openclaw - 💡(How to fix) Fix [Bug]: 2026.5.12-beta.6 isolated agent cron jobs failing at attempt-dispatch [2 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

My cron jobs used to work as of 2026.5.6, but in the 2026.5.12 beta series the isolated cron jobs in particular are failing at attempt-dispatch.

Steps to reproduce

1. Codex harness, openai/gpt-5.5, isolated agent cron jobs set up.
2. Attempt to run a cron job that takes at least 60s (or allow it to run by time)

Expected behavior

Cron job completes successfully as per previous version.

Actual behavior

Cron job fails with diagnostic: cron: isolated agent run stalled before execution start (last phase: attempt-dispatch)

OpenClaw version

2026.5.12-beta.6

Operating system

macOS 15.7.5

Install method

npm global

Model

openai/gpt-5.5

Provider / routing chain

openclaw local gateway, codex harness

Additional provider/model setup details

openclaw-cron-repro-48ffa2e7-86f0-4407-9a36-2920c1eb06ba.tgz

Logs, screenshots, and evidence

attached a bunch of log and session files suggested by Krill bot in Discord

Impact and severity

No response

Additional information

It's possible main-session cron jobs are running? Or just manually running them seems more successful.

Codex run on my local openclaw setup tells me:

For the fresh openclaw cron run 48ffa2e7-86f0-4407-9a36-2920c1eb06ba, computer-use was no longer the direct failure. The run actually did start:

• Session 9da62f94-6a6c-4213-b10d-6d838352df9d started at 2026-05-13 18:14:05 PDT.
• The model completed at 18:15:05.304 PDT.
• Cron then cleaned it up as stalled at 18:15:05.345 PDT.

So it lost a race by about 40ms. The cron watchdog still thought execution had not started because its “execution started” detector only clears on phases like model_call_started, tool_execution_started, or assistant_output_started. This Codex/app-server path recorded prompt.submitted and model.completed, but did not emit the phase event cron is waiting for before the 60s pre-execution watchdog fired.

Why the tiny probe worked: it completed in ~6s, well under the 60s watchdog. This real job uses gpt-5.5 with default high thinking and a large prompt/context, so the first visible model result can land right at the 60s edge.

Important detail: increasing the job’s timeoutSeconds probably won’t fix this, because the pre-execution watchdog is capped at 60s in server-cron-YWxalur2.js.

Actual code fix: cron should treat model.completed/prompt submission or Codex app-server turn accepted as execution progress, or the Codex app-server path should emit model_call_started as soon as it submits the request. The current attempt-dispatch error is misleading for this path.