Triggers and limits
Wake conditions can include elapsed time, impulses, drives, conflicts, or synthesized desires. Silence thresholds and cooldowns reduce overlap with active conversation. Optional noise-triggered waking is separately configurable. The checked-in configuration enables autonomy with a 600-second minimum gap, at most four cycles per hour, and at most two outgoing messages per cycle. These limits concern the wake engine; independent body and memory mechanisms can still make model requests.Sustained wake sessions
wake_session_max_rounds controls how many cognition rounds a wake may chain. It defaults to one. Wall-time, per-round message, and additional tool-step budgets constrain longer sessions. The Entity can end a wake session before reaching its cap.
Longer sessions can complete more work but also consume more inference and produce more external effects. Enable them after confirming tool permissions and delivery routes.
Per-Entity autonomy overrides
The Entity loader merges top-levelautonomy settings onto the harness configuration:
Delivery and observability
Wake output uses configured presence routes. The worker can record structured events and a detailed transcript, while chat settings decide how much activity appears to the user. Telegram/wakequiet on suppresses visible wake activity. It does not disable the wake engine or model calls.
Disabling autonomous wake
Setautonomy.enabled: false to disable the wake engine. In v1, a legacy initiative path can still run when this flag is false, and body, memory, and routine jobs are separate.
Use Pause inference or Telegram /pause when your intent is to stop all new model requests for the Entity. The shared provider control covers those other paths too. See usage and controls.
Source: presence/wake_cycle.py and presence/daemon.py.