diff --git a/mateclaw-server/src/main/java/vip/mate/llm/chatmodel/OpenAiCompatibleHttpInterceptor.java b/mateclaw-server/src/main/java/vip/mate/llm/chatmodel/OpenAiCompatibleHttpInterceptor.java
index fef8c8c9..4c8f6c25 100644
--- a/mateclaw-server/src/main/java/vip/mate/llm/chatmodel/OpenAiCompatibleHttpInterceptor.java
+++ b/mateclaw-server/src/main/java/vip/mate/llm/chatmodel/OpenAiCompatibleHttpInterceptor.java
@@ -59,19 +59,23 @@ final class OpenAiCompatibleHttpInterceptor implements Interceptor {
                 rewritten.path("messages").size(),
                 rewritten.path("tools").size(),
                 rewritten.path("stream").asBoolean());
-        // fix：https://github.com/mateaix/mateclaw/issues/640
-        // The endpoint (provider baseUrl + completions path) is the authoritative URL.
-        // Carry over SDK-attached query params selectively: openai-java classifies
-        // *.openai.azure.com hosts whose baseUrl does not end in /openai/v1 as
-        // AZURE_LEGACY and appends a default `api-version` query — which the unified
-        // /openai/v1 surface rejects with 400 "API version not supported". Keep
-        // `api-version` only when the endpoint itself is a legacy deployments URL,
-        // where Azure requires it.
+        // Only Azure's unified v1 surface rejects the legacy SDK api-version.
+        // Keep explicit endpoint query values authoritative and avoid duplicating
+        // parameters already present there when copying SDK query parameters.
         HttpUrl parsedEndpoint = HttpUrl.get(endpoint);
-        boolean legacyDeploymentsEndpoint = parsedEndpoint.encodedPath().contains("/deployments/");
+        String host = parsedEndpoint.host();
+        boolean azureUnifiedEndpoint =
+                (host.endsWith(".openai.azure.com") || host.endsWith(".cognitiveservices.azure.com"))
+                        && parsedEndpoint.encodedPath().startsWith("/openai/v1/");
         HttpUrl.Builder url = parsedEndpoint.newBuilder();
+        if (azureUnifiedEndpoint) {
+            for (String name : parsedEndpoint.queryParameterNames()) {
+                if ("api-version".equalsIgnoreCase(name)) url.removeAllQueryParameters(name);
+            }
+        }
         for (String name : original.url().queryParameterNames()) {
-            if ("api-version".equalsIgnoreCase(name) && !legacyDeploymentsEndpoint) continue;
+            if (azureUnifiedEndpoint && "api-version".equalsIgnoreCase(name)) continue;
+            if (parsedEndpoint.queryParameterNames().contains(name)) continue;
             for (String value : original.url().queryParameterValues(name)) {
                 url.addQueryParameter(name, value);
             }
diff --git a/mateclaw-server/src/test/java/vip/mate/llm/chatmodel/OpenAiAzureWirePolicyTest.java b/mateclaw-server/src/test/java/vip/mate/llm/chatmodel/OpenAiAzureWirePolicyTest.java
new file mode 100644
index 00000000..df565a44
--- /dev/null
+++ b/mateclaw-server/src/test/java/vip/mate/llm/chatmodel/OpenAiAzureWirePolicyTest.java
@@ -0,0 +1,97 @@
+package vip.mate.llm.chatmodel;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.databind.node.ObjectNode;
+import okhttp3.*;
+import org.junit.jupiter.api.Test;
+import vip.mate.llm.model.ModelProviderEntity;
+
+import java.time.Duration;
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.atomic.AtomicReference;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+class OpenAiAzureWirePolicyTest {
+    private static final ObjectMapper JSON = new ObjectMapper();
+
+    @Test
+    void stripsOrphanToolOptionsAndEmptyToolsWithoutMutatingSource() throws Exception {
+        for (String tools : List.of("", ",\"tools\":[]")) {
+            ObjectNode source = (ObjectNode) JSON.readTree("{\"model\":\"gpt-4o\",\"messages\":[],"
+                    + "\"tool_choice\":\"none\",\"parallel_tool_calls\":false" + tools + "}");
+            var out = OpenAiRequestRewriter.rewrite(source, provider(), false);
+            assertFalse(out.has("tools"));
+            assertFalse(out.has("tool_choice"));
+            assertFalse(out.has("parallel_tool_calls"));
+            assertTrue(source.has("tool_choice"));
+        }
+    }
+
+    @Test
+    void preservesToolDisableGuardWhenToolsExist() throws Exception {
+        ObjectNode source = (ObjectNode) JSON.readTree("""
+                {"model":"gpt-4o","messages":[],"tools":[{"type":"function","function":{"name":"lookup"}}],
+                 "tool_choice":"none","parallel_tool_calls":false}
+                """);
+        assertEquals(source, OpenAiRequestRewriter.rewrite(source, provider(), false));
+    }
+
+    @Test
+    void keepsOptionsForInjectedKimiSearchTool() throws Exception {
+        ObjectNode source = (ObjectNode) JSON.readTree("{\"model\":\"kimi-k2\",\"messages\":[],\"tool_choice\":\"none\"}");
+        var out = OpenAiRequestRewriter.rewrite(source, provider(), true);
+        assertEquals("none", out.path("tool_choice").asText());
+        assertEquals(1, out.path("tools").size());
+    }
+
+    @Test
+    void unifiedAzureDropsSdkVersionButPreservesOtherParameters() throws Exception {
+        var url = wireUrl("https://demo.openai.azure.com/openai/v1/chat/completions",
+                "https://demo.openai.azure.com/openai/deployments/gpt/chat/completions?api-version=2025-01-01&trace=a&trace=b");
+        assertNull(url.queryParameter("api-version"));
+        assertEquals(List.of("a", "b"), url.queryParameterValues("trace"));
+    }
+
+    @Test
+    void legacyAzureKeepsVersion() throws Exception {
+        var url = wireUrl("https://demo.openai.azure.com/openai/deployments/gpt/chat/completions",
+                "https://demo.openai.azure.com/openai/deployments/gpt/chat/completions?api-version=2025-01-01");
+        assertEquals("2025-01-01", url.queryParameter("api-version"));
+    }
+
+    @Test
+    void legacyConfiguredVersionWinsWithoutDuplicate() throws Exception {
+        var url = wireUrl("https://demo.openai.azure.com/openai/deployments/gpt/chat/completions?api-version=custom",
+                "https://demo.openai.azure.com/openai/deployments/gpt/chat/completions?api-version=sdk");
+        assertEquals(List.of("custom"), url.queryParameterValues("api-version"));
+    }
+
+    @Test
+    void nonAzureGatewayKeepsVersionParameter() throws Exception {
+        var url = wireUrl("https://gateway.example/v1/chat/completions",
+                "https://gateway.example/v1/chat/completions?api-version=custom");
+        assertEquals("custom", url.queryParameter("api-version"));
+    }
+
+    private static HttpUrl wireUrl(String endpoint, String original) throws Exception {
+        AtomicReference<HttpUrl> sent = new AtomicReference<>();
+        OkHttpClient client = new OkHttpClient.Builder()
+                .addInterceptor(new OpenAiCompatibleHttpInterceptor(provider(), endpoint, Map.of(), false, Duration.ofSeconds(3)))
+                .addInterceptor(chain -> {
+                    sent.set(chain.request().url());
+                    return new Response.Builder().request(chain.request()).protocol(Protocol.HTTP_1_1).code(200)
+                            .message("OK").body(ResponseBody.create("{}", MediaType.get("application/json"))).build();
+                }).build();
+        Request request = new Request.Builder().url(original)
+                .post(RequestBody.create("{\"model\":\"gpt-4o\",\"messages\":[]}", MediaType.get("application/json"))).build();
+        try (Response ignored = client.newCall(request).execute()) { return sent.get(); }
+    }
+
+    private static ModelProviderEntity provider() {
+        var provider = new ModelProviderEntity();
+        provider.setProviderId("azure-openai");
+        return provider;
+    }
+}