Description:

When using TensorZero Gateway to proxy LLM requests to GCP Vertex AI Gemini (gcp_vertex_gemini provider), tool/function declarations that contain nullable parameter types cause a 400 Bad Request error
from the Gemini API.

Root Cause:

Some tool parameters use the JSON Schema convention "type": ["string", "null"] to express nullable types (common in OpenAI-compatible schemas). TensorZero passes these array-typed type fields directly
to the Gemini Vertex AI API without conversion. Gemini's protobuf-based API expects type to be a single string enum value (e.g. "STRING"), not an array.

Error:

Invalid JSON payload received. Unknown name "type" at
'tools[0].function_declarations[4].parameters.properties[0].value':
Proto field is not repeating, cannot start list.

Expected Behavior:

TensorZero should normalize nullable type declarations before sending to Gemini. For example:

• "type": ["string", "null"] → "type": "string" (with "nullable": true if supported)
• Or strip the null variant and use only the primary type

Steps to Reproduce:

1. Configure TensorZero with a gcp_vertex_gemini provider
2. Send a chat completion request that includes tools with nullable parameters (e.g. "type": ["string", "null"])
3. Observe 400 INVALID_ARGUMENT from Vertex AI

Environment:

• TensorZero Gateway: 2026.2.1 (commit: 2f139c9)
• Provider: gcp_vertex_gemini
• Model: gemini-3.1-pro-preview

Workaround:

Bypass TensorZero and call Gemini directly using a client library (e.g. genai crate with Gemini adapter) that properly converts tool schemas to Gemini's expected format.