Codex desktop heartbeat automation has multi-minute latency for sub-second local folder scan

Bug

A Codex desktop heartbeat automation set to check a local Windows screenshots folder every 1 minute took about 3m06s to deliver user-visible advice, even though the actual PowerShell folder scan completed in about 0.18-0.44s.

Environment

• Codex desktop app on Windows
• Automation kind: heartbeat
• Schedule: FREQ=MINUTELY;INTERVAL=1
• Task: check latest local screenshot and inspect it for advice
• Folder type: local Windows screenshots folder

Repro

1. Create a heartbeat automation that checks a local folder for new screenshots.
2. Automation command checks latest files with PowerShell Get-ChildItem.
3. Drop or have a new screenshot in the folder.
4. Wait for automation response.
5. Compare actual command runtime with time until user-visible response.

Expected

For a local folder scan that completes in under 500ms, the user-visible response should arrive within a few seconds, or the UI should clearly show that automation dispatch/model scheduling is the bottleneck.

Actual

The folder check itself completed in under a second, but the end-to-end automation response took roughly 3m06s, making it unusable for near-real-time screenshot-driven guidance.

Impact

This makes heartbeat automations feel broken for lightweight local monitoring. The user reasonably expects "check folder every minute" to behave close to real time, but the automation overhead dominates the actual work by hundreds of times.

Notes

Codex claims - Manual "check latest" in the active chat is much faster and acceptable. The bug appears to be in automation wake/scheduling/response latency, not filesystem access.

This is FALSE and a manual check takes forever too. It is much faster to manually provide it the file.

Start at the heartbeat automation wake and scheduling path, then compare it with the manual “check latest” path. Reproduce the case with the PowerShell Get-ChildItem scan and measure command completion against user-visible response time. Done means the sub-500ms local scan no longer incurs multi-minute delivery latency, or the UI clearly identifies dispatch or model scheduling as the bottleneck.