[Bug] CronCreate registration silently dropped after agent /compact in agent-mode session

Resolved 💬 4 comments Opened Apr 30, 2026 by ThatDragonOverThere Closed May 4, 2026

Summary: A Claude Code agent running in agent-mode (dispatched by a parent session) loses its CronCreate registration when it compacts. After compaction, cron ticks do not resume — the agent goes dark with no warning.

Version: v2.1.121 (not yet tested on v2.1.123)

Steps to reproduce:

  1. Parent session (e.g., Opus, non-agent mode) dispatches a child agent in agent mode
  2. Child agent registers a recurring schedule via CronCreate (e.g., every 15 minutes for polling work)
  3. Child agent accumulates context and compacts
  4. Observe: cron ticks do not fire after compaction completes

Expected: CronCreate registrations persist through compaction — agent resumes receiving scheduled ticks after compact.

Actual: Cron schedule is silently dropped. Agent no longer receives timed ticks. No error, no warning. The only recovery paths are: (a) operator manually restarts the agent window, or (b) parent session re-dispatches inline.

Impact:

  • Autonomous sessions that cross a compaction boundary lose their scheduling silently
  • Parent session waiting for filesystem check-ins from the cron-driven child gets indefinite silence — indistinguishable from the child being alive but idle vs. cron-dead
  • Compounds with the model-downgrade bug (#54426): both result in autonomous agent sessions going dark at the context-limit boundary, both fail silently

Platform: Windows, observed in agent-mode child dispatched by non-agent-mode parent

Workaround: Keep the parent/dispatcher window in non-agent mode so it accumulates less context, compacts less often, and can re-dispatch workers on natural wakes. Reduces exposure but does not fix the underlying cron-state persistence gap through compaction.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗