Skip to main content

Knowledge > Runbooks > Agent Ops > How Claude Code Agents Work in This System

How Claude Code Agents Work in This System

Reference guide for understanding how Claude Code agents are configured, what tools they have access to, and how they coordinate across the ChurchWiseAI portfolio.

Agent Entry Points​

Agents are invoked via the Claude Code CLI (claude) or through the /chief-of-staff skill. The session-start hook at ~/.claude/hooks/session-start.sh automatically injects system context at the start of every session, ensuring agents always have current portfolio context.

Skills (Slash Commands)​

Skills extend agent capabilities for common workflows. Invoke with /skill-name in the conversation:

SkillPurpose
/chief-of-staffEnhanced executive assistant — daily briefing, task dispatch, coaching
/commitStructured git commit with correct format
/review-prPull request review with QA lens
/dispatching-parallel-agentsLaunch multiple Claude agents for concurrent tasks

Tool Access​

MCP Plugins (Always Available)​

PluginKey ToolsUse For
Supabasemcp__plugin_supabase_supabase__execute_sql, apply_migration, list_tablesAll database queries and migrations
Cartesiamcp__cartesia-mcp__text_to_speech, list_voicesVoice testing and TTS generation
Gmailmcp__claude_ai_Gmail__gmail_*Read/search emails as john@churchwiseai.com
Google Calendarmcp__claude_ai_Google_Calendar__gcal_*View and manage calendar
Context7mcp__plugin_context7_context7__*Look up current library documentation

CLI Tools (All Logged In)​

ToolAccountKey Commands
vercelchurchwiseai-5386vercel logs, vercel env ls, echo "val" | vercel env add NAME production, vercel --prod
stripechurchwiseai@gmail.comstripe customers list, stripe subscriptions list — add --api-key $STRIPE_LIVE_SECRET_KEY for live mode
ghJohnMoelkergh pr create, gh issue list, gh run view

LiveKit CLI (Voice Agent Deploy)​

The voice agent deploys to LiveKit Cloud via the LiveKit CLI via the LiveKit CLI:

C:\dev\lk.exe agent deploy --project cwa-voice --silent

The Cartesia CLI is legacy and no longer used for deployments. The voice agent runs on LiveKit Agents v1.5, deployed to LiveKit Cloud. Never attempt pip install cartesia or npm install cartesia on Windows — it will fail.

Memory and Knowledge Architecture​

Persistent Memory​

Agent memory is stored in C:\Users\johnm\.claude\projects\C--dev\memory\ as markdown files. MEMORY.md is the index. Memory persists across sessions and contains project state, feedback, and architectural decisions.

Knowledge Hierarchy​

Always read information sources in this order:

  1. C:\dev\knowledge\ — canonical truth for all business knowledge (pricing, features, policies, brand)
  2. Per-codebase CLAUDE.md — code-level context (file paths, build commands, safety rules)
  3. C:\dev\knowledge\drafts\ — proposals/in-progress, verify before trusting
  4. Other .md files — check if a canonical version in knowledge/ exists first
  5. C:\dev\knowledge\archive\ — contains ⚠️ ARCHIVED header, never use for decisions

Plans​

When agents develop multi-step plans, they save them to C:\Users\johnm\.claude\plans\ as markdown files. This allows work to resume across sessions.

Multi-Agent Coordination​

Multiple Claude agents may be running concurrently across sessions. Coordination happens through:

  • Git branches — each agent creates its own feature branch (see multi-agent-git.md)
  • DECISION_LOG.md — append-only journal of significant decisions
  • FOUNDER_ACTIONS.md — queue of items requiring founder intervention
  • FEATURE_REGISTRY.md — prevents two agents from building the same feature in different places

What Agents Should Do Without Asking​

  • Read any database table (SELECT queries)
  • Read Stripe data (products, prices, customers)
  • Read and SET Vercel env vars
  • Run builds (pnpm build)
  • Git operations on feature branches (commit, push)
  • Search Gmail and Calendar for context
  • MailerLite API operations

What Agents Must Confirm Before Doing​

  • Writing to the production database (INSERT, UPDATE, DELETE)
  • Creating/modifying Stripe products or prices in LIVE mode
  • Deploying to Vercel production (vercel --prod)
  • Pushing to main or master branches
  • Sending emails to customers
  • Voice agent deployment (push to main + cartesia connect)
  • Deleting Vercel env vars

See Also​