From 862bb0f12e9663bc754ae1a8cf54a572acdc4e3a Mon Sep 17 00:00:00 2001
From: aarongxa <aarongriffith06@gmail.com>
Date: Sun, 19 Apr 2026 21:10:31 -0400
Subject: [PATCH] fix(zai): set tool_stream=true to avoid api.z.ai 30s idle
 timeout
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

GLM models (4.6, 4.7, 5, 5.1) batch tool_call argument generation in a
single chunk by default. When the tool args are large (long file writes,
big shell commands), the model goes silent for 30+ seconds while
generating them, and api.z.ai's 30s server-side idle timeout kills the
connection with ECONNRESET — surfacing as a "timeout" in Hermes.

Z.AI's documented fix is to send tool_stream=true alongside stream=true,
which makes the model emit tool args incrementally and keeps chunks
flowing within the 30s window.

This patch injects extra_body.tool_stream=true in _build_api_kwargs
whenever the provider is "zai" or the base URL points at api.z.ai or
open.bigmodel.cn, and only when tools are present (no benefit otherwise).
A user-supplied extra_body.tool_stream wins via setdefault.

Refs:
- https://docs.z.ai/guides/capabilities/stream-tool
- vercel/ai#12949 (30s idle timeout root cause analysis)
- opencode#15350 (ECONNRESET on zai-coding-plan)

Made-with: Cursor
---
 run_agent.py                      | 16 +++++++++
 tests/run_agent/test_run_agent.py | 60 +++++++++++++++++++++++++++++++
 2 files changed, 76 insertions(+)

diff --git a/run_agent.py b/run_agent.py
index 85eaad1b3753..ded412e05fff 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -7286,6 +7286,22 @@ def _build_api_kwargs(self, api_messages: list) -> dict:
         if self._is_qwen_portal():
             extra_body["vl_high_resolution_images"] = True
 
+        # Z.AI / GLM: enable incremental tool-call argument streaming.
+        # GLM models (4.6, 4.7, 5, 5.1) batch tool_call args in one chunk by
+        # default, producing 30+ second silence gaps that trip api.z.ai's
+        # 30s server-side idle timeout (vercel/ai#12949, opencode#15350).
+        # Setting tool_stream=true makes the model stream tool args
+        # incrementally, eliminating the silence gap.
+        # Docs: https://docs.z.ai/guides/capabilities/stream-tool
+        _provider_lower = (self.provider or "").lower()
+        _is_zai_endpoint = (
+            _provider_lower == "zai"
+            or "z.ai" in self._base_url_lower
+            or "bigmodel.cn" in self._base_url_lower
+        )
+        if _is_zai_endpoint and self.tools:
+            extra_body.setdefault("tool_stream", True)
+
         if extra_body:
             api_kwargs["extra_body"] = extra_body
 
diff --git a/tests/run_agent/test_run_agent.py b/tests/run_agent/test_run_agent.py
index 9bc637135c7d..f4788e16a87f 100644
--- a/tests/run_agent/test_run_agent.py
+++ b/tests/run_agent/test_run_agent.py
@@ -247,6 +247,66 @@ def _mock_response(
 # ===================================================================
 
 
+class TestZaiToolStream:
+    """Verify Z.AI / GLM endpoints get tool_stream=true to avoid api.z.ai's
+    30s server-side idle timeout when GLM batches tool_call argument output.
+    Refs: vercel/ai#12949, opencode#15350, https://docs.z.ai/guides/capabilities/stream-tool
+    """
+
+    def _make_zai_agent(self, base_url, provider="zai", model="glm-5.1"):
+        with (
+            patch(
+                "run_agent.get_tool_definitions",
+                return_value=_make_tool_defs("web_search"),
+            ),
+            patch("run_agent.check_toolset_requirements", return_value={}),
+            patch("run_agent.OpenAI"),
+        ):
+            a = AIAgent(
+                model=model,
+                provider=provider,
+                base_url=base_url,
+                api_key="test-key-1234567890",
+                quiet_mode=True,
+                skip_context_files=True,
+                skip_memory=True,
+            )
+            a.client = MagicMock()
+            return a
+
+    def test_zai_provider_injects_tool_stream(self):
+        agent = self._make_zai_agent("https://api.z.ai/api/coding/paas/v4")
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs["extra_body"]["tool_stream"] is True
+
+    def test_zai_global_endpoint_injects_tool_stream(self):
+        agent = self._make_zai_agent(
+            "https://api.z.ai/api/paas/v4", provider="custom"
+        )
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs["extra_body"]["tool_stream"] is True
+
+    def test_bigmodel_cn_endpoint_injects_tool_stream(self):
+        agent = self._make_zai_agent(
+            "https://open.bigmodel.cn/api/coding/paas/v4", provider="custom"
+        )
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs["extra_body"]["tool_stream"] is True
+
+    def test_no_tools_omits_tool_stream(self):
+        agent = self._make_zai_agent("https://api.z.ai/api/coding/paas/v4")
+        agent.tools = []
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert "tool_stream" not in kwargs.get("extra_body", {})
+
+    def test_non_zai_provider_omits_tool_stream(self):
+        agent = self._make_zai_agent(
+            "https://openrouter.ai/api/v1", provider="openrouter", model="x"
+        )
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert "tool_stream" not in kwargs.get("extra_body", {})
+
+
 class TestHasContentAfterThinkBlock:
     def test_none_returns_false(self, agent):
         assert agent._has_content_after_think_block(None) is False