fix: apply thinkingLevel from model suffix metadata for Gemini 3

The previous commit added thinkingLevel support but didn't apply it
when the reasoning effort came from model name suffix (e.g., model(minimal)).

This was because ResolveThinkingConfigFromMetadata returns nil for
level-based models, bypassing the metadata application.

Changes:
- Add ApplyGemini3ThinkingLevelFromMetadata for standard Gemini API
- Add ApplyGemini3ThinkingLevelFromMetadataCLI for CLI API format
- Update gemini_cli_executor to apply Gemini 3 thinkingLevel from metadata
- Update antigravity_executor to apply Gemini 3 thinkingLevel from metadata
- Update aistudio_executor to apply Gemini 3 thinkingLevel from metadata
- Add comprehensive test coverage for Gemini 3 thinkingLevel functions

internal/runtime/executor/aistudio_executor.go

@@ -323,6 +323,7 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	to := sdktranslator.FromString("gemini")
 	payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
 	payload = ApplyThinkingMetadata(payload, req.Metadata, req.Model)
+	payload = util.ApplyGemini3ThinkingLevelFromMetadata(req.Model, req.Metadata, payload)
 	payload = util.ApplyDefaultThinkingIfNeeded(req.Model, payload)
 	payload = util.ConvertThinkingLevelToBudget(payload, req.Model)
 	payload = util.NormalizeGeminiThinkingBudget(req.Model, payload)

internal/runtime/executor/antigravity_executor.go

@@ -89,6 +89,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 
 	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
+	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
 	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
 	translated = normalizeAntigravityThinking(req.Model, translated)
 
@@ -182,6 +183,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
 	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
+	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
 	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
 	translated = normalizeAntigravityThinking(req.Model, translated)
 
@@ -514,6 +516,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
 	translated = applyThinkingMetadataCLI(translated, req.Metadata, req.Model)
+	translated = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, translated)
 	translated = util.ApplyDefaultThinkingIfNeededCLI(req.Model, translated)
 	translated = normalizeAntigravityThinking(req.Model, translated)
 

internal/runtime/executor/gemini_cli_executor.go

@@ -79,6 +79,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	to := sdktranslator.FromString("gemini-cli")
 	basePayload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
 	basePayload = applyThinkingMetadataCLI(basePayload, req.Metadata, req.Model)
+	basePayload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, basePayload)
 	basePayload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, basePayload)
 	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
 	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
@@ -217,6 +218,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	to := sdktranslator.FromString("gemini-cli")
 	basePayload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 	basePayload = applyThinkingMetadataCLI(basePayload, req.Metadata, req.Model)
+	basePayload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, basePayload)
 	basePayload = util.ApplyDefaultThinkingIfNeededCLI(req.Model, basePayload)
 	basePayload = util.NormalizeGeminiCLIThinkingBudget(req.Model, basePayload)
 	basePayload = util.StripThinkingConfigIfUnsupported(req.Model, basePayload)
@@ -418,6 +420,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	for _, attemptModel := range models {
 		payload := sdktranslator.TranslateRequest(from, to, attemptModel, bytes.Clone(req.Payload), false)
 		payload = applyThinkingMetadataCLI(payload, req.Metadata, req.Model)
+		payload = util.ApplyGemini3ThinkingLevelFromMetadataCLI(req.Model, req.Metadata, payload)
 		payload = deleteJSONField(payload, "project")
 		payload = deleteJSONField(payload, "model")
 		payload = deleteJSONField(payload, "request.safetySettings")

internal/util/gemini_thinking.go

@@ -274,6 +274,42 @@ func ApplyDefaultThinkingIfNeeded(model string, body []byte) []byte {
 	return updated
 }
 
+// ApplyGemini3ThinkingLevelFromMetadata applies thinkingLevel from metadata for Gemini 3 models.
+// For standard Gemini API format (generationConfig.thinkingConfig path).
+// This handles the case where reasoning_effort is specified via model name suffix (e.g., model(minimal)).
+func ApplyGemini3ThinkingLevelFromMetadata(model string, metadata map[string]any, body []byte) []byte {
+	if !IsGemini3Model(model) {
+		return body
+	}
+	effort, ok := ReasoningEffortFromMetadata(metadata)
+	if !ok || effort == "" {
+		return body
+	}
+	// Validate and apply the thinkingLevel
+	if level, valid := ValidateGemini3ThinkingLevel(model, effort); valid {
+		return ApplyGeminiThinkingLevel(body, level, nil)
+	}
+	return body
+}
+
+// ApplyGemini3ThinkingLevelFromMetadataCLI applies thinkingLevel from metadata for Gemini 3 models.
+// For Gemini CLI API format (request.generationConfig.thinkingConfig path).
+// This handles the case where reasoning_effort is specified via model name suffix (e.g., model(minimal)).
+func ApplyGemini3ThinkingLevelFromMetadataCLI(model string, metadata map[string]any, body []byte) []byte {
+	if !IsGemini3Model(model) {
+		return body
+	}
+	effort, ok := ReasoningEffortFromMetadata(metadata)
+	if !ok || effort == "" {
+		return body
+	}
+	// Validate and apply the thinkingLevel
+	if level, valid := ValidateGemini3ThinkingLevel(model, effort); valid {
+		return ApplyGeminiCLIThinkingLevel(body, level, nil)
+	}
+	return body
+}
+
 // ApplyDefaultThinkingIfNeededCLI injects default thinkingConfig for models that require it.
 // For Gemini CLI API format (request.generationConfig.thinkingConfig path).
 // Returns the modified body if thinkingConfig was added, otherwise returns the original.