[dhaern]

Problem

Gemini 2.5+ models always have thinking capability, but the translator only set includeThoughts: true when the client explicitly requested it via reasoning_effort or thinking.type.

Without this flag, the upstream does not mark thought parts with thought: true in its response. The response translator (gemini-to-openai.ts) checks part.thought === true to route text to reasoning_content — so without the flag, the model's internal reasoning leaks into visible content instead of being separated into reasoning_content.

This is especially problematic for smaller models (e.g. gemini-3.5-flash-low) that tend to reason extensively about tool names, instructions, and internal logic when their reasoning has no separate channel. Users see the model's internal deliberation (including echoed system prompt fragments) as regular assistant output.

Root Cause

In openai-to-gemini.ts, thinkingConfig is only set in two cases:

1. Client sends reasoning_effort (OpenAI format) → maps to a budget + includeThoughts: true
2. Client sends thinking.type: "enabled" (Claude format) → sets budget + includeThoughts: true

If neither is present (the common case for clients that don't explicitly request thinking), no thinkingConfig is sent at all, and the upstream treats all output as regular content.

Fix

Add a default thinkingConfig with includeThoughts: true for all modern Gemini models (2.5+) when the client hasn't explicitly requested thinking.

if (!result.generationConfig.thinkingConfig) {
  const modelLower = model.toLowerCase();
  if (
    modelLower.includes("gemini") &&
    !modelLower.includes("gemini-1") &&
    !modelLower.includes("gemini-2.0")
  ) {
    result.generationConfig.thinkingConfig = {
      thinkingBudget: getDefaultThinkingBudget(model) || 24576,
      includeThoughts: true,
    };
  }
}

Design decisions

• Excludes legacy models (gemini-1.x, gemini-2.0) that lack thinking capability. All other Gemini models (2.5+, 3.x, future 4.x) are included automatically.
• Uses getDefaultThinkingBudget(model) when available, falling back to 24576 (the highest budget from the 2.5 era). This respects per-model caps while ensuring a sensible default.
• Does not override explicit client requests. The guard if (!result.generationConfig.thinkingConfig) ensures that reasoning_effort and thinking.type still take precedence.

Files changed

• open-sse/translator/request/openai-to-gemini.ts (+20 lines)

Validation

• Built and tested locally against gemini-3.5-flash-low and gemini-3.5-flash-high via Antigravity.
• Confirmed tokens.reasoning is now non-zero in call logs (e.g. 483 reasoning tokens for flash-low), indicating the upstream is generating and returning thought parts.
• Tool calls (glob, grep, read) execute correctly without _ide suffix leaks.
• No <internal_reminder> or internal reasoning echoed in visible content.

Compatibility

No breaking changes. Models that previously received no thinkingConfig now receive a default one, which only affects how the upstream marks thought parts — it does not change the model's behavior or capabilities. Clients that explicitly request thinking are unaffected.

[gemini-code-assist]

Code Review

This pull request introduces default thinking capability configuration for modern Gemini models (2.5+) in open-sse/translator/request/openai-to-gemini.ts to prevent reasoning leakage. The review feedback highlights a style guide violation for missing tests, an issue with the exclusion logic that incorrectly filters out gemini-2.0-flash-thinking-exp, and a recommendation to wrap the fallback thinking budget in capThinkingBudget to prevent potential API errors.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

According to the Repository Style Guide (Rule 2.9), tests must always be included when changing production code under open-sse/. Since this PR modifies open-sse/translator/request/openai-to-gemini.ts but does not include any test changes, it violates this rule.

Please add a test case to tests/fixtures/translation/openai-to-gemini.json to verify that the default thinkingConfig is correctly applied to modern Gemini models. For example:

{
  "name": "openai-to-gemini/modern-gemini-default-thinking",
  "sourceFormat": "openai",
  "targetFormat": "gemini",
  "input": {
    "model": "gemini-2.5-flash",
    "messages": [{ "role": "user", "content": "Hello Gemini!" }]
  }
}

References

1. Always include tests when changing production code (src/, open-sse/, electron/, bin/). ^(link)

[gemini-code-assist]

The current exclusion logic !modelLower.includes("gemini-2.0") will incorrectly exclude gemini-2.0-flash-thinking-exp, which is a valid thinking model and should have includeThoughts: true enabled by default to prevent reasoning leakage.

We should update the condition to allow models that explicitly contain thinking even if they are in the gemini-2.0 family.

Suggested change

	if (
	modelLower.includes("gemini") &&
	!modelLower.includes("gemini-1") &&
	!modelLower.includes("gemini-2.0")
	) {
	if (
	modelLower.includes("gemini") &&
	!modelLower.includes("gemini-1") &&
	(!modelLower.includes("gemini-2.0") || modelLower.includes("thinking"))
	) {

[gemini-code-assist]

To prevent potential 400 Bad Request errors from the Gemini API if a model has a lower thinking budget cap than 24576, we should wrap the fallback budget in capThinkingBudget(model, 24576). This ensures the fallback budget is always safe and within the model's capabilities.

Suggested change

	result.generationConfig.thinkingConfig = {
	thinkingBudget: getDefaultThinkingBudget(model) || 24576,
	includeThoughts: true,
	};
	result.generationConfig.thinkingConfig = {
	thinkingBudget: getDefaultThinkingBudget(model) || capThinkingBudget(model, 24576),
	includeThoughts: true,
	};

[dhaern]

Hi @gemini-code-assist, thanks for the review. All three points addressed in 072b05f:

1. Missing tests (high) — Added test fixture openai-to-gemini/modern-gemini-default-thinking to tests/fixtures/translation/openai-to-gemini.json covering the default thinkingConfig path for modern Gemini models.

2. gemini-2.0-flash-thinking-exp exclusion (medium) — Updated the condition to allow thinking-capable 2.0 models through: (!modelLower.includes("gemini-2.0") || modelLower.includes("thinking")).

3. Wrap fallback in capThinkingBudget (medium) — Changed 24576 to capThinkingBudget(model, 24576) so the fallback is always within the model's capabilities and never triggers a 400.

[gemini-code-assist]

Thanks for the updates, @dhaern. These changes look excellent and correctly address the points raised. The inclusion of the test fixture and the refinement of the thinking budget logic ensure both correctness and robustness for modern Gemini models. The PR is in great shape.

[diegosouzapw]

Merged into release/v3.8.29 🎉 Thanks @dhaern — nice correctness fix; aligning OmniRoute with Gemini 2.5+'s always-on thinking so reasoning lands in reasoning_content is the right call. On review we just regenerated the two existing gemini golden snapshots (they're also 2.5 models, so they now carry the thinkingConfig) and committed the modern-gemini-default-thinking snapshot whose fixture case you added. All 61 translator + golden tests green. 🙏