fix(zai): resolve model list for China API users (endpoint-aware)

plugins/model-providers/zai/__init__.py

@@ -32,6 +32,17 @@
 from providers import register_provider
 from providers.base import ProviderProfile
 
+# Universally-free models available to all Z.AI users (both the
+# international and China platforms), but not listed by the provider's
+# /v1/models endpoint.  All verified with real API calls.
+ZAI_FREE_MODELS = (
+    "glm-4v-flash",
+    "glm-4.6v-flash",
+    "glm-4.1v-thinking-flash",
+    "glm-4.5-flash",
+    "glm-4-flash-250414",
+)
+
 _GLM_VERSION_RE = re.compile(r"^glm-(\d+)(?:\.(\d+))?")
 
 
@@ -89,6 +100,35 @@ def _glm_5_2_reasoning_effort(reasoning_config: dict | None) -> str | None:
 class ZaiProfile(ProviderProfile):
     """Z.AI / GLM — extra_body.thinking on/off + GLM-5.2 reasoning_effort."""
 
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Fetch the live model list and append universally-free models.
+
+        Z.AI's ``/v1/models`` endpoint omits a handful of Flash models that
+        nonetheless accept real API calls (``glm-4v-flash`` etc., verified
+        against both the international and China platforms).  Append them
+        here so China API users — whose keys are rejected by the
+        international endpoint — still get the free models in the picker.
+
+        ``base_url`` is forwarded so the China routing fix in
+        ``provider_model_ids`` (``hermes_cli/models.py``) is not undone.
+        """
+        live = super().fetch_models(
+            api_key=api_key, base_url=base_url, timeout=timeout
+        )
+        models = list(live) if live else []
+        seen = {m.lower() for m in models}
+        for m in ZAI_FREE_MODELS:
+            if m.lower() not in seen:
+                models.append(m)
+                seen.add(m.lower())
+        return models
+
     def build_api_kwargs_extras(
         self, *, reasoning_config: dict | None = None, model: str | None = None, **context
     ) -> tuple[dict[str, Any], dict[str, Any]]:

tests/plugins/model_providers/test_zai_profile.py

@@ -176,3 +176,76 @@ def test_glm_5_2_effort_reaches_top_level(self, zai_profile):
         )
         assert kwargs["reasoning_effort"] == "max"
         assert kwargs["extra_body"]["thinking"] == {"type": "enabled"}
+
+
+class TestZaiFetchModels:
+    """``fetch_models`` appends universally-free models and forwards base_url.
+
+    Z.AI's ``/v1/models`` omits a handful of Flash models that nonetheless
+    accept real API calls.  China API keys are rejected by the international
+    endpoint, so the free-model append must survive an empty/None live
+    result, and the resolved China ``base_url`` must reach the base class.
+    """
+
+    def test_appends_free_models_to_live(self, zai_profile, monkeypatch):
+        from providers.base import ProviderProfile
+        from plugins.model_providers.zai import ZAI_FREE_MODELS
+
+        def fake_fetch(self, *, api_key=None, base_url=None, timeout=8.0):
+            return ["glm-5", "glm-4-9b"]
+
+        monkeypatch.setattr(ProviderProfile, "fetch_models", fake_fetch)
+        models = zai_profile.fetch_models(api_key="k")
+        assert "glm-5" in models
+        assert "glm-4v-flash" in models
+        assert "glm-4.5-flash" in models
+        # Every verified free model is present.
+        assert set(ZAI_FREE_MODELS) <= set(models)
+
+    def test_dedup_when_live_already_lists_free_model(self, zai_profile, monkeypatch):
+        from providers.base import ProviderProfile
+
+        def fake_fetch(self, *, api_key=None, base_url=None, timeout=8.0):
+            return ["glm-5", "glm-4v-flash"]
+
+        monkeypatch.setattr(ProviderProfile, "fetch_models", fake_fetch)
+        models = zai_profile.fetch_models(api_key="k")
+        assert models.count("glm-4v-flash") == 1
+
+    def test_empty_or_none_live_still_returns_free_models(self, zai_profile, monkeypatch):
+        from providers.base import ProviderProfile
+        from plugins.model_providers.zai import ZAI_FREE_MODELS
+
+        expected = sorted(ZAI_FREE_MODELS)
+
+        def fake_fetch_empty(self, *, api_key=None, base_url=None, timeout=8.0):
+            return []
+
+        monkeypatch.setattr(ProviderProfile, "fetch_models", fake_fetch_empty)
+        assert sorted(zai_profile.fetch_models(api_key="k")) == expected
+
+        def fake_fetch_none(self, *, api_key=None, base_url=None, timeout=8.0):
+            return None
+
+        monkeypatch.setattr(ProviderProfile, "fetch_models", fake_fetch_none)
+        # China keys: live fetch rejects → None; free models still surface.
+        assert sorted(zai_profile.fetch_models(api_key="china-key")) == expected
+
+    def test_forwards_base_url_to_super(self, zai_profile, monkeypatch):
+        """The resolved China endpoint must reach the base implementation."""
+        from providers.base import ProviderProfile
+
+        seen = {}
+
+        def fake_fetch(self, *, api_key=None, base_url=None, timeout=8.0):
+            seen["api_key"] = api_key
+            seen["base_url"] = base_url
+            return ["glm-5"]
+
+        monkeypatch.setattr(ProviderProfile, "fetch_models", fake_fetch)
+        zai_profile.fetch_models(
+            api_key="china-key",
+            base_url="https://open.bigmodel.cn/api/paas/v4",
+        )
+        assert seen["api_key"] == "china-key"
+        assert seen["base_url"] == "https://open.bigmodel.cn/api/paas/v4"