diff --git a/backend/internal/service/gateway_request.go b/backend/internal/service/gateway_request.go
index 1cf5728e0c..ff9786431d 100644
--- a/backend/internal/service/gateway_request.go
+++ b/backend/internal/service/gateway_request.go
@@ -1294,10 +1294,10 @@ func RectifyThinkingBudget(body []byte) ([]byte, bool) {
 // NormalizeChineseLLMThinking rewrites the top-level `thinking` object for Chinese
 // LLM providers that use Anthropic-compatible endpoints but have different accepted
 // values for `thinking.type`. Currently scoped to:
-//   - MiniMax M3 / M3.x (`MiniMax-m*`): official docs accept only `thinking.type`
-//     of "adaptive" or "disabled"; "enabled" is not a valid value and may be
-//     rejected/ignored. Pi-ai and other Anthropic-SDK clients default to "enabled"
-//     (Anthropic-original) and never auto-rewrite for non-Anthropic models.
+//   - MiniMax M-series (`MiniMax-m*`, covering M2.x / M3 / M3.x): official docs accept
+//     only `thinking.type` of "adaptive" or "disabled"; "enabled" is not a valid value
+//     and may be rejected/ignored. Pi-ai and other Anthropic-SDK clients default to
+//     "enabled" (Anthropic-original) and never auto-rewrite for non-Anthropic models.
 //
 // Non-MiniMax models (Kimi/GLM/DeepSeek) currently accept "enabled" as-is, so this
 // function is intentionally a no-op for them. New Chinese LLM quirks should be
diff --git a/backend/internal/service/gemini_messages_compat_service.go b/backend/internal/service/gemini_messages_compat_service.go
index f041292e2f..d56df59e4d 100644
--- a/backend/internal/service/gemini_messages_compat_service.go
+++ b/backend/internal/service/gemini_messages_compat_service.go
@@ -832,6 +832,9 @@ func (s *GeminiMessagesCompatService) Forward(ctx context.Context, c *gin.Contex
 
 				var strippedClaudeBody []byte
 				stageName := ""
+				// 路径说明：本处上游是 Gemini，但被剥离的 body 是 Anthropic 格式。传 originalModel
+				// （客户端原 Anthropic model）而非 mappedModel（上游 Gemini model），让剥离逻辑按
+				// 客户端请求的 Anthropic 子协议族判定（详见 ResolveThinkingProtocol 文档）。
 				switch signatureRetryStage {
 				case 0:
 					// Stage 1: disable thinking + thinking->text
diff --git a/backend/internal/service/thinking_protocol.go b/backend/internal/service/thinking_protocol.go
index b7c0f2bc01..686baf467d 100644
--- a/backend/internal/service/thinking_protocol.go
+++ b/backend/internal/service/thinking_protocol.go
@@ -26,9 +26,15 @@ const (
 	ThinkingProtocolPassbackRequired
 )
 
-// ResolveThinkingProtocol 根据「实际发给上游的模型 ID」推断 thinking 协议族。
-// 必须传入映射后的模型 ID（mappedModel），不要传客户端原始请求模型 ID（reqModel），
-// 否则用户配置「claude-sonnet-4-6 → deepseek-v4-pro」时会判错。
+// ResolveThinkingProtocol 根据「作为 thinking block 处理参考的模型 ID」推断 thinking 协议族。
+//
+// 传入参数的语义随调用路径不同：
+//   - **Anthropic gateway**（转发原始 Anthropic 请求）：传 mappedModel（账号级 model mapping
+//     后的上游 model ID）。例：用户配置「claude-sonnet-4-6 → deepseek-v4-pro」后，
+//     传 deepseek-v4-pro 才能被正确判为 passback-required。
+//   - **Gemini messages compat**（Anthropic body → Gemini upstream）：传 originalModel
+//     （客户端 Anthropic 请求的 model ID）。原因：此场景下上游是 Gemini，但被剥
+//     离的 body 是 Anthropic 格式，需按客户端请求的 Anthropic 子协议族判定剥离行为。
 //
 // 匹配规则按厂商前缀硬编码：
 //   - anthropic-strict: claude-* / opus-* / sonnet-* / haiku-*