fix(agent): honor explicit thinking off (#2898)

* fix(agent): honor explicit thinking off

* fix(agent): address thinking off lint failures

* Clarify unset thinking level display

* fix ci

pkg/agent/agent_command.go

@@ -319,6 +319,7 @@ func (al *AgentLoop) buildCommandsRuntime(
 			agent.Provider = nextProvider
 			agent.Candidates = nextCandidates
 			agent.ThinkingLevel = parseThinkingLevel(modelCfg.ThinkingLevel)
+			agent.ThinkingLevelConfigured = isConfiguredThinkingLevel(modelCfg.ThinkingLevel)
 
 			if oldProvider != nil && oldProvider != nextProvider {
 				if stateful, ok := oldProvider.(providers.StatefulProvider); ok {

pkg/agent/agent_test.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"maps"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -138,6 +139,121 @@ func (r *recordingProvider) GetDefaultModel() string {
 	return "mock-model"
 }
 
+type thinkingRecordingProvider struct {
+	lastOptions map[string]any
+}
+
+func (r *thinkingRecordingProvider) Chat(
+	ctx context.Context,
+	messages []providers.Message,
+	tools []providers.ToolDefinition,
+	model string,
+	opts map[string]any,
+) (*providers.LLMResponse, error) {
+	r.lastOptions = maps.Clone(opts)
+	return &providers.LLMResponse{
+		Content:   "Mock response",
+		ToolCalls: []providers.ToolCall{},
+	}, nil
+}
+
+func (r *thinkingRecordingProvider) GetDefaultModel() string {
+	return "mock-model"
+}
+
+func (r *thinkingRecordingProvider) SupportsThinking() bool {
+	return true
+}
+
+type thinkingOptionRecordingProvider struct {
+	lastOptions map[string]any
+}
+
+func (r *thinkingOptionRecordingProvider) Chat(
+	ctx context.Context,
+	messages []providers.Message,
+	tools []providers.ToolDefinition,
+	model string,
+	opts map[string]any,
+) (*providers.LLMResponse, error) {
+	r.lastOptions = maps.Clone(opts)
+	return &providers.LLMResponse{
+		Content:   "Mock response",
+		ToolCalls: []providers.ToolCall{},
+	}, nil
+}
+
+func (r *thinkingOptionRecordingProvider) GetDefaultModel() string {
+	return "mock-model"
+}
+
+type reasoningOptionRecordingProvider struct {
+	lastOptions map[string]any
+}
+
+func (r *reasoningOptionRecordingProvider) Chat(
+	ctx context.Context,
+	messages []providers.Message,
+	tools []providers.ToolDefinition,
+	model string,
+	opts map[string]any,
+) (*providers.LLMResponse, error) {
+	r.lastOptions = maps.Clone(opts)
+	return &providers.LLMResponse{
+		Content:          "final answer",
+		ReasoningContent: "thinking trace",
+		ToolCalls:        []providers.ToolCall{},
+	}, nil
+}
+
+func (r *reasoningOptionRecordingProvider) GetDefaultModel() string {
+	return "mock-model"
+}
+
+type reasoningResponseProvider struct{}
+
+func (p *reasoningResponseProvider) Chat(
+	ctx context.Context,
+	messages []providers.Message,
+	tools []providers.ToolDefinition,
+	model string,
+	opts map[string]any,
+) (*providers.LLMResponse, error) {
+	return &providers.LLMResponse{
+		Content:          "Mock response",
+		ReasoningContent: "thinking trace",
+		ToolCalls:        []providers.ToolCall{},
+	}, nil
+}
+
+func (p *reasoningResponseProvider) GetDefaultModel() string {
+	return "mock-model"
+}
+
+type sideQuestionFallbackTestProvider struct {
+	model string
+}
+
+func (p *sideQuestionFallbackTestProvider) Chat(
+	ctx context.Context,
+	messages []providers.Message,
+	tools []providers.ToolDefinition,
+	model string,
+	opts map[string]any,
+) (*providers.LLMResponse, error) {
+	if p.model == "test-model" {
+		return nil, context.DeadlineExceeded
+	}
+	return &providers.LLMResponse{
+		ReasoningContent: "thinking trace",
+		ToolCalls:        []providers.ToolCall{},
+	}, nil
+}
+
+func (p *sideQuestionFallbackTestProvider) GetDefaultModel() string {
+	return p.model
+}
+
 type modelRewriteHook struct {
 	model string
 }
@@ -386,6 +502,463 @@ func TestProcessMessage_IncludesCurrentSenderInDynamicContext(t *testing.T) {
 	}
 }
 
+func TestProcessMessage_DoesNotPassImplicitThinkingOffToCapableProvider(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	provider := &thinkingRecordingProvider{}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+
+	response, err := al.processMessage(context.Background(), testInboundMessage(bus.InboundMessage{
+		Channel: "pico",
+		ChatID:  "chat-1",
+		Content: "hello",
+	}))
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "Mock response" {
+		t.Fatalf("processMessage() response = %q, want %q", response, "Mock response")
+	}
+	if _, ok := provider.lastOptions["thinking_level"]; ok {
+		t.Fatalf("thinking_level option should be omitted when unset, got %#v", provider.lastOptions["thinking_level"])
+	}
+}
+
+func TestProcessMessage_PassesExplicitThinkingOffToCapableProvider(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{{
+			ModelName:     "test-model",
+			Model:         "test-model",
+			ThinkingLevel: "off",
+		}},
+	}
+
+	provider := &thinkingRecordingProvider{}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+
+	response, err := al.processMessage(context.Background(), testInboundMessage(bus.InboundMessage{
+		Channel: "pico",
+		ChatID:  "chat-1",
+		Content: "hello",
+	}))
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "Mock response" {
+		t.Fatalf("processMessage() response = %q, want %q", response, "Mock response")
+	}
+	if got := provider.lastOptions["thinking_level"]; got != "off" {
+		t.Fatalf("thinking_level option = %#v, want %q", got, "off")
+	}
+}
+
+func TestProcessMessage_PassesExplicitThinkingOffToProviderWithoutThinkingCapability(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{{
+			ModelName:     "test-model",
+			Model:         "test-model",
+			ThinkingLevel: "off",
+		}},
+	}
+
+	provider := &thinkingOptionRecordingProvider{}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+
+	response, err := al.processMessage(context.Background(), testInboundMessage(bus.InboundMessage{
+		Channel: "pico",
+		ChatID:  "chat-1",
+		Content: "hello",
+	}))
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "Mock response" {
+		t.Fatalf("processMessage() response = %q, want %q", response, "Mock response")
+	}
+	if got := provider.lastOptions["thinking_level"]; got != "off" {
+		t.Fatalf("thinking_level option = %#v, want %q", got, "off")
+	}
+}
+
+func TestProcessMessage_SuppressesReasoningWhenThinkingOff(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{{
+			ModelName:     "test-model",
+			Model:         "test-model",
+			ThinkingLevel: "off",
+		}},
+	}
+
+	msgBus := bus.NewMessageBus()
+	al := NewAgentLoop(cfg, msgBus, &reasoningResponseProvider{})
+
+	response, err := al.runAgentLoop(
+		context.Background(),
+		al.GetRegistry().GetDefaultAgent(),
+		processOptions{
+			SessionKey:      "agent:main:pico:chat-1",
+			Channel:         "pico",
+			ChatID:          "chat-1",
+			UserMessage:     "hello",
+			SendResponse:    false,
+			DefaultResponse: defaultResponse,
+			NoHistory:       true,
+		},
+	)
+	if err != nil {
+		t.Fatalf("runAgentLoop() error = %v", err)
+	}
+	if response != "Mock response" {
+		t.Fatalf("response = %q, want %q", response, "Mock response")
+	}
+	select {
+	case outbound := <-msgBus.OutboundChan():
+		t.Fatalf("expected no reasoning outbound when thinking is off, got %+v", outbound)
+	case <-time.After(50 * time.Millisecond):
+	}
+}
+
+func TestProcessMessage_BeforeLLMModelRewriteReevaluatesThinkingOff(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "plain-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "plain-model",
+				Model:     "openai/plain-model",
+			},
+			{
+				ModelName:     "off-model",
+				Model:         "openai/off-model",
+				ThinkingLevel: "off",
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	provider := &reasoningOptionRecordingProvider{}
+	al := NewAgentLoop(cfg, msgBus, provider)
+	if err := al.MountHook(NamedHook("rewrite-model", modelRewriteHook{model: "off-model"})); err != nil {
+		t.Fatalf("MountHook failed: %v", err)
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "pico",
+		SenderID: "user1",
+		ChatID:   "pico:test-session",
+		Content:  "hello",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "final answer" {
+		t.Fatalf("processMessage() response = %q, want final answer", response)
+	}
+	if got := provider.lastOptions["thinking_level"]; got != "off" {
+		t.Fatalf("thinking_level option = %#v, want off after hook model rewrite", got)
+	}
+	select {
+	case outbound := <-msgBus.OutboundChan():
+		t.Fatalf("expected no reasoning outbound after hook rewrote to off model, got %+v", outbound)
+	case <-time.After(50 * time.Millisecond):
+	}
+}
+
+func TestProcessMessage_BeforeLLMModelRewriteDoesNotLeakThinkingOff(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         t.TempDir(),
+				ModelName:         "off-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName:     "off-model",
+				Model:         "openai/off-model",
+				ThinkingLevel: "off",
+			},
+			{
+				ModelName: "plain-model",
+				Model:     "openai/plain-model",
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	provider := &reasoningOptionRecordingProvider{}
+	al := NewAgentLoop(cfg, msgBus, provider)
+	if err := al.MountHook(NamedHook("rewrite-model", modelRewriteHook{model: "plain-model"})); err != nil {
+		t.Fatalf("MountHook failed: %v", err)
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "pico",
+		SenderID: "user1",
+		ChatID:   "pico:test-session",
+		Content:  "hello",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "final answer" {
+		t.Fatalf("processMessage() response = %q, want final answer", response)
+	}
+	if _, ok := provider.lastOptions["thinking_level"]; ok {
+		t.Fatalf(
+			"thinking_level option should be cleared after hook rewrote away from off model, got %#v",
+			provider.lastOptions["thinking_level"],
+		)
+	}
+	select {
+	case outbound := <-msgBus.OutboundChan():
+		if outbound.Content != "thinking trace" {
+			t.Fatalf("reasoning outbound content = %q, want thinking trace", outbound.Content)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("expected reasoning outbound after hook rewrote away from off model")
+	}
+}
+
+func TestProcessMessage_BtwCommandSuppressesReasoningWhenThinkingOff(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				ModelName:         "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{{
+			ModelName:     "test-model",
+			Model:         "openai/test-model",
+			ThinkingLevel: "off",
+		}},
+	}
+
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), &sideQuestionFallbackTestProvider{model: "test-model"})
+	al.providerFactory = func(mc *config.ModelConfig) (providers.LLMProvider, string, error) {
+		model := ""
+		if mc != nil {
+			_, model = providers.ExtractProtocol(mc)
+		}
+		if model == "" {
+			model = "test-model"
+		}
+		return &sideQuestionFallbackTestProvider{model: model}, model, nil
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "telegram:123",
+		ChatID:   "chat-1",
+		Content:  "/btw explain privately",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if strings.Contains(response, "thinking trace") {
+		t.Fatalf("processMessage() response = %q, should not expose reasoning with thinking off", response)
+	}
+}
+
+func TestProcessMessage_BtwHookModelRewriteReevaluatesThinkingOff(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				ModelName:         "plain-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "plain-model",
+				Model:     "openai/plain-model",
+			},
+			{
+				ModelName:     "off-model",
+				Model:         "openai/off-model",
+				ThinkingLevel: "off",
+			},
+		},
+	}
+
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), &sideQuestionFallbackTestProvider{model: "plain-model"})
+	al.providerFactory = func(mc *config.ModelConfig) (providers.LLMProvider, string, error) {
+		model := ""
+		if mc != nil {
+			_, model = providers.ExtractProtocol(mc)
+		}
+		if model == "" {
+			model = "plain-model"
+		}
+		return &sideQuestionFallbackTestProvider{model: model}, model, nil
+	}
+	if err := al.MountHook(NamedHook("rewrite-model", modelRewriteHook{model: "off-model"})); err != nil {
+		t.Fatalf("MountHook failed: %v", err)
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "telegram:123",
+		ChatID:   "chat-1",
+		Content:  "/btw explain privately",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if strings.Contains(response, "thinking trace") {
+		t.Fatalf(
+			"processMessage() response = %q, should not expose reasoning after hook rewrote to off model",
+			response,
+		)
+	}
+}
+
+func TestProcessMessage_BtwHookModelRewriteDoesNotLeakThinkingOff(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				ModelName:         "off-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName:     "off-model",
+				Model:         "openai/off-model",
+				ThinkingLevel: "off",
+			},
+			{
+				ModelName: "plain-model",
+				Model:     "openai/plain-model",
+			},
+		},
+	}
+
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), &sideQuestionFallbackTestProvider{model: "off-model"})
+	al.providerFactory = func(mc *config.ModelConfig) (providers.LLMProvider, string, error) {
+		model := ""
+		if mc != nil {
+			_, model = providers.ExtractProtocol(mc)
+		}
+		if model == "" {
+			model = "off-model"
+		}
+		return &sideQuestionFallbackTestProvider{model: model}, model, nil
+	}
+	if err := al.MountHook(NamedHook("rewrite-model", modelRewriteHook{model: "plain-model"})); err != nil {
+		t.Fatalf("MountHook failed: %v", err)
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "telegram:123",
+		ChatID:   "chat-1",
+		Content:  "/btw explain privately",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "thinking trace" {
+		t.Fatalf("processMessage() response = %q, want reasoning after hook rewrote away from off model", response)
+	}
+}
+
+func TestProcessMessage_BtwFallbackDoesNotInheritPrimaryThinkingOff(t *testing.T) {
+	tmpDir := t.TempDir()
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				ModelName:         "test-model",
+				ModelFallbacks:    []string{"openai/fallback-model"},
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+		ModelList: []*config.ModelConfig{{
+			ModelName:     "test-model",
+			Model:         "openai/test-model",
+			ThinkingLevel: "off",
+		}},
+	}
+
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), &sideQuestionFallbackTestProvider{model: "test-model"})
+	al.providerFactory = func(mc *config.ModelConfig) (providers.LLMProvider, string, error) {
+		model := ""
+		if mc != nil {
+			_, model = providers.ExtractProtocol(mc)
+		}
+		if model == "" {
+			model = "test-model"
+		}
+		return &sideQuestionFallbackTestProvider{model: model}, model, nil
+	}
+
+	response, err := al.processMessage(context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "telegram:123",
+		ChatID:   "chat-1",
+		Content:  "/btw explain fallback reasoning",
+	})
+	if err != nil {
+		t.Fatalf("processMessage() error = %v", err)
+	}
+	if response != "thinking trace" {
+		t.Fatalf("processMessage() response = %q, want fallback reasoning when fallback has no off", response)
+	}
+}
+
 func TestProcessMessage_UseCommandLoadsRequestedSkill(t *testing.T) {
 	tmpDir := t.TempDir()
 	skillDir := filepath.Join(tmpDir, "skills", "shell")
@@ -3178,6 +3751,300 @@ func TestProcessMessage_FallbackUsesPerCandidateProvider(t *testing.T) {
 	}
 }
 
+func TestProcessMessage_FallbackReceivesExplicitThinkingOff(t *testing.T) {
+	workspace := t.TempDir()
+
+	primaryServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusTooManyRequests)
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"error": map[string]any{
+				"message": "rate limit exceeded",
+				"type":    "rate_limit_error",
+			},
+		})
+	}))
+	defer primaryServer.Close()
+
+	fallbackCalls := 0
+	fallbackServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		fallbackCalls++
+		if r.URL.Path != "/chat/completions" {
+			t.Fatalf("fallback server path = %q, want /chat/completions", r.URL.Path)
+		}
+		defer r.Body.Close()
+
+		var req map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatalf("decode fallback request: %v", err)
+		}
+		if got := req["model"]; got != "doubao-seed-1-6-flash-250828" {
+			t.Fatalf("fallback request model = %#v, want doubao-seed-1-6-flash-250828", got)
+		}
+		thinking, ok := req["thinking"].(map[string]any)
+		if !ok {
+			t.Fatalf("fallback request thinking = %#v, want map", req["thinking"])
+		}
+		if got := thinking["type"]; got != "disabled" {
+			t.Fatalf("fallback request thinking.type = %#v, want disabled", got)
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		if err := json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{
+				{
+					"message":       map[string]any{"content": "fallback reply"},
+					"finish_reason": "stop",
+				},
+			},
+		}); err != nil {
+			t.Fatalf("encode fallback response: %v", err)
+		}
+	}))
+	defer fallbackServer.Close()
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         workspace,
+				ModelName:         "primary-model",
+				ModelFallbacks:    []string{"doubao-fallback"},
+				MaxTokens:         4096,
+				MaxToolIterations: 3,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "primary-model",
+				Model:     "openrouter/primary-model",
+				APIBase:   primaryServer.URL,
+				APIKeys:   config.SimpleSecureStrings("primary-key"),
+				Workspace: workspace,
+			},
+			{
+				ModelName:     "doubao-fallback",
+				Model:         "openai/doubao-seed-1-6-flash-250828",
+				APIBase:       fallbackServer.URL,
+				APIKeys:       config.SimpleSecureStrings("fallback-key"),
+				ThinkingLevel: "off",
+				Workspace:     workspace,
+			},
+		},
+	}
+
+	provider, _, err := providers.CreateProvider(cfg)
+	if err != nil {
+		t.Fatalf("CreateProvider() error = %v", err)
+	}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+	helper := testHelper{al: al}
+
+	resp := helper.executeAndGetResponse(t, context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "user1",
+		ChatID:   "chat1",
+		Content:  "hi",
+	})
+
+	if resp != "fallback reply" {
+		t.Fatalf("response = %q, want fallback reply", resp)
+	}
+	if fallbackCalls != 1 {
+		t.Fatalf("fallback server calls = %d, want 1", fallbackCalls)
+	}
+}
+
+func TestProcessMessage_PrimaryThinkingOffDoesNotLeakToFallback(t *testing.T) {
+	workspace := t.TempDir()
+
+	primaryServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusTooManyRequests)
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"error": map[string]any{
+				"message": "rate limit exceeded",
+				"type":    "rate_limit_error",
+			},
+		})
+	}))
+	defer primaryServer.Close()
+
+	fallbackServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/chat/completions" {
+			t.Fatalf("fallback server path = %q, want /chat/completions", r.URL.Path)
+		}
+		defer r.Body.Close()
+
+		var req map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatalf("decode fallback request: %v", err)
+		}
+		if _, ok := req["thinking"]; ok {
+			t.Fatalf("fallback request should not inherit primary thinking off, got thinking=%#v", req["thinking"])
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		if err := json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{
+				{
+					"message":       map[string]any{"content": "fallback reply"},
+					"finish_reason": "stop",
+				},
+			},
+		}); err != nil {
+			t.Fatalf("encode fallback response: %v", err)
+		}
+	}))
+	defer fallbackServer.Close()
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         workspace,
+				ModelName:         "primary-model",
+				ModelFallbacks:    []string{"doubao-fallback"},
+				MaxTokens:         4096,
+				MaxToolIterations: 3,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName:     "primary-model",
+				Model:         "openrouter/primary-model",
+				APIBase:       primaryServer.URL,
+				APIKeys:       config.SimpleSecureStrings("primary-key"),
+				ThinkingLevel: "off",
+				Workspace:     workspace,
+			},
+			{
+				ModelName: "doubao-fallback",
+				Model:     "openai/doubao-seed-1-6-flash-250828",
+				APIBase:   fallbackServer.URL,
+				APIKeys:   config.SimpleSecureStrings("fallback-key"),
+				Workspace: workspace,
+			},
+		},
+	}
+
+	provider, _, err := providers.CreateProvider(cfg)
+	if err != nil {
+		t.Fatalf("CreateProvider() error = %v", err)
+	}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+	helper := testHelper{al: al}
+
+	resp := helper.executeAndGetResponse(t, context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "user1",
+		ChatID:   "chat1",
+		Content:  "hi",
+	})
+	if resp != "fallback reply" {
+		t.Fatalf("response = %q, want fallback reply", resp)
+	}
+}
+
+func TestProcessMessage_FallbackThinkingOffUsesCandidateIdentity(t *testing.T) {
+	workspace := t.TempDir()
+
+	primaryServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusTooManyRequests)
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"error": map[string]any{
+				"message": "rate limit exceeded",
+				"type":    "rate_limit_error",
+			},
+		})
+	}))
+	defer primaryServer.Close()
+
+	fallbackServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/chat/completions" {
+			t.Fatalf("fallback server path = %q, want /chat/completions", r.URL.Path)
+		}
+		defer r.Body.Close()
+
+		var req map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatalf("decode fallback request: %v", err)
+		}
+		thinking, ok := req["thinking"].(map[string]any)
+		if !ok {
+			t.Fatalf("fallback request thinking = %#v, want map", req["thinking"])
+		}
+		if got := thinking["type"]; got != "disabled" {
+			t.Fatalf("fallback request thinking.type = %#v, want disabled", got)
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		if err := json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{
+				{
+					"message":       map[string]any{"content": "fallback reply"},
+					"finish_reason": "stop",
+				},
+			},
+		}); err != nil {
+			t.Fatalf("encode fallback response: %v", err)
+		}
+	}))
+	defer fallbackServer.Close()
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         workspace,
+				ModelName:         "primary-model",
+				ModelFallbacks:    []string{"doubao-off"},
+				MaxTokens:         4096,
+				MaxToolIterations: 3,
+			},
+		},
+		ModelList: []*config.ModelConfig{
+			{
+				ModelName: "primary-model",
+				Model:     "openrouter/primary-model",
+				APIBase:   primaryServer.URL,
+				APIKeys:   config.SimpleSecureStrings("primary-key"),
+				Workspace: workspace,
+			},
+			{
+				ModelName: "doubao-default",
+				Model:     "openai/doubao-seed-1-6-flash-250828",
+				APIBase:   fallbackServer.URL,
+				APIKeys:   config.SimpleSecureStrings("fallback-key"),
+				Workspace: workspace,
+			},
+			{
+				ModelName:     "doubao-off",
+				Model:         "openai/doubao-seed-1-6-flash-250828",
+				APIBase:       fallbackServer.URL,
+				APIKeys:       config.SimpleSecureStrings("fallback-key"),
+				ThinkingLevel: "off",
+				Workspace:     workspace,
+			},
+		},
+	}
+
+	provider, _, err := providers.CreateProvider(cfg)
+	if err != nil {
+		t.Fatalf("CreateProvider() error = %v", err)
+	}
+	al := NewAgentLoop(cfg, bus.NewMessageBus(), provider)
+	helper := testHelper{al: al}
+
+	resp := helper.executeAndGetResponse(t, context.Background(), bus.InboundMessage{
+		Channel:  "telegram",
+		SenderID: "user1",
+		ChatID:   "chat1",
+		Content:  "hi",
+	})
+	if resp != "fallback reply" {
+		t.Fatalf("response = %q, want fallback reply", resp)
+	}
+}
+
 // TestProcessMessage_FallbackUsesActiveProviderWhenCandidateNotRegistered verifies
 // that when a candidate has no model_list entry it is absent from CandidateProviders
 // and the fallback closure falls back to activeProvider instead of panicking.

pkg/agent/agent_utils.go

@@ -560,6 +560,7 @@ func sideQuestionModelName(agent *AgentInstance, usedLight bool) string {
 }
 
 func modelNameFromIdentityKey(identityKey string) string {
+	identityKey = strings.TrimSpace(identityKey)
 	if identityKey == "" {
 		return ""
 	}

pkg/agent/instance.go

@@ -31,6 +31,7 @@ type AgentInstance struct {
 	MaxTokens                 int
 	Temperature               float64
 	ThinkingLevel             ThinkingLevel
+	ThinkingLevelConfigured   bool
 	ContextWindow             int
 	SummarizeMessageThreshold int
 	SummarizeTokenPercent     int
@@ -184,6 +185,7 @@ func NewAgentInstance(
 		thinkingLevelStr = mc.ThinkingLevel
 	}
 	thinkingLevel := parseThinkingLevel(thinkingLevelStr)
+	thinkingLevelConfigured := isConfiguredThinkingLevel(thinkingLevelStr)
 
 	summarizeMessageThreshold := defaults.SummarizeMessageThreshold
 	if summarizeMessageThreshold == 0 {
@@ -251,6 +253,7 @@ func NewAgentInstance(
 		MaxTokens:                 maxTokens,
 		Temperature:               temperature,
 		ThinkingLevel:             thinkingLevel,
+		ThinkingLevelConfigured:   thinkingLevelConfigured,
 		ContextWindow:             contextWindow,
 		SummarizeMessageThreshold: summarizeMessageThreshold,
 		SummarizeTokenPercent:     summarizeTokenPercent,

pkg/agent/pipeline_llm.go

@@ -73,14 +73,7 @@ func (p *Pipeline) CallLLM(
 	if exec.useNativeSearch {
 		exec.llmOpts["native_search"] = true
 	}
-	if ts.agent.ThinkingLevel != ThinkingOff {
+	applyTurnThinkingOptions(exec, ts.agent, exec.activeProvider, true)
-		if tc, ok := ts.agent.Provider.(providers.ThinkingCapable); ok && tc.SupportsThinking() {
-			exec.llmOpts["thinking_level"] = string(ts.agent.ThinkingLevel)
-		} else {
-			logger.WarnCF("agent", "thinking_level is set but current provider does not support it, ignoring",
-				map[string]any{"agent_id": ts.agent.ID, "thinking_level": string(ts.agent.ThinkingLevel)})
-		}
-	}
 
 	exec.llmModel = exec.activeModel
 
@@ -105,6 +98,7 @@ func (p *Pipeline) CallLLM(
 				exec.llmOpts = llmReq.Options
 				if strings.TrimSpace(exec.llmModel) != "" && exec.llmModel != prevModel {
 					p.applyBeforeLLMModelRewrite(ts, exec)
+					applyTurnThinkingOptions(exec, ts.agent, exec.activeProvider, true)
 				}
 			}
 		case HookActionAbortTurn:
@@ -172,21 +166,33 @@ func (p *Pipeline) CallLLM(
 		}
 
 		if len(exec.activeCandidates) > 1 && p.Fallback != nil {
-			fbResult, fbErr := p.Fallback.Execute(
+			fbResult, fbErr := p.Fallback.ExecuteCandidate(
 				providerCtx,
 				exec.activeCandidates,
-				func(ctx context.Context, provider, model string) (*providers.LLMResponse, error) {
+				func(ctx context.Context, candidate providers.FallbackCandidate) (*providers.LLMResponse, error) {
 					candidateProvider, err := providerForFallbackCandidate(
 						ts.agent,
 						exec.activeProvider,
 						exec.activeCandidates,
-						provider,
+						candidate.Provider,
-						model,
+						candidate.Model,
 					)
 					if err != nil {
 						return nil, err
 					}
-					return candidateProvider.Chat(ctx, messagesForCall, toolDefsForCall, model, exec.llmOpts)
+					callOpts := shallowCloneLLMOptions(exec.llmOpts)
+					delete(callOpts, "thinking_level")
+					candidateCfg := resolveActiveModelConfig(
+						p.Cfg,
+						ts.agent.Workspace,
+						[]providers.FallbackCandidate{candidate},
+						candidate.Model,
+						p.Cfg.Agents.Defaults.Provider,
+					)
+					candidateThinking := thinkingSettingsFromModelConfig(candidateCfg)
+					applyThinkingOption(callOpts, candidateProvider, candidateThinking, true, ts.agent.ID)
+					exec.suppressReasoning = shouldSuppressReasoningFor(candidateThinking)
+					return candidateProvider.Chat(ctx, messagesForCall, toolDefsForCall, candidate.Model, callOpts)
 				},
 			)
 			if fbErr != nil {
@@ -469,6 +475,11 @@ func (p *Pipeline) CallLLM(
 		}
 	}
 
+	if exec.suppressReasoning {
+		exec.response.Reasoning = ""
+		exec.response.ReasoningContent = ""
+		exec.response.ReasoningDetails = nil
+	}
 	reasoningContent := responseReasoningContent(exec.response)
 	shouldPublishPicoToolCallInterim := ts.channel == "pico" && len(exec.response.ToolCalls) > 0
 	if shouldPublishPicoToolCallInterim {

pkg/agent/pipeline_streaming.go

@@ -85,7 +85,7 @@ func (p *Pipeline) tryConfiguredStreamingLLM(
 			exec.llmOpts,
 			func(chunk providers.StreamChunk) {
 				recordChunk()
-				if strings.TrimSpace(chunk.ReasoningContent) != "" {
+				if !exec.suppressReasoning && strings.TrimSpace(chunk.ReasoningContent) != "" {
 					publisher.UpdateReasoning(ctx, chunk.ReasoningContent)
 				}
 				if strings.TrimSpace(chunk.Content) != "" {

pkg/agent/pipeline_streaming_test.go

@@ -517,6 +517,46 @@ func TestConfiguredStreamingStreamsPicoReasoningBeforeAnswerContent(t *testing.T
 	}
 }
 
+func TestConfiguredStreamingSuppressesPicoReasoningWhenThinkingOff(t *testing.T) {
+	cfg := newConfiguredStreamingTestConfig(t, true, true, nil)
+	cfg.ModelList[0].ThinkingLevel = "off"
+	streamer := &recordingStreamer{}
+	msgBus := bus.NewMessageBus()
+	msgBus.SetStreamDelegate(configuredStreamingDelegate{streamer: streamer})
+	provider := &configuredStreamingProvider{
+		eventPlan: []configuredStreamingEventCall{{
+			chunks: []providers.StreamChunk{
+				{ReasoningContent: "thinking"},
+				{Content: "answer"},
+			},
+			response: &providers.LLMResponse{
+				Content:          "answer",
+				ReasoningContent: "thinking",
+			},
+		}},
+	}
+	al := NewAgentLoop(cfg, msgBus, provider)
+
+	got := runConfiguredStreamingTurn(t, al, "pico")
+	if got != "answer" {
+		t.Fatalf("response = %q, want answer", got)
+	}
+	if len(streamer.reasoningUpdates) != 0 {
+		t.Fatalf("reasoning updates = %v, want none when thinking is off", streamer.reasoningUpdates)
+	}
+	if len(streamer.reasoningFinalized) != 0 {
+		t.Fatalf("reasoning finalized = %v, want none when thinking is off", streamer.reasoningFinalized)
+	}
+	if len(streamer.updates) != 1 || streamer.updates[0] != "answer" {
+		t.Fatalf("content updates = %v, want [answer]", streamer.updates)
+	}
+	select {
+	case outbound := <-msgBus.OutboundChan():
+		t.Fatalf("expected no reasoning outbound when thinking is off, got %+v", outbound)
+	case <-time.After(50 * time.Millisecond):
+	}
+}
+
 func TestConfiguredStreamingFinalFlushFailureAfterVisibleOutputReturnsErrorWithoutFallbackOrCancel(t *testing.T) {
 	cfg := newConfiguredStreamingTestConfig(t, true, true, nil)
 	streamer := &failingFinalizeStreamer{err: errors.New("final failed")}

pkg/agent/thinking.go

@@ -1,6 +1,12 @@
 package agent
 
-import "strings"
+import (
+	"strings"
+
+	"github.com/sipeed/picoclaw/pkg/config"
+	"github.com/sipeed/picoclaw/pkg/logger"
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
 
 // ThinkingLevel controls how the provider sends thinking parameters.
 //
@@ -37,3 +43,87 @@ func parseThinkingLevel(level string) ThinkingLevel {
 		return ThinkingOff
 	}
 }
+
+func isConfiguredThinkingLevel(level string) bool {
+	switch strings.ToLower(strings.TrimSpace(level)) {
+	case "off", "low", "medium", "high", "xhigh", "adaptive":
+		return true
+	default:
+		return false
+	}
+}
+
+type thinkingSettings struct {
+	level      ThinkingLevel
+	configured bool
+}
+
+func thinkingSettingsFromModelConfig(mc *config.ModelConfig) thinkingSettings {
+	if mc == nil || !isConfiguredThinkingLevel(mc.ThinkingLevel) {
+		return thinkingSettings{}
+	}
+	return thinkingSettings{
+		level:      parseThinkingLevel(mc.ThinkingLevel),
+		configured: true,
+	}
+}
+
+func activeThinkingSettings(agent *AgentInstance, modelCfg *config.ModelConfig) thinkingSettings {
+	if settings := thinkingSettingsFromModelConfig(modelCfg); settings.configured {
+		return settings
+	}
+	if modelCfg == nil && agent != nil {
+		return thinkingSettings{
+			level:      agent.ThinkingLevel,
+			configured: agent.ThinkingLevelConfigured,
+		}
+	}
+	return thinkingSettings{}
+}
+
+func applyThinkingOption(
+	opts map[string]any,
+	provider providers.LLMProvider,
+	settings thinkingSettings,
+	warnUnsupported bool,
+	agentID string,
+) {
+	if opts == nil || !settings.configured {
+		return
+	}
+	if settings.level == ThinkingOff {
+		opts["thinking_level"] = string(settings.level)
+		return
+	}
+	if tc, ok := provider.(providers.ThinkingCapable); ok && tc.SupportsThinking() {
+		opts["thinking_level"] = string(settings.level)
+		return
+	}
+	if warnUnsupported {
+		logger.WarnCF("agent", "thinking_level is set but current provider does not support it, ignoring",
+			map[string]any{"agent_id": agentID, "thinking_level": string(settings.level)})
+	}
+}
+
+func applyTurnThinkingOptions(
+	exec *turnExecution,
+	agent *AgentInstance,
+	provider providers.LLMProvider,
+	warnUnsupported bool,
+) {
+	if exec == nil || exec.llmOpts == nil {
+		return
+	}
+	delete(exec.llmOpts, "thinking_level")
+	settings := activeThinkingSettings(agent, exec.activeModelConfig)
+	agentID := ""
+	if agent != nil {
+		agentID = agent.ID
+	}
+	applyThinkingOption(exec.llmOpts, provider, settings, warnUnsupported, agentID)
+	exec.suppressReasoning = shouldSuppressReasoningFor(settings)
+}
+
+func shouldSuppressReasoningFor(settings thinkingSettings) bool {
+	return settings.configured && settings.level == ThinkingOff
+}

pkg/agent/turn_coord.go

@@ -423,6 +423,7 @@ func (al *AgentLoop) askSideQuestion(
 	}
 
 	hookModelChanged := false
+	sideSuppressReasoning := false
 	callProvider := func(
 		ctx context.Context,
 		candidate providers.FallbackCandidate,
@@ -430,7 +431,15 @@ func (al *AgentLoop) askSideQuestion(
 		forceModel bool,
 		callMessages []providers.Message,
 	) (*providers.LLMResponse, error) {
-		provider, providerModel, cleanup, err := al.isolatedSideQuestionProvider(agent, selectedModelName, candidate)
+		baseModelName := selectedModelName
+		if forceModel && strings.TrimSpace(model) != "" {
+			baseModelName = model
+		}
+		provider, providerModel, modelCfg, cleanup, err := al.isolatedSideQuestionProvider(
+			agent,
+			baseModelName,
+			candidate,
+		)
 		if err != nil {
 			return nil, err
 		}
@@ -439,10 +448,12 @@ func (al *AgentLoop) askSideQuestion(
 			model = providerModel
 		}
 		callOpts := llmOpts
-		if _, exists := callOpts["thinking_level"]; !exists && agent.ThinkingLevel != ThinkingOff {
+		settings := thinkingSettingsFromModelConfig(modelCfg)
-			if tc, ok := provider.(providers.ThinkingCapable); ok && tc.SupportsThinking() {
+		sideSuppressReasoning = shouldSuppressReasoningFor(settings)
+		if _, exists := callOpts["thinking_level"]; !exists {
+			if settings.configured {
 				callOpts = shallowCloneLLMOptions(llmOpts)
-				callOpts["thinking_level"] = string(agent.ThinkingLevel)
+				applyThinkingOption(callOpts, provider, settings, false, agent.ID)
 			}
 		}
 		return provider.Chat(ctx, callMessages, nil, model, callOpts)
@@ -500,18 +511,11 @@ func (al *AgentLoop) askSideQuestion(
 
 	callSideLLM := func(callMessages []providers.Message) (*providers.LLMResponse, error) {
 		if len(activeCandidates) > 1 && al.fallback != nil {
-			fbResult, err := al.fallback.Execute(
+			fbResult, err := al.fallback.ExecuteCandidate(
 				ctx,
 				activeCandidates,
-				func(ctx context.Context, providerName, model string) (*providers.LLMResponse, error) {
+				func(ctx context.Context, candidate providers.FallbackCandidate) (*providers.LLMResponse, error) {
-					candidate := providers.FallbackCandidate{Provider: providerName, Model: model}
+					return callProvider(ctx, candidate, candidate.Model, false, callMessages)
-					for _, activeCandidate := range activeCandidates {
-						if activeCandidate.Provider == providerName && activeCandidate.Model == model {
-							candidate = activeCandidate
-							break
-						}
-					}
-					return callProvider(ctx, candidate, model, false, callMessages)
 				},
 			)
 			if err != nil {
@@ -584,6 +588,11 @@ func (al *AgentLoop) askSideQuestion(
 			return "", fmt.Errorf("hook aborted turn during after_llm: %s", reason)
 		}
 	}
+	if sideSuppressReasoning {
+		resp.Reasoning = ""
+		resp.ReasoningContent = ""
+		resp.ReasoningDetails = nil
+	}
 
 	return sideQuestionResponseContent(resp), nil
 }
@@ -592,14 +601,14 @@ func (al *AgentLoop) isolatedSideQuestionProvider(
 	agent *AgentInstance,
 	baseModelName string,
 	candidate providers.FallbackCandidate,
-) (providers.LLMProvider, string, func(), error) {
+) (providers.LLMProvider, string, *config.ModelConfig, func(), error) {
 	if agent == nil {
-		return nil, "", func() {}, fmt.Errorf("isolatedSideQuestionProvider: no agent available for /btw")
+		return nil, "", nil, func() {}, fmt.Errorf("isolatedSideQuestionProvider: no agent available for /btw")
 	}
 
 	modelCfg, err := al.sideQuestionModelConfig(agent, baseModelName, candidate)
 	if err != nil {
-		return nil, "", func() {}, fmt.Errorf("isolatedSideQuestionProvider: %w", err)
+		return nil, "", nil, func() {}, fmt.Errorf("isolatedSideQuestionProvider: %w", err)
 	}
 
 	factory := al.providerFactory
@@ -608,13 +617,13 @@ func (al *AgentLoop) isolatedSideQuestionProvider(
 	}
 	provider, modelID, err := factory(modelCfg)
 	if err != nil {
-		return nil, "", func() {}, fmt.Errorf("isolatedSideQuestionProvider: %w", err)
+		return nil, "", nil, func() {}, fmt.Errorf("isolatedSideQuestionProvider: %w", err)
 	}
 
 	cleanup := func() {
 		closeProviderIfStateful(provider)
 	}
-	return provider, modelID, cleanup, nil
+	return provider, modelID, modelCfg, cleanup, nil
 }
 
 func (al *AgentLoop) sideQuestionModelConfig(
@@ -626,7 +635,15 @@ func (al *AgentLoop) sideQuestionModelConfig(
 		return nil, fmt.Errorf("sideQuestionModelConfig: no agent available for /btw")
 	}
 
-	// If candidate has an identity key, use that
+	if name := modelAliasFromCandidateIdentityKey(candidate.IdentityKey); name != "" {
+		modelCfg, err := resolvedModelConfig(al.GetConfig(), name, agent.Workspace)
+		if err == nil {
+			return modelCfg, nil
+		}
+		// Fallback: create a minimal config if lookup fails
+	}
+
+	// Older identity keys used provider/model; keep resolving those by model.
 	if name := modelNameFromIdentityKey(candidate.IdentityKey); name != "" {
 		modelCfg, err := resolvedModelConfig(al.GetConfig(), name, agent.Workspace)
 		if err == nil {
@@ -635,6 +652,18 @@ func (al *AgentLoop) sideQuestionModelConfig(
 		// Fallback: create a minimal config if lookup fails
 	}
 
+	if candidate.Provider != "" && candidate.Model != "" {
+		candidateRef := providers.NormalizeProvider(candidate.Provider) + "/" + candidate.Model
+		if modelCfg, err := resolvedModelConfig(al.GetConfig(), candidateRef, agent.Workspace); err == nil {
+			return modelCfg, nil
+		}
+		return &config.ModelConfig{
+			ModelName: candidateRef,
+			Model:     candidateRef,
+			Workspace: agent.Workspace,
+		}, nil
+	}
+
 	// Otherwise, clean up the base model name and use it
 	baseModelName = strings.TrimSpace(baseModelName)
 	modelCfg, err := resolvedModelConfig(al.GetConfig(), baseModelName, agent.Workspace)
@@ -658,8 +687,5 @@ func (al *AgentLoop) sideQuestionModelConfig(
 
 	// If candidate specifies a different provider/model, override
 	clone := *modelCfg
-	if candidate.Provider != "" && candidate.Model != "" {
-		clone.Model = providers.NormalizeProvider(candidate.Provider) + "/" + candidate.Model
-	}
 	return &clone, nil
 }

pkg/agent/turn_state.go

@@ -138,6 +138,7 @@ type turnExecution struct {
 	allResponsesHandled bool
 	streamingPublisher  *streamingChunkPublisher
 	streamingFallback   bool
+	suppressReasoning   bool
 	callMessages        []providers.Message
 	providerToolDefs    []providers.ToolDefinition
 	llmModel            string

pkg/providers/fallback.go

@@ -118,6 +118,22 @@ func (fc *FallbackChain) Execute(
 	ctx context.Context,
 	candidates []FallbackCandidate,
 	run func(ctx context.Context, provider, model string) (*LLMResponse, error),
+) (*FallbackResult, error) {
+	return fc.ExecuteCandidate(
+		ctx,
+		candidates,
+		func(ctx context.Context, candidate FallbackCandidate) (*LLMResponse, error) {
+			return run(ctx, candidate.Provider, candidate.Model)
+		},
+	)
+}
+
+// ExecuteCandidate runs the fallback chain and passes the complete candidate
+// to the caller so model-list identity metadata remains available.
+func (fc *FallbackChain) ExecuteCandidate(
+	ctx context.Context,
+	candidates []FallbackCandidate,
+	run func(ctx context.Context, candidate FallbackCandidate) (*LLMResponse, error),
 ) (*FallbackResult, error) {
 	if len(candidates) == 0 {
 		return nil, fmt.Errorf("fallback: no candidates configured")
@@ -181,7 +197,7 @@ func (fc *FallbackChain) Execute(
 
 		// Execute the run function.
 		start := time.Now()
-		resp, err := run(ctx, candidate.Provider, candidate.Model)
+		resp, err := run(ctx, candidate)
 		elapsed := time.Since(start)
 
 		if err == nil {

pkg/providers/openai_compat/provider.go

@@ -193,6 +193,8 @@ func (p *Provider) buildRequestBody(
 		}
 	}
 
+	p.applyThinkingControl(requestBody, model, options)
+
 	// Merge extra body fields configured per-provider/model.
 	// These are injected last so they take precedence over defaults.
 	maps.Copy(requestBody, p.extraBody)
@@ -200,6 +202,81 @@ func (p *Provider) buildRequestBody(
 	return requestBody
 }
 
+func (p *Provider) applyThinkingControl(requestBody map[string]any, model string, options map[string]any) {
+	level, ok := normalizedThinkingLevel(options)
+	if !ok || level != "off" {
+		return
+	}
+
+	switch p.thinkingControlKind(model) {
+	case "thinking_type":
+		requestBody["thinking"] = map[string]any{"type": "disabled"}
+	case "enable_thinking":
+		requestBody["enable_thinking"] = false
+	}
+}
+
+func normalizedThinkingLevel(options map[string]any) (string, bool) {
+	raw, ok := options["thinking_level"].(string)
+	if !ok {
+		return "", false
+	}
+	switch strings.ToLower(strings.TrimSpace(raw)) {
+	case "off", "low", "medium", "high", "xhigh", "adaptive":
+		return strings.ToLower(strings.TrimSpace(raw)), true
+	default:
+		return "", false
+	}
+}
+
+func (p *Provider) thinkingControlKind(model string) string {
+	providerName := strings.ToLower(strings.TrimSpace(p.providerName))
+	lowerModel := strings.ToLower(strings.TrimSpace(model))
+
+	switch providerName {
+	case "volcengine":
+		return "thinking_type"
+	case "zhipu", "zai":
+		return "thinking_type"
+	case "qwen", "qwen-portal", "qwen-intl", "qwen-international", "dashscope-intl", "qwen-us", "dashscope-us":
+		return "enable_thinking"
+	case "modelscope":
+		if strings.Contains(lowerModel, "qwen") {
+			return "enable_thinking"
+		}
+	}
+
+	if providerName == "openai" || providerName == "" {
+		if isVolcengineHost(p.apiBase) || strings.Contains(lowerModel, "doubao") {
+			return "thinking_type"
+		}
+		if isDashScopeHost(p.apiBase) || strings.Contains(lowerModel, "qwen") {
+			return "enable_thinking"
+		}
+	}
+
+	return ""
+}
+
+func isVolcengineHost(apiBase string) bool {
+	host := normalizedHostname(apiBase)
+	return host == "volcengine.com" || strings.HasSuffix(host, ".volcengine.com") ||
+		host == "volces.com" || strings.HasSuffix(host, ".volces.com")
+}
+
+func isDashScopeHost(apiBase string) bool {
+	host := normalizedHostname(apiBase)
+	return host == "dashscope.aliyuncs.com" || strings.HasSuffix(host, ".dashscope.aliyuncs.com")
+}
+
+func normalizedHostname(rawURL string) string {
+	parsed, err := url.Parse(strings.TrimSpace(rawURL))
+	if err != nil {
+		return ""
+	}
+	return strings.ToLower(strings.TrimSpace(parsed.Hostname()))
+}
+
 func (p *Provider) applyCustomHeaders(req *http.Request) {
 	for k, v := range p.customHeaders {
 		if strings.TrimSpace(k) == "" {

pkg/providers/openai_compat/provider_test.go

@@ -62,6 +62,69 @@ func TestProviderChat_UsesMaxCompletionTokensForGLM(t *testing.T) {
 	}
 }
 
+func TestBuildRequestBody_DisablesDoubaoThinkingWhenThinkingLevelOff(t *testing.T) {
+	p := NewProvider("key", "https://ark.cn-beijing.volces.com/api/v3", "")
+	p.SetProviderName("openai")
+
+	body := p.buildRequestBody(
+		[]Message{{Role: "user", Content: "hi"}},
+		nil,
+		"doubao-seed-1-6-flash-250828",
+		map[string]any{"thinking_level": "off"},
+	)
+
+	thinking, ok := body["thinking"].(map[string]any)
+	if !ok {
+		t.Fatalf("thinking = %#v, want map", body["thinking"])
+	}
+	if got := thinking["type"]; got != "disabled" {
+		t.Fatalf("thinking.type = %#v, want %q", got, "disabled")
+	}
+}
+
+func TestBuildRequestBody_DisablesModelDependentQwenThinkingWhenThinkingLevelOff(t *testing.T) {
+	p := NewProvider("key", "https://api-inference.modelscope.cn/v1", "")
+	p.SetProviderName("modelscope")
+
+	body := p.buildRequestBody(
+		[]Message{{Role: "user", Content: "hi"}},
+		nil,
+		"qwen3-coder-plus",
+		map[string]any{"thinking_level": "off"},
+	)
+
+	if got := body["enable_thinking"]; got != false {
+		t.Fatalf("enable_thinking = %#v, want false", got)
+	}
+}
+
+func TestBuildRequestBody_PreservesDoubaoRequestWhenThinkingLevelIsNotOff(t *testing.T) {
+	p := NewProvider("key", "https://ark.cn-beijing.volces.com/api/v3", "")
+	p.SetProviderName("openai")
+
+	for _, level := range []string{"low", "adaptive", "unexpected"} {
+		t.Run(level, func(t *testing.T) {
+			body := p.buildRequestBody(
+				[]Message{{Role: "user", Content: "hi"}},
+				nil,
+				"doubao-seed-1-6-flash-250828",
+				map[string]any{"thinking_level": level},
+			)
+
+			if _, ok := body["thinking"]; ok {
+				t.Fatalf(
+					"thinking should be omitted for %q to preserve existing behavior, got %#v",
+					level,
+					body["thinking"],
+				)
+			}
+			if _, ok := body["enable_thinking"]; ok {
+				t.Fatalf("enable_thinking should be omitted for %q, got %#v", level, body["enable_thinking"])
+			}
+		})
+	}
+}
+
 func TestProviderChat_ParsesToolCalls(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		resp := map[string]any{

web/frontend/src/components/models/add-model-sheet.tsx

@@ -720,7 +720,7 @@ export function AddModelSheet({
                   <Input
                     value={form.thinkingLevel}
                     onChange={setField("thinkingLevel")}
-                    placeholder="off"
+                    placeholder={t("models.field.providerDefault")}
                   />
                 </Field>
 

web/frontend/src/components/models/edit-model-sheet.tsx

@@ -690,7 +690,7 @@ export function EditModelSheet({
                   <Input
                     value={form.thinkingLevel}
                     onChange={setField("thinkingLevel")}
-                    placeholder="off"
+                    placeholder={t("models.field.providerDefault")}
                   />
                 </Field>
 

web/frontend/src/i18n/locales/en.json

@@ -313,7 +313,8 @@
       "rpm": "Rate Limit (RPM)",
       "rpmHint": "Maximum requests per minute. 0 = no limit.",
       "thinkingLevel": "Thinking Level",
-      "thinkingLevelHint": "Extended thinking budget: off, low, medium, high, xhigh, adaptive.",
+      "thinkingLevelHint": "Leave blank to omit thinking_level and use the provider default. Values: off, low, medium, high, xhigh, adaptive.",
+      "providerDefault": "provider default",
       "maxTokensField": "Max Tokens Field",
       "maxTokensFieldHint": "Override the request field name for max tokens, e.g. max_completion_tokens.",
       "toolSchemaTransform": "Tool Schema Transform",

web/frontend/src/i18n/locales/pt-br.json

@@ -312,7 +312,8 @@
       "rpm": "Limite de Taxa (RPM)",
       "rpmHint": "Máximo de requisições por minuto. 0 = sem limite.",
       "thinkingLevel": "Nível de Pensamento",
-      "thinkingLevelHint": "Orçamento de pensamento estendido: off, low, medium, high, xhigh, adaptive.",
+      "thinkingLevelHint": "Deixe em branco para omitir thinking_level e usar o padrão do provider. Valores: off, low, medium, high, xhigh, adaptive.",
+      "providerDefault": "padrão do provider",
       "maxTokensField": "Campo de Max Tokens",
       "maxTokensFieldHint": "Sobrescreve o nome do campo de max tokens na requisição, ex: max_completion_tokens.",
       "toolSchemaTransform": "Transformação de Schema de Ferramentas",

web/frontend/src/i18n/locales/zh.json

@@ -313,7 +313,8 @@
       "rpm": "速率限制（RPM）",
       "rpmHint": "每分钟最大请求数，0 表示不限制。",
       "thinkingLevel": "思考级别",
-      "thinkingLevelHint": "扩展思考预算：off、low、medium、high、xhigh、adaptive。",
+      "thinkingLevelHint": "留空则不传 thinking_level，使用 Provider 默认值。可选值：off、low、medium、high、xhigh、adaptive。",
+      "providerDefault": "Provider 默认",
       "maxTokensField": "Max Tokens 字段名",
       "maxTokensFieldHint": "覆盖请求中 max_tokens 的字段名，例如 max_completion_tokens。",
       "toolSchemaTransform": "工具 Schema 转换",