fix(agent): continue ollama glm truncation replies

run_agent.py

@@ -2103,6 +2103,59 @@ class AIAgent:
         content = re.sub(r'</?(?:think|thinking|reasoning|thought|REASONING_SCRATCHPAD)>\s*', '', content, flags=re.IGNORECASE)
         return content
 
+    @staticmethod
+    def _has_natural_response_ending(content: str) -> bool:
+        """Heuristic: does visible assistant text look intentionally finished?"""
+        if not content:
+            return False
+        stripped = content.rstrip()
+        if not stripped:
+            return False
+        if stripped.endswith("```"):
+            return True
+        return stripped[-1] in '.!?:)"\']}。！？：）】」』》'
+
+    def _is_ollama_glm_backend(self) -> bool:
+        """Detect the narrow backend family affected by Ollama/GLM stop misreports."""
+        model_lower = (self.model or "").lower()
+        provider_lower = (self.provider or "").lower()
+        if "glm" not in model_lower and provider_lower != "zai":
+            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))
+
+    def _should_treat_stop_as_truncated(
+        self,
+        finish_reason: str,
+        assistant_message,
+        messages: Optional[list] = None,
+    ) -> bool:
+        """Detect conservative stop->length misreports for Ollama-hosted GLM models."""
+        if finish_reason != "stop" or self.api_mode != "chat_completions":
+            return False
+        if not self._is_ollama_glm_backend():
+            return False
+        if not any(
+            isinstance(msg, dict) and msg.get("role") == "tool"
+            for msg in (messages or [])
+        ):
+            return False
+        if assistant_message is None or getattr(assistant_message, "tool_calls", None):
+            return False
+
+        content = getattr(assistant_message, "content", None)
+        if not isinstance(content, str):
+            return False
+
+        visible_text = self._strip_think_blocks(content).strip()
+        if not visible_text:
+            return False
+        if len(visible_text) < 20 or not re.search(r"\s", visible_text):
+            return False
+
+        return not self._has_natural_response_ending(visible_text)
+
     def _looks_like_codex_intermediate_ack(
         self,
         user_message: str,
@@ -9038,6 +9091,17 @@ class AIAgent:
                         finish_reason = stop_reason_map.get(response.stop_reason, "stop")
                     else:
                         finish_reason = response.choices[0].finish_reason
+                        assistant_message = response.choices[0].message
+                        if self._should_treat_stop_as_truncated(
+                            finish_reason,
+                            assistant_message,
+                            messages,
+                        ):
+                            self._vprint(
+                                f"{self.log_prefix}⚠️  Treating suspicious Ollama/GLM stop response as truncated",
+                                force=True,
+                            )
+                            finish_reason = "length"
 
                     if finish_reason == "length":
                         self._vprint(f"{self.log_prefix}⚠️  Response truncated (finish_reason='length') - model hit max output tokens", force=True)

tests/run_agent/test_run_agent.py

@@ -2202,6 +2202,114 @@ class TestRunConversation:
         assert second_call_messages[-1]["role"] == "user"
         assert "truncated by the output length limit" in second_call_messages[-1]["content"]
 
+    def test_ollama_glm_stop_after_tools_without_terminal_boundary_requests_continuation(self, agent):
+        """Ollama-hosted GLM responses can misreport truncated output as stop."""
+        self._setup_agent(agent)
+        agent.base_url = "http://localhost:11434/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.model = "glm-5.1:cloud"
+
+        tool_turn = _mock_response(
+            content="",
+            finish_reason="tool_calls",
+            tool_calls=[_mock_tool_call(name="web_search", arguments="{}", call_id="c1")],
+        )
+        misreported_stop = _mock_response(
+            content="Based on the search results, the best next",
+            finish_reason="stop",
+        )
+        continued = _mock_response(
+            content=" step is to update the config.",
+            finish_reason="stop",
+        )
+        agent.client.chat.completions.create.side_effect = [
+            tool_turn,
+            misreported_stop,
+            continued,
+        ]
+
+        with (
+            patch("run_agent.handle_function_call", return_value="search result"),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("hello")
+
+        assert result["completed"] is True
+        assert result["api_calls"] == 3
+        assert (
+            result["final_response"]
+            == "Based on the search results, the best next step is to update the config."
+        )
+
+        third_call_messages = agent.client.chat.completions.create.call_args_list[2].kwargs["messages"]
+        assert third_call_messages[-1]["role"] == "user"
+        assert "truncated by the output length limit" in third_call_messages[-1]["content"]
+
+    def test_ollama_glm_stop_with_terminal_boundary_does_not_continue(self, agent):
+        """Complete Ollama/GLM responses should not be reclassified as truncated."""
+        self._setup_agent(agent)
+        agent.base_url = "http://localhost:11434/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.model = "glm-5.1:cloud"
+
+        tool_turn = _mock_response(
+            content="",
+            finish_reason="tool_calls",
+            tool_calls=[_mock_tool_call(name="web_search", arguments="{}", call_id="c1")],
+        )
+        complete_stop = _mock_response(
+            content="Based on the search results, the best next step is to update the config.",
+            finish_reason="stop",
+        )
+        agent.client.chat.completions.create.side_effect = [tool_turn, complete_stop]
+
+        with (
+            patch("run_agent.handle_function_call", return_value="search result"),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("hello")
+
+        assert result["completed"] is True
+        assert result["api_calls"] == 2
+        assert (
+            result["final_response"]
+            == "Based on the search results, the best next step is to update the config."
+        )
+
+    def test_non_ollama_stop_without_terminal_boundary_does_not_continue(self, agent):
+        """The stop->length workaround should stay scoped to Ollama/GLM backends."""
+        self._setup_agent(agent)
+        agent.base_url = "https://api.openai.com/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.model = "gpt-4o-mini"
+
+        tool_turn = _mock_response(
+            content="",
+            finish_reason="tool_calls",
+            tool_calls=[_mock_tool_call(name="web_search", arguments="{}", call_id="c1")],
+        )
+        normal_stop = _mock_response(
+            content="Based on the search results, the best next",
+            finish_reason="stop",
+        )
+        agent.client.chat.completions.create.side_effect = [tool_turn, normal_stop]
+
+        with (
+            patch("run_agent.handle_function_call", return_value="search result"),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("hello")
+
+        assert result["completed"] is True
+        assert result["api_calls"] == 2
+        assert result["final_response"] == "Based on the search results, the best next"
+
     def test_length_thinking_exhausted_skips_continuation(self, agent):
         """When finish_reason='length' but content is only thinking, skip retries."""
         self._setup_agent(agent)