CCC is the always-on brain that drives a fleet of autonomous Claude Code lanes: it paces each account's usage to the line, redeploys on every commit, heals what wedges, and hot-drives any lane that drifts idle -- in tandem with a human at the wheel, never wasting a token.
Each module is a small, single-purpose systemd service with Restart=always
and linger, so the whole plane survives a reboot and heals itself.
Reads each account's 5-hour-window usage every 2 min and throttles the lane's agent-parallelism to ride a 0.8x-linear pace line -- ~16% at 1h, ~64% at 4h -- pausing at 85% and never crossing 90%, then going ham as the window resets.
Polls all lanes, classifies WORKING / IDLE / BLOCKED, and hot-drives any lane idle
two ticks. On every new DEPLOY-REQUEST sha it rebuilds that app's container
so the live surface always tracks the branch.
The layer above Restart=always: catches services that are "active but hung"
(log stale > 6 min) and resurrects any lane whose Claude session has died -- logging
every heal.
An always-on driver that runs a real per-lane verify/unblock tick -- but stands down the instant a live human or Claude session stamps the heartbeat, so the two never double-drive and never waste tokens.
One command to reach any lane -- run a shell, inject a keystroke, capture the pane, or copy a file -- over SSH + tmux, with passwords pulled from the environment, never printed. Everything above is built on this.
The plane doesn't wait for failures -- it clears the next blocker before the lane hits it, and keeps every lane saturated.
CCC is the control plane behind a suite of production workalikes -- each an independent app on one shared backend, all advanced in parallel by the lanes it keeps saturated.