From 8ac2936a3d8cc9c61b02e3f92f10de343afbed78 Mon Sep 17 00:00:00 2001
From: SavitarC <71271200+SavitarC@users.noreply.github.com>
Date: Sun, 27 Sep 2026 11:08:41 +0800
Subject: [PATCH] =?UTF-8?q?refactor(gateway):=20=E4=B8=89=E4=B8=AA?=
 =?UTF-8?q?=E5=85=A5=E7=AB=99=E5=85=A5=E5=8F=A3=E7=9A=84=E4=B8=8A=E6=B8=B8?=
 =?UTF-8?q?=E5=8D=8F=E8=AE=AE=E5=88=86=E6=B5=81=E6=94=B6=E6=95=9B=E4=B8=BA?=
 =?UTF-8?q?=20resolveUpstreamProtocol?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

此前 /v1/responses、/v1/chat/completions、/v1/messages 各自用一串 if 决定
上游协议：OpenCode 按模型规则、国产 adaptive 按入站、显式 api_protocol、
Responses 探针四个来源的优先级靠语句先后隐式决定，且每个入口都有
IsOpenCodeGo 特判。

- provider profile 新增 Routing（by_inbound / by_model），OpenCode 登记为
  by_model；routesByModel 取代转发路径与 GetAPIProtocol 默认值中的
  IsOpenCodeGo 判断。
- 新增 upstream_protocol_routing.go：resolveUpstreamProtocol 以单一函数给出
  上游协议（按模型分流 → adaptive 按入站 → 显式 anthropic → CC 直转条件
  → Responses），resolveModelRoutedProtocol 以 profile 规则泛化 OpenCode
  的规则解析；三个入口改为 switch 该结果选择既有转换路径，Responses 形状
  请求体的转换合并为 convertResponsesShapedChatBody。

转换函数本身不变。新增分流矩阵测试经真实入口遍历 平台 × api_protocol ×
探针 × 账号类型 × 接入模式 × protocol_rules × 入站 × 模型（2045 组），
与改动前逐条 diff 仅 6 条不同：智谱 adaptive 的非 API Key 账号（UI 无法
创建）在 /v1/responses 入站时由 Responses 改为回落 CC，与
/v1/chat/completions 入口一致；这 6 组前后都在发出上游请求前报错。

协议规则与映射模型相关的类型、函数已被所有按模型分流的供应商共用，同时去掉
OpenCode 前缀（OpenCodeGoProtocolRule → ProtocolRule、matchOpenCodeGoProtocolRules
→ matchProtocolRules、NormalizeOpenCodeGoProtocolRulesCredentials →
NormalizeProtocolRulesCredentials、resolveOpenCodeGoMappedModel →
resolveMappedUpstreamModel 等），并删除收敛后已无调用方的
defaultOpenCodeProtocolRules。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 backend/internal/service/account.go           |   3 +-
 backend/internal/service/admin_account.go     |   8 +-
 .../openai_gateway_chat_completions.go        |  99 ++---
 .../service/openai_gateway_forward.go         |  41 +--
 .../service/openai_gateway_messages.go        |  39 +-
 ...penai_gateway_messages_anthropic_native.go |   2 +-
 backend/internal/service/opencode_go.go       |  84 ++---
 backend/internal/service/opencode_go_test.go  |  30 +-
 backend/internal/service/provider_profile.go  |  29 +-
 .../internal/service/provider_profile_test.go |   9 +-
 .../service/upstream_protocol_routing.go      |  92 +++++
 .../upstream_protocol_routing_matrix_test.go  | 344 ++++++++++++++++++
 .../service/upstream_protocol_routing_test.go | 119 ++++++
 13 files changed, 704 insertions(+), 195 deletions(-)
 create mode 100644 backend/internal/service/upstream_protocol_routing.go
 create mode 100644 backend/internal/service/upstream_protocol_routing_matrix_test.go
 create mode 100644 backend/internal/service/upstream_protocol_routing_test.go

diff --git a/backend/internal/service/account.go b/backend/internal/service/account.go
index 783fffc349..97a797580c 100644
--- a/backend/internal/service/account.go
+++ b/backend/internal/service/account.go
@@ -1444,7 +1444,8 @@ func (a *Account) GetAPIProtocol() string {
 	case APIProtocolChatCompletions:
 		return APIProtocolChatCompletions
 	}
-	if a.IsOpenCodeGo() {
+	// 按模型分流的供应商（多模型聚合平台）未显式配置时默认 adaptive。
+	if a.routesByModel() {
 		return APIProtocolAdaptive
 	}
 	return APIProtocolChatCompletions
diff --git a/backend/internal/service/admin_account.go b/backend/internal/service/admin_account.go
index ef4a47fa1f..256e488d09 100644
--- a/backend/internal/service/admin_account.go
+++ b/backend/internal/service/admin_account.go
@@ -323,7 +323,7 @@ func (s *adminServiceImpl) DuplicateAccount(ctx context.Context, id int64, actor
 	if err := NormalizeHeaderOverrideCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
-	if err := NormalizeOpenCodeGoProtocolRulesCredentials(input.Credentials); err != nil {
+	if err := NormalizeProtocolRulesCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
 	duplicate, err := buildAccountForCreate(input, accountExtra)
@@ -522,7 +522,7 @@ func (s *adminServiceImpl) CreateAccount(ctx context.Context, input *CreateAccou
 	if err := NormalizeHeaderOverrideCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
-	if err := NormalizeOpenCodeGoProtocolRulesCredentials(input.Credentials); err != nil {
+	if err := NormalizeProtocolRulesCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
 	// Never persist ephemeral SSO/password secrets after OAuth conversion.
@@ -655,7 +655,7 @@ func (s *adminServiceImpl) UpdateAccount(ctx context.Context, id int64, input *U
 		if err := NormalizeHeaderOverrideCredentials(account.Credentials); err != nil {
 			return nil, err
 		}
-		if err := NormalizeOpenCodeGoProtocolRulesCredentials(account.Credentials); err != nil {
+		if err := NormalizeProtocolRulesCredentials(account.Credentials); err != nil {
 			return nil, err
 		}
 		// Strip SSO/password residue that must never sit next to OAuth tokens.
@@ -1102,7 +1102,7 @@ func (s *adminServiceImpl) BulkUpdateAccounts(ctx context.Context, input *BulkUp
 	if err := NormalizeHeaderOverrideCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
-	if err := NormalizeOpenCodeGoProtocolRulesCredentials(input.Credentials); err != nil {
+	if err := NormalizeProtocolRulesCredentials(input.Credentials); err != nil {
 		return nil, err
 	}
 	// Bulk may mix platforms; always drop ephemeral SSO/password keys (cookie
diff --git a/backend/internal/service/openai_gateway_chat_completions.go b/backend/internal/service/openai_gateway_chat_completions.go
index 7ffd20916f..66c2471ae7 100644
--- a/backend/internal/service/openai_gateway_chat_completions.go
+++ b/backend/internal/service/openai_gateway_chat_completions.go
@@ -110,88 +110,43 @@ func (s *OpenAIGatewayService) forwardAsChatCompletions(
 	}
 
 	// Cursor compatibility: some clients send a Responses-shaped body to the
-	// /v1/chat/completions URL. Detect it before adaptive routing so adaptive
+	// /v1/chat/completions URL. Detect it before protocol routing so adaptive
 	// accounts never forward the body unchanged to a Chat Completions endpoint.
 	isResponsesShape := !gjson.GetBytes(body, "messages").Exists() && gjson.GetBytes(body, "input").Exists()
-
-	// OpenCode Go：按模型原生协议分流（与 inbound 协议正交）。
-	// 规则未命中一律兜底 Chat Completions，只有显式 Responses 才走下方转换链。
-	if account.IsOpenCodeGo() {
-		mapped := resolveOpenCodeGoMappedModel(account, body, defaultMappedModel)
-		if IsOpenCodeUnsupportedModel(mapped) {
-			return nil, writeOpenCodeUnsupportedModelError(c, false, mapped)
-		}
-		proto := openCodeGoNativeProtocol(account, mapped)
-		if proto != APIProtocolResponses {
-			if isResponsesShape {
-				if proto == APIProtocolAnthropic {
-					return s.forwardResponsesViaNativeAnthropic(ctx, c, account, body, "")
-				}
-				var responsesReq apicompat.ResponsesRequest
-				if err := json.Unmarshal(body, &responsesReq); err != nil {
-					return nil, fmt.Errorf("parse responses-shaped chat completions request: %w", err)
-				}
-				chatReq, err := apicompat.ResponsesToChatCompletionsRequestWithOptions(
-					&responsesReq,
-					&apicompat.ResponsesToChatOptions{ReasoningContentByID: s.reasoningContentByID},
-				)
-				if err != nil {
-					return nil, fmt.Errorf("convert responses-shaped chat completions request: %w", err)
-				}
-				chatBody, err := json.Marshal(chatReq)
-				if err != nil {
-					return nil, fmt.Errorf("marshal converted chat completions request: %w", err)
-				}
-				return s.forwardAsRawChatCompletions(ctx, c, account, chatBody, defaultMappedModel)
-			}
-			if proto == APIProtocolAnthropic {
-				return s.forwardChatCompletionsViaNativeAnthropic(ctx, c, account, body, defaultMappedModel)
-			}
-			return s.forwardAsRawChatCompletions(ctx, c, account, body, defaultMappedModel)
-		}
+	inbound := APIProtocolChatCompletions
+	if isResponsesShape {
+		inbound = APIProtocolResponses
 	}
+	// 按模型分流与 adaptive 账号按请求体形状处理 Responses 形状入站；其余账号
+	// 维持原样把请求体交给所选路径。
+	convertResponsesShape := isResponsesShape && (account.routesByModel() || account.IsAdaptiveAPIProtocol())
 
-	// 自适应账号的标准 Chat Completions 入站使用供应商原生 CC 端点。
-	// Responses 形状下，DeepSeek / Kimi 继续走下方原生 Responses 链；GLM
-	// 没有 Responses 端点，先转换成 Chat Completions 再直转。
-	if account.IsAdaptiveAPIProtocol() && !account.IsOpenCodeGo() {
-		if !isResponsesShape {
-			return s.forwardAsRawChatCompletions(ctx, c, account, body, defaultMappedModel)
+	// 上游协议统一由 resolveUpstreamProtocol 判定。Anthropic 分流必须先于
+	// ShouldUseResponsesAPI：Anthropic 协议账号经 probe 落标
+	// openai_responses_supported=false，否则会命中 CC 直转。
+	routingModel := upstreamRoutingModel(account, body, defaultMappedModel)
+	if account.IsOpenCodeGo() && IsOpenCodeUnsupportedModel(routingModel) {
+		return nil, writeOpenCodeUnsupportedModelError(c, false, routingModel)
+	}
+	switch resolveUpstreamProtocol(account, inbound, routingModel) {
+	case APIProtocolAnthropic:
+		if convertResponsesShape {
+			return s.forwardResponsesViaNativeAnthropic(ctx, c, account, body, "")
 		}
-		if !account.SupportsNativeCNResponses() {
-			var responsesReq apicompat.ResponsesRequest
-			if err := json.Unmarshal(body, &responsesReq); err != nil {
-				return nil, fmt.Errorf("parse responses-shaped chat completions request: %w", err)
-			}
-			chatReq, err := apicompat.ResponsesToChatCompletionsRequestWithOptions(
-				&responsesReq,
-				&apicompat.ResponsesToChatOptions{ReasoningContentByID: s.reasoningContentByID},
-			)
+		// CC 入站经 CC→Responses→Anthropic 转换链直通供应商原生 Anthropic 端点。
+		return s.forwardChatCompletionsViaNativeAnthropic(ctx, c, account, body, defaultMappedModel)
+	case APIProtocolChatCompletions:
+		if convertResponsesShape {
+			chatBody, err := s.convertResponsesShapedChatBody(body)
 			if err != nil {
-				return nil, fmt.Errorf("convert responses-shaped chat completions request: %w", err)
-			}
-			chatBody, err := json.Marshal(chatReq)
-			if err != nil {
-				return nil, fmt.Errorf("marshal converted chat completions request: %w", err)
+				return nil, err
 			}
 			return s.forwardAsRawChatCompletions(ctx, c, account, chatBody, defaultMappedModel)
 		}
-		// DeepSeek / Kimi 原生 Responses 请求继续走下方 Responses→Chat 回程转换。
-	}
-
-	// 入口分流（国产供应商 Anthropic 协议）：上游为供应商原生 Anthropic 端点，
-	// CC 入站请求经 CC→Responses→Anthropic 转换链直通该端点。必须先于
-	// ShouldUseResponsesAPI 分流：该类账号经 probe 落标
-	// openai_responses_supported=false，会先命中下方的 CC 直转分支。
-	if account.IsAnthropicProtocol() {
-		return s.forwardChatCompletionsViaNativeAnthropic(ctx, c, account, body, defaultMappedModel)
-	}
-
-	// 固定 chat_completions 的 CN 账号，以及强制或已探测确认不支持 Responses
-	// 的其他 APIKey 账号，均走 CC 直转。
-	if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
 		return s.forwardAsRawChatCompletions(ctx, c, account, body, defaultMappedModel)
 	}
+	// 上游为 Responses（含 adaptive 下 Responses 形状入站且供应商有原生
+	// Responses 端点）：走下方 Chat Completions ↔ Responses 转换链。
 	SetActualOpenAIUpstreamEndpoint(c, openAIResponsesUpstreamEndpoint)
 
 	startTime := time.Now()
@@ -424,7 +379,7 @@ func (s *OpenAIGatewayService) forwardAsChatCompletions(
 			return s.forwardAsChatCompletions(markAgentIdentityTaskRecoveryTried(ctx), c, account, body, promptCacheKey, defaultMappedModel, compatPromptCacheTenantIsolated)
 		}
 		if account.Type == AccountTypeAPIKey &&
-			!account.IsOpenCodeGo() &&
+			!account.routesByModel() &&
 			openai_compat.ResolveResponsesSupport(account.Extra) == openai_compat.ResponsesSupportUnknown &&
 			!isResponsesEndpointSupportedByStatus(resp.StatusCode) {
 			logger.L().Info("openai chat_completions: /responses unsupported, falling back to raw chat completions",
diff --git a/backend/internal/service/openai_gateway_forward.go b/backend/internal/service/openai_gateway_forward.go
index 78dd47afe6..03d239774b 100644
--- a/backend/internal/service/openai_gateway_forward.go
+++ b/backend/internal/service/openai_gateway_forward.go
@@ -159,27 +159,6 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 		return s.forwardGrokResponses(ctx, c, account, body, originalModel, reqStream, startTime)
 	}
 
-	if account.IsOpenCodeGo() {
-		mapped := resolveOpenCodeGoMappedModel(account, body, "")
-		if IsOpenCodeUnsupportedModel(mapped) {
-			return nil, writeOpenCodeUnsupportedModelError(c, false, mapped)
-		}
-		switch openCodeGoNativeProtocol(account, mapped) {
-		case APIProtocolAnthropic:
-			return s.forwardResponsesViaNativeAnthropic(ctx, c, account, body, "")
-		case APIProtocolResponses:
-			break
-		default:
-			return s.forwardResponsesViaRawChatCompletions(ctx, c, account, body)
-		}
-	}
-
-	// CN 供应商 anthropic 协议账号：/v1/responses 入站是交叉协议组合
-	// （Responses 客户端 × Anthropic 上游），转成 Anthropic 请求走原生端点。
-	// 不能落到下面的 raw-CC 分支——其 URL 构造会把 anthropic base 当 CC base 用。
-	if account.IsAnthropicProtocol() {
-		return s.forwardResponsesViaNativeAnthropic(ctx, c, account, body, reqModel)
-	}
 	if account.IsOpenAIApiKey() {
 		if normalized, changed, normalizeErr := normalizeOpenAIParallelToolCallsWithoutTools(body, responsesLite); normalizeErr != nil {
 			return nil, normalizeErr
@@ -198,7 +177,23 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 		originalModel = reqModel
 	}
 
-	if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+	// 上游协议统一由 resolveUpstreamProtocol 判定。OpenAI API Key 账号只会落到
+	// Responses / Chat Completions，上面的归一化对两条路径都生效。
+	routingModel := upstreamRoutingModel(account, body, "")
+	if account.IsOpenCodeGo() && IsOpenCodeUnsupportedModel(routingModel) {
+		return nil, writeOpenCodeUnsupportedModelError(c, false, routingModel)
+	}
+	switch resolveUpstreamProtocol(account, APIProtocolResponses, routingModel) {
+	case APIProtocolAnthropic:
+		// Responses 客户端 × Anthropic 上游：转成 Anthropic 请求走原生端点。不能落到
+		// raw-CC 分支——其 URL 构造会把 anthropic base 当 CC base 用。
+		// 按模型分流的账号不以请求模型兜底模型映射。
+		defaultMappedModel := reqModel
+		if account.routesByModel() {
+			defaultMappedModel = ""
+		}
+		return s.forwardResponsesViaNativeAnthropic(ctx, c, account, body, defaultMappedModel)
+	case APIProtocolChatCompletions:
 		return s.forwardResponsesViaRawChatCompletions(ctx, c, account, body)
 	}
 	SetActualOpenAIUpstreamEndpoint(c, openAIResponsesUpstreamEndpoint)
@@ -1355,7 +1350,7 @@ func shouldForwardOpenAIResponsesViaRawChatCompletions(account *Account) bool {
 	if account == nil || account.Type != AccountTypeAPIKey {
 		return false
 	}
-	if account.IsOpenCodeGo() {
+	if account.routesByModel() {
 		// Model protocol_rules are the authority. Probe Extra must not collapse
 		// Grok/GPT/Muse into Chat Completions.
 		return false
diff --git a/backend/internal/service/openai_gateway_messages.go b/backend/internal/service/openai_gateway_messages.go
index a45dd5e565..9594a79a73 100644
--- a/backend/internal/service/openai_gateway_messages.go
+++ b/backend/internal/service/openai_gateway_messages.go
@@ -53,32 +53,21 @@ func (s *OpenAIGatewayService) ForwardAsAnthropic(
 		return nil, err
 	}
 
-	// OpenCode Go：按模型原生协议分流。规则未命中兜底 Chat Completions。
-	if account.IsOpenCodeGo() {
-		mapped := resolveOpenCodeGoMappedModel(account, body, defaultMappedModel)
-		if IsOpenCodeUnsupportedModel(mapped) {
-			return nil, writeOpenCodeUnsupportedModelError(c, true, mapped)
-		}
-		switch openCodeGoNativeProtocol(account, mapped) {
-		case APIProtocolAnthropic:
-			return s.forwardAnthropicViaNativeAnthropicEndpoint(ctx, c, account, body, defaultMappedModel)
-		case APIProtocolResponses:
-			break
-		default:
-			return s.forwardAnthropicViaRawChatCompletions(ctx, c, account, body, defaultMappedModel)
-		}
-	} else if account.IsAnthropicProtocol() || account.IsAdaptiveAPIProtocol() {
-		// 入口分流（国产供应商 Anthropic 协议）：上游为供应商原生 Anthropic 端点时，
-		// /v1/messages 请求零转换直通（仅模型名映射 + 少量 body 清洗），完整保留
-		// thinking / tool_use / cache 语义，适配 Claude Code 等原生客户端。
-		// 必须先于 ShouldUseResponsesAPI 分流：Anthropic 协议账号经 probe 落标
-		// openai_responses_supported=false，会先命中下方的 CC 直转分支。
-		return s.forwardAnthropicViaNativeAnthropicEndpoint(ctx, c, account, body, defaultMappedModel)
+	// 上游协议统一由 resolveUpstreamProtocol 判定。Anthropic 分流必须先于
+	// ShouldUseResponsesAPI：Anthropic 协议账号经 probe 落标
+	// openai_responses_supported=false，否则会命中 CC 直转。
+	routingModel := upstreamRoutingModel(account, body, defaultMappedModel)
+	if account.IsOpenCodeGo() && IsOpenCodeUnsupportedModel(routingModel) {
+		return nil, writeOpenCodeUnsupportedModelError(c, true, routingModel)
 	}
-
-	// 固定 chat_completions 的 CN 账号，以及不支持 Responses 的其他 APIKey
-	// 账号，均将 Messages 转为 CC；固定 responses 的 CN 账号不受探针旧值覆盖。
-	if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+	switch resolveUpstreamProtocol(account, APIProtocolAnthropic, routingModel) {
+	case APIProtocolAnthropic:
+		// 上游为供应商原生 Anthropic 端点：/v1/messages 零转换直通（仅模型名映射 +
+		// 少量 body 清洗），完整保留 thinking / tool_use / cache 语义。
+		return s.forwardAnthropicViaNativeAnthropicEndpoint(ctx, c, account, body, defaultMappedModel)
+	case APIProtocolChatCompletions:
+		// 固定 chat_completions 的多协议账号、规则命中 CC 的按模型分流账号，以及
+		// 不支持 Responses 的其他 APIKey 账号，均将 Messages 转为 CC。
 		return s.forwardAnthropicViaRawChatCompletions(ctx, c, account, body, defaultMappedModel)
 	}
 	SetActualOpenAIUpstreamEndpoint(c, openAIResponsesUpstreamEndpoint)
diff --git a/backend/internal/service/openai_gateway_messages_anthropic_native.go b/backend/internal/service/openai_gateway_messages_anthropic_native.go
index 0026046a58..eb061c9750 100644
--- a/backend/internal/service/openai_gateway_messages_anthropic_native.go
+++ b/backend/internal/service/openai_gateway_messages_anthropic_native.go
@@ -140,7 +140,7 @@ func (s *OpenAIGatewayService) nativeAnthropicTargetURL(account *Account) (strin
 	return strings.TrimRight(validatedURL, "/") + "/v1/messages", nil
 }
 
-func resolveOpenCodeGoMappedModel(account *Account, body []byte, defaultMappedModel string) string {
+func resolveMappedUpstreamModel(account *Account, body []byte, defaultMappedModel string) string {
 	original := strings.TrimSpace(gjson.GetBytes(body, "model").String())
 	billing := resolveOpenAIForwardModel(account, original, defaultMappedModel)
 	return normalizeOpenAIModelForUpstream(account, billing)
diff --git a/backend/internal/service/opencode_go.go b/backend/internal/service/opencode_go.go
index 073a7bd116..d052451a56 100644
--- a/backend/internal/service/opencode_go.go
+++ b/backend/internal/service/opencode_go.go
@@ -19,9 +19,9 @@ const (
 	// the UI does not pick a model. glm-5.3 is a Chat Completions catalog ID.
 	DefaultOpenCodeGoTestModel = "glm-5.3"
 
-	openCodeGoProtocolRulesKey         = "protocol_rules"
-	maxOpenCodeGoProtocolRules         = 64
-	maxOpenCodeGoProtocolPatternLength = 128
+	protocolRulesCredentialKey   = "protocol_rules"
+	maxProtocolRules             = 64
+	maxProtocolRulePatternLength = 128
 )
 
 // DefaultOpenCodeGoModelIDs 是官方文档当前公开的模型 ID 目录，
@@ -101,16 +101,16 @@ func writeOpenCodeUnsupportedModelError(c *gin.Context, isAnthropic bool, model
 	return fmt.Errorf("opencode unsupported model: %s", model)
 }
 
-// OpenCodeGoProtocolRule is one model-pattern → native protocol mapping.
+// ProtocolRule is one model-pattern → native protocol mapping.
 // Pattern is an exact ID or a suffix glob (foo* / *). First match wins.
-type OpenCodeGoProtocolRule struct {
+type ProtocolRule struct {
 	Pattern  string `json:"pattern"`
 	Protocol string `json:"protocol"`
 }
 
 // DefaultOpenCodeGoProtocolRules is the built-in adaptive routing table for Go.
-func DefaultOpenCodeGoProtocolRules() []OpenCodeGoProtocolRule {
-	return []OpenCodeGoProtocolRule{
+func DefaultOpenCodeGoProtocolRules() []ProtocolRule {
+	return []ProtocolRule{
 		{Pattern: "grok-*", Protocol: APIProtocolResponses},
 		{Pattern: "gpt-*", Protocol: APIProtocolResponses},
 		{Pattern: "muse-spark-*", Protocol: APIProtocolResponses},
@@ -122,8 +122,8 @@ func DefaultOpenCodeGoProtocolRules() []OpenCodeGoProtocolRule {
 // DefaultOpenCodeZenProtocolRules 对齐 https://opencode.ai/docs/zen/ 端点表：
 // GPT/Grok/Muse Spark → Responses，Claude/Qwen(除 qwen3.8-max) → Anthropic，
 // qwen3.8-max 及其余模型 → Chat Completions。
-func DefaultOpenCodeZenProtocolRules() []OpenCodeGoProtocolRule {
-	return []OpenCodeGoProtocolRule{
+func DefaultOpenCodeZenProtocolRules() []ProtocolRule {
+	return []ProtocolRule{
 		{Pattern: "grok-*", Protocol: APIProtocolResponses},
 		{Pattern: "gpt-*", Protocol: APIProtocolResponses},
 		{Pattern: "muse-spark-*", Protocol: APIProtocolResponses},
@@ -133,12 +133,7 @@ func DefaultOpenCodeZenProtocolRules() []OpenCodeGoProtocolRule {
 	}
 }
 
-// defaultOpenCodeProtocolRules 从 provider profile 取该接入模式的内置分流表。
-func defaultOpenCodeProtocolRules(mode string) []OpenCodeGoProtocolRule {
-	return LookupProviderProfile(PlatformOpenCodeGo).Endpoints(mode).ProtocolRules
-}
-
-func isNativeOpenCodeGoProtocol(protocol string) bool {
+func isNativeUpstreamProtocol(protocol string) bool {
 	switch protocol {
 	case APIProtocolChatCompletions, APIProtocolAnthropic, APIProtocolResponses:
 		return true
@@ -147,7 +142,7 @@ func isNativeOpenCodeGoProtocol(protocol string) bool {
 	}
 }
 
-func openCodeGoPatternMatches(pattern, model string) bool {
+func protocolRulePatternMatches(pattern, model string) bool {
 	pattern = strings.ToLower(strings.TrimSpace(pattern))
 	model = normalizeOpenCodeGoModelID(model)
 	if pattern == "" || model == "" {
@@ -162,12 +157,12 @@ func openCodeGoPatternMatches(pattern, model string) bool {
 	return model == pattern
 }
 
-func matchOpenCodeGoProtocolRules(model string, rules []OpenCodeGoProtocolRule) string {
+func matchProtocolRules(model string, rules []ProtocolRule) string {
 	for _, rule := range rules {
-		if !isNativeOpenCodeGoProtocol(rule.Protocol) {
+		if !isNativeUpstreamProtocol(rule.Protocol) {
 			continue
 		}
-		if openCodeGoPatternMatches(rule.Pattern, model) {
+		if protocolRulePatternMatches(rule.Pattern, model) {
 			return rule.Protocol
 		}
 	}
@@ -178,33 +173,33 @@ func matchOpenCodeGoProtocolRules(model string, rules []OpenCodeGoProtocolRule)
 // 未命中时为 Chat Completions。账号自定义 protocol_rules 见
 // ResolveOpenCodeGoUpstreamProtocol。
 func OpenCodeGoModelProtocol(model string) string {
-	return matchOpenCodeGoProtocolRules(model, DefaultOpenCodeGoProtocolRules())
+	return matchProtocolRules(model, DefaultOpenCodeGoProtocolRules())
 }
 
-func (a *Account) openCodeGoProtocolRules() ([]OpenCodeGoProtocolRule, bool) {
+func (a *Account) configuredProtocolRules() ([]ProtocolRule, bool) {
 	if a == nil || a.Credentials == nil {
 		return nil, false
 	}
-	raw, ok := a.Credentials[openCodeGoProtocolRulesKey]
+	raw, ok := a.Credentials[protocolRulesCredentialKey]
 	if !ok || raw == nil {
 		return nil, false
 	}
-	rules, err := parseOpenCodeGoProtocolRules(raw)
+	rules, err := parseProtocolRules(raw)
 	if err != nil {
 		return nil, false
 	}
 	return rules, true
 }
 
-func parseOpenCodeGoProtocolRules(raw any) ([]OpenCodeGoProtocolRule, error) {
-	items, err := openCodeGoProtocolRuleItems(raw)
+func parseProtocolRules(raw any) ([]ProtocolRule, error) {
+	items, err := protocolRuleItems(raw)
 	if err != nil {
 		return nil, err
 	}
-	if len(items) > maxOpenCodeGoProtocolRules {
-		return nil, fmt.Errorf("protocol_rules supports at most %d entries", maxOpenCodeGoProtocolRules)
+	if len(items) > maxProtocolRules {
+		return nil, fmt.Errorf("protocol_rules supports at most %d entries", maxProtocolRules)
 	}
-	rules := make([]OpenCodeGoProtocolRule, 0, len(items))
+	rules := make([]ProtocolRule, 0, len(items))
 	for i, item := range items {
 		entry, ok := item.(map[string]any)
 		if !ok {
@@ -212,20 +207,20 @@ func parseOpenCodeGoProtocolRules(raw any) ([]OpenCodeGoProtocolRule, error) {
 		}
 		pattern, _ := entry["pattern"].(string)
 		protocol, _ := entry["protocol"].(string)
-		pattern, err := normalizeOpenCodeGoProtocolPattern(pattern)
+		pattern, err := normalizeProtocolRulePattern(pattern)
 		if err != nil {
 			return nil, fmt.Errorf("protocol_rules[%d]: %w", i, err)
 		}
 		protocol = strings.TrimSpace(protocol)
-		if !isNativeOpenCodeGoProtocol(protocol) {
+		if !isNativeUpstreamProtocol(protocol) {
 			return nil, fmt.Errorf("protocol_rules[%d]: protocol must be chat_completions, anthropic, or responses", i)
 		}
-		rules = append(rules, OpenCodeGoProtocolRule{Pattern: pattern, Protocol: protocol})
+		rules = append(rules, ProtocolRule{Pattern: pattern, Protocol: protocol})
 	}
 	return rules, nil
 }
 
-func openCodeGoProtocolRuleItems(raw any) ([]any, error) {
+func protocolRuleItems(raw any) ([]any, error) {
 	switch items := raw.(type) {
 	case []any:
 		return items, nil
@@ -240,12 +235,12 @@ func openCodeGoProtocolRuleItems(raw any) ([]any, error) {
 	}
 }
 
-func normalizeOpenCodeGoProtocolPattern(pattern string) (string, error) {
+func normalizeProtocolRulePattern(pattern string) (string, error) {
 	pattern = strings.ToLower(strings.TrimSpace(pattern))
 	if pattern == "" {
 		return "", fmt.Errorf("pattern is required")
 	}
-	if len(pattern) > maxOpenCodeGoProtocolPatternLength {
+	if len(pattern) > maxProtocolRulePatternLength {
 		return "", fmt.Errorf("pattern is too long")
 	}
 	if strings.ContainsAny(pattern, " \t") {
@@ -258,17 +253,17 @@ func normalizeOpenCodeGoProtocolPattern(pattern string) (string, error) {
 	return pattern, nil
 }
 
-// NormalizeOpenCodeGoProtocolRulesCredentials 校验并原地规范化 credentials.protocol_rules。
+// NormalizeProtocolRulesCredentials 校验并原地规范化 credentials.protocol_rules。
 // 未携带该字段时为 no-op，使旧账号继续使用内置默认表。
-func NormalizeOpenCodeGoProtocolRulesCredentials(credentials map[string]any) error {
+func NormalizeProtocolRulesCredentials(credentials map[string]any) error {
 	if credentials == nil {
 		return nil
 	}
-	raw, ok := credentials[openCodeGoProtocolRulesKey]
+	raw, ok := credentials[protocolRulesCredentialKey]
 	if !ok || raw == nil {
 		return nil
 	}
-	rules, err := parseOpenCodeGoProtocolRules(raw)
+	rules, err := parseProtocolRules(raw)
 	if err != nil {
 		return infraerrors.New(http.StatusBadRequest, "INVALID_OPENCODE_GO_PROTOCOL_RULES", err.Error())
 	}
@@ -279,7 +274,7 @@ func NormalizeOpenCodeGoProtocolRulesCredentials(credentials map[string]any) err
 			"protocol": rule.Protocol,
 		})
 	}
-	credentials[openCodeGoProtocolRulesKey] = encoded
+	credentials[protocolRulesCredentialKey] = encoded
 	return nil
 }
 
@@ -327,19 +322,12 @@ func openCodeGoNativeProtocol(account *Account, model string) string {
 // ResolveOpenCodeGoUpstreamProtocol 按账号协议配置与模型规则决定上游协议。
 // 显式 pinned 协议优先；adaptive（默认）先走 credentials.protocol_rules，
 // 未配置时回落内置默认表；已配置但未命中则走 Chat Completions。
+// 规则见 resolveModelRoutedProtocol。
 func (a *Account) ResolveOpenCodeGoUpstreamProtocol(model string) string {
 	if a == nil || !a.IsOpenCodeGo() {
 		return ""
 	}
-	switch a.GetAPIProtocol() {
-	case APIProtocolChatCompletions, APIProtocolAnthropic, APIProtocolResponses:
-		return a.GetAPIProtocol()
-	default:
-		if rules, present := a.openCodeGoProtocolRules(); present {
-			return matchOpenCodeGoProtocolRules(model, rules)
-		}
-		return matchOpenCodeGoProtocolRules(model, defaultOpenCodeProtocolRules(a.GetOpenCodeAccountMode()))
-	}
+	return a.resolveModelRoutedProtocol(model)
 }
 
 // openCodeGoQuotaURL 根据 base_url 解析 OpenCode Go 额度端点。
diff --git a/backend/internal/service/opencode_go_test.go b/backend/internal/service/opencode_go_test.go
index 2583d30ac1..a8dbce1003 100644
--- a/backend/internal/service/opencode_go_test.go
+++ b/backend/internal/service/opencode_go_test.go
@@ -67,7 +67,7 @@ func TestResolveOpenCodeGoUpstreamProtocolUsesAccountRules(t *testing.T) {
 		Platform: PlatformOpenCodeGo,
 		Credentials: map[string]any{
 			"api_protocol": APIProtocolAdaptive,
-			openCodeGoProtocolRulesKey: []any{
+			protocolRulesCredentialKey: []any{
 				map[string]any{"pattern": "grok-*", "protocol": APIProtocolChatCompletions},
 				map[string]any{"pattern": "deepseek-v4-flash", "protocol": APIProtocolResponses},
 				map[string]any{"pattern": "qwen*", "protocol": APIProtocolAnthropic},
@@ -86,7 +86,7 @@ func TestResolveOpenCodeGoUpstreamProtocolEmptyRulesAreAllChatCompletions(t *tes
 	account := &Account{
 		Platform: PlatformOpenCodeGo,
 		Credentials: map[string]any{
-			openCodeGoProtocolRulesKey: []any{},
+			protocolRulesCredentialKey: []any{},
 		},
 	}
 	require.Equal(t, APIProtocolChatCompletions, account.ResolveOpenCodeGoUpstreamProtocol("grok-4.6"))
@@ -98,7 +98,7 @@ func TestResolveOpenCodeGoUpstreamProtocolFirstMatchWins(t *testing.T) {
 	account := &Account{
 		Platform: PlatformOpenCodeGo,
 		Credentials: map[string]any{
-			openCodeGoProtocolRulesKey: []any{
+			protocolRulesCredentialKey: []any{
 				map[string]any{"pattern": "gpt-5.6-luna", "protocol": APIProtocolChatCompletions},
 				map[string]any{"pattern": "gpt-*", "protocol": APIProtocolResponses},
 			},
@@ -114,7 +114,7 @@ func TestOpenCodeGoNativeProtocolUnmatchedFallsBackToChatCompletions(t *testing.
 		Platform: PlatformOpenCodeGo,
 		Credentials: map[string]any{
 			"api_protocol": APIProtocolAdaptive,
-			openCodeGoProtocolRulesKey: []any{
+			protocolRulesCredentialKey: []any{
 				map[string]any{"pattern": "grok-*", "protocol": APIProtocolResponses},
 				map[string]any{"pattern": "gpt-*", "protocol": APIProtocolResponses},
 				map[string]any{"pattern": "qwen*", "protocol": APIProtocolAnthropic},
@@ -136,12 +136,12 @@ func TestOpenCodeGoNativeProtocolUnmatchedFallsBackToChatCompletions(t *testing.
 
 func TestParseOpenCodeGoProtocolRulesAcceptsTypedMaps(t *testing.T) {
 	t.Parallel()
-	rules, err := parseOpenCodeGoProtocolRules([]map[string]any{
+	rules, err := parseProtocolRules([]map[string]any{
 		{"pattern": "grok-*", "protocol": APIProtocolResponses},
 		{"pattern": "qwen*", "protocol": APIProtocolAnthropic},
 	})
 	require.NoError(t, err)
-	require.Equal(t, []OpenCodeGoProtocolRule{
+	require.Equal(t, []ProtocolRule{
 		{Pattern: "grok-*", Protocol: APIProtocolResponses},
 		{Pattern: "qwen*", Protocol: APIProtocolAnthropic},
 	}, rules)
@@ -178,31 +178,31 @@ func TestStampOpenAIResponsesUpstreamEndpoint(t *testing.T) {
 
 func TestNormalizeOpenCodeGoProtocolRulesCredentials(t *testing.T) {
 	t.Parallel()
-	require.NoError(t, NormalizeOpenCodeGoProtocolRulesCredentials(nil))
-	require.NoError(t, NormalizeOpenCodeGoProtocolRulesCredentials(map[string]any{}))
+	require.NoError(t, NormalizeProtocolRulesCredentials(nil))
+	require.NoError(t, NormalizeProtocolRulesCredentials(map[string]any{}))
 
 	creds := map[string]any{
-		openCodeGoProtocolRulesKey: []any{
+		protocolRulesCredentialKey: []any{
 			map[string]any{"pattern": " Grok-* ", "protocol": APIProtocolResponses},
 		},
 	}
-	require.NoError(t, NormalizeOpenCodeGoProtocolRulesCredentials(creds))
-	rules, ok := creds[openCodeGoProtocolRulesKey].([]any)
+	require.NoError(t, NormalizeProtocolRulesCredentials(creds))
+	rules, ok := creds[protocolRulesCredentialKey].([]any)
 	require.True(t, ok)
 	require.NotEmpty(t, rules)
 	entry, ok := rules[0].(map[string]any)
 	require.True(t, ok)
 	require.Equal(t, "grok-*", entry["pattern"])
 
-	err := NormalizeOpenCodeGoProtocolRulesCredentials(map[string]any{
-		openCodeGoProtocolRulesKey: []any{
+	err := NormalizeProtocolRulesCredentials(map[string]any{
+		protocolRulesCredentialKey: []any{
 			map[string]any{"pattern": "*grok", "protocol": APIProtocolResponses},
 		},
 	})
 	require.Error(t, err)
 
-	err = NormalizeOpenCodeGoProtocolRulesCredentials(map[string]any{
-		openCodeGoProtocolRulesKey: []any{
+	err = NormalizeProtocolRulesCredentials(map[string]any{
+		protocolRulesCredentialKey: []any{
 			map[string]any{"pattern": "grok-*", "protocol": "adaptive"},
 		},
 	})
diff --git a/backend/internal/service/provider_profile.go b/backend/internal/service/provider_profile.go
index 7fee5d96df..111cf968fb 100644
--- a/backend/internal/service/provider_profile.go
+++ b/backend/internal/service/provider_profile.go
@@ -11,9 +11,21 @@ type ProviderEndpoints struct {
 	BaseURLs map[string]string
 	// ProtocolRules 是按模型分流的内置默认表（首条命中生效）；为空表示该模式
 	// 不按模型分流，由账号 api_protocol 决定上游协议。
-	ProtocolRules []OpenCodeGoProtocolRule
+	ProtocolRules []ProtocolRule
 }
 
+// ProviderRouting 决定多协议供应商账号如何选择上游协议。
+type ProviderRouting string
+
+const (
+	// ProviderRoutingByInbound 适用于单一厂商：模型在各原生端点都可用，adaptive 账号
+	// 按入站协议选同协议端点，供应商不提供该协议时回落 Chat Completions。
+	ProviderRoutingByInbound ProviderRouting = "by_inbound"
+	// ProviderRoutingByModel 适用于多模型聚合平台：每个模型只在某一协议端点提供，
+	// 按 protocol_rules 以模型名选协议，与入站协议无关。
+	ProviderRoutingByModel ProviderRouting = "by_model"
+)
+
 // ProviderProfile 描述一个多协议 API Key 供应商（国产厂商或多模型聚合平台）的
 // 静态数据：各接入模式的默认端点、内置分流规则与端点路径约定。
 // 额度解析、会话头等行为差异不在这里，仍由各自的代码路径处理。
@@ -22,6 +34,8 @@ type ProviderProfile struct {
 	// DefaultMode 在 credentials.account_mode 缺失或不在 Modes 中时使用。
 	DefaultMode string
 	Modes       map[string]ProviderEndpoints
+	// Routing 为空时按 ProviderRoutingByInbound 处理。
+	Routing ProviderRouting
 	// ResponsesPath 覆盖 Responses 端点路径（按 buildOpenAIEndpointURL 的版本感知
 	// 规则拼接），空值为 /v1/responses。对自定义 base_url 的账号同样生效。
 	ResponsesPath string
@@ -50,12 +64,13 @@ func (p *ProviderProfile) SupportsProtocol(mode, protocol string) bool {
 
 // providerProfiles 是多协议 API Key 供应商的单一数据来源。新增同类供应商时在此
 // 登记一条即可获得默认端点、原生 Responses 能力判定（有 responses 条目即支持）、
-// Responses 路径约定与 IsMultiProtocolAPIKeyProvider 归属。
+// 上游协议分流方式、Responses 路径约定与 IsMultiProtocolAPIKeyProvider 归属。
 // 默认端点需与前端 credentialsBuilder.ts 的预设保持一致。
 var providerProfiles = map[string]*ProviderProfile{
 	PlatformKimi: {
 		Platform:    PlatformKimi,
 		DefaultMode: AccountModePayG,
+		Routing:     ProviderRoutingByInbound,
 		Modes: map[string]ProviderEndpoints{
 			AccountModePayG: {
 				BaseURLs: map[string]string{
@@ -76,6 +91,7 @@ var providerProfiles = map[string]*ProviderProfile{
 	PlatformZhipu: {
 		Platform:    PlatformZhipu,
 		DefaultMode: AccountModePayG,
+		Routing:     ProviderRoutingByInbound,
 		// 智谱没有原生 Responses 端点，故不登记 responses 条目。
 		Modes: map[string]ProviderEndpoints{
 			AccountModePayG: {
@@ -95,6 +111,7 @@ var providerProfiles = map[string]*ProviderProfile{
 	PlatformDeepseek: {
 		Platform:    PlatformDeepseek,
 		DefaultMode: AccountModePayG,
+		Routing:     ProviderRoutingByInbound,
 		// DeepSeek 官方 Responses 端点为 /responses（无 /v1 前缀，适配 Codex）。
 		ResponsesPath: "/responses",
 		Modes: map[string]ProviderEndpoints{
@@ -110,6 +127,7 @@ var providerProfiles = map[string]*ProviderProfile{
 	PlatformMiniMax: {
 		Platform:    PlatformMiniMax,
 		DefaultMode: AccountModePayG,
+		Routing:     ProviderRoutingByInbound,
 		Modes: map[string]ProviderEndpoints{
 			// 按量付费与 Coding/Token Plan 共用推理域名，靠 API Key 区分套餐。
 			AccountModePayG: {
@@ -131,6 +149,7 @@ var providerProfiles = map[string]*ProviderProfile{
 	PlatformOpenCodeGo: {
 		Platform:    PlatformOpenCodeGo,
 		DefaultMode: AccountModeGo,
+		Routing:     ProviderRoutingByModel,
 		Modes: map[string]ProviderEndpoints{
 			AccountModeGo: {
 				BaseURLs: map[string]string{
@@ -179,3 +198,9 @@ func (a *Account) providerSupportsProtocol(protocol string) bool {
 	profile := a.providerProfile()
 	return profile != nil && profile.SupportsProtocol(a.GetCredential("account_mode"), protocol)
 }
+
+// routesByModel 报告账号所在供应商是否按模型名分流上游协议（多模型聚合平台）。
+func (a *Account) routesByModel() bool {
+	profile := a.providerProfile()
+	return profile != nil && profile.Routing == ProviderRoutingByModel
+}
diff --git a/backend/internal/service/provider_profile_test.go b/backend/internal/service/provider_profile_test.go
index 1dc9f94175..e9fc2994bd 100644
--- a/backend/internal/service/provider_profile_test.go
+++ b/backend/internal/service/provider_profile_test.go
@@ -150,7 +150,7 @@ func TestProviderProfile_BaseURLKeysAreNativeProtocols(t *testing.T) {
 		for mode, endpoints := range profile.Modes {
 			require.NotEmpty(t, endpoints.BaseURLs[APIProtocolChatCompletions], "%s/%s must have a chat_completions base", platform, mode)
 			for protocol, baseURL := range endpoints.BaseURLs {
-				require.True(t, isNativeOpenCodeGoProtocol(protocol), "%s/%s has unknown protocol key %q", platform, mode, protocol)
+				require.True(t, isNativeUpstreamProtocol(protocol), "%s/%s has unknown protocol key %q", platform, mode, protocol)
 				require.NotEmpty(t, baseURL, "%s/%s/%s", platform, mode, protocol)
 			}
 		}
@@ -180,8 +180,9 @@ func TestProviderProfile_ResponsesPathMatchesLegacy(t *testing.T) {
 func TestProviderProfile_OpenCodeProtocolRulesByMode(t *testing.T) {
 	t.Parallel()
 
-	require.Equal(t, DefaultOpenCodeGoProtocolRules(), defaultOpenCodeProtocolRules(AccountModeGo))
-	require.Equal(t, DefaultOpenCodeGoProtocolRules(), defaultOpenCodeProtocolRules(""))
-	require.Equal(t, DefaultOpenCodeZenProtocolRules(), defaultOpenCodeProtocolRules(AccountModeZen))
+	openCode := LookupProviderProfile(PlatformOpenCodeGo)
+	require.Equal(t, DefaultOpenCodeGoProtocolRules(), openCode.Endpoints(AccountModeGo).ProtocolRules)
+	require.Equal(t, DefaultOpenCodeGoProtocolRules(), openCode.Endpoints("").ProtocolRules)
+	require.Equal(t, DefaultOpenCodeZenProtocolRules(), openCode.Endpoints(AccountModeZen).ProtocolRules)
 	require.Empty(t, LookupProviderProfile(PlatformKimi).Endpoints(AccountModeCoding).ProtocolRules)
 }
diff --git a/backend/internal/service/upstream_protocol_routing.go b/backend/internal/service/upstream_protocol_routing.go
new file mode 100644
index 0000000000..c0c6df333c
--- /dev/null
+++ b/backend/internal/service/upstream_protocol_routing.go
@@ -0,0 +1,92 @@
+package service
+
+import (
+	"encoding/json"
+	"fmt"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
+)
+
+// resolveUpstreamProtocol 决定 OpenAI 网关本次请求发往上游的协议
+// （chat_completions / responses / anthropic），是 /v1/responses、
+// /v1/chat/completions、/v1/messages 三个入站入口共用的唯一判定。
+// 只做判断，不改写请求体、不发请求；各入口再按「入站 × 上游」选择转换路径。
+//
+// inbound 为入站协议；/v1/chat/completions 收到 Responses 形状请求体时按
+// responses 计。model 仅按模型分流的账号使用，取 upstreamRoutingModel 的结果。
+//
+// 优先级：
+//  1. 按模型分流的供应商（多模型聚合平台，如 OpenCode）：显式 api_protocol →
+//     账号 protocol_rules → profile 内置规则 → Chat Completions，与入站协议无关。
+//  2. adaptive（按入站协议分流）：供应商提供入站协议的原生端点则走同协议，
+//     否则回落 Chat Completions。
+//  3. 显式 anthropic 协议 → Anthropic。
+//  4. shouldForwardOpenAIResponsesViaRawChatCompletions（显式 chat_completions、
+//     探针确认不支持 Responses 等）→ Chat Completions。
+//  5. 其余 → Responses。
+//
+// Grok 在 Responses / Chat Completions 入口有专属链路，须在调用本函数前分流。
+func resolveUpstreamProtocol(account *Account, inbound, model string) string {
+	if account.routesByModel() {
+		return account.resolveModelRoutedProtocol(model)
+	}
+	if account.IsAdaptiveAPIProtocol() {
+		if account.providerSupportsProtocol(inbound) {
+			return inbound
+		}
+		return APIProtocolChatCompletions
+	}
+	if account.IsAnthropicProtocol() {
+		return APIProtocolAnthropic
+	}
+	if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+		return APIProtocolChatCompletions
+	}
+	return APIProtocolResponses
+}
+
+// resolveModelRoutedProtocol 按模型名解析按模型分流账号的上游协议。
+// 显式 pinned 协议优先；adaptive 先走 credentials.protocol_rules，未配置时回落
+// profile 对应接入模式的内置规则；规则未命中或结果未知一律 Chat Completions，
+// 避免落入 Responses 转换链。
+func (a *Account) resolveModelRoutedProtocol(model string) string {
+	switch proto := a.GetAPIProtocol(); proto {
+	case APIProtocolChatCompletions, APIProtocolAnthropic, APIProtocolResponses:
+		return proto
+	}
+	rules, present := a.configuredProtocolRules()
+	if !present {
+		rules = a.providerProfile().Endpoints(a.GetCredential("account_mode")).ProtocolRules
+	}
+	return matchProtocolRules(model, rules)
+}
+
+// upstreamRoutingModel 返回参与协议分流的模型名：仅按模型分流的账号需要，按账号
+// 模型映射与上游归一后的名字匹配规则；其余账号的分流与模型无关，返回空串。
+func upstreamRoutingModel(account *Account, body []byte, defaultMappedModel string) string {
+	if !account.routesByModel() {
+		return ""
+	}
+	return resolveMappedUpstreamModel(account, body, defaultMappedModel)
+}
+
+// convertResponsesShapedChatBody 把发到 /v1/chat/completions 的 Responses 形状
+// 请求体（Cursor 兼容）转换为 Chat Completions 请求体。
+func (s *OpenAIGatewayService) convertResponsesShapedChatBody(body []byte) ([]byte, error) {
+	var responsesReq apicompat.ResponsesRequest
+	if err := json.Unmarshal(body, &responsesReq); err != nil {
+		return nil, fmt.Errorf("parse responses-shaped chat completions request: %w", err)
+	}
+	chatReq, err := apicompat.ResponsesToChatCompletionsRequestWithOptions(
+		&responsesReq,
+		&apicompat.ResponsesToChatOptions{ReasoningContentByID: s.reasoningContentByID},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("convert responses-shaped chat completions request: %w", err)
+	}
+	chatBody, err := json.Marshal(chatReq)
+	if err != nil {
+		return nil, fmt.Errorf("marshal converted chat completions request: %w", err)
+	}
+	return chatBody, nil
+}
diff --git a/backend/internal/service/upstream_protocol_routing_matrix_test.go b/backend/internal/service/upstream_protocol_routing_matrix_test.go
new file mode 100644
index 0000000000..fc3a5165a1
--- /dev/null
+++ b/backend/internal/service/upstream_protocol_routing_matrix_test.go
@@ -0,0 +1,344 @@
+//go:build unit
+
+package service
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"sort"
+	"strings"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/openai_compat"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+// 上游协议分流矩阵：经真实入口（Forward / ForwardAsChatCompletions /
+// ForwardAsAnthropic）遍历 平台 × api_protocol × 探针 × 账号类型 × 接入模式 ×
+// protocol_rules × 入站 × 模型，从录制到的上游 URL 与请求体推断实际走的协议
+// 与是否做了 Responses 形状转换，并与 legacyUpstreamRouting（统一判定前三个
+// 入口各自的分流顺序）逐条比对。
+//
+// 设置 ROUTING_MATRIX_DUMP=<path> 时把全部观测（含 URL、上游模型、错误）按行
+// 写入文件，便于在两个版本间直接 diff。
+
+type routingMatrixIngress struct {
+	name          string
+	path          string
+	responsesBody bool
+	forward       func(*OpenAIGatewayService, *gin.Context, *Account, []byte) error
+}
+
+func routingMatrixIngresses() []routingMatrixIngress {
+	return []routingMatrixIngress{
+		{
+			name: "responses", path: "/v1/responses", responsesBody: true,
+			forward: func(svc *OpenAIGatewayService, c *gin.Context, account *Account, body []byte) error {
+				_, err := svc.Forward(context.Background(), c, account, body)
+				return err
+			},
+		},
+		{
+			name: "chat", path: "/v1/chat/completions",
+			forward: func(svc *OpenAIGatewayService, c *gin.Context, account *Account, body []byte) error {
+				_, err := svc.ForwardAsChatCompletions(context.Background(), c, account, body, "", "")
+				return err
+			},
+		},
+		{
+			name: "chat_responses_shape", path: "/v1/chat/completions", responsesBody: true,
+			forward: func(svc *OpenAIGatewayService, c *gin.Context, account *Account, body []byte) error {
+				_, err := svc.ForwardAsChatCompletions(context.Background(), c, account, body, "", "")
+				return err
+			},
+		},
+		{
+			name: "messages", path: "/v1/messages",
+			forward: func(svc *OpenAIGatewayService, c *gin.Context, account *Account, body []byte) error {
+				_, err := svc.ForwardAsAnthropic(context.Background(), c, account, body, "", "")
+				return err
+			},
+		},
+	}
+}
+
+type routingMatrixCase struct {
+	platform    string
+	apiProtocol string
+	mode        string
+	accountType string
+	probe       string
+	rules       bool
+	ingress     routingMatrixIngress
+	model       string
+}
+
+func (tc routingMatrixCase) key() string {
+	return fmt.Sprintf("%s|proto=%s|mode=%s|type=%s|probe=%s|rules=%t|%s|%s",
+		tc.platform, tc.apiProtocol, tc.mode, tc.accountType, tc.probe, tc.rules, tc.ingress.name, tc.model)
+}
+
+func (tc routingMatrixCase) account() *Account {
+	credentials := map[string]any{
+		"api_key":  "sk-test",
+		"base_url": "http://base.example",
+		"api_base_urls": map[string]any{
+			APIProtocolChatCompletions: "http://cc.example",
+			APIProtocolResponses:       "http://responses.example",
+			APIProtocolAnthropic:       "http://anthropic.example",
+		},
+	}
+	if tc.apiProtocol != "" {
+		credentials["api_protocol"] = tc.apiProtocol
+	}
+	if tc.mode != "" {
+		credentials["account_mode"] = tc.mode
+	}
+	if tc.rules {
+		credentials["protocol_rules"] = []any{
+			map[string]any{"pattern": "glm-*", "protocol": APIProtocolAnthropic},
+			map[string]any{"pattern": "gpt-*", "protocol": APIProtocolResponses},
+		}
+	}
+	extra := map[string]any{}
+	switch tc.probe {
+	case "yes":
+		extra[openai_compat.ExtraKeyResponsesSupported] = true
+	case "no":
+		extra[openai_compat.ExtraKeyResponsesSupported] = false
+	}
+	return &Account{
+		ID:          801,
+		Name:        "routing-matrix",
+		Platform:    tc.platform,
+		Type:        tc.accountType,
+		Concurrency: 1,
+		Credentials: credentials,
+		Extra:       extra,
+	}
+}
+
+func (tc routingMatrixCase) body() []byte {
+	switch {
+	case tc.ingress.responsesBody:
+		return []byte(fmt.Sprintf(`{"model":%q,"input":"hello","max_output_tokens":32,"stream":false}`, tc.model))
+	case tc.ingress.name == "messages":
+		return []byte(fmt.Sprintf(`{"model":%q,"max_tokens":32,"messages":[{"role":"user","content":"hello"}],"stream":false}`, tc.model))
+	default:
+		return []byte(fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":"hello"}],"stream":false}`, tc.model))
+	}
+}
+
+func routingMatrixCases() []routingMatrixCase {
+	var cases []routingMatrixCase
+	protocols := []string{"", APIProtocolAdaptive, APIProtocolChatCompletions, APIProtocolAnthropic, APIProtocolResponses}
+	types := []string{AccountTypeAPIKey, AccountTypeUpstream, AccountTypeOAuth}
+	probes := []string{"", "yes", "no"}
+	add := func(platform string, modes, models []string, rulesOptions []bool, probeOptions, typeOptions []string, ingresses []routingMatrixIngress) {
+		for _, proto := range protocols {
+			for _, mode := range modes {
+				for _, accountType := range typeOptions {
+					for _, probe := range probeOptions {
+						for _, rules := range rulesOptions {
+							for _, ingress := range ingresses {
+								for _, model := range models {
+									cases = append(cases, routingMatrixCase{
+										platform: platform, apiProtocol: proto, mode: mode, accountType: accountType,
+										probe: probe, rules: rules, ingress: ingress, model: model,
+									})
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+	}
+	all := routingMatrixIngresses()
+	add(PlatformKimi, []string{AccountModePayG}, []string{"kimi-k2"}, []bool{false}, probes, types, all)
+	add(PlatformZhipu, []string{AccountModePayG}, []string{"glm-4.7"}, []bool{false}, probes, types, all)
+	add(PlatformDeepseek, []string{AccountModePayG}, []string{"deepseek-chat"}, []bool{false}, probes, types, all)
+	add(PlatformOpenAI, []string{""}, []string{"gpt-5"}, []bool{false}, probes, types, all)
+	add(PlatformOpenCodeGo, []string{AccountModeGo, AccountModeZen},
+		[]string{"gpt-5.6-luna", "minimax-m3", "glm-5.3", "claude-sonnet-4"},
+		[]bool{false, true}, []string{"", "no"}, []string{AccountTypeAPIKey, AccountTypeUpstream}, all)
+	// Grok 仅 /v1/messages 经统一判定；另两个入口在判定前走 Grok 专属链路。
+	add(PlatformGrok, []string{""}, []string{"grok-4"}, []bool{false}, probes, types, all[3:])
+	return cases
+}
+
+type routingObservation struct {
+	urls      []string
+	lastBody  []byte
+	errText   string
+	panicText string
+}
+
+func (o routingObservation) protocol() string {
+	if len(o.urls) == 0 {
+		return ""
+	}
+	last := o.urls[len(o.urls)-1]
+	switch {
+	case strings.HasSuffix(last, "/chat/completions"):
+		return APIProtocolChatCompletions
+	case strings.HasSuffix(last, "/messages"):
+		return APIProtocolAnthropic
+	case strings.HasSuffix(last, "/responses"):
+		return APIProtocolResponses
+	default:
+		return "other:" + last
+	}
+}
+
+// bodyKind 报告最后一个上游请求体的形状：Chat Completions 上游收到 messages
+// 说明 Responses 形状请求体已被转换，收到 input 说明原样透传。
+func (o routingObservation) bodyKind() string {
+	switch {
+	case gjson.GetBytes(o.lastBody, "messages").Exists():
+		return "messages"
+	case gjson.GetBytes(o.lastBody, "input").Exists():
+		return "input"
+	default:
+		return "other"
+	}
+}
+
+func (o routingObservation) String() string {
+	if o.panicText != "" {
+		return "panic: " + o.panicText
+	}
+	if len(o.urls) == 0 {
+		return "no-upstream: " + o.errText
+	}
+	return fmt.Sprintf("urls=%s body=%s model=%s",
+		strings.Join(o.urls, ","), o.bodyKind(), gjson.GetBytes(o.lastBody, "model").String())
+}
+
+func observeRouting(tc routingMatrixCase) (obs routingObservation) {
+	upstream := &httpUpstreamRecorder{err: errors.New("stop after capture")}
+	svc := &OpenAIGatewayService{cfg: rawChatCompletionsTestConfig(), httpUpstream: upstream}
+	body := tc.body()
+	defer func() {
+		if r := recover(); r != nil {
+			obs.panicText = fmt.Sprint(r)
+		}
+		for _, req := range upstream.requests {
+			obs.urls = append(obs.urls, req.URL.String())
+		}
+		obs.lastBody = upstream.lastBody
+	}()
+	err := tc.ingress.forward(svc, adaptiveProtocolTestContext(tc.ingress.path, body), tc.account(), body)
+	if err != nil {
+		obs.errText = err.Error()
+	}
+	return obs
+}
+
+// legacyUpstreamRouting 复述统一判定前三个入口各自的分流顺序，返回预期的上游协议
+// 以及 Chat Completions 上游是否先把 Responses 形状请求体转换成 messages。
+func legacyUpstreamRouting(account *Account, ingress routingMatrixIngress, model string) (string, bool) {
+	openCode := account.IsOpenCodeGo()
+	var openCodeProto string
+	if openCode {
+		openCodeProto = openCodeGoNativeProtocol(account, model)
+	}
+	switch ingress.name {
+	case "responses":
+		if openCode && openCodeProto != APIProtocolResponses {
+			return openCodeProto, false
+		}
+		if account.IsAnthropicProtocol() {
+			return APIProtocolAnthropic, false
+		}
+		// 有意差异：adaptive 且供应商无原生 Responses 端点时，统一判定对任何账号类型
+		// 都回落 Chat Completions（旧 /v1/responses 入口对非 API Key 账号会误发到
+		// 不存在的 /responses；Chat Completions 入口本就如此处理）。
+		if account.IsAdaptiveAPIProtocol() && !account.SupportsNativeCNResponses() {
+			return APIProtocolChatCompletions, false
+		}
+		if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+			return APIProtocolChatCompletions, false
+		}
+		return APIProtocolResponses, false
+	case "chat", "chat_responses_shape":
+		shape := ingress.responsesBody
+		if openCode && openCodeProto != APIProtocolResponses {
+			return openCodeProto, shape
+		}
+		if account.IsAdaptiveAPIProtocol() && !openCode {
+			if !shape {
+				return APIProtocolChatCompletions, false
+			}
+			if !account.SupportsNativeCNResponses() {
+				return APIProtocolChatCompletions, true
+			}
+		}
+		if account.IsAnthropicProtocol() {
+			return APIProtocolAnthropic, false
+		}
+		if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+			return APIProtocolChatCompletions, false
+		}
+		return APIProtocolResponses, false
+	case "messages":
+		if openCode {
+			if openCodeProto != APIProtocolResponses {
+				return openCodeProto, false
+			}
+		} else if account.IsAnthropicProtocol() || account.IsAdaptiveAPIProtocol() {
+			return APIProtocolAnthropic, false
+		}
+		if shouldForwardOpenAIResponsesViaRawChatCompletions(account) {
+			return APIProtocolChatCompletions, false
+		}
+		return APIProtocolResponses, false
+	}
+	panic("unknown ingress " + ingress.name)
+}
+
+func TestUpstreamProtocolRoutingMatrixMatchesLegacy(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	cases := routingMatrixCases()
+	observations := make([]routingObservation, len(cases))
+	dump := make([]string, 0, len(cases))
+	for i, tc := range cases {
+		observations[i] = observeRouting(tc)
+		dump = append(dump, tc.key()+" => "+observations[i].String())
+	}
+	if path := os.Getenv("ROUTING_MATRIX_DUMP"); path != "" {
+		sort.Strings(dump)
+		require.NoError(t, os.WriteFile(path, []byte(strings.Join(dump, "\n")+"\n"), 0o644))
+	}
+
+	compared := 0
+	for i, tc := range cases {
+		obs := observations[i]
+		if tc.accountType == AccountTypeAPIKey && tc.platform != PlatformGrok {
+			require.NotEmpty(t, obs.urls, "API Key 账号应到达上游：%s\n%s", tc.key(), obs)
+		}
+		if obs.panicText != "" || len(obs.urls) == 0 {
+			// 请求在到达上游前就失败（凭证/类型不适配），无法从 URL 推断协议；
+			// 这类组合由 ROUTING_MATRIX_DUMP 的跨版本 diff 覆盖。
+			continue
+		}
+		compared++
+		account := tc.account()
+		model := resolveMappedUpstreamModel(account, tc.body(), "")
+		wantProto, wantConverted := legacyUpstreamRouting(account, tc.ingress, model)
+		require.Equal(t, wantProto, obs.protocol(), "%s\n%s", tc.key(), obs)
+		if wantProto == APIProtocolChatCompletions && tc.ingress.name == "chat_responses_shape" {
+			wantKind := "input"
+			if wantConverted {
+				wantKind = "messages"
+			}
+			require.Equal(t, wantKind, obs.bodyKind(), "%s\n%s", tc.key(), obs)
+		}
+	}
+	require.Positive(t, compared)
+}
diff --git a/backend/internal/service/upstream_protocol_routing_test.go b/backend/internal/service/upstream_protocol_routing_test.go
new file mode 100644
index 0000000000..df487b1f2a
--- /dev/null
+++ b/backend/internal/service/upstream_protocol_routing_test.go
@@ -0,0 +1,119 @@
+//go:build unit
+
+package service
+
+import (
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/openai_compat"
+	"github.com/stretchr/testify/require"
+)
+
+func routingTestAccount(platform, accountType string, credentials, extra map[string]any) *Account {
+	if credentials == nil {
+		credentials = map[string]any{}
+	}
+	credentials["api_key"] = "sk-test"
+	return &Account{Platform: platform, Type: accountType, Credentials: credentials, Extra: extra}
+}
+
+func TestResolveUpstreamProtocol(t *testing.T) {
+	t.Parallel()
+
+	const (
+		cc        = APIProtocolChatCompletions
+		responses = APIProtocolResponses
+		anthropic = APIProtocolAnthropic
+	)
+	adaptive := func(platform string) *Account {
+		return routingTestAccount(platform, AccountTypeAPIKey, map[string]any{"api_protocol": APIProtocolAdaptive}, nil)
+	}
+	pinned := func(platform, protocol string) *Account {
+		return routingTestAccount(platform, AccountTypeAPIKey, map[string]any{"api_protocol": protocol}, nil)
+	}
+	openCode := func(mode string, credentials map[string]any) *Account {
+		if credentials == nil {
+			credentials = map[string]any{}
+		}
+		credentials["account_mode"] = mode
+		return routingTestAccount(PlatformOpenCodeGo, AccountTypeAPIKey, credentials, nil)
+	}
+	openAIWithProbe := func(accountType string, supported any) *Account {
+		extra := map[string]any{}
+		if supported != nil {
+			extra[openai_compat.ExtraKeyResponsesSupported] = supported
+		}
+		return routingTestAccount(PlatformOpenAI, accountType, nil, extra)
+	}
+
+	cases := []struct {
+		name    string
+		account *Account
+		inbound string
+		model   string
+		want    string
+	}{
+		// 按入站协议分流（国产单一厂商 adaptive）
+		{"kimi adaptive anthropic inbound", adaptive(PlatformKimi), anthropic, "", anthropic},
+		{"kimi adaptive cc inbound", adaptive(PlatformKimi), cc, "", cc},
+		{"kimi adaptive responses inbound", adaptive(PlatformKimi), responses, "", responses},
+		{"zhipu adaptive responses inbound falls back to cc", adaptive(PlatformZhipu), responses, "", cc},
+		{"zhipu adaptive anthropic inbound", adaptive(PlatformZhipu), anthropic, "", anthropic},
+
+		// 显式固定协议，与入站无关
+		{"kimi pinned anthropic", pinned(PlatformKimi, anthropic), cc, "", anthropic},
+		{"kimi pinned cc", pinned(PlatformKimi, cc), responses, "", cc},
+		{"kimi pinned responses", pinned(PlatformKimi, responses), anthropic, "", responses},
+		{"zhipu pinned responses is unsupported and stays cc", pinned(PlatformZhipu, responses), responses, "", cc},
+
+		// 按模型分流（多模型聚合平台），与入站无关
+		{"opencode go gpt", openCode(AccountModeGo, nil), anthropic, "gpt-5.6-luna", responses},
+		{"opencode go minimax", openCode(AccountModeGo, nil), responses, "minimax-m3", anthropic},
+		{"opencode go glm falls back to cc", openCode(AccountModeGo, nil), anthropic, "glm-5.3", cc},
+		{"opencode go claude has no rule", openCode(AccountModeGo, nil), anthropic, "claude-sonnet-4", cc},
+		{"opencode zen claude", openCode(AccountModeZen, nil), cc, "claude-sonnet-4", anthropic},
+		{"opencode account rules override defaults", openCode(AccountModeGo, map[string]any{
+			"protocol_rules": []any{map[string]any{"pattern": "glm-*", "protocol": anthropic}},
+		}), cc, "glm-5.3", anthropic},
+		{"opencode account rules miss falls back to cc", openCode(AccountModeGo, map[string]any{
+			"protocol_rules": []any{map[string]any{"pattern": "glm-*", "protocol": anthropic}},
+		}), cc, "gpt-5.6-luna", cc},
+		{"opencode pinned protocol beats rules", openCode(AccountModeGo, map[string]any{"api_protocol": cc}), anthropic, "gpt-5.6-luna", cc},
+		{"opencode ignores responses probe", func() *Account {
+			account := openCode(AccountModeGo, nil)
+			account.Extra = map[string]any{openai_compat.ExtraKeyResponsesSupported: false}
+			return account
+		}(), cc, "gpt-5.6-luna", responses},
+
+		// 无 profile 的账号按探针决定 Responses / CC
+		{"openai apikey unknown probe", openAIWithProbe(AccountTypeAPIKey, nil), cc, "", responses},
+		{"openai apikey probe unsupported", openAIWithProbe(AccountTypeAPIKey, false), anthropic, "", cc},
+		{"openai oauth ignores probe", openAIWithProbe(AccountTypeOAuth, false), cc, "", responses},
+	}
+	for _, tc := range cases {
+		require.Equal(t, tc.want, resolveUpstreamProtocol(tc.account, tc.inbound, tc.model), tc.name)
+	}
+}
+
+func TestRoutesByModelOnlyForAggregators(t *testing.T) {
+	t.Parallel()
+
+	for platform, profile := range providerProfiles {
+		account := &Account{Platform: platform}
+		require.Equal(t, platform == PlatformOpenCodeGo, account.routesByModel(), platform)
+		require.Equal(t, account.IsOpenCodeGo(), profile.Routing == ProviderRoutingByModel, platform)
+	}
+	require.False(t, (&Account{Platform: PlatformOpenAI}).routesByModel())
+	require.False(t, (*Account)(nil).routesByModel())
+}
+
+func TestUpstreamRoutingModelOnlyForModelRoutedAccounts(t *testing.T) {
+	t.Parallel()
+
+	body := []byte(`{"model":"gpt-5.6-luna","input":"hi"}`)
+	openCode := routingTestAccount(PlatformOpenCodeGo, AccountTypeAPIKey, map[string]any{
+		"model_mapping": map[string]any{"gpt-5.6-luna": "glm-5.3"},
+	}, nil)
+	require.Equal(t, "glm-5.3", upstreamRoutingModel(openCode, body, ""))
+	require.Empty(t, upstreamRoutingModel(routingTestAccount(PlatformKimi, AccountTypeAPIKey, nil, nil), body, ""))
+}