My Cron Said OK But Did Nothing
Last Tuesday, my OpenClaw agent ran a security audit.
The dashboard showed a green light. It said status: ok. There were no errors and no alerts.
But the agent did nothing.
The agent crashed during the task. A MiniMax overload error occurred. The outer framework did not catch it. The framework saw a successful completion even though the agent failed.
I only found this error three days later when I checked the session transcript manually.
I needed a way to find these silent crashes. I built a 30-line review script to solve it.
The Problem Frameworks detect network timeouts and auth failures. They do not detect what happens inside an agent turn. When a sub-agent crashes, the system often produces a specific message: "[assistant turn failed before producing content]".
To the framework, this looks like a normal message. The status remains "ok". This is a silent failure. It is the hardest type of error to find.
The Solution I added a script to check the actual transcript content instead of just the status code.
The script looks for that specific failure string. It also uses a regular expression to pull the exact error message from the text.
This allows the script to show the real cause, such as:
- overloaded_error
- rate_limit_exceeded
- context_length_exceeded
Once I saw the error details, I found the root cause. The crashes happened because of a model fallback chain. I removed the free fallback model that was causing cascading failures. Removing it made my crons faster and more reliable.
The Result The script now runs every night. It checks the previous day's transcripts. If it finds a silent crash, it sends an alert to my Telegram.
I no longer wait days to find errors. I see them every morning.
The Lesson A green dashboard does not mean your agent worked. Framework status and agent output are different things.
If you run automated agents, do not rely on status codes alone. Check the transcripts. Build a tool to check the transcripts for you. Silent failures are the ones that cause the most damage.
Optional learning community: https://t.me/GyaanSetuAi