From 421fd3cd61001bedbc7b7285b0b58f0bc0452dc0 Mon Sep 17 00:00:00 2001
From: matevip <matevip@mate.vip>
Date: Mon, 29 Jun 2026 14:48:44 +0800
Subject: [PATCH] fix(llm): stop assuming DeepSeek is vision-capable

---
 .../llm/service/ModelCapabilityService.java   |  9 ++++---
 .../service/ModelCapabilityServiceTest.java   | 27 ++++++++++++-------
 2 files changed, 23 insertions(+), 13 deletions(-)

diff --git a/mateclaw-server/src/main/java/vip/mate/llm/service/ModelCapabilityService.java b/mateclaw-server/src/main/java/vip/mate/llm/service/ModelCapabilityService.java
index 3eaba40e..ca1ee69c 100644
--- a/mateclaw-server/src/main/java/vip/mate/llm/service/ModelCapabilityService.java
+++ b/mateclaw-server/src/main/java/vip/mate/llm/service/ModelCapabilityService.java
@@ -97,9 +97,12 @@ public class ModelCapabilityService {
         m.put("claude-haiku",  EnumSet.of(Modality.VISION));
 
         // ===== DeepSeek =====
-        // V4 (Apr 2026) is the first DeepSeek line with native multimodal — image + video.
-        // V3 and earlier are text-only (no entry → defaults to text only).
-        m.put("deepseek-v4",   EnumSet.of(Modality.VISION, Modality.VIDEO));
+        // All released DeepSeek chat models (deepseek-chat / deepseek-reasoner /
+        // deepseek-v3.x) are text-only, so there is no entry and they default to
+        // text. Do NOT assume a future line is multimodal here: a wrong vision
+        // assumption makes the router send image_url to a text model, which the
+        // provider rejects with a 400. A genuinely multimodal model should declare
+        // its modalities on the model config instead.
 
         // ===== ByteDance Doubao / Seed =====
         // Seed 2.0 Pro (Feb 2026) handles hour-long videos. Seed1.5-VL also supports video.
diff --git a/mateclaw-server/src/test/java/vip/mate/llm/service/ModelCapabilityServiceTest.java b/mateclaw-server/src/test/java/vip/mate/llm/service/ModelCapabilityServiceTest.java
index 0c485548..3b231e4a 100644
--- a/mateclaw-server/src/test/java/vip/mate/llm/service/ModelCapabilityServiceTest.java
+++ b/mateclaw-server/src/test/java/vip/mate/llm/service/ModelCapabilityServiceTest.java
@@ -98,17 +98,24 @@ class ModelCapabilityServiceTest {
     }
 
     @Test
-    @DisplayName("DeepSeek V4 / V4-Pro → VIDEO; V3 (text-only) gets nothing")
-    void deepseekV4_supportsVideo() {
-        // DeepSeek V4 (Apr 2026) introduced native multimodal incl. video to the line.
-        // V3 and earlier remain text-only and must NOT match the V4 entry.
-        assertTrue(service.supports("deepseek-v4", null, Modality.VIDEO));
-        assertTrue(service.supports("deepseek-v4-pro", null, Modality.VIDEO));
-        assertTrue(service.supports("deepseek-v4-flash", null, Modality.VIDEO));
-        assertFalse(service.supports("deepseek-v3", null, Modality.VIDEO),
-                "V3 must NOT inherit V4 capabilities — text-only base differs from V4 entirely");
+    @DisplayName("DeepSeek is text-only by default — no speculative vision entry (issue #288)")
+    void deepseek_textOnlyByDefault() {
+        // DeepSeek's released chat models (deepseek-chat / deepseek-reasoner /
+        // deepseek-v3.x) are text-only. A hardcoded deepseek-v4 → vision/video entry
+        // made the router send image_url to a text model, which DeepSeek rejects with
+        // a 400 (and it never fell back to the vision sidecar). Default must be
+        // text-only; a genuinely multimodal model is opted in via the DB modalities
+        // declaration, not assumed here.
+        assertFalse(service.supports("deepseek-v4", null, Modality.VISION),
+                "deepseek must not be assumed vision-capable (issue #288)");
+        assertFalse(service.supports("deepseek-v4", null, Modality.VIDEO));
+        assertFalse(service.supports("deepseek-chat", null, Modality.VISION));
+        assertFalse(service.supports("deepseek-reasoner", null, Modality.VISION));
         assertFalse(service.supports("deepseek-v3.2", null, Modality.VIDEO));
-        assertFalse(service.supports("deepseek-r1", null, Modality.VIDEO));
+        assertEquals(EnumSet.of(Modality.TEXT), service.resolve("deepseek-v4", null));
+        // A real multimodal model can still be declared explicitly via DB modalities.
+        assertTrue(service.supports("deepseek-v4", "[\"vision\"]", Modality.VISION),
+                "an explicit DB declaration must still grant vision when the model truly has it");
     }
 
     @Test