fixes azure-firework model output token capping fix

.playwright-mcp/page-2026-09-14T07-09-24-129Z.yml

@@ -0,0 +1,69 @@
+- generic [ref=e2]:
+  - navigation [ref=e3]:
+    - link "About" [ref=e4] [cursor=pointer]:
+      - /url: https://about.google/?fg=1&utm_source=google-IN&utm_medium=referral&utm_campaign=hp-header
+    - link "Store" [ref=e5] [cursor=pointer]:
+      - /url: https://store.google.com/IN?utm_source=hp_header&utm_medium=google_ooo&utm_campaign=GS100042&hl=en-IN
+    - generic [ref=e7]:
+      - generic [ref=e8]:
+        - link "Gmail" [ref=e10] [cursor=pointer]:
+          - /url: https://mail.google.com/mail/&ogbl
+        - link "Search for Images" [ref=e12] [cursor=pointer]:
+          - /url: https://www.google.com/imghp?hl=en&ogbl
+          - text: Images
+      - button "Google apps" [ref=e15] [cursor=pointer]
+      - link "Sign in" [ref=e20] [cursor=pointer]:
+        - /url: https://accounts.google.com/ServiceLogin?hl=en&passive=true&continue=https://www.google.com/&ec=futura_exp_og_so_72776762_e
+  - img "Google" [ref=e24]
+  - search [ref=e27]:
+    - generic [ref=e29]:
+      - generic [ref=e31]:
+        - button "Add files and tools" [ref=e36] [cursor=pointer]
+        - combobox "Search" [active] [ref=e41]
+        - generic [ref=e42]:
+          - generic [ref=e43]:
+            - button "Search by voice" [ref=e46] [cursor=pointer]
+            - button "Search by image" [ref=e51] [cursor=pointer]
+          - link "AI Mode" [ref=e54] [cursor=pointer]
+      - generic [ref=e68]:
+        - button "Create images" [ref=e69] [cursor=pointer]:
+          - generic [aria-hidden] [ref=e70]: 🍌
+        - button "Ask about files" [ref=e72] [cursor=pointer]
+        - button "Brainstorm" [ref=e77] [cursor=pointer]
+        - button "I am feeling lucky" [ref=e82] [cursor=pointer]
+  - generic [ref=e86]:
+    - text: "Google offered in:"
+    - link "हिन्दी" [ref=e87] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=hi&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCCk
+    - link "বাংলা" [ref=e88] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=bn&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCCo
+    - link "తెలుగు" [ref=e89] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=te&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCCs
+    - link "मराठी" [ref=e90] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=mr&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCCw
+    - link "தமிழ்" [ref=e91] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=ta&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCC0
+    - link "ગુજરાતી" [ref=e92] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=gu&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCC4
+    - link "ಕನ್ನಡ" [ref=e93] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=kn&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCC8
+    - link "മലയാളം" [ref=e94] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=ml&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCDA
+    - link "ਪੰਜਾਬੀ" [ref=e95] [cursor=pointer]:
+      - /url: https://www.google.com/setprefs?sig=0_lN96UW5bm3syzUlMoUoVzv-YH2Y%3D&hl=pa&source=homepage&sa=X&ved=0ahUKEwiSt_Towe2WAxXRbmwGHcj9CdAQ2ZgBCDE
+  - contentinfo [ref=e97]:
+    - generic [ref=e98]: India
+    - generic [ref=e99]:
+      - generic [ref=e100]:
+        - link "Advertising" [ref=e101] [cursor=pointer]:
+          - /url: https://www.google.com/intl/en_in/ads/?subid=ww-ww-et-g-awa-a-g_hpafoot1_1!o2&utm_source=google.com&utm_medium=referral&utm_campaign=google_hpafooter&fg=1
+        - link "Business" [ref=e102] [cursor=pointer]:
+          - /url: https://www.google.com/services/?subid=ww-ww-et-g-awa-a-g_hpbfoot1_1!o2&utm_source=google.com&utm_medium=referral&utm_campaign=google_hpbfooter&fg=1
+        - link "How Search works" [ref=e103] [cursor=pointer]:
+          - /url: https://google.com/search/howsearchworks/?fg=1
+      - generic [ref=e104]:
+        - link "Privacy" [ref=e105] [cursor=pointer]:
+          - /url: https://policies.google.com/privacy?hl=en-IN&fg=1
+        - link "Terms" [ref=e106] [cursor=pointer]:
+          - /url: https://policies.google.com/terms?hl=en-IN&fg=1
+        - button "Settings" [ref=e110] [cursor=pointer]

core/providers/anthropic/responses.go

@@ -4982,9 +4982,14 @@ func ToAnthropicResponsesResponse(ctx *schemas.BifrostContext, bifrostResp *sche
 		break
 	}
 
-	// Map stop reason from Bifrost response if available, otherwise infer from content
+	// Stop reason precedence: StopReason > IncompleteDetails > tool_use inference > end_turn.
 	if bifrostResp.StopReason != nil {
 		anthropicResp.StopReason = ConvertBifrostFinishReasonToAnthropic(*bifrostResp.StopReason)
+	} else if reason := anthropicStopReasonFromIncompleteDetails(bifrostResp.IncompleteDetails); reason != "" {
+		// OpenAI-shaped providers never send stop_reason; a turn truncated by the
+		// output cap or a content filter carries only incomplete_details. Reporting
+		// end_turn there hides the truncation from the client (#6782).
+		anthropicResp.StopReason = reason
 	} else {
 		anthropicResp.StopReason = AnthropicStopReasonEndTurn
 		for _, block := range contentBlocks {

core/providers/anthropic/responses_test.go

@@ -1,6 +1,7 @@
 package anthropic
 
 import (
+	"context"
 	"testing"
 	"time"
 
@@ -468,3 +469,80 @@ func TestAnthropicIngressLiftsServerSideToolOptIn(t *testing.T) {
 			bifrostReq.Params.IncludeServerSideToolInvocations)
 	}
 }
+
+// A non-streaming Responses turn cut short by the output-token cap arrives from
+// OpenAI-shaped providers (Azure, OpenAI, chat-completions fallbacks) with
+// status "incomplete" and incomplete_details.reason set, but no stop_reason:
+// that field is Anthropic/Bedrock-only. The Anthropic egress must derive
+// stop_reason from incomplete_details, never report end_turn for a truncated
+// turn (#6782). Mirrors the streaming precedence StopReason > IncompleteDetails
+// > tool_use inference > end_turn.
+func TestToAnthropicResponsesResponse_IncompleteReportsTruncationStopReason(t *testing.T) {
+	cases := []struct {
+		name       string
+		status     string
+		stopReason *string
+		incomplete *schemas.ResponsesResponseIncompleteDetails
+		want       AnthropicStopReason
+	}{
+		{
+			name:       "StopReasonLength",
+			status:     schemas.ResponsesResponseStatusIncomplete,
+			stopReason: schemas.Ptr("length"),
+			incomplete: &schemas.ResponsesResponseIncompleteDetails{Reason: schemas.ResponsesResponseIncompleteReasonMaxOutputTokens},
+			want:       AnthropicStopReasonMaxTokens,
+		},
+		{
+			// The reported shape: Azure /openai/v1/responses sets no stop_reason.
+			name:       "MaxTokensFromIncompleteDetailsOnly",
+			status:     schemas.ResponsesResponseStatusIncomplete,
+			incomplete: &schemas.ResponsesResponseIncompleteDetails{Reason: schemas.ResponsesResponseIncompleteReasonMaxOutputTokens},
+			want:       AnthropicStopReasonMaxTokens,
+		},
+		{
+			name:       "ContentFilterFromIncompleteDetailsOnly",
+			status:     schemas.ResponsesResponseStatusIncomplete,
+			incomplete: &schemas.ResponsesResponseIncompleteDetails{Reason: schemas.ResponsesResponseIncompleteReasonContentFilter},
+			want:       AnthropicStopReasonRefusal,
+		},
+		{
+			// Control: a completed text turn with neither field keeps end_turn.
+			name:   "CompletedTextIsEndTurn",
+			status: schemas.ResponsesResponseStatusCompleted,
+			want:   AnthropicStopReasonEndTurn,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			ctx := schemas.NewBifrostContext(context.Background(), time.Time{})
+			resp := ToAnthropicResponsesResponse(ctx, &schemas.BifrostResponsesResponse{
+				ID:                schemas.Ptr("resp_1"),
+				Model:             "azure-glm-5.2",
+				Status:            schemas.Ptr(tc.status),
+				StopReason:        tc.stopReason,
+				IncompleteDetails: tc.incomplete,
+				Output: []schemas.ResponsesMessage{{
+					Type: schemas.Ptr(schemas.ResponsesMessageTypeMessage),
+					Role: schemas.Ptr(schemas.ResponsesInputMessageRoleAssistant),
+					Content: &schemas.ResponsesMessageContent{
+						ContentBlocks: []schemas.ResponsesMessageContentBlock{{
+							Type: schemas.ResponsesOutputMessageContentTypeText,
+							Text: schemas.Ptr("1\n2\n3"),
+						}},
+					},
+				}},
+				Usage: &schemas.ResponsesResponseUsage{InputTokens: 55974, OutputTokens: 4096, TotalTokens: 60070},
+			})
+			if resp == nil {
+				t.Fatal("ToAnthropicResponsesResponse returned nil")
+			}
+			if resp.StopReason != tc.want {
+				t.Errorf("stop_reason = %q, want %q", resp.StopReason, tc.want)
+			}
+			if resp.Usage == nil || resp.Usage.OutputTokens != 4096 {
+				t.Errorf("usage.output_tokens not carried: %+v", resp.Usage)
+			}
+		})
+	}
+}

core/providers/azure/azure.go

@@ -537,6 +537,20 @@ func (provider *AzureProvider) Responses(ctx *schemas.BifrostContext, key schema
 		)
 	}
 
+	// Models whose datasheet row lists supported_endpoints without /v1/responses
+	// (Fireworks-hosted Foundry models, which Microsoft types as chat-completions
+	// models) are served through chat completions: their /openai/v1/responses
+	// route caps output at 4096 tokens regardless of max_output_tokens (#6782).
+	// A silent or missing row keeps the Responses route, like Bedrock Mantle.
+	canonicalModel := schemas.ResolveCanonicalModel(ctx, request.Model)
+	if !schemas.ResolveModelCaps(provider.GetProviderKey(), canonicalModel).SupportsResponsesEndpoint(true) {
+		chatResponse, bifrostErr := provider.ChatCompletion(ctx, key, request.ToChatRequest())
+		if bifrostErr != nil {
+			return nil, bifrostErr
+		}
+		return chatResponse.ToBifrostResponsesResponse(), nil
+	}
+
 	// OpenAI-family models use the OpenAI-compatible Azure endpoint via the shared handler.
 	authHeader, bifrostErr := provider.getAzureAuthHeaders(ctx, key, false)
 	if bifrostErr != nil {
@@ -610,6 +624,14 @@ func (provider *AzureProvider) ResponsesStream(ctx *schemas.BifrostContext, post
 			postHookSpanFinalizer,
 		)
 	} else {
+		// Same datasheet gate as Responses; the shared chat streaming handler
+		// re-assembles Responses events from the chat chunks when this flag is set.
+		canonicalModel := schemas.ResolveCanonicalModel(ctx, request.Model)
+		if !schemas.ResolveModelCaps(provider.GetProviderKey(), canonicalModel).SupportsResponsesEndpoint(true) {
+			ctx.SetValue(schemas.BifrostContextKeyIsResponsesToChatCompletionFallback, true)
+			return provider.ChatCompletionStream(ctx, postHookRunner, postHookSpanFinalizer, key, request.ToChatRequest())
+		}
+
 		authHeader, err := provider.getAzureAuthHeaders(ctx, key, false)
 		if err != nil {
 			return nil, err