fix: count Gemini thoughtsTokenCount as output tokens

internal/transformer/convert/claude_gemini.go

@@ -290,7 +290,7 @@ func GeminiRespToClaude(geminiResp []byte) ([]byte, error) {
 	var inputTokens, outputTokens int
 	if resp.UsageMetadata != nil {
 		inputTokens = resp.UsageMetadata.PromptTokenCount
-		outputTokens = resp.UsageMetadata.CandidatesTokenCount
+		outputTokens = geminiOutputTokens(&resp)
 	}
 
 	claudeResp := map[string]interface{}{

internal/transformer/convert/common.go

@@ -87,11 +87,20 @@ func syncGeminiUsageMetadata(resp *transformer.GeminiResponse, ctx *transformer.
 	if resp.UsageMetadata.PromptTokenCount > 0 {
 		ctx.InputTokens = resp.UsageMetadata.PromptTokenCount
 	}
-	if resp.UsageMetadata.CandidatesTokenCount > 0 {
-		ctx.OutputTokens = resp.UsageMetadata.CandidatesTokenCount
+	if output := geminiOutputTokens(resp); output > 0 {
+		ctx.OutputTokens = output
 	}
 }
 
+// geminiOutputTokens returns the billed output tokens: Gemini reports reasoning
+// tokens in thoughtsTokenCount, outside candidatesTokenCount.
+func geminiOutputTokens(resp *transformer.GeminiResponse) int {
+	if resp == nil || resp.UsageMetadata == nil {
+		return 0
+	}
+	return resp.UsageMetadata.CandidatesTokenCount + resp.UsageMetadata.ThoughtsTokenCount
+}
+
 func currentOpenAIUsage(ctx *transformer.StreamContext) map[string]interface{} {
 	if ctx == nil || (ctx.InputTokens == 0 && ctx.OutputTokens == 0) {
 		return nil

internal/transformer/convert/gemini_thoughts_tokens_test.go

@@ -0,0 +1,45 @@
+package convert
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/lich0821/ccNexus/internal/transformer"
+)
+
+// Gemini reports reasoning tokens in thoughtsTokenCount, outside candidatesTokenCount,
+// and bills them as output.
+const geminiThinkingResp = `{"candidates":[{"content":{"role":"model","parts":[{"text":"4"}]},"finishReason":"STOP"}],` +
+	`"usageMetadata":{"promptTokenCount":5,"candidatesTokenCount":3,"thoughtsTokenCount":7,"totalTokenCount":15}}`
+
+func TestGeminiRespToClaudeCountsThoughtTokensAsOutput(t *testing.T) {
+	out, err := GeminiRespToClaude([]byte(geminiThinkingResp))
+	if err != nil {
+		t.Fatalf("GeminiRespToClaude failed: %v", err)
+	}
+
+	assertContains(t, string(out), `"output_tokens":10`, "Expected output_tokens to include the 7 thought tokens")
+}
+
+func TestGeminiRespToOpenAICountsThoughtTokensAsCompletion(t *testing.T) {
+	out, err := GeminiRespToOpenAI([]byte(geminiThinkingResp), "gemini-3.8-flash")
+	if err != nil {
+		t.Fatalf("GeminiRespToOpenAI failed: %v", err)
+	}
+
+	assertContains(t, string(out), `"completion_tokens":10`, "Expected completion_tokens to include the 7 thought tokens")
+}
+
+func TestGeminiStreamToClaudeCountsThoughtTokensAsOutput(t *testing.T) {
+	ctx := transformer.NewStreamContext()
+	ctx.ModelName = "gemini-3.8-flash"
+
+	out, err := GeminiStreamToClaude([]byte("data: "+geminiThinkingResp+"\n\n"), ctx)
+	if err != nil {
+		t.Fatalf("GeminiStreamToClaude failed: %v", err)
+	}
+
+	if !strings.Contains(string(out), `"output_tokens":10`) {
+		t.Errorf("Expected output_tokens 10 (3 candidate + 7 thought) in the stream usage, got: %s", out)
+	}
+}

internal/transformer/convert/openai2_gemini.go

@@ -125,7 +125,7 @@ func GeminiRespToOpenAI2(geminiResp []byte) ([]byte, error) {
 	if resp.UsageMetadata != nil {
 		usage = map[string]interface{}{
 			"input_tokens":  resp.UsageMetadata.PromptTokenCount,
-			"output_tokens": resp.UsageMetadata.CandidatesTokenCount,
+			"output_tokens": geminiOutputTokens(&resp),
 			"total_tokens":  resp.UsageMetadata.TotalTokenCount,
 		}
 	}

internal/transformer/convert/openai_gemini.go

@@ -164,7 +164,7 @@ func GeminiRespToOpenAI(geminiResp []byte, model string) ([]byte, error) {
 	if resp.UsageMetadata != nil {
 		usage = map[string]interface{}{
 			"prompt_tokens":     resp.UsageMetadata.PromptTokenCount,
-			"completion_tokens": resp.UsageMetadata.CandidatesTokenCount,
+			"completion_tokens": geminiOutputTokens(&resp),
 			"total_tokens":      resp.UsageMetadata.TotalTokenCount,
 		}
 	}

internal/transformer/types.go

@@ -308,6 +308,7 @@ type GeminiResponse struct {
 	UsageMetadata *struct {
 		PromptTokenCount     int `json:"promptTokenCount"`
 		CandidatesTokenCount int `json:"candidatesTokenCount"`
+		ThoughtsTokenCount   int `json:"thoughtsTokenCount"`
 		TotalTokenCount      int `json:"totalTokenCount"`
 	} `json:"usageMetadata,omitempty"`
 }
@@ -325,6 +326,7 @@ type GeminiStreamChunk struct {
 	UsageMetadata *struct {
 		PromptTokenCount     int `json:"promptTokenCount"`
 		CandidatesTokenCount int `json:"candidatesTokenCount"`
+		ThoughtsTokenCount   int `json:"thoughtsTokenCount"`
 		TotalTokenCount      int `json:"totalTokenCount"`
 	} `json:"usageMetadata,omitempty"`
 }