GUIDE / OPERATING THE LOOP

Run a loop you can audit.

Install to first loop is four commands (they are on the home page). This page covers what runs after that: the roles, the console, and the files that make the loop recoverable.

The loop, end to end

Advisor plans → queue records → executor works → review gate → human decides. Every handoff lands as a file first: advisor output is saved before it is acted on, queue entries are an append-only ledger, and decisions wait for a person. A crash at any point recovers from disk, not from memory.

Roles and the master session

Sessions self-report a role: master (your main agent), project-loop (owns one project's state), queue-manager (serializes browser work), executor, and governor (quota + reserve policy). arthur init --master-agent claude-code designates the master; the console warns when no live master session is reporting.

The console

arthur web serves the localhost console: a projects sidebar, per-project tabs (Loop, Tasks, Logs, Prompts), the pan-and-zoom loop canvas with a live log tail, and a stats rail of the role sessions. Tasks can be imported read-only from Atlas Tasker and converted to jobs one by one — always by a human, never in bulk.

Prompts, adapters, trackers

Role prompt packs live in prompts/; per-edge prompts (prompts/edges/advisor--queue.md and friends) attach guidance to a specific handoff, with per-project overrides. Advisors, executors, and trackers are adapters — mix a ChatGPT browser advisor with a Claude Code executor and an Atlas Tasker tracker, or run everything manual. See the adapter contract and the workflow runbook for the full picture.