Hi everyone,
I’m experiencing a persistent blocking issue when using the Agent Chat View and Agent Prompter nodes with the latest Gemini models (specifically the 2.x and 3.x series). Despite several workarounds, the agent fails to execute tool calls correctly.
The Problems:
-
Missing
thought_signature(Error 400): When using reasoning models (like Gemini 2.5 Pro or Gemini 3), the API returns a400 Bad Requeststating: “Function call is missing a thought_signature in functionCall parts.” It seems the current KNIME nodes are not capturing or re-sending the reasoning tokens required by Google’s new agentic infrastructure. -
Integer-to-Float Conversion: Even when using “Lite” models that don’t always trigger the signature error, the models tend to output Tool IDs as floats (e.g.,
1.0) instead of strict integers (1). This causes aTypeErrorin downstream Python-based tools that expect integer indices.
What we have already tried:
-
Model Variety: We tested multiple models from the Gemini 2.0, 2.5, and 3.0 series (Pro, Flash, and Lite).
-
Prompt Engineering (System Message): We added strict instructions in the System Message to force “STRICT INTEGER” formatting and to “strip decimals” from IDs.
-
Data Message Prefix: We modified the Data Message Prefix to explicitly state that the Data Repository IDs are integers and that any float input would break the system.
-
Node Testing: The issue persists across both the Agent Chat View and the Agent Prompter nodes.
I have KNIME 5.8.2 LTS.
