feat(providers): add streaming support for Google Gemini provider

[rabi]

Summary

• Add stream() and supports_streaming() to GoogleProvider
• Use streamGenerateContent?alt=sse endpoint for SSE streaming
• Implement response_to_streaming_message for SSE format parsing
• Handle text chunks, function calls, thinking content, and usage
• Add mid-stream error detection and propagation
• Trust thoughtSignature for thinking content classification
• Add tests for streaming text, function calls, and error handling
• Fix fragmented streaming output in desktop UI by generating a consistent message ID for all chunks within a streaming session

Type of Change

• Feature
• Bug fix
• Refactor / Code quality
• Performance improvement
• Documentation
• Tests
• Security fix
• Build / Release
• Other (specify below)

AI Assistance

• This PR was created or reviewed with AI assistance

Testing

Unit and manual testing with provider

[rabi]

@DOsinga @alexhancock Hi, Apologies for the direct ping for review. I don't seem to have permissions to request a review. Probably need an invite to join the github organization.

[pkos98]

Geese admirer here:

• Testing this PR immediately shows a regression: one LLM response is now split into several messages. Have you actually tested your PR?
• Also, looking at the code, it destroys basic formatting.

IMO more work is needed before one of the main contributors spends their valuable time reviewing this.

[rabi]

Testing this PR immediately shows a regression: one LLM response is now split into several messages.

What model are you testing with? I don't see any such issue.

[pkos98]

I've tested using gemini-2.0-flash (as well as gemini-3-flash-preview after merging #6187)

[rabi]

I've tested using gemini-2.0-flash (as well as gemini-3-flash-preview after merging #6187)

Here is what I see with gemini-2.0 flash. Are you using cli, web or some third-party UI?

$ GOOSE_PROVIDER=google GOOSE_MODEL=gemini-2.0-flash  ./target/debug/goose session
starting session | provider: google model: gemini-2.0-flash
    session id: 20251222_55
    working directory: /home/ramishra/work/LLM/goose

goose is running! Enter your instructions, or try asking what goose can do.

Context: ○○○○○○○○○○ 0% (0/1000000 tokens)
( O)> can you write a 100 line poem about anything?

The clock ticks on, a steady beat,
In shadows deep, where secrets meet.
A world unseen, a whispered dream,
Reflects within a moonlit stream.

The wind it sighs through barren trees,
A symphony of rustling pleas.
The stars ignite, a distant fire,
A cosmic dance of pure desire.

The mountains rise, in silent grace,
A timeless watch upon this place.
The valleys sleep, in velvet dark,
Awaiting dawn's ascending spark.

The ocean roars, a restless soul,
Its waves crash on, beyond control.
The sandy shore, a fragile line,
Where water meets the fading shine.

A single bird, in lonely flight,
Across the canvas of the night.
Its mournful call, a haunting sound,
As echoes softly drift around.

The city sleeps, a concrete maze,
Where neon lights begin to blaze.
A hidden world, of hopes and fears,
Reflected in a million tears.

The poet's pen, it starts to flow,
With words that dance, and softly glow.
A tapestry of thoughts unfolds,
A story that the heart beholds.

The moon ascends, a silver sphere,
Dispelling darkness, calming fear.
A gentle light, upon the land,
A guiding hand, a helping hand.

The world awakes, in gentle hues,
As morning light the dark subdues.
A new beginning, fresh and bright,
A promise whispered in the light.

The poem ends, the words take flight,
A fleeting glimpse, into the night.
A memory held, a moment shared,
A world of wonder, now declared.


⏱️  Elapsed time: 4.09s
Context: ○○○○○○○○○○ 1% (5118/1000000 tokens)

[pkos98]

Only tested the desktop electron app (just run-ui # or just make-ui).
Thanks for having a look!

[rabi]

Only tested the desktop electron app (just run-ui # or just make-ui). Thanks for having a look!

Right, I noticed that desktop UI aggregates streaming chunks into a single message bubble with metadata (timestamp, copy button, etc.) before rendering. It does this by matching message IDs which does not come in gemini response chunks (API does not povide it unlike other providers). I did not see the issue in cli as cli does not do any aggregation. I've fixed this by generating a consistent UUID at the start of each streaming session.

Also, can you elaborate what section of code you were referring to when you said "it destroys basic formatting"? This is all rustfmt checked.

[pkos98]

Looks like it's indeed working now (using gemini-3-flash-preview on desktop app).

Also, can you elaborate what section of code you were referring to when you said "it destroys basic formatting"? This is all rustfmt checked.

I just quickly compared some indention in the diff and it looked off, I re-compared and it's looking better now

[rabi]

Folks PTAL. This would help us implement streaming in other providers like Vertex AI and OpenRouter etc.

[jamadeo]

Looks great, thanks for the contribution!

I also ran test_providers.sh for gemini, and it passed.

[alexhancock]

Nice