A watchdog for Qwen3.6-27B tool failures and loops
When runs long tasks locally, it can repeat the same behavior or stop after inventing an invalid tool call. This makes it necessary to watch the model closely during large jobs.
A custom extension for the Pi works around the problem by monitoring the JSON stream for loops. It also checks whether the user's goal has been completed.
If work remains but the main model has stopped, the extension automatically adds prompts telling it to continue. In firsthand use, this has brought the model back on course and made it more practical to leave large tasks running unattended.
Key points
- can loop or stop after producing an invalid tool call during long tasks.
- The Pi extension watches the JSON stream for repeated behavior.
- It checks whether the user's requested goal has actually been reached.
- If unfinished work remains after the model stops, it automatically prompts the model to continue.
- rates and extra token use should be measured before claiming a cost saving.