fix(qwen): correct context lengths for qwen3-coder models and send max_tokens to portal

Based on PR #7285 by @kshitijk4poor.

Two bugs affecting Qwen OAuth users:

1. Wrong context window — qwen3-coder-plus showed 128K instead of 1M.
   Added specific entries before the generic qwen catch-all:
   - qwen3-coder-plus: 1,000,000 (corrected from PR's 1,048,576 per
     official Alibaba Cloud docs and OpenRouter)
   - qwen3-coder: 262,144

2. Random stopping — max_tokens was suppressed for Qwen Portal, so the
   server applied its own low default. Reasoning models exhaust that on
   thinking tokens. Now: honor explicit max_tokens, default to 65536
   when unset.

Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>

agent/model_metadata.py

@@ -113,7 +113,10 @@ DEFAULT_CONTEXT_LENGTHS = {
     "deepseek": 128000,
     # Meta
     "llama": 131072,
-    # Qwen
+    # Qwen — specific model families before the catch-all.
+    # Official docs: https://help.aliyun.com/zh/model-studio/developer-reference/
+    "qwen3-coder-plus": 1000000,  # 1M context
+    "qwen3-coder": 262144,        # 256K context
     "qwen": 131072,
     # MiniMax — official docs: 204,800 context for all models
     # https://platform.minimax.io/docs/api-reference/text-anthropic-api

run_agent.py

@@ -5888,8 +5888,16 @@ class AIAgent:
             api_kwargs["tools"] = self.tools
 
         if self.max_tokens is not None:
-            if not self._is_qwen_portal():
             api_kwargs.update(self._max_tokens_param(self.max_tokens))
+        elif self._is_qwen_portal():
+            # Qwen Portal defaults to a very low max_tokens when omitted.
+            # Reasoning models (qwen3-coder-plus) exhaust that budget on
+            # thinking tokens alone, causing the portal to return
+            # finish_reason="stop" with truncated output — the agent sees
+            # this as an intentional stop and exits the loop.  Send 65536
+            # (the documented max output for qwen3-coder models) so the
+            # model has adequate output budget for tool calls.
+            api_kwargs.update(self._max_tokens_param(65536))
         elif (self._is_openrouter_url() or "nousresearch" in self._base_url_lower) and "claude" in (self.model or "").lower():
             # OpenRouter and Nous Portal translate requests to Anthropic's
             # Messages API, which requires max_tokens as a mandatory field.

tests/agent/test_model_metadata.py

@@ -222,6 +222,24 @@ class TestGetModelContextLength:
         mock_fetch.return_value = {}
         assert get_model_context_length("openai/gpt-4o") == 128000
 
+    @patch("agent.model_metadata.fetch_model_metadata")
+    def test_qwen3_coder_plus_context_length(self, mock_fetch):
+        """qwen3-coder-plus has a 1M context window, not the generic 128K Qwen default."""
+        mock_fetch.return_value = {}
+        assert get_model_context_length("qwen3-coder-plus") == 1000000
+
+    @patch("agent.model_metadata.fetch_model_metadata")
+    def test_qwen3_coder_context_length(self, mock_fetch):
+        """qwen3-coder has a 256K context window, not the generic 128K Qwen default."""
+        mock_fetch.return_value = {}
+        assert get_model_context_length("qwen3-coder") == 262144
+
+    @patch("agent.model_metadata.fetch_model_metadata")
+    def test_qwen_generic_context_length(self, mock_fetch):
+        """Generic qwen models still get the 128K default."""
+        mock_fetch.return_value = {}
+        assert get_model_context_length("qwen3-plus") == 131072
+
     @patch("agent.model_metadata.fetch_model_metadata")
     def test_api_missing_context_length_key(self, mock_fetch):
         """Model in API but without context_length → defaults to 128000."""

tests/run_agent/test_run_agent.py

@@ -953,14 +953,24 @@ class TestBuildApiKwargs:
         assert kwargs["messages"][0]["content"][0]["text"] == "hi"
         assert "cache_control" not in kwargs["messages"][0]["content"][0]
 
-    def test_qwen_portal_omits_max_tokens(self, agent):
+    def test_qwen_portal_sends_explicit_max_tokens(self, agent):
+        """When the user explicitly sets max_tokens, it should be sent to Qwen Portal."""
         agent.base_url = "https://portal.qwen.ai/v1"
         agent._base_url_lower = agent.base_url.lower()
         agent.max_tokens = 4096
         messages = [{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}]
         kwargs = agent._build_api_kwargs(messages)
-        assert "max_tokens" not in kwargs
+        assert kwargs["max_tokens"] == 4096
-        assert "max_completion_tokens" not in kwargs
+
+    def test_qwen_portal_default_max_tokens(self, agent):
+        """When max_tokens is None, Qwen Portal gets a default of 65536
+        to prevent reasoning models from exhausting their output budget."""
+        agent.base_url = "https://portal.qwen.ai/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.max_tokens = None
+        messages = [{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        assert kwargs["max_tokens"] == 65536
 
 
 class TestBuildAssistantMessage: