debug: log candidatesTokensDetails presence in extractGeminiUsage

backend/internal/service/gemini_messages_compat_service.go

@@ -2705,6 +2705,12 @@ func extractGeminiUsage(data []byte) *ClaudeUsage {
 			return true
 		})
 	}
+	if cand > 0 && imageTokens == 0 {
+		// 调试日志：上游是否返回了 candidatesTokensDetails
+		logger.LegacyPrintf("service.gemini_compat",
+			"extractGeminiUsage: candidatesTokensDetails_exists=%t cand=%d imageTokens=%d raw_usage=%s",
+			candidateDetails.Exists(), cand, imageTokens, truncateForLog([]byte(usage.Raw), 500))
+	}
 
 	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
 	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去

test_gemini.py

@@ -0,0 +1,32 @@
+from google import genai
+from google.genai import types
+
+client = genai.Client(
+    api_key="sk-d0678689ab0000696efa7fc0679026bf7bedb954d98ef5da9aa8161f99227ef8",
+    http_options={"api_version": "v1beta", "base_url": "https://clicodeplus.com"}
+)
+
+try:
+    response = client.models.generate_content(
+        model="gemini-3.1-flash-image-preview",
+        contents="Draw a red circle",
+        config=types.GenerateContentConfig(
+            response_modalities=["TEXT", "IMAGE"]
+        )
+    )
+    print("Response received!")
+    for part in response.candidates[0].content.parts:
+        if part.text:
+            print("Text:", part.text[:200])
+        if part.inline_data:
+            print("Image: mime=%s, size=%d bytes" % (part.inline_data.mime_type, len(part.inline_data.data)))
+    um = response.usage_metadata
+    if um:
+        print("Prompt tokens:", um.prompt_token_count)
+        print("Candidates tokens:", um.candidates_token_count)
+        print("Total tokens:", um.total_token_count)
+        if hasattr(um, "candidates_tokens_details") and um.candidates_tokens_details:
+            for d in um.candidates_tokens_details:
+                print("  Detail: modality=%s tokens=%d" % (d.modality, d.token_count))
+except Exception as e:
+    print("Error:", type(e).__name__, e)