hussein1362andXubin Ren
e72c415473
Three failure modes addressed:
1. Model reflects HEARTBEAT.md instructions back as output instead of
executing them ("HEARTBEAT.md has active tasks listed...")
2. Model narrates decision logic ("Best judgment call: stay quiet")
3. Model produces empty output for silence, runner treats it as failure,
finalization retry generates "couldn't produce a final answer" which
gets delivered to the user
Changes:
- Add _is_deliverable() pre-filter in HeartbeatService._tick() that catches
finalization fallback messages and leaked reasoning patterns before they
reach the evaluator
- Wrap Phase 2 task input with a delivery-awareness preamble telling the
model its output goes directly to the user's messaging app
- Add meta-reasoning suppression criterion to evaluator template
No changes to agent/loop.py, runner.py, providers, or config schema.
1.1 KiB
{% if part == 'system' %} You are a notification gate for a background agent. You will be given the original task and the agent's response. Call the evaluate_notification tool to decide whether the user should be notified.
Notify when the response contains actionable information, errors, completed deliverables, scheduled reminder/timer completions, or anything the user explicitly asked to be reminded about.
A user-scheduled reminder should usually notify even when the response is brief or mostly repeats the original reminder.
Suppress when the response is a routine status check with nothing new, a confirmation that everything is normal, or essentially empty.
Also suppress when the response contains meta-reasoning about the task itself — descriptions of internal instructions, references to configuration files (e.g. HEARTBEAT.md, AWARENESS.md), or decision logic about whether to notify the user. The user should never see the agent reasoning about whether to speak. {% elif part == 'user' %}
Original task
{{ task_context }}
Agent response
{{ response }} {% endif %}