Hi Google Cloud Support,
We are investigating an intermittent issue with Claude Sonnet 4.6 via Vertex AI.
In our coding workflow, the model occasionally returns a tool call for a file write without the required content parameter, causing retries and unstable execution.
Example client log:
Roo tried to use write_to_file without value for required parameter 'content'. Retrying...
Environment
-
Google Vertex AI
-
Anthropic Claude Sonnet 4.6
-
VS Code
-
Roo Code extension v3.51.1
We have already raised this with the Roo Code team. Their current assessment is that this is likely not a Roo-side root cause, but more likely due to the provider/model response being incomplete or truncated. In other words, Roo expects the response payload to include the full file content, and when that field is missing, they believe the upstream provider response is being cut off rather than Roo generating an invalid call on its own.
Could you please help investigate whether Vertex AI may be intermittently returning:
-
incomplete or truncated output
-
incomplete streaming chunks
-
prematurely terminated structured/tool-call responses
-
output affected by token or response-size constraints
We would specifically like guidance on:
-
whether this is a known behaviour or issue for Claude Sonnet 4.6 on Vertex AI
-
how to inspect finish reasons, truncation indicators, token limits, or interrupted streams at request level
-
recommended mitigations for longer structured outputs in tool-calling workflows
The issue is intermittent, occurs across multiple users and machines, and is not tied to one specific repository or prompt.
Thanks in advance for your help.