Merge branch 'main' into fix/responses-synthesized-created-at

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge.go

@@ -249,8 +249,8 @@ func anthropicUserToChatMessages(raw json.RawMessage) ([]ChatMessage, error) {
 
 // anthropicAssistantToChatMessages handles an Anthropic assistant message.
 // Text content → assistant message content; tool_use blocks → tool_calls on the
-// same assistant message; thinking blocks are dropped (Chat Completions has no
-// inbound thinking field, matching anthropicAssistantToResponses).
+// same assistant message; thinking blocks → reasoning_content, but only on a
+// message that carries tool calls (see anthropicThinkingToReasoningContent).
 func anthropicAssistantToChatMessages(raw json.RawMessage) ([]ChatMessage, error) {
 	// Plain string → single assistant message.
 	var s string
@@ -289,9 +289,40 @@ func anthropicAssistantToChatMessages(raw json.RawMessage) ([]ChatMessage, error
 		})
 	}
 
+	msg.ReasoningContent = anthropicThinkingToReasoningContent(blocks, len(msg.ToolCalls) > 0)
+
 	return []ChatMessage{msg}, nil
 }
 
+// anthropicThinkingToReasoningContent folds thinking blocks back into the
+// Chat Completions reasoning_content field.
+//
+// chatMessageToAnthropicBlocks emits the upstream's reasoning_content as a
+// thinking block on the way out, so a multi-turn client echoes it back on the
+// next request; dropping it here made the bridge lose exactly what it had just
+// produced. DeepSeek's thinking mode requires the reasoning_content that
+// produced a tool call to be replayed on that assistant message and answers
+// 400 otherwise, which is why buildChatMessagesFromItems already carries
+// pendingReasoning onto assistant tool-call messages in the Responses→Chat
+// bridge. hasToolCalls keeps the scope identical to that sibling: reasoning
+// rides along with tool calls only, never on a plain assistant text turn.
+//
+// redacted_thinking blocks and signature-only placeholders carry no plaintext
+// and contribute nothing. Multiple blocks join with "\n", matching
+// extractResponsesReasoningText.
+func anthropicThinkingToReasoningContent(blocks []AnthropicContentBlock, hasToolCalls bool) string {
+	if !hasToolCalls {
+		return ""
+	}
+	var parts []string
+	for _, b := range blocks {
+		if b.Type == "thinking" && b.Thinking != "" {
+			parts = append(parts, b.Thinking)
+		}
+	}
+	return strings.Join(parts, "\n")
+}
+
 // anthropicToolsToChatTools maps Anthropic tool definitions to Chat Completions
 // function tools. Server-side tools (web_search_*) are dropped — they have no
 // Chat Completions equivalent.

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge_test.go

@@ -143,8 +143,11 @@ func TestAnthropicToChatCompletionsRequest_ThinkingDropped(t *testing.T) {
 	out, err := AnthropicToChatCompletionsRequest(req)
 	require.NoError(t, err)
 	require.Len(t, out.Messages, 1)
-	// Only text survives; thinking is dropped
+	// Only text survives. Thinking is dropped because this turn carries no tool
+	// calls — reasoning rides along with tool calls only, matching the
+	// Responses→Chat bridge (see anthropicThinkingToReasoningContent).
 	require.Equal(t, `"answer"`, string(out.Messages[0].Content))
+	require.Empty(t, out.Messages[0].ReasoningContent)
 }
 
 func TestAnthropicToChatCompletionsRequest_ToolChoiceAuto(t *testing.T) {

backend/internal/pkg/apicompat/chatcompletions_anthropic_reasoning_passback_test.go

@@ -0,0 +1,229 @@
+package apicompat
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// issue #5528：/v1/messages 客户端(Claude Code 等)打到只会 Chat Completions 的
+// OpenAI 兼容上游时，历史 assistant 消息里的 thinking 块被整块丢弃。DeepSeek 的
+// thinking mode 要求产生工具调用的 reasoning_content 随该 assistant 消息回传，
+// 于是「单轮正常、一进多轮工具对话必现 400」。
+
+func anthropicAssistantMsg(t *testing.T, blocks string) *AnthropicRequest {
+	t.Helper()
+	return &AnthropicRequest{
+		Model:     "deepseek-v4-flash",
+		MaxTokens: 256,
+		Messages: []AnthropicMessage{
+			{Role: "user", Content: json.RawMessage(`"what's the weather?"`)},
+			{Role: "assistant", Content: json.RawMessage(blocks)},
+			{Role: "user", Content: json.RawMessage(`[{"type":"tool_result","tool_use_id":"toolu_1","content":"sunny"}]`)},
+		},
+	}
+}
+
+const anthropicThinkingToolTurn = `[
+	{"type":"thinking","thinking":"user wants weather, call the tool"},
+	{"type":"text","text":"checking"},
+	{"type":"tool_use","id":"toolu_1","name":"get_weather","input":{"city":"SF"}}
+]`
+
+func TestAnthropicToChatCompletionsRequest_ThinkingBecomesReasoningContentOnToolTurn(t *testing.T) {
+	out, err := AnthropicToChatCompletionsRequest(anthropicAssistantMsg(t, anthropicThinkingToolTurn))
+	require.NoError(t, err)
+
+	var assistant *ChatMessage
+	for i := range out.Messages {
+		if out.Messages[i].Role == "assistant" {
+			assistant = &out.Messages[i]
+			break
+		}
+	}
+	require.NotNil(t, assistant, "assistant message must survive the bridge")
+	require.Equal(t, "user wants weather, call the tool", assistant.ReasoningContent,
+		"产生工具调用的 thinking 必须作为 reasoning_content 回传，否则 DeepSeek 400")
+	require.Len(t, assistant.ToolCalls, 1)
+	require.Equal(t, `"checking"`, string(assistant.Content), "text/tool_use 处理保持不变")
+}
+
+// 上游线格式才是上游看到的东西：字段没序列化出去，等于没修。
+func TestAnthropicToChatCompletionsRequest_ReasoningContentSerializesOnWire(t *testing.T) {
+	out, err := AnthropicToChatCompletionsRequest(anthropicAssistantMsg(t, anthropicThinkingToolTurn))
+	require.NoError(t, err)
+
+	payload, err := json.Marshal(out)
+	require.NoError(t, err)
+	require.Contains(t, string(payload), `"reasoning_content":"user wants weather, call the tool"`)
+}
+
+// 闭环不变式：thinking 块本来就是本桥出站时用上游 reasoning_content 生成的
+// (chatMessageToAnthropicBlocks)，客户端只是原样回传。出站造、入站丢 = 自己丢自己的东西。
+func TestAnthropicChatBridge_ReasoningSurvivesOutboundInboundRoundTrip(t *testing.T) {
+	upstream := ChatMessage{
+		Role:             "assistant",
+		ReasoningContent: "step 1: need the weather tool",
+		Content:          json.RawMessage(`"checking"`),
+		ToolCalls: []ChatToolCall{{
+			ID:       "call_1",
+			Type:     "function",
+			Function: ChatFunctionCall{Name: "get_weather", Arguments: `{"city":"SF"}`},
+		}},
+	}
+
+	// 出站：Chat 响应 → Anthropic content blocks
+	blocks := chatMessageToAnthropicBlocks(upstream)
+	require.Equal(t, "thinking", blocks[0].Type)
+	require.Equal(t, upstream.ReasoningContent, blocks[0].Thinking)
+
+	// 客户端下一轮把同一组 blocks 原样回传
+	raw, err := json.Marshal(blocks)
+	require.NoError(t, err)
+
+	// 入站：Anthropic content blocks → Chat 请求
+	back, err := anthropicAssistantToChatMessages(raw)
+	require.NoError(t, err)
+	require.Len(t, back, 1)
+	require.Equal(t, upstream.ReasoningContent, back[0].ReasoningContent,
+		"出站生成的 thinking 必须能原样还原回 reasoning_content")
+	require.Len(t, back[0].ToolCalls, 1)
+}
+
+// 兄弟不变式：Responses→Chat 桥(buildChatMessagesFromItems 的 pendingReasoning)
+// 早就把 reasoning 挂到带 tool_calls 的 assistant 消息上了。等价历史下两条桥必须一致。
+func TestAnthropicChatBridge_MatchesResponsesChatBridgeReasoningPlacement(t *testing.T) {
+	responsesReq := &ResponsesRequest{
+		Model: "deepseek-v4-flash",
+		Input: json.RawMessage(`[
+			{"type":"message","role":"user","content":[{"type":"input_text","text":"what's the weather?"}]},
+			{"type":"reasoning","summary":[{"type":"summary_text","text":"call the tool"}]},
+			{"type":"function_call","call_id":"call_1","name":"get_weather","arguments":"{\"city\":\"SF\"}"},
+			{"type":"function_call_output","call_id":"call_1","output":"sunny"}
+		]`),
+	}
+	viaResponses, err := ResponsesToChatCompletionsRequest(responsesReq)
+	require.NoError(t, err)
+
+	viaAnthropic, err := AnthropicToChatCompletionsRequest(anthropicAssistantMsg(t, `[
+		{"type":"thinking","thinking":"call the tool"},
+		{"type":"tool_use","id":"toolu_1","name":"get_weather","input":{"city":"SF"}}
+	]`))
+	require.NoError(t, err)
+
+	reasoningOnToolCallMessage := func(msgs []ChatMessage) string {
+		for _, m := range msgs {
+			if m.Role == "assistant" && len(m.ToolCalls) > 0 {
+				return m.ReasoningContent
+			}
+		}
+		return ""
+	}
+	require.Equal(t, "call the tool", reasoningOnToolCallMessage(viaResponses.Messages),
+		"前置条件：兄弟桥本来就带 reasoning_content")
+	require.Equal(t, reasoningOnToolCallMessage(viaResponses.Messages),
+		reasoningOnToolCallMessage(viaAnthropic.Messages),
+		"两条桥对等价历史必须产出同样的 reasoning_content 位置")
+}
+
+// 作用域守卫：不带工具调用的纯文本轮次维持现状(与兄弟桥一致 —— reasoning 只随
+// 工具调用回传)，避免把 reasoning_content 撒到不需要它的上游请求上。
+func TestAnthropicToChatCompletionsRequest_ThinkingWithoutToolCallsStaysDropped(t *testing.T) {
+	req := &AnthropicRequest{
+		Model:     "deepseek-v4-flash",
+		MaxTokens: 100,
+		Messages: []AnthropicMessage{
+			{Role: "assistant", Content: json.RawMessage(
+				`[{"type":"thinking","thinking":"secret thoughts"},{"type":"text","text":"answer"}]`)},
+		},
+	}
+
+	out, err := AnthropicToChatCompletionsRequest(req)
+	require.NoError(t, err)
+	require.Len(t, out.Messages, 1)
+	require.Empty(t, out.Messages[0].ReasoningContent)
+	require.Equal(t, `"answer"`, string(out.Messages[0].Content))
+
+	payload, err := json.Marshal(out)
+	require.NoError(t, err)
+	require.NotContains(t, string(payload), "reasoning_content")
+}
+
+func TestAnthropicThinkingToReasoningContent(t *testing.T) {
+	blocksOf := func(t *testing.T, raw string) []AnthropicContentBlock {
+		t.Helper()
+		var blocks []AnthropicContentBlock
+		require.NoError(t, json.Unmarshal([]byte(raw), &blocks))
+		return blocks
+	}
+
+	cases := []struct {
+		name         string
+		raw          string
+		hasToolCalls bool
+		want         string
+	}{
+		{
+			name:         "single_thinking_block",
+			raw:          `[{"type":"thinking","thinking":"a"}]`,
+			hasToolCalls: true,
+			want:         "a",
+		},
+		{
+			// 多个 thinking 块用 "\n" 连接，与 extractResponsesReasoningText 一致。
+			name:         "multiple_blocks_join_with_newline",
+			raw:          `[{"type":"thinking","thinking":"a"},{"type":"text","text":"x"},{"type":"thinking","thinking":"b"}]`,
+			hasToolCalls: true,
+			want:         "a\nb",
+		},
+		{
+			// redacted_thinking 没有明文可回传。
+			name:         "redacted_thinking_has_no_plaintext",
+			raw:          `[{"type":"redacted_thinking","signature":"abc"}]`,
+			hasToolCalls: true,
+			want:         "",
+		},
+		{
+			// 只带 signature 的 thinking 占位块(xAI/Codex 密文回放形态)同样无明文。
+			name:         "signature_only_thinking",
+			raw:          `[{"type":"thinking","thinking":"","signature":"gAAAAxxx"}]`,
+			hasToolCalls: true,
+			want:         "",
+		},
+		{
+			name:         "no_tool_calls_returns_empty",
+			raw:          `[{"type":"thinking","thinking":"a"}]`,
+			hasToolCalls: false,
+			want:         "",
+		},
+		{
+			name:         "no_thinking_blocks",
+			raw:          `[{"type":"text","text":"x"}]`,
+			hasToolCalls: true,
+			want:         "",
+		},
+		{
+			name:         "empty_blocks",
+			raw:          `[]`,
+			hasToolCalls: true,
+			want:         "",
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			require.Equal(t, tc.want,
+				anthropicThinkingToReasoningContent(blocksOf(t, tc.raw), tc.hasToolCalls))
+		})
+	}
+}
+
+// 纯字符串形态的 assistant content 没有 blocks 可读，走早返回分支，不得 panic。
+func TestAnthropicAssistantToChatMessages_PlainStringContentUnaffected(t *testing.T) {
+	msgs, err := anthropicAssistantToChatMessages(json.RawMessage(`"just text"`))
+	require.NoError(t, err)
+	require.Len(t, msgs, 1)
+	require.Empty(t, msgs[0].ReasoningContent)
+	require.Equal(t, `"just text"`, string(msgs[0].Content))
+}

backend/internal/pkg/apicompat/responses_client_tools.go

@@ -358,6 +358,9 @@ func normalizeClientToolOutput(item map[string]any) {
 	if _, ok := output.(string); ok {
 		return
 	}
+	if isResponsesToolOutputContent(output) {
+		return
+	}
 	if output == nil {
 		item["output"] = ""
 		return
@@ -370,6 +373,25 @@ func normalizeClientToolOutput(item map[string]any) {
 	item["output"] = string(encoded)
 }
 
+func isResponsesToolOutputContent(output any) bool {
+	parts, ok := output.([]any)
+	if !ok || len(parts) == 0 {
+		return false
+	}
+	for _, part := range parts {
+		typed, ok := part.(map[string]any)
+		if !ok {
+			return false
+		}
+		switch stringValue(typed["type"]) {
+		case "input_text", "input_image", "input_file":
+		default:
+			return false
+		}
+	}
+	return true
+}
+
 // normalizeToolSearchOutput converts both tool_search output wire shapes into
 // the string output required by function_call_output. Older clients send an
 // output field directly; newer Codex clients return discovered definitions in

backend/internal/pkg/apicompat/responses_client_tools_test.go

@@ -460,7 +460,55 @@ func TestAdaptResponsesClientToolsWithInheritedMapping_LowersFollowupHistoryWith
 	output := requireResponsesClientToolValue[map[string]any](t, items[1])
 	require.Equal(t, "function_call_output", output["type"])
 	require.NotContains(t, output, "id")
-	require.JSONEq(t, `[{"text":"ok","type":"input_text"}]`, requireResponsesClientToolValue[string](t, output["output"]))
+	require.Equal(t, []any{map[string]any{"type": "input_text", "text": "ok"}}, output["output"])
+}
+
+func TestAdaptResponsesClientTools_NormalizesCustomToolOutput(t *testing.T) {
+	tests := []struct {
+		name       string
+		output     any
+		wantOutput any
+	}{
+		{
+			name: "supported content parts remain an array",
+			output: []any{
+				map[string]any{"type": "input_text", "text": "ok"},
+				map[string]any{"type": "input_image", "image_url": "https://example.com/image.png"},
+				map[string]any{"type": "input_file", "file_id": "file_123"},
+			},
+			wantOutput: []any{
+				map[string]any{"type": "input_text", "text": "ok"},
+				map[string]any{"type": "input_image", "image_url": "https://example.com/image.png"},
+				map[string]any{"type": "input_file", "file_id": "file_123"},
+			},
+		},
+		{name: "ordinary object is stringified", output: map[string]any{"ok": true}, wantOutput: `{"ok":true}`},
+		{name: "arbitrary array is stringified", output: []any{"ok"}, wantOutput: `["ok"]`},
+		{name: "empty array is stringified", output: []any{}, wantOutput: `[]`},
+		{name: "mixed array is stringified", output: []any{map[string]any{"type": "input_text", "text": "ok"}, "bad"}, wantOutput: `[{"text":"ok","type":"input_text"},"bad"]`},
+		{name: "unknown content type is stringified", output: []any{map[string]any{"type": "output_text", "text": "bad"}}, wantOutput: `[{"text":"bad","type":"output_text"}]`},
+		{name: "whitespace-padded content type is stringified", output: []any{map[string]any{"type": " input_text ", "text": "bad"}}, wantOutput: `[{"text":"bad","type":" input_text "}]`},
+		{name: "missing content type is stringified", output: []any{map[string]any{"text": "bad"}}, wantOutput: `[{"text":"bad"}]`},
+		{name: "non-string content type is stringified", output: []any{map[string]any{"type": 1}}, wantOutput: `[{"type":1}]`},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			req := map[string]any{
+				"tools": []any{map[string]any{"type": "custom", "name": "exec"}},
+				"input": []any{map[string]any{
+					"type": "custom_tool_call_output", "call_id": "call_1", "output": tc.output,
+				}},
+			}
+
+			_, changed, err := AdaptResponsesClientTools(req)
+			require.NoError(t, err)
+			require.True(t, changed)
+			item := requireResponsesClientToolValue[map[string]any](t, requireResponsesClientToolValue[[]any](t, req["input"])[0])
+			require.Equal(t, "function_call_output", item["type"])
+			require.Equal(t, tc.wantOutput, item["output"])
+		})
+	}
 }
 
 func TestAdaptResponsesClientToolsWithInheritedMapping_PromotesOmittedToolsDiscoveryIntoEffectiveDeclarations(t *testing.T) {

backend/internal/service/openai_account_runtime_block_fastpath.go

@@ -133,6 +133,17 @@ func (s *OpenAIGatewayService) handleOpenAIAccountUpstreamError(ctx context.Cont
 		return false
 	}
 
+	// Self-built images requests always carry a matching image_generation tool, so a
+	// "tool choice not found in 'tools'" 400 means upstream revoked this account's
+	// image capability. Gated on the self-built marker: passthrough clients control
+	// their own tools/tool_choice and could otherwise poison a healthy account.
+	if isOpenAIImagesSelfBuiltRequest(ctx) && isOpenAIImageCapabilityLossError(statusCode, responseBody) {
+		if s != nil && s.rateLimitService != nil {
+			_ = s.rateLimitService.HandleOpenAIImageCapabilityLoss(stateCtx, account, statusCode, responseBody)
+		}
+		return false
+	}
+
 	if s == nil || account == nil {
 		return false
 	}

backend/internal/service/openai_gateway_forward.go

@@ -121,7 +121,7 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 	}
 	if shouldStripOpenAIResponsesInputNamespaces(account, wsDecision.Transport, passthroughEnabled) {
 		keepToolCallNamespaces := shouldKeepOpenAIResponsesToolCallNamespaces(
-			account, wsDecision.Transport, passthroughEnabled, compactPath,
+			account, wsDecision.Transport, passthroughEnabled, compactPath, body,
 		)
 		body, err = stripOpenAIResponsesInputNamespaces(body, keepToolCallNamespaces)
 		if err != nil {
@@ -182,6 +182,17 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 		return s.forwardResponsesViaRawChatCompletions(ctx, c, account, body)
 	}
 	if account.IsOpenAI() && (account.IsOpenAIApiKey() || account.IsOpenAIOAuthLike()) {
+		normalizedReasoningBody, reasoningChanged, reasoningErr := normalizeOpenAIResponsesReasoningContentReplay(body)
+		if reasoningErr != nil {
+			return nil, fmt.Errorf("normalize OpenAI Responses reasoning content replay: %w", reasoningErr)
+		}
+		if reasoningChanged {
+			body = normalizedReasoningBody
+			originalBody = normalizedReasoningBody
+			requestView = newOpenAIRequestView(normalizedReasoningBody)
+			reqModel, reqStream, promptCacheKey = requestView.Model, requestView.Stream, requestView.PromptCacheKey
+			originalModel = reqModel
+		}
 		sanitizedBody, changed, sanitizeErr := sanitizeOpenAIResponsesInputItemIDs(body)
 		if sanitizeErr != nil {
 			return nil, fmt.Errorf("sanitize OpenAI Responses input item IDs: %w", sanitizeErr)
@@ -472,13 +483,29 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 		if decodeErr != nil {
 			return nil, decodeErr
 		}
+		// Responses OAuth 与 Chat 兼容入口保持一致：纯文本 system 可以无损提升后删除，
+		// JSON object 模式仍需在 input 中保留 JSON 指令供上游兼容校验。
+		omitPromotedSystemMessages := !strings.EqualFold(
+			strings.TrimSpace(gjson.GetBytes(body, "text.format.type").String()),
+			"json_object",
+		)
 		codexResult := codexTransformResult{}
 		if compatMessagesBridge {
-			codexResult = applyCodexOAuthTransformWithOptions(decoded, codexOAuthTransformOptions{IsCodexCLI: isCodexCLI, IsCompact: isCompactRequest, SkipDefaultInstructions: true, PreserveToolCallIDs: true})
+			codexResult = applyCodexOAuthTransformWithOptions(decoded, codexOAuthTransformOptions{
+				IsCodexCLI:                          isCodexCLI,
+				IsCompact:                           isCompactRequest,
+				SkipDefaultInstructions:             true,
+				PreserveToolCallIDs:                 true,
+				OmitPromotedSystemMessagesFromInput: omitPromotedSystemMessages,
+			})
 			ensureCodexOAuthInstructionsField(decoded)
 			markDecodedModified()
 		} else {
-			codexResult = applyCodexOAuthTransform(decoded, isCodexCLI, isCompactRequest)
+			codexResult = applyCodexOAuthTransformWithOptions(decoded, codexOAuthTransformOptions{
+				IsCodexCLI:                          isCodexCLI,
+				IsCompact:                           isCompactRequest,
+				OmitPromotedSystemMessagesFromInput: omitPromotedSystemMessages,
+			})
 		}
 		if codexResult.Error != nil {
 			c.JSON(http.StatusBadRequest, gin.H{"error": gin.H{"type": "invalid_request_error", "message": codexResult.Error.Error()}})

backend/internal/service/openai_gateway_request_body.go

@@ -458,6 +458,67 @@ func openAIRequestBodyHasTools(body []byte) bool {
 	return false
 }
 
+// normalizeOpenAIResponsesReasoningContentReplay removes non-portable
+// reasoning.content arrays before history is sent to a real OpenAI Responses
+// endpoint. Compatible providers may return visible reasoning blocks there,
+// while OpenAI accepts only an empty array when the item is replayed.
+//
+// Keep the reasoning item and its portable fields (summary, encrypted_content,
+// ids, and opaque extensions). Callers scope this normalization to OpenAI
+// destinations; compatible providers may still consume their own content.
+func normalizeOpenAIResponsesReasoningContentReplay(body []byte) ([]byte, bool, error) {
+	input := gjson.GetBytes(body, "input")
+	if !input.IsArray() {
+		return body, false, nil
+	}
+
+	needsNormalization := false
+	input.ForEach(func(_, item gjson.Result) bool {
+		if strings.TrimSpace(item.Get("type").String()) != "reasoning" {
+			return true
+		}
+		content := item.Get("content")
+		if content.IsArray() && len(content.Array()) > 0 {
+			needsNormalization = true
+			return false
+		}
+		return true
+	})
+	if !needsNormalization {
+		return body, false, nil
+	}
+
+	var reqBody map[string]any
+	if err := decodeOpenAIJSONUseNumber(body, &reqBody); err != nil {
+		return body, false, fmt.Errorf("normalize OpenAI reasoning content replay: %w", err)
+	}
+	items, ok := reqBody["input"].([]any)
+	if !ok {
+		return body, false, nil
+	}
+	changed := false
+	for _, rawItem := range items {
+		item, ok := rawItem.(map[string]any)
+		if !ok || strings.TrimSpace(firstNonEmptyString(item["type"])) != "reasoning" {
+			continue
+		}
+		content, ok := item["content"].([]any)
+		if !ok || len(content) == 0 {
+			continue
+		}
+		delete(item, "content")
+		changed = true
+	}
+	if !changed {
+		return body, false, nil
+	}
+	normalized, err := marshalOpenAIUpstreamJSON(reqBody)
+	if err != nil {
+		return body, false, fmt.Errorf("serialize normalized OpenAI reasoning content replay: %w", err)
+	}
+	return normalized, true, nil
+}
+
 func normalizeOpenAIAPIKeyStoreFalseReasoningReplay(body []byte, knownStoreFalse bool) ([]byte, bool, error) {
 	if !knownStoreFalse && gjson.GetBytes(body, "store").Type != gjson.False {
 		return body, false, nil
@@ -1018,6 +1079,12 @@ func normalizeOpenAIResponsesWebSocketCompatibilityBody(body []byte, account *Ac
 			return body, false, err
 		}
 	}
+	if next, normalizedReasoningContent, err := normalizeOpenAIResponsesReasoningContentReplay(normalized); err != nil {
+		return body, false, err
+	} else if normalizedReasoningContent {
+		normalized = next
+		changed = true
+	}
 	if account.IsOpenAIApiKey() {
 		if next, normalizedParallel, err := normalizeOpenAIParallelToolCallsWithoutTools(normalized, responsesLite); err != nil {
 			return body, false, err

backend/internal/service/openai_gateway_request_body_reasoning_test.go

@@ -353,3 +353,60 @@ func TestNormalizeOpenAIParallelToolCallsWithoutTools_KeepsResponsesLiteAddition
 	require.False(t, changed)
 	require.Equal(t, gjson.False, gjson.GetBytes(normalized, "parallel_tool_calls").Type)
 }
+
+func TestNormalizeOpenAIResponsesReasoningContentReplayStripsCrossProviderArray(t *testing.T) {
+	body := []byte(`{"model":"gpt-5.6-sol","input":[` +
+		`{"type":"message","role":"user","content":"one"},` +
+		`{"type":"message","role":"assistant","content":"two"},` +
+		`{"type":"function_call","call_id":"call_1","name":"lookup","arguments":"{}"},` +
+		`{"type":"function_call_output","call_id":"call_1","output":"ok"},` +
+		`{"type":"message","role":"user","content":"five"},` +
+		`{"type":"reasoning","id":"rs_provider","summary":[{"type":"summary_text","text":"portable"}],"content":[{"type":"reasoning_text","text":"visible reasoning"}],"opaque":9007199254740993},` +
+		`{"type":"message","role":"assistant","content":[{"type":"output_text","text":"answer"}]}` +
+		`]}`)
+
+	normalized, changed, err := normalizeOpenAIResponsesReasoningContentReplay(body)
+
+	require.NoError(t, err)
+	require.True(t, changed)
+	require.Equal(t, "reasoning", gjson.GetBytes(normalized, "input.5.type").String())
+	require.False(t, gjson.GetBytes(normalized, "input.5.content").Exists())
+	require.Equal(t, "portable", gjson.GetBytes(normalized, "input.5.summary.0.text").String())
+	require.Equal(t, "9007199254740993", gjson.GetBytes(normalized, "input.5.opaque").Raw)
+	require.Equal(t, "answer", gjson.GetBytes(normalized, "input.6.content.0.text").String())
+}
+
+func TestNormalizeOpenAIResponsesReasoningContentReplayKeepsPortableShapes(t *testing.T) {
+	for _, body := range []string{
+		`{"input":[{"type":"reasoning","summary":[]}]}`,
+		`{"input":[{"type":"reasoning","content":[],"summary":[]}]}`,
+		`{"input":[{"type":"message","content":[{"type":"input_text","text":"keep"}]}]}`,
+	} {
+		normalized, changed, err := normalizeOpenAIResponsesReasoningContentReplay([]byte(body))
+		require.NoError(t, err)
+		require.False(t, changed)
+		require.JSONEq(t, body, string(normalized))
+	}
+}
+
+func TestNormalizeOpenAIResponsesWebSocketCompatibilityBodyStripsReasoningContentOnlyForOpenAI(t *testing.T) {
+	body := []byte(`{"type":"response.create","model":"gpt-5.6-sol","store":true,"input":[{"type":"reasoning","summary":[{"type":"summary_text","text":"keep"}],"content":[{"type":"reasoning_text","text":"remove"}]}]}`)
+	for _, accountType := range []string{AccountTypeAPIKey, AccountTypeOAuth} {
+		normalized, changed, err := normalizeOpenAIResponsesWebSocketCompatibilityBody(body, &Account{
+			Platform: PlatformOpenAI,
+			Type:     accountType,
+		}, false)
+		require.NoError(t, err)
+		require.True(t, changed)
+		require.False(t, gjson.GetBytes(normalized, "input.0.content").Exists())
+		require.Equal(t, "keep", gjson.GetBytes(normalized, "input.0.summary.0.text").String())
+	}
+
+	normalized, changed, err := normalizeOpenAIResponsesWebSocketCompatibilityBody(body, &Account{
+		Platform: PlatformZhipu,
+		Type:     AccountTypeAPIKey,
+	}, false)
+	require.NoError(t, err)
+	require.False(t, changed)
+	require.JSONEq(t, string(body), string(normalized))
+}

backend/internal/service/openai_gateway_responses_client_tools_test.go

@@ -250,6 +250,34 @@ func TestOpenAIPassthroughAPIKeyRestoresClientToolsNonStreaming(t *testing.T) {
 	require.Equal(t, "*** Begin Patch", gjson.Get(recorder.Body.String(), "output.1.input").String())
 }
 
+func TestOpenAIPassthroughAPIKeyPreservesCustomToolOutputContentParts(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	body := []byte(`{"model":"gpt-5.4","stream":false,"tools":[{"type":"custom","name":"exec"}],"input":[{"type":"custom_tool_call_output","call_id":"call_1","output":[{"type":"input_text","text":"result"},{"type":"input_file","file_id":"file_123"}]}]}`)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", bytes.NewReader(body))
+
+	upstream := &httpUpstreamRecorder{resp: &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+		Body:       io.NopCloser(strings.NewReader(`{"id":"resp_tools","status":"completed","output":[],"usage":{}}`)),
+	}}
+	svc := openAIClientToolsTestService(upstream)
+	account := &Account{ID: 6240, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Credentials: map[string]any{"api_key": "test-key"}}
+
+	result, err := svc.forwardOpenAIPassthrough(context.Background(), c, account, body, body, "gpt-5.4", false, nil, false, time.Now())
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.Equal(t, "function_call_output", gjson.GetBytes(upstream.lastBody, "input.0.type").String())
+	output := gjson.GetBytes(upstream.lastBody, "input.0.output")
+	require.True(t, output.IsArray(), "native Responses content parts must reach the upstream as an array")
+	require.Equal(t, "input_text", output.Get("0.type").String())
+	require.Equal(t, "result", output.Get("0.text").String())
+	require.Equal(t, "input_file", output.Get("1.type").String())
+	require.Equal(t, "file_123", output.Get("1.file_id").String())
+}
+
 func TestOpenAIPassthroughAPIKeyRestoresClientToolsStreaming(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	body := openAIClientToolsRequest(true)

backend/internal/service/openai_images_responses.go

@@ -37,6 +37,13 @@ type OpenAIImagesUpstreamError struct {
 	Message           string
 	Param             string
 	UpstreamRequestID string
+
+	// SynthesizedFromModelText marks an error the gateway inferred from the
+	// model's plain-text output instead of reading it off a structured upstream
+	// error frame. Such a verdict describes this one turn ("the model answered
+	// with words instead of an image"), not the account — see
+	// shouldCoolOpenAIImagesToolForError.
+	SynthesizedFromModelText bool
 }
 
 func (e *OpenAIImagesUpstreamError) Error() string {
@@ -328,6 +335,26 @@ func openAIImageUploadToDataURL(upload OpenAIImagesUpload) (string, error) {
 	return "data:" + contentType + ";base64," + base64.StdEncoding.EncodeToString(upload.Data), nil
 }
 
+// openAIImagesSelfBuiltRequestContextKey marks a request whose upstream body was
+// fully constructed by buildOpenAIImagesResponsesRequest, i.e. tool_choice and the
+// matching image_generation tool are always both present and never client-controlled.
+type openAIImagesSelfBuiltRequestContextKey struct{}
+
+func withOpenAIImagesSelfBuiltRequest(ctx context.Context) context.Context {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, openAIImagesSelfBuiltRequestContextKey{}, true)
+}
+
+func isOpenAIImagesSelfBuiltRequest(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	selfBuilt, _ := ctx.Value(openAIImagesSelfBuiltRequestContextKey{}).(bool)
+	return selfBuilt
+}
+
 func buildOpenAIImagesResponsesRequest(parsed *OpenAIImagesRequest, toolModel string) ([]byte, error) {
 	if parsed == nil {
 		return nil, fmt.Errorf("parsed images request is required")
@@ -711,6 +738,10 @@ func openAIImagesTextFallbackErrorForText(text string) *OpenAIImagesUpstreamErro
 		ErrorType:  "upstream_error",
 		Code:       "image_generation_unavailable",
 		Message:    "Upstream did not execute image generation",
+		// Inferred from the model's own words, not from an upstream error frame:
+		// good enough to fail this turn over to another account, not evidence that
+		// this account's image tool is down for the next 30 minutes.
+		SynthesizedFromModelText: true,
 	}
 }
 
@@ -1775,6 +1806,7 @@ func (s *OpenAIGatewayService) forwardOpenAIImagesOAuth(
 	if err != nil {
 		return nil, err
 	}
+	upstreamCtx = withOpenAIImagesSelfBuiltRequest(upstreamCtx)
 	upstreamReq, err := s.buildUpstreamRequest(upstreamCtx, c, account, responsesBody, token, true, parsed.StickySessionSeed(), false)
 	if err != nil {
 		return nil, err
@@ -1922,6 +1954,26 @@ const (
 	openAIImagesOAuthUnavailableReason   = "openai_images_oauth_tool_unavailable"
 )
 
+// shouldCoolOpenAIImagesToolForError decides whether an image_generation_unavailable
+// verdict is durable enough to park the account's image tool for
+// openAIImagesOAuthUnavailableCooldown.
+//
+// Only an upstream error frame that names the condition qualifies. A verdict the
+// gateway synthesized from the model's plain-text reply does not: it merely says
+// this prompt produced words instead of an image, which is prompt-dependent and
+// happens on healthy accounts. Writing a 30-minute account-level cooldown from it
+// is doubly wrong because the very same error is classified retryable
+// (IsOpenAIImagesRetryableUpstreamError: status >= 500) and drives
+// newOpenAIAccountFailoverError — so one such reply walks the pool and cools every
+// account the retry touches.
+//
+// This mirrors the rule the alpha/search path already states in words: a
+// tool-endpoint failure "仍允许本次请求换号，但不修改任何账号状态"
+// (see shouldApplyOpenAIAlphaSearchAccountErrorSideEffects).
+func shouldCoolOpenAIImagesToolForError(upstreamErr *OpenAIImagesUpstreamError) bool {
+	return upstreamErr != nil && !upstreamErr.SynthesizedFromModelText
+}
+
 func (s *OpenAIGatewayService) coolOpenAIImagesOAuthTool(ctx context.Context, account *Account) {
 	if s == nil || s.accountRepo == nil || account == nil || account.Platform != PlatformOpenAI {
 		return
@@ -2017,7 +2069,9 @@ func (s *OpenAIGatewayService) handleOpenAIImagesOAuthResponseError(
 
 	responseBody := openAIImagesUpstreamErrorResponseBody(upstreamErr)
 	if upstreamErr.Code == "image_generation_unavailable" {
-		s.coolOpenAIImagesOAuthTool(ctx, account)
+		if shouldCoolOpenAIImagesToolForError(upstreamErr) {
+			s.coolOpenAIImagesOAuthTool(ctx, account)
+		}
 		if responseWritten {
 			return err
 		}

backend/internal/service/openai_images_tool_cooldown_test.go

@@ -0,0 +1,177 @@
+//go:build unit
+
+package service
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+// issue #6171：v0.1.181 起，/v1/images/generations 只要上游"回文字没回图"，账号就被
+// 写 30 分钟 openai:image_generation 模型级冷却。该判据是**请求级**的（这个 prompt
+// 这一轮模型选择了说话），却被当成**账号级**能力失效；又因为同一个错误被判为
+// 可重试（502）并驱动 failover，一次闲聊回复会沿着号池逐个把账号冷却掉。
+
+// countingModelRateLimitRepo 记录 SetModelRateLimit 调用，用于断言"没写账号状态"。
+type countingModelRateLimitRepo struct {
+	accountRepoStub
+	calls  int
+	scopes []string
+}
+
+func (r *countingModelRateLimitRepo) SetModelRateLimit(_ context.Context, _ int64, scope string, _ time.Time, _ ...string) error {
+	r.calls++
+	r.scopes = append(r.scopes, scope)
+	return nil
+}
+
+func newImagesCooldownContext(t *testing.T) (*gin.Context, *httptest.ResponseRecorder) {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/images/generations", nil)
+	return c, rec
+}
+
+func imagesCooldownAccount() *Account {
+	return &Account{ID: 77, Platform: PlatformOpenAI, Type: AccountTypeOAuth, Name: "img-oauth"}
+}
+
+func TestShouldCoolOpenAIImagesToolForError(t *testing.T) {
+	cases := []struct {
+		name string
+		err  *OpenAIImagesUpstreamError
+		want bool
+	}{
+		{
+			name: "nil_error",
+			err:  nil,
+			want: false,
+		},
+		{
+			// 网关从模型文字里推断出来的判据：只说明这一轮没出图。
+			name: "synthesized_from_model_text",
+			err: &OpenAIImagesUpstreamError{
+				StatusCode:               http.StatusBadGateway,
+				Code:                     "image_generation_unavailable",
+				SynthesizedFromModelText: true,
+			},
+			want: false,
+		},
+		{
+			// 上游自己在 error 帧里点名该状态：这才是账号级证据，保持冷却。
+			name: "structured_upstream_error_frame",
+			err: &OpenAIImagesUpstreamError{
+				StatusCode: http.StatusBadGateway,
+				Code:       "image_generation_unavailable",
+			},
+			want: true,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			require.Equal(t, tc.want, shouldCoolOpenAIImagesToolForError(tc.err))
+		})
+	}
+}
+
+// 主复现：文字兜底判据不得写账号级冷却。
+func TestHandleOpenAIImagesOAuthResponseError_TextFallbackDoesNotCoolAccount(t *testing.T) {
+	c, _ := newImagesCooldownContext(t)
+	repo := &countingModelRateLimitRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := imagesCooldownAccount()
+
+	upstreamErr := openAIImagesTextFallbackErrorForText("Here's a polished image prompt for your request.")
+	require.NotNil(t, upstreamErr)
+	require.Equal(t, "image_generation_unavailable", upstreamErr.Code)
+
+	err := svc.handleOpenAIImagesOAuthResponseError(
+		context.Background(), c, account, "gpt-image-2", "https://upstream.example/v1/responses",
+		&http.Response{StatusCode: http.StatusOK, Header: http.Header{}},
+		OpenAIImagesJSONKeepaliveAdjustedWrittenSize(c), upstreamErr,
+	)
+
+	require.Zero(t, repo.calls, "模型闲聊不构成账号级证据，不得写 30 分钟冷却")
+
+	// 换号行为必须原样保留：本 PR 只撤销账号状态写入，不动 failover。
+	var failover *UpstreamFailoverError
+	require.True(t, errors.As(err, &failover), "仍应触发换号，got %T", err)
+}
+
+// 对照不变式：上游 error 帧点名该状态时仍然冷却，否则等于把功能整个废掉。
+func TestHandleOpenAIImagesOAuthResponseError_StructuredUnavailableStillCoolsAccount(t *testing.T) {
+	c, _ := newImagesCooldownContext(t)
+	repo := &countingModelRateLimitRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := imagesCooldownAccount()
+
+	upstreamErr := &OpenAIImagesUpstreamError{
+		StatusCode: http.StatusBadGateway,
+		ErrorType:  "upstream_error",
+		Code:       "image_generation_unavailable",
+		Message:    "image generation tool is not available for this account",
+	}
+
+	_ = svc.handleOpenAIImagesOAuthResponseError(
+		context.Background(), c, account, "gpt-image-2", "https://upstream.example/v1/responses",
+		&http.Response{StatusCode: http.StatusOK, Header: http.Header{}},
+		OpenAIImagesJSONKeepaliveAdjustedWrittenSize(c), upstreamErr,
+	)
+
+	require.Equal(t, 1, repo.calls, "结构化上游证据仍须写冷却")
+	require.Equal(t, []string{openAIImageGenerationRateLimitKey}, repo.scopes)
+}
+
+// 标记必须打在文字兜底的两个入口上，且不影响违规拦截分支的判定。
+func TestOpenAIImagesTextFallback_MarksSynthesizedVerdicts(t *testing.T) {
+	t.Run("plain_text_reply_is_synthesized", func(t *testing.T) {
+		err := openAIImagesTextFallbackErrorForText("Here's a polished image prompt for your request.")
+		require.NotNil(t, err)
+		require.True(t, err.SynthesizedFromModelText)
+		require.Equal(t, "image_generation_unavailable", err.Code)
+		require.Equal(t, http.StatusBadGateway, err.StatusCode)
+	})
+
+	t.Run("body_entrypoint_is_synthesized", func(t *testing.T) {
+		body := []byte("event: response.completed\n" +
+			`data: {"type":"response.completed","response":{"id":"r","status":"completed",` +
+			`"output":[{"type":"message","content":[{"type":"output_text","text":"I drafted a prompt for you."}]}]}}` +
+			"\n\n")
+		err := openAIImagesTextFallbackError(body)
+		require.NotNil(t, err)
+		require.True(t, err.SynthesizedFromModelText)
+	})
+
+	t.Run("content_policy_branch_unchanged", func(t *testing.T) {
+		err := openAIImagesTextFallbackErrorForText("Blocked by our content policy.")
+		require.NotNil(t, err)
+		require.Equal(t, "content_policy_violation", err.Code)
+		require.Equal(t, http.StatusBadRequest, err.StatusCode)
+		// 该分支本来就不走冷却（Code 不匹配），标记与否都不改变行为；
+		// 断言它没有被顺手打标，避免语义漂移。
+		require.False(t, err.SynthesizedFromModelText)
+	})
+
+	t.Run("empty_text_yields_no_error", func(t *testing.T) {
+		require.Nil(t, openAIImagesTextFallbackErrorForText("   "))
+	})
+}
+
+// 级联的前提条件：该错误确实是可重试的，所以会带着"已写冷却"的副作用换号。
+// 这条用例把前提钉死，避免以后有人把 502 改成非重试后误以为本修复多余。
+func TestOpenAIImagesTextFallback_RemainsRetryableAndThusCascades(t *testing.T) {
+	err := openAIImagesTextFallbackErrorForText("Here's a polished image prompt for your request.")
+	require.NotNil(t, err)
+	require.True(t, IsOpenAIImagesRetryableUpstreamError(err),
+		"文字兜底判据是可重试的——正因如此，写账号冷却会沿号池级联")
+}

backend/internal/service/openai_oauth_passthrough_test.go

@@ -132,6 +132,44 @@ func TestOpenAIGatewayService_ResponsesUnknownModelDoesNotFallbackToGPT54(t *tes
 	require.True(t, rec.Code >= http.StatusBadRequest)
 }
 
+func TestOpenAIGatewayService_OAuthResponsesPromotesSystemMessageWithoutDuplication(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	const systemPrompt = "Unique system prefix for Responses token accounting."
+	const existingInstructions = "Existing instructions."
+	body := []byte(`{"model":"gpt-5.4","stream":false,"instructions":"` + existingInstructions + `","input":[{"role":"system","content":"` + systemPrompt + `"},{"role":"user","content":"hello"}]}`)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	upstream := &httpUpstreamRecorder{err: errors.New("stop after capture")}
+	svc := &OpenAIGatewayService{cfg: &config.Config{}, httpUpstream: upstream}
+	account := &Account{
+		ID:          124,
+		Name:        "openai-oauth",
+		Platform:    PlatformOpenAI,
+		Type:        AccountTypeOAuth,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"access_token":       "oauth-token",
+			"chatgpt_account_id": "chatgpt-acc",
+		},
+		Status:      StatusActive,
+		Schedulable: true,
+	}
+
+	result, err := svc.Forward(context.Background(), c, account, body)
+
+	require.Error(t, err)
+	require.Nil(t, result)
+	require.NotEmpty(t, upstream.lastBody)
+	require.Equal(t, systemPrompt+"\n\n"+existingInstructions, gjson.GetBytes(upstream.lastBody, "instructions").String())
+	require.Equal(t, int64(1), gjson.GetBytes(upstream.lastBody, "input.#").Int())
+	require.Equal(t, "user", gjson.GetBytes(upstream.lastBody, "input.0.role").String())
+	require.Equal(t, 1, strings.Count(string(upstream.lastBody), systemPrompt))
+}
+
 func TestOpenAIGatewayService_NativeResponsesBodyModificationPreservesHTMLChars(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 

backend/internal/service/openai_responses_namespace.go

@@ -77,25 +77,48 @@ func shouldStripOpenAIResponsesInputNamespaces(account *Account, transport OpenA
 //     故 OAuth 非 compact 请求必须保留。
 //   - compact 端点的 schema 不含该字段，携带即 400 `Unknown parameter:
 //     input[N].namespace`（issue #4761 正文），故 compact 一律清理。
-//   - API Key 出口是标准 Responses API（api.openai.com 或自定义 base_url），同样
-//     不认识该字段，维持全量清理；否则只能退化成
-//     openai_responses_rejected_field_retry 的逐项删除，6 次上限根本盖不住长历史。
+//   - API Key 出口默认按标准 Responses API 处理并清理该字段；但当请求本身声明
+//     namespace 工具时，上游显然使用了 namespace 扩展，此时必须保留调用项上的
+//     namespace，否则声明与历史调用会失配并触发 Missing namespace。
 //   - 摊平模式下调用项已被改写成平名，残留 namespace 指向的声明已不存在，一律清理。
 func shouldKeepOpenAIResponsesToolCallNamespaces(
 	account *Account,
 	transport OpenAIUpstreamTransport,
 	passthroughEnabled bool,
 	compactPath bool,
+	body []byte,
 ) bool {
-	if account == nil || !account.IsOpenAIOAuthLike() {
+	if account == nil {
 		return false
 	}
 	if compactPath {
 		return false
 	}
+	if account.IsOpenAIApiKey() {
+		return hasOpenAIResponsesNamespaceToolDeclaration(body)
+	}
+	if !account.IsOpenAIOAuthLike() {
+		return false
+	}
 	return !shouldFlattenOpenAIResponsesNamespaces(account, transport, passthroughEnabled, compactPath)
 }
 
+func hasOpenAIResponsesNamespaceToolDeclaration(body []byte) bool {
+	tools := gjson.GetBytes(body, "tools")
+	if !tools.IsArray() {
+		return false
+	}
+	found := false
+	tools.ForEach(func(_, tool gjson.Result) bool {
+		if strings.EqualFold(strings.TrimSpace(tool.Get("type").String()), "namespace") {
+			found = true
+			return false
+		}
+		return true
+	})
+	return found
+}
+
 // openAIResponsesToolCallItemTypes 是携带 namespace 的调用项类型集合。与
 // removeOpenAIResponsesRejectedNamespaceAtIndex 的反应式白名单保持一致；codex-rs
 // protocol/src/models.rs 中只有 FunctionCall 与 CustomToolCall 序列化 namespace，

backend/internal/service/openai_responses_namespace_forward_test.go

@@ -66,6 +66,29 @@ func TestOpenAIGatewayService_OAuthPreservesCodexNamespaceTools(t *testing.T) {
 	require.Empty(t, openAIResponsesNamespaceNames(c))
 }
 
+// API Key 自定义上游若接受 namespace 工具声明，也要求历史 function_call 原样携带
+// namespace。声明仍为命名空间工具却清掉调用项字段，会触发 Missing namespace。
+func TestOpenAIGatewayService_APIKeyPreservesDeclaredNamespaceToolCalls(t *testing.T) {
+	body := []byte(codexNamespaceRequestBody)
+	upstream := &httpUpstreamRecorder{responses: []*http.Response{
+		newOpenAIRejectedFieldTestResponse(http.StatusOK, namespaceForwardOKResponse),
+	}}
+	c := newOpenAIRejectedFieldTestContext(body)
+
+	result, err := newOpenAIRejectedFieldTestService(upstream).Forward(
+		context.Background(), c, newOpenAIRejectedFieldTestAccount(), body,
+	)
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.Len(t, upstream.bodies, 1)
+	forwarded := upstream.bodies[0]
+
+	require.True(t, gjson.GetBytes(forwarded, `tools.#(type=="namespace")`).Exists())
+	require.Equal(t, "collaboration", gjson.GetBytes(forwarded, "input.0.namespace").String())
+	require.False(t, gjson.GetBytes(forwarded, "input.1.namespace").Exists())
+}
+
 // compact 端点 schema 更窄：input[].namespace 会 400 Unknown parameter（issue #4761），
 // 且没有证据表明它接受 namespace 工具声明。compact 只做历史摘要、不需要模型寻址工具，
 // 因此保持既有的摊平 + 全量清理行为，不随默认值翻转扩大风险面。

backend/internal/service/openai_responses_namespace_test.go

@@ -78,6 +78,7 @@ func TestShouldKeepOpenAIResponsesToolCallNamespaces(t *testing.T) {
 		transport          OpenAIUpstreamTransport
 		passthroughEnabled bool
 		compactPath        bool
+		body               []byte
 		want               bool
 	}{
 		// 上游按 namespace 解析历史调用，缺字段会 400 "Missing namespace for function_call"。
@@ -92,15 +93,20 @@ func TestShouldKeepOpenAIResponsesToolCallNamespaces(t *testing.T) {
 		// WSv2 + compact 是唯一「不摊平但仍必须清理」的组合，钉住 compact 判定本身，
 		// 使其不会被误当成可由 shouldFlatten 推导出的冗余分支。
 		{name: "oauth_compact_wsv2_strips", account: oauth, transport: OpenAIUpstreamTransportResponsesWebsocketV2, compactPath: true, want: false},
-		// API Key 出口是标准 Responses API，不认识该字段。
-		{name: "apikey_strips", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, want: false},
+		// API Key 默认按标准 Responses API 清理；请求显式声明 namespace 工具时，
+		// 自定义上游需要原样接收对应的历史调用。
+		{name: "apikey_without_namespace_tool_strips", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, want: false},
+		{name: "apikey_with_namespace_tool_keeps", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, body: []byte(`{"tools":[{"type":"namespace","name":"mcp__codex_app","tools":[]}]}`), want: true},
+		{name: "apikey_with_mixed_case_namespace_tool_keeps", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, body: []byte(`{"tools":[{"type":" Namespace ","name":"mcp__codex_app","tools":[]}]}`), want: true},
+		{name: "apikey_function_tool_with_namespace_field_strips", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, body: []byte(`{"tools":[{"type":"function","name":"automation_update","namespace":"mcp__codex_app"}]}`), want: false},
+		{name: "apikey_compact_with_namespace_tool_strips", account: apiKey, transport: OpenAIUpstreamTransportHTTPSSE, compactPath: true, body: []byte(`{"tools":[{"type":"namespace","name":"mcp__codex_app","tools":[]}]}`), want: false},
 		{name: "setup_token_keeps", account: setupToken, transport: OpenAIUpstreamTransportHTTPSSE, want: true},
 		{name: "nil_account", account: nil, transport: OpenAIUpstreamTransportHTTPSSE, want: false},
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			require.Equal(t, tt.want, shouldKeepOpenAIResponsesToolCallNamespaces(
-				tt.account, tt.transport, tt.passthroughEnabled, tt.compactPath,
+				tt.account, tt.transport, tt.passthroughEnabled, tt.compactPath, tt.body,
 			))
 		})
 	}

backend/internal/service/openai_responses_rejected_field_retry_test.go

@@ -538,6 +538,34 @@ func TestOpenAIGatewayService_APIKeyStripsAllIndexedNamespacesBeforeFirstForward
 	require.False(t, gjson.GetBytes(upstream.bodies[0], "input.1.namespace").Exists())
 }
 
+func TestOpenAIGatewayServiceProactivelyStripsCrossProviderReasoningContent(t *testing.T) {
+	body := []byte(`{"model":"gpt-5.5","stream":false,"store":true,"input":[` +
+		`{"type":"message","role":"user","content":"one"},` +
+		`{"type":"message","role":"assistant","content":"two"},` +
+		`{"type":"function_call","call_id":"call_1","name":"lookup","arguments":"{}"},` +
+		`{"type":"function_call_output","call_id":"call_1","output":"ok"},` +
+		`{"type":"message","role":"user","content":"five"},` +
+		`{"type":"reasoning","summary":[{"type":"summary_text","text":"keep"}],"content":[{"type":"reasoning_text","text":"remove"}]}` +
+		`]}`)
+	upstream := &httpUpstreamRecorder{responses: []*http.Response{
+		newOpenAIRejectedFieldTestResponse(http.StatusOK, `{"output":[],"usage":{"input_tokens":1,"output_tokens":1,"input_tokens_details":{"cached_tokens":0}}}`),
+	}}
+
+	result, err := newOpenAIRejectedFieldTestService(upstream).Forward(
+		context.Background(),
+		newOpenAIRejectedFieldTestContext(body),
+		newOpenAIRejectedFieldTestAccount(),
+		body,
+	)
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.Len(t, upstream.bodies, 1, "reasoning content should be normalized before the first upstream request")
+	require.Equal(t, "reasoning", gjson.GetBytes(upstream.bodies[0], "input.5.type").String())
+	require.False(t, gjson.GetBytes(upstream.bodies[0], "input.5.content").Exists())
+	require.Equal(t, "keep", gjson.GetBytes(upstream.bodies[0], "input.5.summary.0.text").String())
+}
+
 func TestOpenAIGatewayService_OpenAIHTTPStripsInputNamespacesBeforeFirstForward(t *testing.T) {
 	accounts := []struct {
 		name    string

backend/internal/service/ratelimit_service.go

@@ -75,6 +75,8 @@ const (
 const (
 	openAIImageRateLimitDefaultCooldown = time.Minute
 	openAIImageRateLimitReason          = "openai_image_rate_limited"
+	openAIImageCapabilityLossCooldown   = 30 * time.Minute
+	openAIImageCapabilityLossReason     = "openai_image_capability_lost"
 )
 
 var openAIImageTryAgainPattern = regexp.MustCompile(`(?i)try again in\s+([0-9]+(?:\.[0-9]+)?)\s*(ms|s|sec|secs|second|seconds|m|min|mins|minute|minutes)`)
@@ -2190,6 +2192,44 @@ func (s *RateLimitService) HandleOpenAIImageRateLimit(ctx context.Context, accou
 	return true
 }
 
+func (s *RateLimitService) HandleOpenAIImageCapabilityLoss(ctx context.Context, account *Account, statusCode int, responseBody []byte) bool {
+	if s == nil || account == nil || s.accountRepo == nil {
+		return false
+	}
+	if account.Platform != PlatformOpenAI {
+		return false
+	}
+	if !account.ShouldHandleErrorCode(statusCode) {
+		slog.Info("openai_image_capability_loss_skipped_by_error_code_policy", "account_id", account.ID, "status_code", statusCode)
+		return false
+	}
+	if !isOpenAIImageCapabilityLossError(statusCode, responseBody) {
+		return false
+	}
+
+	resetAt := time.Now().Add(openAIImageCapabilityLossCooldown)
+	if err := s.accountRepo.SetModelRateLimit(ctx, account.ID, openAIImageGenerationRateLimitKey, resetAt, openAIImageCapabilityLossReason); err != nil {
+		slog.Warn("openai_image_capability_loss_set_model_rate_limit_failed", "account_id", account.ID, "scope", openAIImageGenerationRateLimitKey, "error", err)
+		return true
+	}
+	slog.Info("openai_image_capability_lost", "account_id", account.ID, "scope", openAIImageGenerationRateLimitKey, "reset_at", resetAt, "reset_in", time.Until(resetAt).Truncate(time.Second))
+	return true
+}
+
+// isOpenAIImageCapabilityLossError reports whether upstream rejected the
+// image_generation tool choice that sub2api itself put into the request body.
+// Only meaningful for self-built images requests, where tools always carries a
+// matching image_generation entry — upstream saying otherwise means the account
+// lost the capability.
+func isOpenAIImageCapabilityLossError(statusCode int, body []byte) bool {
+	if statusCode != http.StatusBadRequest || len(body) == 0 {
+		return false
+	}
+	lower := strings.ToLower(string(body))
+	return strings.Contains(lower, "image_generation") &&
+		strings.Contains(lower, "not found in 'tools' parameter")
+}
+
 func isOpenAIImageRateLimitError(statusCode int, body []byte) bool {
 	if statusCode != http.StatusTooManyRequests || len(body) == 0 {
 		return false

backend/internal/service/ratelimit_service_openai_image_test.go

@@ -120,7 +120,11 @@ func TestOpenAIGatewayServiceForwardImages_ImageRateLimitReturnsFailoverAndCools
 	require.Equal(t, openAIImageGenerationRateLimitKey, repo.modelRateLimitCalls[0].scope)
 }
 
-func TestOpenAIGatewayServiceForwardImages_TextFallbackCoolsImageCapability(t *testing.T) {
+// issue #6171：上游"回文字没回图"是**这一轮**的结果（模型选择了说话），不是账号能力
+// 失效。它同时被判为可重试（502）并驱动 failover，若还写 30 分钟账号级冷却，一次闲聊
+// 回复就会沿号池把每个被重试到的账号依次冷却掉。冷却仍保留给结构化上游证据，见
+// TestOpenAIGatewayServiceForwardImages_StructuredUnavailableCoolsImageCapability。
+func TestOpenAIGatewayServiceForwardImages_TextFallbackDoesNotCoolImageCapability(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	repo := &modelNotFoundAccountRepoStub{}
 	body := []byte(`{"model":"gpt-image-2","prompt":"draw a cat"}`)
@@ -154,7 +158,6 @@ func TestOpenAIGatewayServiceForwardImages_TextFallbackCoolsImageCapability(t *t
 		},
 	}
 
-	before := time.Now()
 	result, err := svc.ForwardImages(context.Background(), c, account, body, parsed, "")
 
 	require.Nil(t, result)
@@ -162,6 +165,56 @@ func TestOpenAIGatewayServiceForwardImages_TextFallbackCoolsImageCapability(t *t
 	var failoverErr *UpstreamFailoverError
 	require.ErrorAs(t, err, &failoverErr)
 	require.False(t, failoverErr.RetryableOnSameAccount)
+	// 换号行为不变：该判据仍足以放弃本账号重试这一次请求……
+	require.Equal(t, http.StatusBadGateway, failoverErr.StatusCode)
+	// ……但不再写任何账号级状态，否则重试会把冷却一路刷到整个号池。
+	require.Empty(t, repo.modelRateLimitCalls,
+		"模型回文字只说明这一轮没出图，不构成账号 30 分钟不可用的证据")
+}
+
+// 对照不变式：上游 error 帧点名 image_generation_unavailable 时仍写冷却，
+// 保证 #6171 的修复没有把这项能力保护整个废掉。
+func TestOpenAIGatewayServiceForwardImages_StructuredUnavailableCoolsImageCapability(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	repo := &modelNotFoundAccountRepoStub{}
+	body := []byte(`{"model":"gpt-image-2","prompt":"draw a cat"}`)
+	upstreamSSE := "data: {\"type\":\"response.failed\",\"response\":{\"id\":\"r\",\"error\":" +
+		"{\"type\":\"upstream_error\",\"code\":\"image_generation_unavailable\"," +
+		"\"message\":\"image generation tool is not available for this account\"}}}\n\n"
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/images/generations", bytes.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = req
+
+	svc := &OpenAIGatewayService{
+		accountRepo: repo,
+		httpUpstream: &httpUpstreamRecorder{
+			resp: &http.Response{
+				StatusCode: http.StatusOK,
+				Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+				Body:       io.NopCloser(strings.NewReader(upstreamSSE)),
+			},
+		},
+	}
+	parsed, err := svc.ParseOpenAIImagesRequest(c, body)
+	require.NoError(t, err)
+	account := &Account{
+		ID:       206,
+		Name:     "openai-oauth",
+		Platform: PlatformOpenAI,
+		Type:     AccountTypeOAuth,
+		Credentials: map[string]any{
+			"access_token": "token-123",
+		},
+	}
+
+	before := time.Now()
+	result, err := svc.ForwardImages(context.Background(), c, account, body, parsed, "")
+
+	require.Nil(t, result)
+	require.Error(t, err)
 	require.Len(t, repo.modelRateLimitCalls, 1)
 	call := repo.modelRateLimitCalls[0]
 	require.Equal(t, account.ID, call.accountID)
@@ -169,3 +222,120 @@ func TestOpenAIGatewayServiceForwardImages_TextFallbackCoolsImageCapability(t *t
 	require.Equal(t, openAIImagesOAuthUnavailableReason, call.reason)
 	require.WithinDuration(t, before.Add(openAIImagesOAuthUnavailableCooldown), call.resetAt, time.Second)
 }
+
+func TestOpenAIGatewayServiceForwardImages_CapabilityLossCoolsImageScope(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	repo := &modelNotFoundAccountRepoStub{}
+	body := []byte(`{"model":"gpt-image-2","prompt":"draw a cat"}`)
+	errorBody := `{"error":{"message":"Tool choice 'image_generation' not found in 'tools' parameter.","param":"tool_choice","type":"invalid_request_error"}}`
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/images/generations", bytes.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = req
+
+	svc := &OpenAIGatewayService{
+		rateLimitService: &RateLimitService{accountRepo: repo},
+		httpUpstream: &httpUpstreamRecorder{
+			resp: &http.Response{
+				StatusCode: http.StatusBadRequest,
+				Header:     http.Header{"X-Request-Id": []string{"req_img_capability_lost"}},
+				Body:       io.NopCloser(strings.NewReader(errorBody)),
+			},
+		},
+	}
+	parsed, err := svc.ParseOpenAIImagesRequest(c, body)
+	require.NoError(t, err)
+	account := &Account{
+		ID:       205,
+		Name:     "openai-oauth",
+		Platform: PlatformOpenAI,
+		Type:     AccountTypeOAuth,
+		Credentials: map[string]any{
+			"access_token": "token-123",
+		},
+	}
+
+	before := time.Now()
+	result, err := svc.ForwardImages(context.Background(), c, account, body, parsed, "")
+
+	require.Nil(t, result)
+	require.Error(t, err)
+	require.Len(t, repo.modelRateLimitCalls, 1)
+	call := repo.modelRateLimitCalls[0]
+	require.Equal(t, account.ID, call.accountID)
+	require.Equal(t, openAIImageGenerationRateLimitKey, call.scope)
+	require.Equal(t, openAIImageCapabilityLossReason, call.reason)
+	require.WithinDuration(t, before.Add(openAIImageCapabilityLossCooldown), call.resetAt, time.Second)
+}
+
+func TestOpenAIGatewayServiceHandleUpstreamError_PassthroughCapabilityLossDoesNotCool(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &OpenAIGatewayService{rateLimitService: &RateLimitService{accountRepo: repo}}
+	account := &Account{ID: 206, Platform: PlatformOpenAI, Type: AccountTypeOAuth}
+	body := []byte(`{"error":{"message":"Tool choice 'image_generation' not found in 'tools' parameter.","param":"tool_choice","type":"invalid_request_error"}}`)
+
+	disabled := svc.handleOpenAIAccountUpstreamError(context.Background(), account, http.StatusBadRequest, http.Header{}, body, "gpt-5.5")
+
+	require.False(t, disabled)
+	require.Empty(t, repo.modelRateLimitCalls)
+	_, wholeAccountBlocked := svc.openaiAccountRuntimeBlockUntil.Load(account.ID)
+	require.False(t, wholeAccountBlocked)
+}
+
+func TestRateLimitServiceHandleOpenAIImageCapabilityLoss_IgnoresGenericBadRequest(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := &Account{ID: 207, Platform: PlatformOpenAI, Type: AccountTypeOAuth}
+	body := []byte(`{"error":{"message":"Invalid type for input[0].arguments"}}`)
+
+	handled := svc.HandleOpenAIImageCapabilityLoss(context.Background(), account, http.StatusBadRequest, body)
+
+	require.False(t, handled)
+	require.Empty(t, repo.modelRateLimitCalls)
+}
+
+func TestRateLimitServiceHandleOpenAIImageCapabilityLoss_RespectsPlatformAndErrorCodePolicy(t *testing.T) {
+	body := []byte(`{"error":{"message":"Tool choice 'image_generation' not found in 'tools' parameter.","param":"tool_choice","type":"invalid_request_error"}}`)
+
+	t.Run("non_openai_platform", func(t *testing.T) {
+		repo := &modelNotFoundAccountRepoStub{}
+		svc := &RateLimitService{accountRepo: repo}
+		account := &Account{ID: 208, Platform: PlatformAnthropic, Type: AccountTypeOAuth}
+
+		handled := svc.HandleOpenAIImageCapabilityLoss(context.Background(), account, http.StatusBadRequest, body)
+
+		require.False(t, handled)
+		require.Empty(t, repo.modelRateLimitCalls)
+	})
+
+	t.Run("custom_error_code_policy_excludes_400", func(t *testing.T) {
+		repo := &modelNotFoundAccountRepoStub{}
+		svc := &RateLimitService{accountRepo: repo}
+		account := &Account{
+			ID:       209,
+			Platform: PlatformOpenAI,
+			Type:     AccountTypeAPIKey,
+			Credentials: map[string]any{
+				"custom_error_codes_enabled": true,
+				"custom_error_codes":         []any{float64(http.StatusTooManyRequests)},
+			},
+		}
+
+		require.False(t, account.ShouldHandleErrorCode(http.StatusBadRequest))
+		handled := svc.HandleOpenAIImageCapabilityLoss(context.Background(), account, http.StatusBadRequest, body)
+
+		require.False(t, handled)
+		require.Empty(t, repo.modelRateLimitCalls)
+	})
+}
+
+func TestIsOpenAIImageCapabilityLossError(t *testing.T) {
+	capabilityLossBody := []byte(`{"error":{"message":"Tool choice 'image_generation' not found in 'tools' parameter.","param":"tool_choice","type":"invalid_request_error"}}`)
+	genericBadRequestBody := []byte(`{"error":{"message":"Invalid type for input[0].arguments"}}`)
+
+	require.True(t, isOpenAIImageCapabilityLossError(http.StatusBadRequest, capabilityLossBody))
+	require.False(t, isOpenAIImageCapabilityLossError(http.StatusBadRequest, genericBadRequestBody))
+	require.False(t, isOpenAIImageCapabilityLossError(http.StatusTooManyRequests, capabilityLossBody))
+}