fix: restore WorkBuddy Deepseek V4.1 Flash thinking

CHANGELOG.md

@@ -7,6 +7,18 @@ Write each change in both `### English` and `### 中文` under `## Unreleased`.
 
 ### English
 
+- Send WorkBuddy Deepseek V4.1 Flash thinking as official top-level reasoning fields so streamed thinking comes back
+- Show WorkBuddy catalog context budgets (default and optional window) on the model list without inventing a Trae Max switch
+
+### 中文
+
+- WorkBuddy 的 Deepseek V4.1 Flash 改为发送官方顶层思考字段，流式思考内容可以返回
+- 模型列表展示 WorkBuddy 目录里的默认和可选上下文窗口，不套用 Trae 的更大上下文开关
+
+## 0.4.8 - 2026-09-11
+
+### English
+
 - Preserve cached model routes when the dynamic model catalog is temporarily unavailable, while refreshing genuine model misses
 - Let WorkBuddy accounts choose their automatic daily check-in time during creation, with a 21:00 retry for failed attempts
 - Enlarge the account-name input in the create-account wizard

frontend/src/components/ModelDetailsModal.tsx

@@ -41,6 +41,7 @@ export function ModelDetailsModal({ model, t, onClose }: Props) {
                 <div className="text-xs font-medium text-muted">{t('contextWindowCol')}</div>
                 <div className="mt-1 text-sm">{formatTokens(windowDev)}{windowMax && windowMax !== windowDev ? ` → ${formatTokens(windowMax)}` : ''}</div>
                 {model.supports_max_mode ? <p className="mt-1 text-[11px] leading-5 text-muted">{t('maxModeHint')}</p> : null}
+                {!model.supports_max_mode && windowMax && windowMax !== windowDev ? <p className="mt-1 text-[11px] leading-5 text-muted">{t('workbuddyContextHint')}</p> : null}
                 {model.prompt_max_tokens ? <div className="mt-1 text-[11px] text-muted">{t('promptMaxTokens')}: {formatTokens(model.prompt_max_tokens)}</div> : null}
                 {model.max_output_tokens ? <div className="text-[11px] text-muted">{t('maxOutputTokens')}: {formatTokens(model.max_output_tokens)}</div> : null}
               </section>

frontend/src/i18n/messages.ts

@@ -413,6 +413,7 @@ export const messages: Record<Lang, Dict> = {
     maxModeHint: 'Trae larger context window, often 200k to 1M. Not a plan or quota. Used only when the request omits is_max_mode; an explicit request still wins. Unsupported models never send this switch.',
     catalogWindow: 'Catalog size',
     catalogWindowHint: 'This Trae model has no Max-context switch. The number is the catalog window.',
+    workbuddyContextHint: 'WorkBuddy desktop default and optional context budgets from the catalog. This is a local compact budget, not Trae is_max_mode. Chat requests do not send a window switch.',
     reasoningLevels: 'Reasoning intensity',
     reasoningHint: 'Default thinking effort when the request omits reasoning_effort and related flags. This does not lock every call to that level, and it does not cap a higher client value. A level the model does not allow falls back to an allowed option.',
     reasoningFixed: 'Fixed thinking: {type}',
@@ -1021,6 +1022,7 @@ export const messages: Record<Lang, Dict> = {
     maxModeHint: 'Trae 的更大上下文窗口，常见是 200k 提到 1M，不是套餐也不是额度。只在请求没带 is_max_mode 时使用；请求里写了仍以请求为准。模型不支持时不会发送这个开关。',
     catalogWindow: '目录默认窗口',
     catalogWindowHint: '这个 Trae 模型没有更大上下文开关，数字是目录里的默认窗口。',
+    workbuddyContextHint: 'WorkBuddy 桌面端目录里的默认窗口和可选窗口，用于本地压缩预算，不是 Trae 的 is_max_mode。聊天请求不会发送窗口开关。',
     reasoningLevels: '推理强度',
     reasoningHint: '请求没指定推理强度时使用的默认档位。选了不会强制每次都走这一档，也不会限制客户端改用其他允许档位。不在列表里的值会回退到可用档位，而不是锁死。',
     reasoningFixed: '固定思考：{type}',

frontend/src/pages/ProvidersPage.tsx

@@ -94,7 +94,15 @@ function ModelContextControls({
   }
   return (
     <div className="flex min-w-0 flex-wrap items-center gap-3">
-      <span className="mono text-xs text-muted">{formatTokens(model.catalog_context_length || model.context_length)}</span>
+      <span className="mono text-xs text-muted">
+        {formatTokens(model.catalog_context_length || model.context_length)}
+        {model.catalog_context_length_max && model.catalog_context_length_max !== (model.catalog_context_length || model.context_length)
+          ? ` → ${formatTokens(model.catalog_context_length_max)}`
+          : ''}
+      </span>
+      {provider === 'workbuddy' && model.catalog_context_length_max && model.catalog_context_length_max !== (model.catalog_context_length || model.context_length) ? (
+        <HintLabel label={t('catalogWindow')} hint={t('workbuddyContextHint')} />
+      ) : null}
       {provider === 'trae' && model.supports_max_mode ? (
         <div className="flex items-center gap-2">
           <CompactSwitch

internal/api/modelsettings.go

@@ -72,7 +72,7 @@ func (s *Server) decorateProviderSettings(ctx context.Context, item map[string]a
 	dev, _ := asInt(item["catalog_context_length"])
 	max, _ := asInt(item["catalog_context_length_max"])
 	supportsMax, _ := item["supports_max_mode"].(bool)
-	if !supportsMax && max > 0 && max != dev {
+	if provider == "trae" && !supportsMax && max > 0 && max != dev {
 		supportsMax = true
 		item["supports_max_mode"] = true
 	}

internal/providers/workbuddy/catalog.go

@@ -15,17 +15,23 @@ type catalogReasoning struct {
 	Summary            string   `json:"summary"`
 }
 
+type catalogContextWindow struct {
+	DefaultLength    int   `json:"defaultLength"`
+	SupportedLengths []int `json:"supportedLengths"`
+}
+
 type catalogModelEntry struct {
-	ID                string           `json:"id"`
-	Name              string           `json:"name"`
-	MaxInputTokens    int              `json:"maxInputTokens"`
-	MaxOutputTokens   int              `json:"maxOutputTokens"`
-	Disabled          bool             `json:"disabled"`
-	OnlyReasoning     bool             `json:"onlyReasoning"`
-	SupportsReasoning bool             `json:"supportsReasoning"`
-	SupportsImages    bool             `json:"supportsImages"`
-	SupportsToolCall  bool             `json:"supportsToolCall"`
-	Reasoning         catalogReasoning `json:"reasoning"`
+	ID                string               `json:"id"`
+	Name              string               `json:"name"`
+	MaxInputTokens    int                  `json:"maxInputTokens"`
+	MaxOutputTokens   int                  `json:"maxOutputTokens"`
+	Disabled          bool                 `json:"disabled"`
+	OnlyReasoning     bool                 `json:"onlyReasoning"`
+	SupportsReasoning bool                 `json:"supportsReasoning"`
+	SupportsImages    bool                 `json:"supportsImages"`
+	SupportsToolCall  bool                 `json:"supportsToolCall"`
+	ContextWindow     catalogContextWindow `json:"contextWindow"`
+	Reasoning         catalogReasoning     `json:"reasoning"`
 }
 
 func catalogModel(model catalogModelEntry) providers.ModelInfo {
@@ -49,12 +55,14 @@ func catalogModel(model catalogModelEntry) providers.ModelInfo {
 	if canDisable && (len(options) > 0 || model.SupportsReasoning) && !containsLevel(options, "none") {
 		options = append([]string{"none"}, options...)
 	}
+	window, windowMax := contextWindows(model)
 	return providers.ModelInfo{
 		NativeModel: model.ID,
 		PublicModel: model.ID,
 		DisplayName: model.Name,
 		Capabilities: providers.ModelCapabilities{
-			ContextWindow:      model.MaxInputTokens,
+			ContextWindow:      window,
+			ContextWindowMax:   windowMax,
 			MaxOutput:          model.MaxOutputTokens,
 			Tools:              true,
 			Images:             model.SupportsImages,
@@ -66,6 +74,54 @@ func catalogModel(model catalogModelEntry) providers.ModelInfo {
 	}
 }
 
+func contextWindows(model catalogModelEntry) (int, int) {
+	cap := model.MaxInputTokens
+	seen := map[int]struct{}{}
+	var lengths []int
+	for _, value := range model.ContextWindow.SupportedLengths {
+		if value <= 0 || (cap > 0 && value > cap) {
+			continue
+		}
+		if _, dup := seen[value]; dup {
+			continue
+		}
+		seen[value] = struct{}{}
+		lengths = append(lengths, value)
+	}
+	for i := 0; i < len(lengths); i++ {
+		for j := i + 1; j < len(lengths); j++ {
+			if lengths[j] < lengths[i] {
+				lengths[i], lengths[j] = lengths[j], lengths[i]
+			}
+		}
+	}
+	dev := model.ContextWindow.DefaultLength
+	if dev <= 0 || (cap > 0 && dev > cap) || (len(lengths) > 0 && !containsWindow(lengths, dev)) {
+		if len(lengths) > 0 {
+			dev = lengths[0]
+		} else {
+			dev = cap
+		}
+	}
+	max := cap
+	if len(lengths) > 0 {
+		max = lengths[len(lengths)-1]
+	}
+	if max == dev {
+		max = 0
+	}
+	return dev, max
+}
+
+func containsWindow(values []int, want int) bool {
+	for _, value := range values {
+		if value == want {
+			return true
+		}
+	}
+	return false
+}
+
 func containsLevel(options []string, level string) bool {
 	for _, option := range options {
 		if option == level {

internal/providers/workbuddy/client_test.go

@@ -306,8 +306,15 @@ func TestModelsParsesReasoningOptions(t *testing.T) {
 					"onlyReasoning": true, "supportsReasoning": true,
 					"reasoning": map[string]any{"effort": "medium"},
 				},
+				{
+					"id": "deepseek-v4.1-flash", "name": "Deepseek-V4.1-Flash",
+					"maxInputTokens": 1000000, "maxOutputTokens": 128000,
+					"onlyReasoning": true, "supportsReasoning": true,
+					"contextWindow": map[string]any{"defaultLength": 300000, "supportedLengths": []int{300000, 1000000}},
+					"reasoning":     map[string]any{"effort": "high", "summary": "auto"},
+				},
 			},
-			"agents": []map[string]any{{"name": "cli", "models": []string{"glm-5.3", "glm-5.2"}}},
+			"agents": []map[string]any{{"name": "cli", "models": []string{"glm-5.3", "glm-5.2", "deepseek-v4.1-flash"}}},
 		}})
 	}))
 	defer server.Close()
@@ -318,7 +325,7 @@ func TestModelsParsesReasoningOptions(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(models) != 2 {
+	if len(models) != 3 {
 		t.Fatalf("models=%+v", models)
 	}
 	glm53 := models[0]
@@ -332,6 +339,13 @@ func TestModelsParsesReasoningOptions(t *testing.T) {
 	if glm52.Capabilities.CanDisableThinking || glm52.Capabilities.ReasoningDefault != "medium" || strings.Join(glm52.Capabilities.ReasoningOptions, ",") != "medium" {
 		t.Fatalf("glm-5.2 caps=%+v", glm52.Capabilities)
 	}
+	if glm52.Capabilities.ContextWindow != 1000000 || glm52.Capabilities.ContextWindowMax != 0 {
+		t.Fatalf("glm-5.2 window=%+v", glm52.Capabilities)
+	}
+	ds := models[2]
+	if ds.NativeModel != "deepseek-v4.1-flash" || ds.Capabilities.ContextWindow != 300000 || ds.Capabilities.ContextWindowMax != 1000000 || ds.Capabilities.MaxMode {
+		t.Fatalf("deepseek window=%+v", ds.Capabilities)
+	}
 }
 
 func TestChatRequestSendsCatalogReasoningEffort(t *testing.T) {
@@ -369,6 +383,43 @@ func TestChatRequestSendsCatalogReasoningEffort(t *testing.T) {
 	}
 }
 
+func TestChatRequestSendsOfficialReasoningFieldsForDeepseekFlash(t *testing.T) {
+	payload, _ := Credential{AccessToken: "at", UID: "u1", Domain: "codebuddy.cn", ExpiresAt: 4102444800}.Encode()
+	store := &memStore{items: map[string][]byte{"acc1": payload}}
+	var got map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == pathModelsCN {
+			_ = json.NewEncoder(w).Encode(map[string]any{"code": 0, "data": map[string]any{
+				"models": []map[string]any{{
+					"id": "deepseek-v4.1-flash", "name": "Deepseek V4.1 Flash",
+					"onlyReasoning": true, "supportsReasoning": true,
+					"reasoning": map[string]any{"effort": "high", "summary": "auto"},
+				}},
+				"agents": []map[string]any{{"name": "cli", "models": []string{"deepseek-v4.1-flash"}}},
+			}})
+			return
+		}
+		_ = json.NewDecoder(r.Body).Decode(&got)
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(chatSSE))
+	}))
+	defer server.Close()
+	client := NewClient(store)
+	client.http = server.Client()
+	client.http.Transport = rewriteTransport{server: server.URL, round: server.Client().Transport}
+	if _, err := client.ChatNonStream(context.Background(), "acc1", translate.ChatRequest{
+		Model: "deepseek-v4.1-flash", Messages: []translate.ChatMessage{{Role: "user", Content: "hi"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := got["reasoning"]; ok {
+		t.Fatalf("deepseek kept nested reasoning: %v", got["reasoning"])
+	}
+	if got["reasoning_effort"] != "high" || got["reasoning_summary"] != "auto" || got["verbosity"] != "high" {
+		t.Fatalf("deepseek fields=%v", got)
+	}
+}
+
 // Regression: the console stores provider settings under the canonical model
 // key (lowercase, _ folded to -); chat-time lookups must use the same form.
 func TestChatRequestFindsStoredReasoningByCanonicalKey(t *testing.T) {

internal/providers/workbuddy/mapping.go

@@ -2,11 +2,20 @@ package workbuddy
 
 import (
 	"encoding/json"
+	"strings"
 
+	"github.com/caigee-cmd/cli2api/internal/accounts"
 	"github.com/caigee-cmd/cli2api/internal/providers"
 	"github.com/caigee-cmd/cli2api/internal/translate"
 )
 
+// officialTopLevelReasoningModels need the 2.132.0 CLI field names. The
+// nested reasoning object that GLM / Hy4 still accept returns no thinking
+// tokens for these ids.
+var officialTopLevelReasoningModels = map[string]struct{}{
+	"deepseek-v4.1-flash": {},
+}
+
 func requestedReasoningLevel(req translate.ChatRequest) string {
 	if len(req.ReasoningEffort) > 0 {
 		var value any
@@ -58,23 +67,49 @@ func applyChatReasoning(obj map[string]any, req translate.ChatRequest, storedLev
 	}
 	level = providers.ResolveReasoningLevel(level, caps)
 	if level == "" {
-		delete(obj, "reasoning")
+		clearChatReasoning(obj)
 		return
 	}
 	if level == "none" {
 		if !caps.CanDisableThinking {
 			level = providers.ResolveReasoningLevel(caps.ReasoningDefault, caps)
 			if level == "" || level == "none" {
-				delete(obj, "reasoning")
+				clearChatReasoning(obj)
 				return
 			}
 		} else {
-			delete(obj, "reasoning")
+			clearChatReasoning(obj)
 			return
 		}
 	}
 	if level == "max" {
 		level = "xhigh"
 	}
+	if usesTopLevelReasoningFields(req.Model) {
+		delete(obj, "reasoning")
+		obj["reasoning_effort"] = level
+		obj["reasoning_summary"] = "auto"
+		obj["verbosity"] = "high"
+		return
+	}
+	delete(obj, "reasoning_effort")
+	delete(obj, "reasoning_summary")
+	delete(obj, "verbosity")
 	obj["reasoning"] = map[string]any{"effort": level, "summary": "auto"}
 }
+
+func clearChatReasoning(obj map[string]any) {
+	delete(obj, "reasoning")
+	delete(obj, "reasoning_effort")
+	delete(obj, "reasoning_summary")
+	delete(obj, "verbosity")
+}
+
+func usesTopLevelReasoningFields(model string) bool {
+	key := accounts.CanonicalModelID(model)
+	if i := strings.LastIndex(key, "/"); i >= 0 {
+		key = key[i+1:]
+	}
+	_, ok := officialTopLevelReasoningModels[key]
+	return ok
+}

internal/providers/workbuddy/mapping_test.go

@@ -0,0 +1,55 @@
+package workbuddy
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/caigee-cmd/cli2api/internal/providers"
+	"github.com/caigee-cmd/cli2api/internal/translate"
+)
+
+func TestApplyChatReasoningKeepsNestedObjectForGLM(t *testing.T) {
+	obj := map[string]any{}
+	applyChatReasoning(obj, translate.ChatRequest{Model: "glm-5.3", ReasoningEffort: json.RawMessage(`"high"`)}, "", providers.ModelCapabilities{
+		ReasoningOptions: []string{"none", "low", "high", "xhigh"},
+		ReasoningDefault: "high",
+	})
+	reasoning, _ := obj["reasoning"].(map[string]any)
+	if reasoning["effort"] != "high" || reasoning["summary"] != "auto" {
+		t.Fatalf("glm reasoning=%v", obj["reasoning"])
+	}
+	for _, key := range []string{"reasoning_effort", "reasoning_summary", "verbosity"} {
+		if _, ok := obj[key]; ok {
+			t.Fatalf("glm unexpectedly set %s=%v", key, obj[key])
+		}
+	}
+}
+
+func TestApplyChatReasoningUsesOfficialFieldsForDeepseekFlash(t *testing.T) {
+	caps := providers.ModelCapabilities{
+		ReasoningOptions:   []string{"high"},
+		ReasoningDefault:   "high",
+		CanDisableThinking: false,
+	}
+	for _, model := range []string{"deepseek-v4.1-flash", "DeepSeek_V4.1_Flash", "workbuddy/deepseek-v4.1-flash"} {
+		obj := map[string]any{"reasoning": map[string]any{"effort": "stale"}}
+		applyChatReasoning(obj, translate.ChatRequest{Model: model}, "", caps)
+		if _, ok := obj["reasoning"]; ok {
+			t.Fatalf("%s kept nested reasoning: %v", model, obj["reasoning"])
+		}
+		if obj["reasoning_effort"] != "high" || obj["reasoning_summary"] != "auto" || obj["verbosity"] != "high" {
+			t.Fatalf("%s official fields=%v", model, obj)
+		}
+	}
+}
+
+func TestApplyChatReasoningDeepseekUsesCatalogDefault(t *testing.T) {
+	obj := map[string]any{}
+	applyChatReasoning(obj, translate.ChatRequest{Model: "deepseek-v4.1-flash"}, "", providers.ModelCapabilities{
+		ReasoningOptions: []string{"high"},
+		ReasoningDefault: "high",
+	})
+	if obj["reasoning_effort"] != "high" {
+		t.Fatalf("default effort=%v", obj["reasoning_effort"])
+	}
+}

internal/webui/static/index.html

@@ -22,7 +22,7 @@
     <link rel="preconnect" href="https://fonts.googleapis.com" />
     <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
     <link href="https://fonts.googleapis.com/css2?family=IBM+Plex+Mono:wght@400;500&family=Outfit:wght@400;500;600;700&display=swap" rel="stylesheet" />
-    <script type="module" crossorigin src="/assets/index-sgsO79xb.js"></script>
+    <script type="module" crossorigin src="/assets/index-DyfuUCZL.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-DTnXvv8L.css">
   </head>
   <body>