fix: address 3 Anthropic-specific schema-compatibility bugs

core/providers/anthropic/passthrough_usage_test.go

@@ -74,9 +74,9 @@ func TestExtractAnthropicPassthroughUsage(t *testing.T) {
 			},
 		},
 		{
-			name:  "messages zero usage -> nil",
-			path:  "/v1/messages",
-			body:  `{"usage":{"input_tokens":0,"output_tokens":0}}`,
+			name: "messages zero usage -> nil",
+			path: "/v1/messages",
+			body: `{"usage":{"input_tokens":0,"output_tokens":0}}`,
 			check: func(t *testing.T, u *schemas.BifrostPassthroughUsage) {
 				if u != nil {
 					t.Fatalf("expected nil, got %+v", u)

core/providers/anthropic/reasoningtoolcall_test.go

@@ -0,0 +1,90 @@
+package anthropic
+
+import (
+	"testing"
+	"time"
+
+	"github.com/bytedance/sonic"
+	"github.com/maximhq/bifrost/core/schemas"
+)
+
+// rawExtendedThinkingToolCallResponse mirrors issue #3802: an Anthropic Messages API
+// response with extended thinking enabled, where the assistant turn contains both a
+// thinking block and a tool_use block. The bug report claimed reasoning_content is
+// dropped when this turn is routed onward to a custom Anthropic-base provider (Kimi)
+// that requires it to be echoed back on the next turn — reproduced here by chaining
+// Bifrost's real Anthropic ingest conversion into the real Responses->Chat bridge that
+// such a custom OpenAI-compatible provider would consume.
+const rawExtendedThinkingToolCallResponse = `{
+  "model": "claude-opus-4-8",
+  "id": "msg_01ExtendedThinking",
+  "type": "message",
+  "role": "assistant",
+  "content": [
+    { "type": "thinking", "thinking": "I should look up the current weather before answering.", "signature": "sig_abc123" },
+    { "type": "tool_use", "id": "toolu_01Weather", "name": "get_weather", "input": {"location": "SF"} }
+  ],
+  "stop_reason": "tool_use",
+  "usage": { "input_tokens": 120, "output_tokens": 45 }
+}`
+
+// TestExtendedThinkingToolCallTurn_ReasoningSurvivesResponsesToChatBridge is the
+// end-to-end regression test for #3802. It reproduces the full reported pipeline:
+// Anthropic response (thinking + tool_use in the same turn) -> Bifrost's Responses-shaped
+// intermediate (ToBifrostResponsesResponse) -> the Responses->Chat fallback bridge
+// (schemas.ToChatMessages) that a custom Anthropic-base provider like Kimi would receive
+// its history through. If reasoning is dropped anywhere in this chain, the assistant
+// tool-call message that reaches Kimi on the next turn would be missing reasoning_content.
+func TestExtendedThinkingToolCallTurn_ReasoningSurvivesResponsesToChatBridge(t *testing.T) {
+	var resp AnthropicMessageResponse
+	if err := sonic.Unmarshal([]byte(rawExtendedThinkingToolCallResponse), &resp); err != nil {
+		t.Fatalf("unmarshal raw: %v", err)
+	}
+
+	ctx := schemas.NewBifrostContext(nil, time.Time{})
+
+	bifrostResp := resp.ToBifrostResponsesResponse(ctx)
+	if bifrostResp == nil {
+		t.Fatal("ToBifrostResponsesResponse returned nil")
+	}
+
+	var sawReasoning, sawFunctionCall bool
+	for _, out := range bifrostResp.Output {
+		if out.Type == nil {
+			continue
+		}
+		switch *out.Type {
+		case schemas.ResponsesMessageTypeReasoning:
+			sawReasoning = true
+		case schemas.ResponsesMessageTypeFunctionCall:
+			sawFunctionCall = true
+		}
+	}
+	if !sawReasoning {
+		t.Fatal("expected a reasoning output item from the Anthropic thinking block, got none")
+	}
+	if !sawFunctionCall {
+		t.Fatal("expected a function_call output item from the Anthropic tool_use block, got none")
+	}
+
+	// Feed the Responses-shaped output through the same Responses->Chat bridge a
+	// custom OpenAI-compatible provider (Kimi) would use to build its wire history.
+	chatMessages := schemas.ToChatMessages(bifrostResp.Output)
+
+	var found bool
+	for _, cm := range chatMessages {
+		if cm.ChatAssistantMessage == nil || len(cm.ChatAssistantMessage.ToolCalls) == 0 {
+			continue
+		}
+		found = true
+		if cm.ChatAssistantMessage.Reasoning == nil {
+			t.Fatal("reasoning_content dropped on the assistant tool-call turn (#3802) — Reasoning is nil")
+		}
+		if *cm.ChatAssistantMessage.Reasoning == "" {
+			t.Fatal("reasoning_content dropped on the assistant tool-call turn (#3802) — Reasoning is empty")
+		}
+	}
+	if !found {
+		t.Fatal("expected an assistant message with tool calls in the bridged chat messages")
+	}
+}