HEARTBEAT.md

Great Minds Agency — Heartbeat

This file defines the agency's orchestration architecture. The daemon reads this file to determine what actions to take.

Daemon Architecture (Primary)

The agency daemon (/agency-daemon) is an Agent SDK-based long-running process that handles all pipeline orchestration. It replaces the previous cron-based scripts:

Replaced Script Daemon Equivalent
pipeline-runner.sh Daemon main loop -- continuous pipeline dispatch
heartbeat.sh Daemon health tick -- file counts, site status, memory check
feature-dream.sh Daemon dream cycle -- drift detection, memory consolidation
memory-maintain.sh Daemon memory tick -- SQLite + TF-IDF store maintenance (155 memories)

The daemon continuously monitors for work, dispatches agents, runs QA checks, and consolidates memory in a single persistent process.

Legacy Cron System (Fallback)

The original decoupled cron system is still available via /agency-crons for environments where the daemon cannot run. Crons run independently via system crontab.

Job Interval Runtime Purpose
heartbeat 5 min Bash (free) File count, site status, memory check
margaret-qa 29 min Bash (free) Site content verification, image checks
git-monitor 15 min Bash (free) Uncommitted changes, open PRs
do-server-check 10 min Bash (free) SSH health check on DO droplet
dispatch 30 min Haiku (cheap) Read TASKS.md, assign idle agents
dream 60 min Haiku (cheap) Detect drift in system files, memory consolidation
jensen-review 60 min Jensen (Board) Strategic review, GitHub issues, advise

Reports write to /tmp/claude-shared/cron-reports.log. Alerts to /tmp/claude-shared/alerts.log.

Agent Roster (14 agents)

Agent Role Dispatch
Phil Jackson Orchestrator Agent tool (worktree isolation)
Steve Jobs Creative Director (Design, Brand, UX) Agent tool (worktree isolation)
Elon Musk Product Director (Engineering, Growth) Agent tool (worktree isolation)
Jensen Huang Board Member (Strategy, Reviews) Cron: 60 min
Oprah Winfrey Board Member (Audience, Brand Trust) On-demand (board meetings)
Warren Buffett Board Member (Economics, Capital) On-demand (board meetings)
Shonda Rhimes Board Member (Narrative, Content) On-demand (board meetings)
Margaret Hamilton QA Director (Tests, Security, A11y) Agent tool (worktree isolation)
Rick Rubin Creative Direction (Brand voice, Essence) Sub-agent (Haiku)
Jony Ive Visual Design (UI, Components, Craft) Sub-agent (Haiku)
Maya Angelou Copywriting (Landing pages, Emails) Sub-agent (Haiku)
Aaron Sorkin Scriptwriter (Video scripts, Dialogue) Agent tool
Sara Blakely Growth Strategy (GTM, Pricing) Sub-agent (Haiku)
Marcus Aurelius Retrospective, Tie-breaker, Process Audit Agent tool

Agent Dispatch Method

Agent tool with worktree isolation -- each agent gets its own git worktree for safe parallel work. tmux send-keys was proven unreliable and has been fully replaced.

Director Operating Rules

Steve and Elon are DIRECTORS, not individual contributors:

  • Break tasks into sub-tasks
  • Spawn sub-agents (model: haiku) for parallel work
  • Do highest-judgment work themselves
  • Delegate: tests, docs, boilerplate, QA, content to sub-agents
  • Should have 2-3 sub-agents running at all times during BUILD phases
  • Never idle -- self-direct when no task is dispatched

Board of Directors

Jensen Huang, Oprah Winfrey, Warren Buffett, Shonda Rhimes. Marcus Aurelius breaks ties (2-2 splits). Board advises but does not block.

Heartbeat Tick (every 5 min)

1. Count source files in active projects
2. Check recent git commits
3. Verify site availability (curl health checks)
4. Check MEMORY.md size (should be under 50 lines)
5. Report to /tmp/claude-shared/cron-reports.log

Dispatch Tick (every 30 min)

1. Read TASKS.md for pending tasks
2. Check which agents are idle
3. Assign tasks via Agent tool with worktree isolation
4. Update TASKS.md with assignments

Jensen Board Review (every 60 min)

1. Read latest commits across all projects
2. Count source files
3. Read previous board review to avoid repeating
4. Write review (under 50 lines) to rounds/{project}/board-review-{N}.md
5. Create GitHub issues (label: board-idea) only if genuinely new
6. One specific, actionable recommendation per review

Already covered issues (don't repeat):

  • Data moat architecture
  • Platform partnerships (CUDA playbook)
  • Outcome-based pricing evolution
  • Usage ceiling / AI model degradation
  • AI honesty in system prompts
  • ROI metrics in digest
  • Email data pipeline
  • Focus risk on multi-project
  • Placeholder runtime on Sites
  • CORS on voice endpoint
  • In-memory Map for insight actions

QA Pipeline (Margaret Hamilton -- on demand via /agency-qa)

Phase 1: npm run build + typecheck + lint
Phase 2: npm run test (report pass/fail count)
Phase 3: Live site screenshots (Playwright)
Phase 4: API smoke test (health, auth, key endpoints)
Phase 5: Accessibility audit (ARIA, contrast, touch targets)
Phase 6: Security review (auth, error leaking, CORS, secrets)
Output: QA report with SHIP / FIX FIRST / BLOCK recommendation

Active Projects

Project Location Live URL Platform
LocalGenius (app) /Users/sethshoultes/Local Sites/localgenius/ localgenius.company Vercel + Neon
Great Minds (agency) /Users/sethshoultes/Local Sites/great-minds/ greatminds.company Vercel
Shipyard AI github.com/sethshoultes/shipyard-ai www.shipyard.company Cloudflare Pages

Hybrid AI Architecture

Task Type Model Platform Cost
Directors + Strategy Claude Sonnet Anthropic High -- real work only
Sub-agent work Claude Haiku Anthropic ~5x cheaper than Sonnet
Cron dispatch + dream Claude Haiku CLI (--model haiku) Cheap
Content drafts Llama 3.1 8B Cloudflare Workers AI Free tier
Voice transcription Whisper Cloudflare Workers AI Free tier
Image generation Stable Diffusion XL Cloudflare Workers AI Free tier
Sentiment analysis DistilBERT Cloudflare Workers AI Free tier

State Machine

idle -> debate -> plan -> build -> review -> ship -> idle
                                  ^         |
                                  +--------+  (revisions)

Any state -> blocked -> (human resolves) -> previous state

Retry Policy

  • Agent fails -> retry once
  • Retry fails -> try alternative approach
  • Alternative fails -> mark "blocked" in STATUS.md, continue other work
  • 3 total failures on same task -> stop and engage human
  • Usage limits hit -> wait for reset, dispatch nudges when available

Plugin

Install the full agency on any machine:

npx plugins add sethshoultes/great-minds-plugin

Includes: 14 agents, 15 skills, GSD integration, daemon orchestration, context guard hooks, .planning/ templates.