gemini-3-pro-preview context window incorrectly limited to 128k (Expected 1M)

[StarCat077]

Plugin Type

VSCode Extension

App Version

4.143.2

Description

I am experiencing a critical issue where the context window for gemini-3-pro-preview (Google Gemini provider) is being incorrectly throttled or downgraded.

Initially, the context window correctly shows 1,000,000 tokens. However, after using the extension for some time, the window size automatically drops to 128,000 tokens. This has happened multiple times.

Most importantly, after clicking "Reset Data" (重置数据) in an attempt to fix this, the context window is now stuck at 128,000 tokens and will not revert to the expected 1M capacity. This suggests that the downgraded state is being incorrectly cached or written back to the persistent configuration.

Expected Behavior: The context window should remain at 1,000,000 tokens and "Reset Data" should restore the model's default full capacity.

Actual Behavior: The capacity drops to 128,000 tokens during use and becomes permanently stuck at that limit even after a data reset.

Reproduction steps

1. Use Kilo Code with gemini-3-pro-preview (initially at 1M context).
2. Use the chat/coding features for a period of time.
3. Observe the context window display in settings suddenly changing to 128,000.
4. Click "Reset Data" in settings.
5. The context window remains at 128,000 tokens.

Provider

google gemini

Model

gemini 3 pro & flash

System Information

Windows11 & wsl2

[LeHaroun]

Similar behavior is observed in the API from AI Studio.

[Shisuki]

Similar behavior observed with Amazon Bedrock and custom ARNs

[johnnyeric]

Thanks for taking the time to report this!

We've just released a new version of the Kilo Code extension. As part of this transition, we're automatically closing issues that were filed against the legacy extension.

If you're still experiencing this issue with the new extension, feel free to reopen this or open a fresh issue with updated details - we'd love to hear from you. We're actively working to make Kilo Code better, and your feedback is a big part of that.

Thanks again for being part of the community!