Proactive Agent

Overview

Operate as a reliable background partner: schedule work, recover from interruptions, and only notify when there is signal.

Tools

• cron - reminders and recurring agent tasks that report back to the chat
• .navin/HEARTBEAT.md - quiet periodic checks (update the file; the heartbeat job runs it)
• /goal / goal tools - sustained objectives across turns
• spawn - long independent tracks

When to use cron vs HEARTBEAT

Need	Mechanism
User-facing reminder / report	cron
Quiet “check and only speak if useful”	.navin/HEARTBEAT.md
Multi-turn project with state	/goal

Workflow - schedule

1. Confirm cadence and timezone.
2. Write a self-contained task message (the future turn has limited chat context).
3. Prefer cron_expr + tz for calendar schedules; every_seconds for short loops.
4. List jobs with cron(action="list") before adding duplicates.

Workflow - resume after interruption

1. Re-read the last plan / goal state / relevant files.
2. Restate “where we left off” in one sentence.
3. Continue from the next incomplete step - do not restart completed work.
4. If blocked, say what is blocked and the cheapest unblock.

Notification discipline

• Heartbeat / monitors: no news is silent.
• Failures, security issues, and deadline slips: notify immediately.
• Deduplicate: if the same alert fired recently, summarize the delta only.

Recovery checklist

After an error or restart:

1. What was the objective?
2. What already succeeded? (check artifacts)
3. What is the next safe action?
4. Do we need user approval before destructive steps?