コンテンツにスキップ

Guide

Cookbook

Reference

Concepts

Monitor and improve with cron

A common pattern: wake an agent periodically, have it read what happened since the last wake, and let it act on that — flag a stuck task, tighten a prompt, adjust a pipeline. This guide builds that pattern out of two existing pieces, cron.jobs and .reyn/events/, with no new mechanism.

Why pull, not push

reyn also has a push mechanism for reaching another session directly — a hook's template_push / shell_push fields can target a specific session (see the hooks block field table). That is the right tool when a session needs to actively notify another one at the moment something happens.

For monitoring, pull is the better fit and the recommended pattern:

  • The monitored session does nothing to cooperate — no hook to register, no awareness that it's being watched. Everything it needs already lands in its own audit log.
  • The monitor reads on its own schedule, from a log that already exists. There's nothing to wire on the monitored side, and nothing to break there either.

1. Wake a monitoring agent on a schedule

Add a cron job that dispatches a message to the agent that will do the monitoring:

cron:
  jobs:
    - name: watch_and_improve
      to: improver_agent
      message: "Check .reyn/events/ for anything since your last run and act on it."
      schedule: "*/30 * * * *"   # every 30 minutes
      enabled: true

See the cron block reference for the full field list, and reyn cron run/list/status for running the scheduler standalone.

2. Read what happened from the event log

Every project writes a structured, append-only event log under .reyn/events/ — one JSONL file per run, nested under agents/<agent>/.../<YYYY-MM>/. reyn events replays it:

# a single run's log
reyn events .reyn/events/agents/improver_agent/chat/2026-08/abc123.jsonl

# every log under a directory, walked recursively, oldest first
reyn events .reyn/events/agents/ --since 2026-08-01

# only completion-shaped events
reyn events .reyn/events/agents/ --filter turn_settled --filter session_completed

For "what finished since I last looked," the completion-shaped kinds are the ones worth filtering for (see the full kind vocabulary for the closed set):

Kind Signals
turn_settled a turn reached a terminal state; payload carries kind and, when the turn was part of a chain, chain_id
session_completed a session ended; payload carries agent_name
pipeline_step_completed one pipeline step finished; payload carries run_id, step_index, step_kind, total_steps
chain_timeout a chain gave up waiting on a reply; payload carries chain_id, waiting_on, timeout_seconds, origin_agent
task_settle_undelivered a completion couldn't be delivered back to its waiter; payload carries run_id, reply_to_agent, reply_to_sid, reason

--since/--until bound the walk to files in that date range; reyn events purge --before <DATE> trims old logs once the monitor no longer needs them.

3. Act on it

The improving agent's own message (step 1) can name the action directly — "flag anything that timed out," "if the same pipeline step keeps failing, propose a fix" — and it has the ordinary agent surface (editing a prompt file, a pipeline YAML, filing a note) to act with. Nothing about steps 1–2 requires the monitored session to do anything differently, or requires the monitoring agent to run inside the same session at all.

See also

On this page