feat(ai_studio): add thinking configuration in request Options

[OrriMandarin]

This pull request adds support for a new "thinking" configuration to the Google AI Studio provider, allowing users to specify options such as IncludeThoughts and Budget in their requests.
To disable thinking mode, set the Budget to 0

Support for "thinking" configuration:

• Added a new ThinkingConfig struct to request_options.go, and extended RequestOptions to include an optional Thinking field. This allows users to specify IncludeThoughts and Budget when making requests.
• Updated the request adaptation logic in aistudio.go to include the thinkingConfig in the request payload when the Thinking option is provided.

Testing enhancements:

• Added a new test function TestGoogleAiRequestWithThinking in aistudio_test.go to verify that the thinkingConfig fields are correctly included or omitted in the payload based on the provided options. The tests cover cases for no config, only IncludeThoughts, only Budget, both fields, and disabling thinking.
• Added github.com/samber/lo to the imports in aistudio_test.go to support pointer creation for test values.

[Copilot]

Pull Request Overview

This pull request adds support for a "thinking" configuration to the Google AI Studio provider, enabling users to control AI reasoning behavior through IncludeThoughts and Budget parameters. To disable thinking mode, users can set the Budget to 0.

• Added ThinkingConfig struct with IncludeThoughts (bool) and Budget (*int32) fields
• Extended RequestOptions to include optional Thinking field
• Updated request adaptation logic to include thinking configuration in API payloads

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 9 comments.

File	Description
llms/aistudio/request_options.go	Defines new ThinkingConfig struct and adds Thinking field to RequestOptions
llms/aistudio/aistudio.go	Updates request adaptation to include thinking configuration in API payload
llms/aistudio/aistudio_test.go	Adds comprehensive tests for thinking configuration scenarios including edge cases

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[OrriMandarin]

I've run some tests: with thinking mode disabled, my request completed in 7 seconds, compared to 15 seconds when thinking mode was enabled.

I also checked the metadata—thought token usage was 0 without thinking mode, and a higher value when thinking mode was active.

[OrriMandarin]

We can not disable the thinking mode for gemini-2.5-pro model. The provider returns an 400 error

[apognu]

Before I get into the code, I wonder if we shouldn't add a top-level, provider-agnostic settings to (at the minimum) enable and disable reasoning.

The provider code could then do The Right Thing ™️ to do that with its API (or ignore it if the provider and/or the model does not support it), and accept extra, provider-specific settings if appropriate.

My reasoning (see what I did there?) behind that is that the decision to enable or disable reasoning would be lost when switching provider, which arguably is not what one wants to do.

[OrriMandarin]

Before I get into the code, I wonder if we shouldn't add a top-level, provider-agnostic settings to (at the minimum) enable and disable reasoning.

The provider code could then do The Right Thing ™️ to do that with its API (or ignore it if the provider and/or the model does not support it), and accept extra, provider-specific settings if appropriate.

My reasoning (see what I did there?) behind that is that the decision to enable or disable reasoning would be lost when switching provider, which arguably is not what one wants to do.

@apognu
It's possible to implement a top-level setting to disable reasoning, but the process varies significantly depending on the provider, making it complex to manage across platforms.

For example,
Gemini's API uses the Thinking config and Budget parameters
Mistral requires a special field for prompt-mode
OpenAI doesn't let you fully disable reasoning, only adjust the "effort." If you want to avoid the reasoning layer with OpenAI, you'd need to use an older model lol (see: https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/2)
Ollama supports disabling reasoning via the think field.

For Mistral, OpenAI, and Ollama, they all use the OpenAI provider, but the appropriate configuration depends on the specific model. Since the required settings vary by model, it's difficult to determine in advance which configuration should be applied.

[apognu]

It's possible to implement a top-level setting to disable reasoning, but the process varies significantly depending on the provider, making it complex to manage across platforms.

The handling of that setting would remain the responsibility of the adapters (same as you implemented), but the setting itself could be set on the outer Request (same as the temperature setting).

The adapter would (if applicable) retrieve that setting and do whatever is appropriate for that provider. We can keep the budget in the provider, since it is highly different across the board.

[Copilot]

Pull Request Overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}