The web console
A single-operator control surface for one loop — the browser twin of arthur status, plus the handful of actions that genuinely belong to a human.
arthur web # serves http://127.0.0.1:7433 and opens it
arthur web --root ~/loop --port 8080 --no-open
What you are looking at
- The loop canvas — advisor → queue → in flight → executor → review gate → human, as a live map. Edges light up only where work is actually moving; the human node pulses when the loop is blocked on you.
- Topbar readouts — the state chip, the quota mini-meter (GREEN/YELLOW/RED with percent left), and the machine chip with host memory and CPU pressure.
- Tabs per project — Loop, Tasks, Logs, and Prompts, with a live log tail underneath.
- Needs you — the drawer of open human decisions, answerable in place. Optional browser alerts while the console is open.
- Queue control — create jobs with the same validation as the CLI, inspect, recover, and drag-reorder with an explicit confirm step.
Editing the loop from the console
Each edge on the canvas carries a small "P" badge that deep-links into the prompt workbench — every handoff prompt is readable and editable, with per-project overrides. See Prompts for the tiers.
Security model
- Binds
127.0.0.1only and rejects non-localhostHostheaders. - A per-process session token guards every write; reads and writes journal to the same files the CLI uses.
- No accounts, no cloud, no analytics.
Deliberately absent
No start/stop buttons, no claim/submit/complete, no "approve plan" bypass, and no scheduler controls. Agent-owned transitions stay with agents; lifecycle stays in the terminal (arthur init, arthur web, arthur watch). The console shows the truth and takes only human-shaped actions.
A stale browser lock (an advisor conversation that died mid-flight) can be broken from the topbar. A fresh lock cannot — that would yank the conversation from a live agent.