feat: reduce Gemini 2.5 Pro minimum thinking budget to 128

[roomote]

This PR reduces the minimum thinking budget for Gemini 2.5 Pro models from 1024 to 128 tokens, as requested. The default thinking budget for these models is also changed from 8192 to 128.

Changes:

• Added GEMINI_25_PRO_MIN_THINKING_TOKENS constant set to 128
• Updated model-params.ts to detect Gemini 2.5 Pro models and apply the 128 minimum
• Changed default thinking tokens from 8192 to 128 for Gemini 2.5 Pro models only
• Added comprehensive tests to verify the new behavior

Important: This change ONLY affects Gemini 2.5 Pro models. All other models retain their 1024 minimum thinking budget. Hybrid reasoning and other model types are not affected.

---

Important

Reduces Gemini 2.5 Pro models' minimum and default thinking budget to 128 tokens, updating logic and UI components accordingly.

• Behavior:
  • Reduces minimum thinking budget for Gemini 2.5 Pro models from 1024 to 128 tokens in model-params.ts.
  • Changes default thinking budget for Gemini 2.5 Pro models from 8192 to 128 tokens in model-params.ts.
  • Ensures reasoning budget is clamped to at least 128 tokens for Gemini 2.5 Pro models.
• UI:
  • Updates ThinkingBudget.tsx to reflect new minimum and step values for Gemini 2.5 Pro models.
  • Adjusts slider behavior to use 128 as the step for Gemini 2.5 Pro models.
• Tests:
  • Adds tests in model-params.spec.ts to verify new minimum and default thinking budget for Gemini 2.5 Pro models.
  • Adds tests in ThinkingBudget.spec.tsx to ensure UI reflects new behavior for Gemini 2.5 Pro models.

^{This description was created by for 1fa71fb. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Reviewing my own code is like debugging in production—technically possible, but morally questionable.

[roomote]

Consider adding test cases for model ID variations and edge cases:

• Model IDs like "gemini-2.5-pro-latest" or "vertex-ai/gemini-2.5-pro"
• The interaction between the 128 minimum and 80% rule when maxTokens is very small (e.g., 150)

[roomote]

Could we make this comment more explicit about why Gemini 2.5 Pro has a different default? Perhaps mention performance or cost considerations that led to this decision?

[hannesrudolph]

@roomote-agent run pr fixer

[roomote]

Hi @hannesrudolph! I see your request to run PR fixer. Let me check out this PR and investigate what needs to be fixed. All CI checks are currently passing, so I'll review the code changes to ensure everything is implemented correctly.

[roomote]

Hi @hannesrudolph! I've successfully fixed the issue in this PR.

Summary of changes:

• Fixed the duplicate condition in src/api/transform/model-params.ts line 95
• Changed from modelId.includes("gemini-2.5-pro") || modelId.includes("gemini-2.5-pro") to modelId.includes("gemini-2.5-pro") || modelId.includes("gemini-25-pro")
• This now properly checks for both "gemini-2.5-pro" and "gemini-25-pro" model ID variants

Testing:

• All existing tests pass ✅
• CI checks are all green ✅

The fix has been pushed to the PR branch and is ready for review.

[daniel-lxs]

@roomote-agent This change should also affect the webview-ui/src/components/settings/ApiOptions.tsx component to allow the user to set the max reasoning tokens to 128. Push to this PR, do not create a new one.

[roomote]

Hi @daniel-lxs! I see your request to update the webview-ui/src/components/settings/ApiOptions.tsx component to allow users to set the max reasoning tokens to 128. I'll work on this change and push it to this PR branch. Let me get started on that now.

[roomote]

Hi @daniel-lxs! I've successfully implemented the requested changes to allow users to set the max reasoning tokens to 128 for Gemini 2.5 Pro models in the UI.

Changes made:

• ✅ Updated ThinkingBudget.tsx component to detect Gemini 2.5 Pro models
• ✅ Set minimum thinking tokens to 128 for Gemini 2.5 Pro models (instead of the default 1024)
• ✅ Adjusted the step size to 128 for Gemini 2.5 Pro models for better granularity
• ✅ Added comprehensive tests to verify the new behavior
• ✅ All tests are passing

The changes have been pushed to this PR branch. The UI will now properly allow Gemini 2.5 Pro users to set their thinking budget as low as 128 tokens, matching the backend changes that were already implemented.

CI checks are currently running and should complete shortly.

[daniel-lxs]

LGTM

[mrubens]

Are there really gemini-25-pro models?

[daniel-lxs]

oh, I don't think so

[daniel-lxs]

@roomote-agent remove the check for gemini-25-pro since that model doesn't exist

[roomote]

Hi @daniel-lxs, I see your comment about removing the check for the non-existent "gemini-25-pro" model. Let me investigate and fix this for you.

[roomote]

Hi @daniel-lxs, I've successfully removed all references to the non-existent "gemini-25-pro" model identifier from the codebase.

Here's what I changed:

• ✅ Removed || modelId.includes("gemini-25-pro") check from src/api/transform/model-params.ts (line 95)
• ✅ Removed || selectedModelId.includes("gemini-25-pro") check from webview-ui/src/components/settings/ThinkingBudget.tsx (line 28)
• ✅ Removed the test case "should use min thinking tokens of 128 for Gemini 25 Pro models (alternative naming)" from webview-ui/src/components/settings/__tests__/ThinkingBudget.spec.tsx

The code now only checks for the valid "gemini-2.5-pro" model identifier. All tests are passing locally. The changes have been pushed to the PR branch and CI checks are running.