fix(agent): restrict _is_ollama_glm_backend to explicit Ollama indicators

Remove the is_local_endpoint() catch-all that matched any local address
(localhost, 127.0.0.1, RFC-1918). A LiteLLM proxy on localhost:8000
forwarding to remote Z.AI was falsely triggering truncation continuation.

Now only matches: URL containing "ollama" or ":11434", or provider
explicitly set to "ollama".

run_agent.py

@@ -2636,7 +2636,9 @@ class AIAgent:
             return False
         if "ollama" in self._base_url_lower or ":11434" in self._base_url_lower:
             return True
-        return bool(self.base_url and is_local_endpoint(self.base_url))
+        if provider_lower == "ollama":
+            return True
+        return False
 
     def _should_treat_stop_as_truncated(
         self,

tests/run_agent/test_ollama_glm_backend_detection.py

@@ -0,0 +1,193 @@
+"""Tests for _is_ollama_glm_backend() false-positive fix.
+
+Issue #13971: A local LiteLLM proxy on localhost:8000 forwarding to remote
+Z.AI was incorrectly matched by the is_local_endpoint() catch-all, causing
+false truncation continuation for non-Ollama GLM backends.
+
+The fix restricts detection to explicit Ollama indicators (URL containing
+"ollama" or ":11434", or provider set to "ollama") instead of matching any
+RFC-1918 / localhost endpoint.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+import run_agent
+
+
+def _make_agent_stub(
+    model: str = "",
+    provider: str = "",
+    base_url: str = "",
+) -> MagicMock:
+    """Build a minimal mock with the attributes _is_ollama_glm_backend reads."""
+    stub = MagicMock(spec=[])
+    stub.model = model
+    stub.provider = provider
+    stub.base_url = base_url
+    stub._base_url_lower = base_url.lower() if base_url else ""
+    return stub
+
+
+class TestIsOllamaGlmBackend:
+    """Unit tests for RunAgent._is_ollama_glm_backend."""
+
+    # -- True cases: genuine Ollama GLM backends --
+
+    def test_ollama_default_port_with_glm_model(self):
+        """Ollama on its default port (:11434) with a GLM model -> True."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="",
+            base_url="http://192.168.1.112:11434",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True
+
+    def test_ollama_in_url_with_glm_model(self):
+        """URL containing 'ollama' with a GLM model -> True."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="",
+            base_url="http://localhost:11434/v1",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True
+
+    def test_provider_ollama_with_glm_model(self):
+        """Provider explicitly set to 'ollama' with a GLM model -> True."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="ollama",
+            base_url="http://myhost:9999",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True
+
+    def test_provider_ollama_with_zai_provider(self):
+        """Provider 'ollama' also matches when model has no 'glm' but provider is 'zai'.
+
+        (This is a weird edge case — zai provider + ollama provider simultaneously
+        can't really happen, but the first guard passes for provider=='zai'.)
+        """
+        stub = _make_agent_stub(
+            model="some-model",
+            provider="zai",
+            base_url="http://localhost:11434",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True
+
+    # -- False cases: the fix --
+
+    def test_local_proxy_with_glm_and_zai_returns_false(self):
+        """KEY REGRESSION: localhost:8000 + GLM + zai provider -> must be False.
+
+        Before the fix, is_local_endpoint() matched any local address,
+        causing a LiteLLM proxy forwarding to remote Z.AI to be treated
+        as a local Ollama instance.
+        """
+        stub = _make_agent_stub(
+            model="glm-5.1",
+            provider="zai",
+            base_url="http://localhost:8000",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    def test_local_proxy_with_glm_and_litellm_returns_false(self):
+        """Local LiteLLM proxy with GLM model -> False (not Ollama)."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="litellm",
+            base_url="http://127.0.0.1:4000",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    def test_lan_proxy_with_glm_returns_false(self):
+        """RFC-1918 address on non-Ollama port with GLM -> False."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="",
+            base_url="http://192.168.1.50:8080",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    # -- False cases: non-GLM models --
+
+    def test_non_glm_model_on_ollama_port(self):
+        """Non-GLM model on Ollama port -> False (early exit on model check)."""
+        stub = _make_agent_stub(
+            model="llama3:8b",
+            provider="",
+            base_url="http://localhost:11434",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    def test_non_glm_model_non_zai_provider(self):
+        """Non-GLM model with non-zai provider -> False."""
+        stub = _make_agent_stub(
+            model="gpt-4o",
+            provider="openai",
+            base_url="http://localhost:8000",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    # -- False cases: remote endpoints --
+
+    def test_remote_endpoint_with_glm_model(self):
+        """Remote endpoint (api.z.ai) with GLM model -> False."""
+        stub = _make_agent_stub(
+            model="glm-5.1",
+            provider="zai",
+            base_url="https://api.z.ai/api/coding/paas/v4",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    def test_remote_endpoint_no_base_url(self):
+        """No base_url at all with GLM + zai -> False."""
+        stub = _make_agent_stub(
+            model="glm-5.1",
+            provider="zai",
+            base_url="",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    # -- Edge cases --
+
+    def test_none_model_and_provider(self):
+        """None model and provider -> False (no crash)."""
+        stub = _make_agent_stub()
+        stub.model = None
+        stub.provider = None
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is False
+
+    def test_case_insensitive_glm_match(self):
+        """GLM in mixed case still matches on Ollama port."""
+        stub = _make_agent_stub(
+            model="GLM-4-9B",
+            provider="",
+            base_url="http://localhost:11434",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True
+
+    def test_case_insensitive_provider_ollama(self):
+        """Provider 'Ollama' in mixed case still matches."""
+        stub = _make_agent_stub(
+            model="glm-4-9b",
+            provider="Ollama",
+            base_url="http://myhost:9999",
+        )
+        result = run_agent.AIAgent._is_ollama_glm_backend(stub)
+        assert result is True