Index of workJWLs OS
An AI chief of staff,
run like real infrastructure.
It started as a script that sent me a morning briefing. It is now versioned, multi-tenant infrastructure with two clients on it, running continuously on my own hardware.
Specification
- Release
- 0.8.11, shipped July 14 2026
- Clients in production
- 2, onboarded remotely
- Core suite
- 250 tests, plus per-component suites
- Scheduled agents
- 7, across two machines
- Round trip
- About 5 seconds, text to executed action
- Ingress
- Telegram, iMessage
- Runtime
- Python, Docker, one container per client
- Edge
- Cloudflare Worker in front of per-client queues
What was actually hard
Not the model, and not the prompting. The hard problems were trust and liveness: keeping a system alive on someone else’s behalf, and keeping their data away from everyone else’s.
- Drill
Channel proved live before merging
A text to the bot's dedicated Apple ID routed through the real loop and scheduled an actual calendar event, reply back in about five seconds. A stranger's text on the same channel was silently dropped by the default-deny allowlist.
- Drill
Frozen on purpose, recovered on its own
I froze the watcher process mid-conversation. The watchdog restarted it inside sixty seconds and replayed the message sent during the freeze. Zero loss, no duplicates.
- P0
Layered watchdogs, written after an incident
A real silent hang produced two layers: an in-process monitor that restarts stalled loops, and a host-level monitor that restarts containers and alerts me. Shipped the next morning rather than filed in a backlog.
- Standing
Isolation as a hard rule
Per-client containers, vaults, and secrets. Recorded consent. Default-deny allowlists on every ingress. Development never touches a production client.
On authorship
AI writes much of the code. The judgment about boundaries, about what to test, and about what not to build is the job. Spec first, isolated worktree, offline tests green before anything touches production, and a written handoff on every release.