fix(apicompat): honor disabled thinking in OpenAI bridges

backend/internal/pkg/apicompat/anthropic_responses_test.go

@@ -1136,9 +1136,24 @@ func TestAnthropicToResponses_ThinkingDisabled(t *testing.T) {
 
 	resp, err := AnthropicToResponses(req)
 	require.NoError(t, err)
-	// Default effort applies (medium) even when thinking is disabled.
 	require.NotNil(t, resp.Reasoning)
-	assert.Equal(t, "medium", resp.Reasoning.Effort)
+	assert.Equal(t, "none", resp.Reasoning.Effort)
+	assert.Empty(t, resp.Reasoning.Summary)
+}
+
+func TestAnthropicToResponses_ThinkingDisabledOverridesOutputEffort(t *testing.T) {
+	req := &AnthropicRequest{
+		Model:        "gpt-5.6-sol",
+		MaxTokens:    1024,
+		Messages:     []AnthropicMessage{{Role: "user", Content: json.RawMessage(`"Hello"`)}},
+		Thinking:     &AnthropicThinking{Type: "disabled"},
+		OutputConfig: &AnthropicOutputConfig{Effort: "max"},
+	}
+
+	resp, err := AnthropicToResponses(req)
+	require.NoError(t, err)
+	require.NotNil(t, resp.Reasoning)
+	assert.Equal(t, "none", resp.Reasoning.Effort)
 }
 
 func TestAnthropicToResponses_NoThinking(t *testing.T) {

backend/internal/pkg/apicompat/anthropic_to_responses.go

@@ -57,17 +57,13 @@ func AnthropicToResponses(req *AnthropicRequest) (*ResponsesRequest, error) {
 		out.Tools = convertAnthropicToolsToResponses(req.Tools)
 	}
 
-	// Determine reasoning effort: only output_config.effort controls the
-	// level; thinking.type is ignored. Default follows Codex CLI / airgate's
-	// Anthropic bridge shape, which uses medium when unset.
-	// Anthropic levels map 1:1 to OpenAI: low→low, medium→medium, high→high, max→xhigh.
-	effort := "medium"
-	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
-		effort = req.OutputConfig.Effort
-	}
+	// An explicit thinking disable takes precedence over output_config.effort.
+	effort := anthropicReasoningEffort(req)
 	out.Reasoning = &ResponsesReasoning{
-		Effort:  mapAnthropicEffortToResponses(effort),
-		Summary: "auto",
+		Effort: effort,
+	}
+	if effort != "none" {
+		out.Reasoning.Summary = "auto"
 	}
 
 	// Convert tool_choice
@@ -423,17 +419,22 @@ func extractAnthropicTextFromBlocks(blocks []AnthropicContentBlock) string {
 	return strings.Join(parts, "\n\n")
 }
 
-// mapAnthropicEffortToResponses converts Anthropic reasoning effort levels to
-// OpenAI Responses API effort levels.
-//
-// Both APIs default to "high". The mapping is 1:1 for shared levels;
-// only Anthropic's "max" (Opus 4.6 exclusive) maps to OpenAI's "xhigh"
-// (GPT-5.2+ exclusive) as both represent the highest reasoning tier.
-//
-//	low    → low
-//	medium → medium
-//	high   → high
-//	max    → xhigh
+// anthropicReasoningEffort resolves the Anthropic request preference for both
+// OpenAI bridges. Explicitly disabled thinking overrides output_config.effort;
+// otherwise the bridge keeps its medium default.
+func anthropicReasoningEffort(req *AnthropicRequest) string {
+	if req.Thinking != nil && req.Thinking.Type == "disabled" {
+		return "none"
+	}
+	effort := "medium"
+	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
+		effort = req.OutputConfig.Effort
+	}
+	return mapAnthropicEffortToResponses(effort)
+}
+
+// mapAnthropicEffortToResponses maps shared effort levels directly and maps
+// Anthropic's max to OpenAI's xhigh.
 func mapAnthropicEffortToResponses(effort string) string {
 	if effort == "max" {
 		return "xhigh"

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge.go

@@ -99,13 +99,8 @@ func AnthropicToChatCompletionsRequest(req *AnthropicRequest) (*ChatCompletionsR
 		}
 	}
 
-	// Reasoning effort: output_config.effort maps 1:1 (max→xhigh). thinking.type
-	// itself is ignored (the Responses bridge behaves identically).
-	effort := "medium"
-	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
-		effort = req.OutputConfig.Effort
-	}
-	out.ReasoningEffort = mapAnthropicEffortToResponses(effort)
+	// Match the Responses bridge, including an explicit thinking disable.
+	out.ReasoningEffort = anthropicReasoningEffort(req)
 
 	parallelToolCalls := true
 	out.ParallelToolCalls = &parallelToolCalls

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge_test.go

@@ -150,6 +150,20 @@ func TestAnthropicToChatCompletionsRequest_ThinkingDropped(t *testing.T) {
 	require.Empty(t, out.Messages[0].ReasoningContent)
 }
 
+func TestAnthropicToChatCompletionsRequest_ThinkingDisabledOverridesOutputEffort(t *testing.T) {
+	req := &AnthropicRequest{
+		Model:        "gpt-5.6-luna",
+		MaxTokens:    1024,
+		Messages:     []AnthropicMessage{{Role: "user", Content: json.RawMessage(`"Hello"`)}},
+		Thinking:     &AnthropicThinking{Type: "disabled"},
+		OutputConfig: &AnthropicOutputConfig{Effort: "max"},
+	}
+
+	out, err := AnthropicToChatCompletionsRequest(req)
+	require.NoError(t, err)
+	require.Equal(t, "none", out.ReasoningEffort)
+}
+
 func TestAnthropicToChatCompletionsRequest_ToolChoiceAuto(t *testing.T) {
 	req := &AnthropicRequest{
 		Model:     "claude-sonnet-4-20250514",

backend/internal/service/openai_compat_model.go

@@ -107,6 +107,9 @@ func openAIReasoningEffortToClaudeOutputEffort(effort string) string {
 // normally translated to OpenAI xhigh, but GPT-5.6 accepts the original max
 // value on Responses and Chat Completions.
 func openAICompatAnthropicReasoningEffort(req *apicompat.AnthropicRequest, upstreamModel, convertedEffort string) string {
+	if convertedEffort == "none" {
+		return convertedEffort
+	}
 	if req == nil || req.OutputConfig == nil || !strings.EqualFold(strings.TrimSpace(req.OutputConfig.Effort), "max") {
 		return convertedEffort
 	}

backend/internal/service/openai_compat_model_test.go

@@ -313,6 +313,31 @@ func TestForwardAsAnthropic_PreservesMaxForFinalGPT56ResponsesModel(t *testing.T
 	}
 }
 
+func TestForwardAsAnthropic_DisabledThinkingOverridesMaxForResponses(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	body := []byte(`{"model":"sol","max_tokens":16,"messages":[{"role":"user","content":"hello"}],"thinking":{"type":"disabled"},"output_config":{"effort":"max"},"stream":false}`)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	upstream := &httpUpstreamRecorder{resp: openAICompatSSECompletedResponse("resp_no_thinking", "gpt-5.6-sol")}
+	svc := &OpenAIGatewayService{
+		httpUpstream: upstream,
+		cfg:          &config.Config{Security: config.SecurityConfig{URLAllowlist: config.URLAllowlistConfig{Enabled: false}}},
+	}
+	account := rawGPT56ResponsesOAuthAccount("sol", "gpt-5.6-sol")
+
+	result, err := svc.ForwardAsAnthropic(context.Background(), c, account, body, "", "")
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.Equal(t, http.StatusOK, rec.Code)
+	require.Equal(t, "none", gjson.GetBytes(upstream.lastBody, "reasoning.effort").String())
+	require.NotNil(t, result.ReasoningEffort)
+	require.Equal(t, "none", *result.ReasoningEffort)
+}
+
 func rawGPT56ResponsesAPIKeyAccount(requestedModel, mappedModel string) *Account {
 	return &Account{
 		ID:          501,

backend/internal/service/openai_gateway_messages_chat_fallback_test.go

@@ -91,6 +91,13 @@ func TestForwardAsAnthropic_ForceChatCompletionsPreservesFinalModelReasoningEffo
 			mapped:     "gpt-5.6-luna",
 			wantEffort: "medium",
 		},
+		{
+			name:       "disabled thinking overrides max",
+			model:      "gpt-5.6-luna",
+			mapped:     "gpt-5.6-luna",
+			effortJSON: `,"output_config":{"effort":"max"},"thinking":{"type":"disabled"}`,
+			wantEffort: "none",
+		},
 	}
 
 	for _, tt := range tests {