fix(llm): skip chat-style probe for embedding-prefix models in DashScope discovery

mateclaw-server/src/main/java/vip/mate/llm/service/ModelDiscoveryService.java

@@ -139,6 +139,17 @@ public class ModelDiscoveryService {
             "qwen3-livetranslate-"
     );
 
+    /**
+     * Model-id prefixes that mark an embedding model. These pass the chat-focused
+     * acceptable-id check (so the manual "Add model" form works) but must be
+     * excluded from chat-style runtime probes and from "new chat models" auto
+     * suggestions — they only speak the embeddings protocol, not chat completion.
+     */
+    private static final Set<String> EMBEDDING_MODEL_PREFIXES = Set.of(
+            "text-embedding-",
+            "embedding-"
+    );
+
     /**
      * Allow-list prefixes for DashScope models that are known to work on the native
      * protocol (chat or embedding). An empty set means "no prefix filter" (we still
@@ -183,10 +194,13 @@ public class ModelDiscoveryService {
                 .map(ModelConfigEntity::getModelName)
                 .collect(Collectors.toSet());
 
-        // Only propose models that passed the probe (or were not probed) as "new"
+        // Only propose models that passed the probe (or were not probed) as "new".
+        // Embedding models are catalog-visible but excluded from the chat-discovery
+        // suggestion bucket — adding them as chat models would 400 at runtime.
         List<ModelInfoDTO> newModels = discovered.stream()
                 .filter(m -> !existingIds.contains(m.getId()))
                 .filter(m -> !Boolean.FALSE.equals(m.getProbeOk()))
+                .filter(m -> !isEmbeddingModelId(m.getId()))
                 .toList();
 
         return new DiscoverResult(discovered, newModels, discovered.size(), newModels.size());
@@ -235,6 +249,18 @@ public class ModelDiscoveryService {
         return DASHSCOPE_NATIVE_ALLOW_PREFIXES.stream().anyMatch(lower::startsWith);
     }
 
+    /**
+     * Returns true when a model id looks like an embedding model (e.g.
+     * {@code text-embedding-v3}). Used to skip chat-style probes and to keep
+     * embedding entries out of the "new chat models to add" auto-suggestion.
+     * Package-private for unit tests.
+     */
+    static boolean isEmbeddingModelId(String modelId) {
+        if (modelId == null) return false;
+        String lower = modelId.toLowerCase();
+        return EMBEDDING_MODEL_PREFIXES.stream().anyMatch(lower::startsWith);
+    }
+
     /**
      * Defensive guard for code paths that persist a model id without going through
      * discovery (e.g. the manual "Add model" form). Throws a MateClawException with
@@ -270,6 +296,13 @@ public class ModelDiscoveryService {
         Semaphore sem = new Semaphore(MAX_PROBE_CONCURRENCY);
         List<CompletableFuture<Void>> futures = new ArrayList<>(discovered.size());
         for (ModelInfoDTO dto : discovered) {
+            if (isEmbeddingModelId(dto.getId())) {
+                // Embedding models don't speak chat completion — sendTestPrompt would
+                // 400 with InvalidParameter. Leave probeOk null (no warning badge)
+                // and surface a clear, non-error reason in probeError.
+                dto.setProbeError("Embedding model — not chat-probeable");
+                continue;
+            }
             futures.add(CompletableFuture.runAsync(() -> {
                 try { sem.acquire(); }
                 catch (InterruptedException ie) { Thread.currentThread().interrupt(); return; }

mateclaw-server/src/test/java/vip/mate/llm/service/ModelDiscoveryEmbeddingDetectionTest.java

@@ -0,0 +1,61 @@
+package vip.mate.llm.service;
+
+import org.junit.jupiter.api.Test;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Coverage for {@link ModelDiscoveryService#isEmbeddingModelId(String)} —
+ * the predicate that keeps embedding models out of the chat-style probe and
+ * out of the chat-discovery "new models" suggestion bucket.
+ */
+class ModelDiscoveryEmbeddingDetectionTest {
+
+    @Test
+    void dashscopeEmbeddingVariants_recognised() {
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-v1"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-v3"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-v4"));
+    }
+
+    @Test
+    void openAiEmbeddingVariants_recognised() {
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-ada-002"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-3-small"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("text-embedding-3-large"));
+    }
+
+    @Test
+    void genericEmbeddingPrefix_recognised() {
+        // Some providers ship just "embedding-..." without the "text-" prefix.
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("embedding-001"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("embedding-large"));
+    }
+
+    @Test
+    void chatModels_notRecognisedAsEmbedding() {
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId("qwen-plus"));
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId("gpt-4o"));
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId("claude-sonnet-4-6"));
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId("deepseek-r1"));
+    }
+
+    @Test
+    void detectionIsCaseInsensitive() {
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("Text-Embedding-V4"));
+        assertTrue(ModelDiscoveryService.isEmbeddingModelId("TEXT-EMBEDDING-3-SMALL"));
+    }
+
+    @Test
+    void nullAndBlank_notEmbedding() {
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId(null));
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId(""));
+    }
+
+    @Test
+    void stringContainingEmbeddingMidway_notDetected() {
+        // Only prefix-anchored matches count; arbitrary mentions don't.
+        assertFalse(ModelDiscoveryService.isEmbeddingModelId("qwen-embedding-experimental"));
+    }
+}