diff --git a/pkg/providers/openai_compat/provider.go b/pkg/providers/openai_compat/provider.go
index d25a0fce4..915f209bb 100644
--- a/pkg/providers/openai_compat/provider.go
+++ b/pkg/providers/openai_compat/provider.go
@@ -316,6 +316,7 @@ func parseStreamResponse(
 
 	scanner := bufio.NewScanner(reader)
 	scanner.Buffer(make([]byte, 0, 1024*1024), 10*1024*1024) // 1MB initial, 10MB max
+	var pendingData string
 	for scanner.Scan() {
 		// Check for context cancellation between chunks
 		if err := ctx.Err(); err != nil {
@@ -332,6 +333,13 @@ func parseStreamResponse(
 			break
 		}
 
+		// Some SSE implementations can split a JSON payload across multiple data lines.
+		// Keep incomplete fragments and retry parse when the next fragment arrives.
+		payload := data
+		if pendingData != "" {
+			payload = pendingData + payload
+		}
+
 		var chunk struct {
 			Choices []struct {
 				Delta struct {
@@ -350,9 +358,17 @@ func parseStreamResponse(
 			Usage *UsageInfo `json:"usage"`
 		}
 
-		if err := json.Unmarshal([]byte(data), &chunk); err != nil {
-			continue // skip malformed chunks
+		if err := json.Unmarshal([]byte(payload), &chunk); err != nil {
+			// Keep buffer only for likely truncated JSON.
+			// For other malformed payloads, drop it to avoid poisoning later valid chunks.
+			if strings.Contains(err.Error(), "unexpected end of JSON input") {
+				pendingData = payload
+			} else {
+				pendingData = ""
+			}
+			continue
 		}
+		pendingData = ""
 
 		if chunk.Usage != nil {
 			usage = chunk.Usage
diff --git a/pkg/providers/openai_compat/provider_test.go b/pkg/providers/openai_compat/provider_test.go
index d140d63d6..df8e2e42b 100644
--- a/pkg/providers/openai_compat/provider_test.go
+++ b/pkg/providers/openai_compat/provider_test.go
@@ -2,6 +2,7 @@ package openai_compat
 
 import (
 	"bytes"
+	"context"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -16,6 +17,41 @@ import (
 	"github.com/sipeed/picoclaw/pkg/providers/protocoltypes"
 )
 
+func TestParseStreamResponse_ReassemblesSplitJSONAcrossDataLines(t *testing.T) {
+	stream := strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"hel`,
+		`data: lo"},"finish_reason":"stop"}]}`,
+		`data: [DONE]`,
+		"",
+	}, "\n")
+
+	out, err := parseStreamResponse(context.Background(), strings.NewReader(stream), nil)
+	if err != nil {
+		t.Fatalf("parseStreamResponse() error = %v", err)
+	}
+	if out.Content != "hello" {
+		t.Fatalf("Content = %q, want %q", out.Content, "hello")
+	}
+}
+
+func TestParseStreamResponse_DropsMalformedChunkAndRecovers(t *testing.T) {
+	stream := strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"bad"}}]} garbage`,
+		`data: {"choices":[{"delta":{"content":"ok"},"finish_reason":"stop"}]}`,
+		`data: [DONE]`,
+		"",
+	}, "\n")
+
+	out, err := parseStreamResponse(context.Background(), strings.NewReader(stream), nil)
+	if err != nil {
+		t.Fatalf("parseStreamResponse() error = %v", err)
+	}
+	// The malformed chunk should be skipped, and the next valid chunk should still parse.
+	if out.Content != "ok" {
+		t.Fatalf("Content = %q, want %q", out.Content, "ok")
+	}
+}
+
 func TestProviderChat_UsesMaxCompletionTokensForGLM(t *testing.T) {
 	var requestBody map[string]any