fix(im_llm): 代理绕过Docker内部hostname

backend/python/tests/test_regression_r4.py

@@ -0,0 +1,143 @@
+"""Regression R4: 防止能力路由误判回归。
+
+R4.1: detector_media_intent 对纯文本输入返回 None（不误判为 image/video）。
+R4.2: resolve_auto_model("text") 返回有 text 能力的模型。
+R4.3: 普通多轮对话必须走 text 模型，不可走 image/video。
+R4.4: 所有已注册能力可由 resolve_auto_model 解析到启用模型或 None。
+R4.5: is_default 模型必须存在且启用。
+"""
+import json
+import pytest
+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+from app.api.chat_media import detect_media_intent, detect_ambiguous_creative_intent
+from app.capability_registry import resolve_auto_model, get_capability_defaults, CAPABILITY_TYPES, TEXT_FAMILY
+from app.database import SessionLocal
+from app.models import AIModel
+
+
+@pytest.fixture(scope="module")
+def db():
+    """获得独立测试数据库会话"""
+    db = SessionLocal()
+    yield db
+    db.close()
+
+
+# ─── R4.1: detect_media_intent regression suite ───
+MEDIA_INTENT_CASES = [
+    # (input, expected_intent)
+    ("你好", None),
+    ("今天天气怎么样", None),
+    ("帮我写一篇文章", None),
+    ("平台有哪些技能能", None),          # 历史 bug：曾被误判为 image
+    ("新对话", None),                    # 新建对话入口
+    ("画一只小猫", "image"),
+    ("生成一张海报", "image"),
+    ("给我画个头像", "image"),
+    ("生成一个视频介绍", "video"),
+    ("做一段短片", "video"),
+    ("来段视频", "video"),
+    ("朗读这段文字", "audio"),
+    ("语音合成一段播报", "audio"),
+    ("你能生成视频吗？", None),           # 疑问句不触发
+    ("怎么画一只猫？", None),
+    ("如何生成图片呢？", None),
+    ("画一只小猫，要可爱的", "image"),    # 有明确指令
+    ("Generate a cat image", "image"),   # 英文
+    ("help me create a video", "video"),
+]
+
+
+class TestR4_1_MediaIntent:
+    @pytest.mark.parametrize("text,expected", MEDIA_INTENT_CASES)
+    def test_detect_media_intent(self, text, expected):
+        result = detect_media_intent(text)
+        assert result == expected, f"detect_media_intent({text!r}) = {result!r}, expected {expected!r}"
+
+
+# ─── R4.2: resolve_auto_model 回归 ───
+class TestR4_2_AutoModel:
+
+    def test_text_model_has_text_capability(self, db):
+        """文本对话默认模型必须有 text 能力"""
+        m = resolve_auto_model("text", db)
+        assert m is not None, "resolve_auto_model('text') returned None"
+        caps = [c.lower() for c in (m.capabilities or [])]
+        assert "text" in caps, f"Text model {m.name} has no 'text' capability: {caps}"
+
+    def test_text_model_not_image_video(self, db):
+        """文本默认模型不应触发 image/video 误判"""
+        m = resolve_auto_model("text", db)
+        caps = [c.lower() for c in (m.capabilities or [])]
+        if "image" in caps or "video" in caps:
+            # If text model has image caps, detect_media_intent must catch it
+            assert detect_media_intent("你好") is None, "Normal text detected as media despite model having image caps"
+            assert detect_media_intent("新对话") is None
+
+    def test_thinking_model_exists(self, db):
+        m = resolve_auto_model("thinking", db)
+        assert m is not None, "resolve_auto_model('thinking') returned None"
+
+    def test_all_11_capabilities_resolvable(self, db):
+        """所有 11 个能力至少有一个模型或 None"""
+        for cap_def in CAPABILITY_TYPES:
+            cap = cap_def["id"]
+            m = resolve_auto_model(cap, db)
+            is_text_family = cap in TEXT_FAMILY
+            if is_text_family:
+                assert m is not None, f"{cap} is TEXT_FAMILY but resolve_auto_model returned None"
+            # media/embedding/rerank/asr/ocr can legitimately be None if no model
+
+
+# ─── R4.3: is_default model exists ───
+class TestR4_3_DefaultModel:
+
+    def test_default_model_exists(self, db):
+        m = db.query(AIModel).filter(AIModel.is_default == True, AIModel.is_enabled == True).first()
+        assert m is not None, "No is_default=true and enabled model"
+        caps = [c.lower() for c in (m.capabilities or [])]
+        assert "text" in caps, f"Default model {m.name} must have 'text' capability: {caps}"
+
+    def test_default_model_resolvable(self, db):
+        m = db.query(AIModel).filter(AIModel.is_default == True, AIModel.is_enabled == True).first()
+        assert m.id == resolve_auto_model("text", db).id, "is_default model != resolve_auto_model('text')"
+
+
+# ─── R4.4: capability defaults consistency ───
+class TestR4_4_CapabilityDefaults:
+
+    def test_text_defaults_in_sync_with_auto(self, db):
+        """site_settings text capability defaults must match resolve_auto_model"""
+        defaults = get_capability_defaults(db)
+        for cap in ("text", "thinking"):
+            mid = defaults.get(cap)
+            if mid:
+                m = resolve_auto_model(cap, db)
+                assert m is not None, f"SiteSetting {cap}->{mid} not resolvable"
+                assert m.id == mid, f"Mismatch: site_setting {cap}={mid} but resolve_auto_model returns {m.id}"
+
+    def test_no_disabled_models_in_defaults(self, db):
+        defaults = get_capability_defaults(db)
+        for cap, mid in defaults.items():
+            if mid:
+                m = db.query(AIModel).filter(AIModel.id == mid, AIModel.is_enabled == True).first()
+                assert m is not None, f"Capability {cap} points to disabled/nonexistent model {mid}"
+
+
+# ─── R4.5: detect_ambiguous_creative_intent smoke ───
+class TestR4_5_AmbiguousIntent:
+
+    def test_ambiguous_detected(self):
+        r = detect_ambiguous_creative_intent("帮我做一个东西")
+        assert r is not None, "Ambiguous creative intent not detected"
+
+    def test_explicit_not_ambiguous(self):
+        r = detect_ambiguous_creative_intent("帮我写一篇文章")
+        assert r is None, "Article request should not be ambiguous"
+
+    def test_media_not_ambiguous(self):
+        r = detect_ambiguous_creative_intent("画一只小猫")
+        assert r is None, "Image intent should bypass ambiguous detection"