Architecture diagrams
JARVIS (ProjectHelm) — RETIRED
Retired 2026-07-22. Former Plan → Agent → Debug control plane; archaeology diagrams only. Operating system is now Cursor rules + skills per product repo.
Canonical architecture for JARVIS. Linked from the project essay; not listed in Blog.
Plan → Agent → Debug (CEO loop)
RETIRED 2026-07-22. Do not start. See
RETIRED.md. Replaced by Cursor rules + skills in each product repo.
Former default product path (cockpit / py -3 -m apps.agents). Entry points now refuse to run.
Archaeology only. State was io/logs/agent_loop/{project_id}.json.
Legacy OODA autopilot
perpetual_worker.py Observe → Orient → Decide → Act remains available for milestone / strategy experiments. It is not the CEO default. Prefer Plan → Agent → Debug for product work; do not run both loops as primary drivers on the same project.
Cockpit label: Legacy OODA Autopilot.
Away-from-IDE notifications
Webhook from SLACK_WEBHOOK_URL in .env or cockpit-saved io/logs/runtime_secrets.json. Toggle: slack_enabled in runtime_user_settings.json.
Channel: use a dedicated #jarvis Incoming Webhook (do not reuse the Orbit #career-ops weekly webhook). Create #jarvis in Slack, then add a webhook pointed at that channel.
ProjectHelm control plane
Local foreman for multi-agent coding: you set the charter / north star; JARVIS runs Plan → Agent → Debug via the Cursor SDK, appends outcomes to an event ledger, and exposes status + kill switches in the web cockpit.
Merge stays human. Fully autonomous shipping is out of scope.