KG-561 Add Gemini 3 Pro Preview by mltheuser · Pull Request #1236 · JetBrains/koog

mltheuser · 2025-12-07T14:18:20Z

Motivation and Context

This PR adds support for Google's latest model, Gemini 3 Pro Preview (gemini-3-pro-preview).

While integrating this model, significant changes were required in how "Thinking" (Reasoning) is configured compared to previous versions (Gemini 2.5):

Thinking Configuration: Gemini 3 abandons the token-based limit (thinkingBudget) in favor of abstract levels (thinkingLevel). Currently, LOW and HIGH are supported.
Validation Logic: The GoogleThinkingConfig has been updated to support these levels. An init block was added to enforce mutual exclusion: users cannot configure both a thinkingBudget (legacy/Gemini 2) and a thinkingLevel (Gemini 3) simultaneously.
Conversation Consistency: Gemini 3 is stricter regarding conversation history; it will fail if thought signatures are not reproduced correctly in subsequent requests.

To verify these changes:

Unit Tests: Added tests to ThinkingConfigTest to verify the JSON serialization of the new thinkingLevel and to ensure the mutual exclusion validation works as expected.
Integration Tests: The new model was added to reasoningCapableModels. Crucially, this runs it against integration_testReasoningMultiStep. This validates that multi-turn conversations maintain the correct thought signatures/context, preventing the API failures specific to Gemini 3 when history is mishandled.

Breaking Changes

No breaking changes. The new configuration parameters are optional and specifically targeted at the new model. Existing configurations for Gemini 2 or other providers remain unaffected.

Type of the changes

New feature (non-breaking change which adds functionality)
Bug fix (non-breaking change which fixes an issue)
Breaking change (fix or feature that would cause existing functionality to change)
Documentation update
Tests improvement
Refactoring

Checklist

The pull request has a description of the proposed change
I read the Contributing Guidelines before opening the pull request
The pull request uses develop as the base branch
Tests for the changes have been added
All new and existing tests passed

Additional steps for pull requests adding a new feature

An issue describing the proposed change exists
The pull request includes a link to the issue
The change was discussed and approved in the issue
Docs have been added / updated

aozherelyeva

Thank you! Here are some comments regarding the testing part.

aozherelyeva · 2025-12-08T13:12:05Z

-                maxTokens = 256
-            )
+            is LLMProvider.Google -> {
+                val thinkingConfig = if (model.id.contains("gemini-3")) {


It's better to use the variable storing model ID: if (model.id == GoogleModels.Gemini3_Pro_Preview.id)

aozherelyeva · 2025-12-08T13:18:45Z

    fun googleModels(): Stream<LLModel> {
        return Stream.of(
+            GoogleModels.Gemini3_Pro_Preview,
            GoogleModels.Gemini2_5Pro,


Let's not add Gemini 3 to the googleModels until it's a preview? We can add it a bit later when it's more stable :)

aozherelyeva · 2025-12-08T13:19:29Z

        return Stream.of(
            OpenAIModels.Chat.GPT5_1,
            AnthropicModels.Haiku_4_5,
            GoogleModels.Gemini2_5Pro,


Could you please remove GoogleModels.Gemini2_5Pro? I'd suggest having one model from each family to avoid too much time on high-level tests.

mltheuser · 2025-12-08T16:09:49Z

Thanks for the feedback! No problem, only concern I have is that by removing gemini-2.5 from the reasoningCapableModels you might miss it if someone breaks the old thinkingBudget setting, but probably an acceptable amount of risk. The change also means we no longer have to check the gemini model id in createReasoningParams.

aozherelyeva · 2025-12-09T10:34:20Z

Thanks for the feedback! No problem, only concern I have is that by removing gemini-2.5 from the reasoningCapableModels you might miss it if someone breaks the old thinkingBudget setting, but probably an acceptable amount of risk. The change also means we no longer have to check the gemini model id in createReasoningParams.

That's a fair point, thank you! Let's not remove the Gemini 2.5 then 😅

EugeneTheDev

Looks good, can you rebase please?

mltheuser · 2025-12-09T21:19:49Z

No problem. Rebased and added back gemini-2.5-pro to reasoningCapableModels for integration testing.

mltheuser · 2025-12-09T21:48:38Z

Man, these tests have been keeping me on my toes lately! 😅 I find myself having to squash/unsquash updates a few times on almost every PR just to rerun the pipeline until it passes. Do we have a ticket tracking this flakiness?

EugeneTheDev · 2025-12-10T14:04:46Z

Yes, sorry about that, this is a known problem, and I'm disabling these flaky tests for now until we resolve the flakiness. Hopefully after #1251 it should be stable again

Discussed and resolved all comments

## Motivation and Context This PR adds support for Google's latest model, **Gemini 3 Pro Preview** (`gemini-3-pro-preview`). While integrating this model, significant changes were required in how "Thinking" (Reasoning) is configured compared to previous versions (Gemini 2.5): * **Thinking Configuration:** Gemini 3 abandons the token-based limit (`thinkingBudget`) in favor of abstract levels (`thinkingLevel`). Currently, `LOW` and `HIGH` are supported. * **Validation Logic:** The `GoogleThinkingConfig` has been updated to support these levels. An `init` block was added to enforce mutual exclusion: users cannot configure both a `thinkingBudget` (legacy/Gemini 2) and a `thinkingLevel` (Gemini 3) simultaneously. * **Conversation Consistency:** Gemini 3 is stricter regarding conversation history; it will fail if thought signatures are not reproduced correctly in subsequent requests. To verify these changes: * **Unit Tests:** Added tests to `ThinkingConfigTest` to verify the JSON serialization of the new `thinkingLevel` and to ensure the mutual exclusion validation works as expected. * **Integration Tests:** The new model was added to `reasoningCapableModels`. Crucially, this runs it against `integration_testReasoningMultiStep`. This validates that multi-turn conversations maintain the correct thought signatures/context, preventing the API failures specific to Gemini 3 when history is mishandled. ## Breaking Changes No breaking changes. The new configuration parameters are optional and specifically targeted at the new model. Existing configurations for Gemini 2 or other providers remain unaffected. --- #### Type of the changes - [x] New feature (non-breaking change which adds functionality) - [ ] Bug fix (non-breaking change which fixes an issue) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [ ] Documentation update - [x] Tests improvement - [ ] Refactoring #### Checklist - [x] The pull request has a description of the proposed change - [x] I read the [Contributing Guidelines](https://github.com/JetBrains/koog/blob/main/CONTRIBUTING.md) before opening the pull request - [x] The pull request uses **`develop`** as the base branch - [x] Tests for the changes have been added - [x] All new and existing tests passed ##### Additional steps for pull requests adding a new feature - [x] An issue describing the proposed change exists - [x] The pull request includes a link to the issue - [x] The change was discussed and approved in the issue - [ ] Docs have been added / updated --------- Co-authored-by: Malte Heuser <malte.heuser@ing.com>

mltheuser force-pushed the mltheuser/kg-561 branch from 3237a96 to 82e674c Compare December 7, 2025 14:39

mltheuser changed the title ~~Mltheuser/kg 561~~ KG-561 Let's add Gemini 3 Pro Preview Dec 7, 2025

mltheuser force-pushed the mltheuser/kg-561 branch from 82e674c to 8ac25bd Compare December 7, 2025 15:00

aozherelyeva requested review from EugeneTheDev, Rizzen, antoniibelyshev and devcrocod and removed request for Rizzen December 8, 2025 10:19

aozherelyeva previously requested changes Dec 8, 2025

View reviewed changes

mltheuser force-pushed the mltheuser/kg-561 branch from 0530d45 to e3b6257 Compare December 8, 2025 16:09

mltheuser force-pushed the mltheuser/kg-561 branch from e3b6257 to a1b6986 Compare December 8, 2025 17:21

EugeneTheDev approved these changes Dec 9, 2025

View reviewed changes

Malte Heuser added 4 commits December 9, 2025 22:15

KG-561 Add gemini3-pro-preview support

3771c63

KG-561 Verify gemini client correctly passes back thought signatures

315062a

KG-561 fix klint errors

0bcae84

KG-561 fix windows errors

ed34845

mltheuser force-pushed the mltheuser/kg-561 branch from 28d044d to e368e53 Compare December 9, 2025 21:16

KG-561 resolve PR comments

f04e8c1

mltheuser force-pushed the mltheuser/kg-561 branch from e368e53 to f04e8c1 Compare December 9, 2025 21:33

KG-561 resolve PR comments

a4eeeeb

EugeneTheDev changed the title ~~KG-561 Let's add Gemini 3 Pro Preview~~ KG-561 Add Gemini 3 Pro Preview Dec 10, 2025

EugeneTheDev merged commit 3a14ae7 into JetBrains:develop Dec 10, 2025
10 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

KG-561 Add Gemini 3 Pro Preview#1236

KG-561 Add Gemini 3 Pro Preview#1236
EugeneTheDev merged 6 commits intoJetBrains:developfrom
mltheuser:mltheuser/kg-561

mltheuser commented Dec 7, 2025 •

edited

Loading

Uh oh!

aozherelyeva left a comment

Uh oh!

aozherelyeva Dec 8, 2025

Uh oh!

aozherelyeva Dec 8, 2025

Uh oh!

aozherelyeva Dec 8, 2025

Uh oh!

mltheuser commented Dec 8, 2025

Uh oh!

aozherelyeva commented Dec 9, 2025

Uh oh!

EugeneTheDev left a comment

Uh oh!

mltheuser commented Dec 9, 2025

Uh oh!

mltheuser commented Dec 9, 2025 •

edited

Loading

Uh oh!

EugeneTheDev commented Dec 10, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

mltheuser commented Dec 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation and Context

Breaking Changes

Type of the changes

Checklist

Additional steps for pull requests adding a new feature

Uh oh!

aozherelyeva left a comment

Choose a reason for hiding this comment

Uh oh!

aozherelyeva Dec 8, 2025

Choose a reason for hiding this comment

Uh oh!

aozherelyeva Dec 8, 2025

Choose a reason for hiding this comment

Uh oh!

aozherelyeva Dec 8, 2025

Choose a reason for hiding this comment

Uh oh!

mltheuser commented Dec 8, 2025

Uh oh!

aozherelyeva commented Dec 9, 2025

Uh oh!

EugeneTheDev left a comment

Choose a reason for hiding this comment

Uh oh!

mltheuser commented Dec 9, 2025

Uh oh!

mltheuser commented Dec 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

EugeneTheDev commented Dec 10, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

mltheuser commented Dec 7, 2025 •

edited

Loading

mltheuser commented Dec 9, 2025 •

edited

Loading