fix(workbuddy): stop rewriting deepseek-v4.1-flash to deep-model

internal/providers/workbuddy/client.go

@@ -417,14 +417,6 @@ func (c *Client) Models(ctx context.Context, accountID string) ([]providers.Mode
 }
 
 func (c *Client) chatRequest(ctx context.Context, accountID string, credential Credential, req translate.ChatRequest) (*http.Request, error) {
-	body := map[string]any{
-		"model":       upstreamModelID(req.Model),
-		"messages":    req.Messages,
-		"max_tokens":  req.MaxTokens,
-		"temperature": req.Temperature,
-		"tools":       req.Tools,
-		"tool_choice": req.ToolChoice,
-	}
 	caps := c.capsFor(req.Model)
 	storedLevel := ""
 	if setter, ok := c.store.(interface {
@@ -436,10 +428,21 @@ func (c *Client) chatRequest(ctx context.Context, accountID string, credential C
 			storedLevel = stored.ReasoningEffort
 		}
 	}
-	if len(caps.ReasoningOptions) == 0 && accountID != "" {
+	// Resolve the upstream model after the live catalog is warm. WorkBuddy has
+	// renamed deepseek-v4.1-flash from the old deep-model alias; sending the
+	// stale alias lets upstream silently fall back to another model.
+	if (!c.hasCatalogEntry(req.Model) || len(caps.ReasoningOptions) == 0) && accountID != "" {
 		_, _ = c.Models(ctx, accountID)
 		caps = c.capsFor(req.Model)
 	}
+	body := map[string]any{
+		"model":       c.upstreamModelID(req.Model),
+		"messages":    req.Messages,
+		"max_tokens":  req.MaxTokens,
+		"temperature": req.Temperature,
+		"tools":       req.Tools,
+		"tool_choice": req.ToolChoice,
+	}
 	applyChatReasoning(body, req, storedLevel, caps)
 	payload, err := json.Marshal(body)
 	if err != nil {
@@ -1012,11 +1015,39 @@ var workbuddyModelAliases = map[string]string{
 	"deepseek-v4.1-flash": "deep-model",
 }
 
-func upstreamModelID(model string) string {
+func (c *Client) hasCatalogEntry(model string) bool {
+	model = strings.TrimSpace(model)
+	if model == "" || c == nil {
+		return false
+	}
 	canonical := accounts.CanonicalModelID(model)
-	for alias, nativeModel := range workbuddyModelAliases {
-		if accounts.CanonicalModelID(alias) == canonical {
-			return nativeModel
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if _, ok := c.catalog[model]; ok {
+		return true
+	}
+	_, ok := c.catalog[canonical]
+	return ok
+}
+
+// upstreamModelID returns the native ID to send upstream. For current
+// WorkBuddy catalogs that already expose deepseek-v4.1-flash natively, that
+// ID is used as-is. Older catalogs that only expose deep-model still work
+// because appendAliasModels indexes the public alias onto a ModelInfo whose
+// NativeModel is deep-model. There is intentionally no hardcoded rewrite
+// here: blindly mapping to deep-model on a cold/missing catalog is what made
+// production silently fall back to kimi/glm.
+func (c *Client) upstreamModelID(model string) string {
+	canonical := accounts.CanonicalModelID(model)
+	if c != nil {
+		c.mu.Lock()
+		info, ok := c.catalog[model]
+		if !ok {
+			info, ok = c.catalog[canonical]
+		}
+		c.mu.Unlock()
+		if ok && strings.TrimSpace(info.NativeModel) != "" {
+			return info.NativeModel
 		}
 	}
 	return model

internal/providers/workbuddy/client_test.go

@@ -526,6 +526,9 @@ func TestChatRequestSendsOfficialReasoningFieldsForDeepseekFlash(t *testing.T) {
 	}); err != nil {
 		t.Fatal(err)
 	}
+	if got["model"] != "deepseek-v4.1-flash" {
+		t.Fatalf("upstream model=%v body=%v", got["model"], got)
+	}
 	if _, ok := got["reasoning"]; ok {
 		t.Fatalf("deepseek kept nested reasoning: %v", got["reasoning"])
 	}
@@ -679,6 +682,68 @@ func TestChatRequestMapsDeepseekAliasToNativeModel(t *testing.T) {
 	}
 }
 
+// Regression: once WorkBuddy publishes deepseek-v4.1-flash as a native CLI
+// model, chat must send that ID instead of the stale deep-model alias.
+func TestChatRequestKeepsNativeDeepseekWhenCatalogHasIt(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", "maxInputTokens": 1000000,
+					"supportsReasoning": true, "onlyReasoning": true,
+					"reasoning": map[string]any{"defaultEffort": "high", "supportedEfforts": []string{"high"}},
+				}},
+				"agents": []map[string]any{{"name": "cli", "models": []string{"deepseek-v4.1-flash"}}},
+			}})
+			return
+		}
+		if err := json.NewDecoder(r.Body).Decode(&got); err != nil {
+			t.Fatalf("decode chat body: %v", err)
+		}
+		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 got["model"] != "deepseek-v4.1-flash" {
+		t.Fatalf("upstream model=%v body=%v", got["model"], got)
+	}
+}
+
+func TestUpstreamModelIDDoesNotHardcodeStaleDeepModel(t *testing.T) {
+	client := NewClient(&memStore{items: map[string][]byte{}})
+	if got := client.upstreamModelID("deepseek-v4.1-flash"); got != "deepseek-v4.1-flash" {
+		t.Fatalf("cold catalog rewrite = %q", got)
+	}
+	client.rememberCatalog([]providers.ModelInfo{{
+		NativeModel: "deepseek-v4.1-flash",
+		PublicModel: "deepseek-v4.1-flash",
+		DisplayName: "Deepseek-V4.1-Flash",
+	}})
+	if got := client.upstreamModelID("deepseek-v4.1-flash"); got != "deepseek-v4.1-flash" {
+		t.Fatalf("native catalog rewrite = %q", got)
+	}
+	client.rememberCatalog([]providers.ModelInfo{{
+		NativeModel: "deep-model",
+		PublicModel: "deepseek-v4.1-flash",
+		DisplayName: "Deepseek-V4.1-Flash",
+	}})
+	if got := client.upstreamModelID("deepseek-v4.1-flash"); got != "deep-model" {
+		t.Fatalf("legacy alias rewrite = %q", got)
+	}
+}
+
 func TestModelsAcceptsGlobalCLIAgentNamesAndUsesAccountRegion(t *testing.T) {
 	payload, _ := Credential{AccessToken: "at", UID: "u1", Domain: "codebuddy.cn", ExpiresAt: 4102444800}.Encode()
 	store := &memStore{items: map[string][]byte{"acc1": payload}, region: "global"}