Problem

When using Gemini/Vertex AI models with function calling, the API returnsfinish_reason: "stop" even when the response contains tool calls. This is by design from Google – unlike OpenAI and Anthropic, Gemini does not have a dedicatedTOOL_CALL finish reason.

This inconsistency makes it difficult to build provider-agnostic tool execution loops, as the logic typically relies onfinish_reason to determine whether to execute tools and continue the conversation.

Current behavior (Gemini)

{"variant_name":"gemini-2.5-pro","content": [    {"type":"tool_call","id":"0","raw_name":"get_weather","raw_arguments":"{\"location\":\"London\"}"    }  ],"finish_reason":"stop"}

Gemini returns"stop" even with tool calls present.

Expected behavior (OpenAI/Anthropic)

{"content": [...tool_call...],"finish_reason":"tool_call"}

Other providers return"tool_call" or"tool_use".

Proposed solution

Add an option (opt-in or opt-out) to normalizefinish_reason for Gemini responses:

• Ifcontent contains atool_call block ANDfinish_reason isstop, returnfinish_reason: "tool_call" instead

This would provide consistent behavior across all providers and simplify downstream tool execution logic.

Workaround

Currently, we iterate over thecontent array to detect tool calls and overridefinish_reason manually in our application layer. This works but adds complexity that could be handled at the gateway level.

References

• Google Vertex AI FinishReason documentation – confirms noTOOL_CALL value exists