sglyon-claude-plugins

sglyon-claude-plugins

Spencer Lyon's Claude Code plugin marketplace.

Installation

# Add the marketplace
/plugin marketplace add sglyon/sglyon-claude-plugins

# add a plugin from the marketplace
/plugin install sgldev

Available Plugins

sgldev

Personal development tools with persistent per-agent expertise.

Agents (5 specialized code reviewers with persistent mental models):

• sglyon-ash-reviewer — Ash Framework
• sglyon-elixir-reviewer — Elixir
• sglyon-liveview-reviewer — Phoenix LiveView
• sglyon-python-reviewer — Python
• sglyon-typescript-reviewer — TypeScript

Skills (8):

• chartroom — Chart generation
• showboat — Demo document creation
• team-lead — Team lead workflows
• rodney — Browser automation via Chrome CDP
• create-agent-skills — Skill authoring toolkit
• mental-model — Per-agent expertise lifecycle management
• expertise — Dashboard showing agent knowledge state
• conversational-response — Concise response style for multi-agent workflows

Hooks (5):

• SessionStart — Loads expertise system awareness when .expertise/ exists in a project
• SubagentStart — Injects expertise instructions into every spawned subagent via additionalContext
• SubagentStop — Blocks subagents from stopping until they've updated their mental model (if meaningful work was done)
• PostToolUse (Write|Edit) — Validates YAML syntax and line limits for mental model files
• PreCompact — Preserves expertise awareness across context compaction

Agent Expertise System

Each agent maintains a personal mental model file (.expertise/models/<agent>.yaml) that persists across sessions. Agents accumulate codebase-specific patterns, architectural observations, and learnings over time.

Designed to complement the compound-engineering plugin — sgldev owns per-agent expertise while compound-engineering owns institutional knowledge (docs/solutions/).

Requirements: uv (hook scripts use PEP 723 inline metadata to manage dependencies automatically via uv run)

Setup in a project:

bash $CLAUDE_PLUGIN_ROOT/scripts/init-expertise.sh

This creates .expertise/ with empty model files for each agent. Models are git-tracked and grow as agents work on the project.

How It Works

The expertise system delivers instructions automatically via hooks — no manual setup in agent files needed:

• SessionStart hook — Fires when a Claude Code session begins. Injects expertise lifecycle instructions into the main session agent.
• SubagentStart hook — Fires when any subagent is spawned (e.g., team-lead spawning a reviewer). Injects the same expertise instructions into the subagent's context.

Both hooks run the same script, so every agent — plugin-provided or project-defined, main session or subagent — gets the full instructions automatically.

Using Expertise with Custom Project Agents

Any project can define its own agents in .claude/agents/ and have them participate in the expertise system. No changes to the agent's .md file are needed — the SubagentStart hook handles instruction injection automatically.

1. Add your agent to .expertise/config.yaml:

agents:
  - ash-reviewer
  - elixir-reviewer
  - liveview-reviewer
  - python-reviewer
  - typescript-reviewer
  - team-lead
  - my-custom-agent    # your agent

2. Run the init script to create the model file:

bash $CLAUDE_PLUGIN_ROOT/scripts/init-expertise.sh

That's it. The hooks handle the rest.