins |
personal |
Non-coding tools: /audit-aur-updates, /update-services (homelab Renovate walk) with infisical-deployment, and /setup-statusline |
Codex
Two Codex plugins are registered in .agents/plugins/marketplace.json:
Codex Workflows: the composer tiers, discovery, diagnosis, eleven review lenses, Renovate, audits and retrospectives, plus the Codex-specificforgejo-clihelper.Development Skills: sharedcoding-standardsandcomments, Go/Fiber/GORM and SolidJS/TanStack Query guidance with version supplements.
Install both so Codex Workflows workers can follow the shared development rules. Codex entrypoints are skills/<name>/SKILL.md; bundled worker contracts are explicitly loaded role references. Codex Workflows is maintained independently with native Codex agent dispatch and permission handling.
Layout
.claude-plugin/marketplace.json # Claude plugin registry
<plugin>/
├── .claude-plugin/plugin.json # name, version, description
├── agents/*.md # subagents (YAML frontmatter)
├── commands/*.md # /plugin:command entry points
├── skills/<name>/SKILL.md # model-invoked skills
└── hooks/ # optional
Validate a plugin with claude plugin validate <plugin-dir>.