Gemini causing empty replies after tool use

[angiejones]

Summary

When sending functionResponse back to Gemini after a tool call, the code uses the internal tracking ID (response.id) instead of the original function name. This violates the Gemini API contract and causes Gemini to return empty responses.

Evidence

From a user's diagnostic logs (gemini-2.5-flash-lite):

• 75 out of 76 tool responses resulted in empty assistant replies
• 100% of responses after functionResponse were empty
• 100% of responses after text messages were non-empty

Example from logs:

functionCall.name = "developer__text_editor"   ✓ Correct
functionResponse.name = "d9X7Azly"              ✗ Wrong (random ID)

Root Cause

In crates/goose/src/providers/formats/google.rs (lines 130-131 and 148-149):

function_response.insert("name".to_string(), json!(response.id));

The response.id is a random 8-character alphanumeric string generated internally for tracking, not the function name.

Expected Behavior

Per Google's Gemini API docs, functionResponse.name must match the functionCall.name that was originally sent.

Comparison with Other Providers

Provider	Field	Value Used	Correct?
Anthropic	tool_use_id	response.id	✅ Yes (ID-based)
OpenAI	tool_call_id	response.id	✅ Yes (ID-based)
Google/Gemini	name	response.id	❌ Bug (should be function name)

Proposed Fix

The functionResponse.name should use the original function name. This requires either:

1. Storing the function name in ToolResponse when created (from the corresponding ToolRequest)
2. Looking up the function name from the matching ToolRequest when formatting the response

User Impact

Users experience:

• "No response after tool use"
• Tools execute but Goose goes silent
• Files modified but no acknowledgment from the model

Notes

• This code has been present since V1.0 (1c9a7c0b050)

[cbagwell]

I think there is a good chance this overlaps with #6091. Text responses can also include a thoughtSignature and an update to support gemini-3.0-pro appears to have broken gemini-2.5-* models that can return thoughtSignatures as part of responses but shouldn't be treated as a thought.

[angiejones]

I think there is a good chance this overlaps with #6091. Text responses can also include a thoughtSignature and an update to support gemini-3.0-pro appears to have broken gemini-2.5-* models that can return thoughtSignatures as part of responses but shouldn't be treated as a thought.

thanks for this. just checked and the user's logs show no thoughtSignature on any text parts. the empty responses have content={'role': 'model'} with no parts array at all. Gemini is returning an empty response.