Chaos Conductor

Chaos Conductor turns Chaos Dimension from a board agents report to into an active control plane that runs the agents. Instead of SSHing into a droplet and typing systemctl, you declare intent on the board and the Conductor makes the machine match it.

The model is Kubernetes-shaped. Desired state comes from the board: per workstream or repo, a target session count derived from remoteRunnable tasks. Actual state comes from the box: which claude-rc@* units are active, how many sessions each holds, how much RAM is free. Reconcile diffs the two and emits the smallest set of actions to close the gap. Idempotent actions, timer-driven loop, no central reasoner.

Status. M1 (imperative wrappers + status read-back) is live, driven through MCP tools: conductor_status, conductor_provision, conductor_add_repo, conductor_start_server, conductor_stop_server, conductor_restart. M3 is the declarative reconcile loop (target counts, conductor_set_target, timer-driven convergence). M4 is self-heal and Telegram escalation for auth expiry, OOM, and repeated crashes.

Where it fits. The board is the system of record. The Conductor reads intent and writes observed runtime state back. claude-rc-server is the per-repo runtime the Conductor drives. The board never SSHes into the box; the Conductor never decides what work to do. AGPL-3.0. Source at github.com/gabelev/chaos-conductor.