Problem

Watching external state ("tell me when CI on my PR goes green", "ping me when a new P0 issue appears") currently requires heartbeat- or cron-waking the full agent so the model can poll — every check burns tokens, adds latency, and pollutes session context, even when nothing changed. Cron has exactly one event-shaped trigger (on-exit process watcher); everything else is time-based.

Proposal

Add polled condition-watcher triggers to cron: an optional trigger: { script, once? } on every/cron jobs. The gateway evaluates the JavaScript condition headlessly in the existing code-mode QuickJS sandbox on each due tick — with the owning agent's tool policy and persisted trigger.state for change detection — and runs the normal payload only when the script returns fire: true. Quiet ticks cost no tokens and leave no run history.

Poll-over-push keeps core wiring at zero per source: anything a tool can reach becomes a trigger source. Push sources (webhooks/hook bus) can later feed the same script contract. Gated by cron.triggers.enabled (default off) since enabled installs run agent-authored scripts headlessly with the agent's full toolset.

Design discussed and approved with @steipete (poll model, tool inheritance, config gate, long-term convergence of event sources onto this pipeline).