fix(llm): cap auto-detected max_output_tokens when it fills the entire context window

openhands-sdk/openhands/sdk/llm/llm.py

@@ -1210,6 +1210,27 @@ def _init_model_info_and_caps(self) -> None:
             elif self._model_info is not None:
                 if isinstance(self._model_info.get("max_output_tokens"), int):
                     self.max_output_tokens = self._model_info.get("max_output_tokens")
+                    # Guard: if the registry reports max_output_tokens >= the
+                    # context window, requesting that many output tokens would
+                    # leave zero room for input and the provider will reject
+                    # every call. Cap to half the context window so input has
+                    # headroom (e.g. Nemotron 262 144 / 262 144).
+                    if (
+                        self.max_input_tokens is not None
+                        and self.max_output_tokens is not None
+                        and self.max_output_tokens >= self.max_input_tokens
+                    ):
+                        capped = self.max_input_tokens // 2
+                        logger.debug(
+                            "Capping max_output_tokens from %s to %s "
+                            "for %s (max_output_tokens >= context "
+                            "window %s)",
+                            self.max_output_tokens,
+                            capped,
+                            self.model,
+                            self.max_input_tokens,
+                        )
+                        self.max_output_tokens = capped
                 elif isinstance(self._model_info.get("max_tokens"), int):
                     # 'max_tokens' is ambiguous: some providers use it for total
                     # context window, not output limit. Cap it to avoid requesting

tests/sdk/llm/test_llm.py

@@ -1161,4 +1161,48 @@ def test_explicit_max_output_tokens_not_overridden():
     assert llm.max_output_tokens == 32768
 
 
+@patch("openhands.sdk.llm.llm.get_litellm_model_info")
+def test_max_output_tokens_capped_when_equal_to_context_window(
+    mock_get_model_info,
+):
+    """max_output_tokens == context window leaves zero input headroom.
+
+    Nemotron reports max_output_tokens = max_input_tokens = 262144.
+    Without capping, every LLM call is rejected because the entire context
+    window is reserved for output.
+    """
+    mock_get_model_info.return_value = {
+        "max_output_tokens": 262144,
+        "max_input_tokens": 262144,
+    }
+
+    llm = LLM(
+        model="litellm_proxy/converse-nemotron-super-3-120b",
+        api_key=SecretStr("test-key"),
+        usage_id="test-llm",
+    )
+
+    assert llm.max_output_tokens == 262144 // 2
+    assert llm.max_input_tokens == 262144
+
+
+@patch("openhands.sdk.llm.llm.get_litellm_model_info")
+def test_max_output_tokens_not_capped_when_below_context_window(
+    mock_get_model_info,
+):
+    """max_output_tokens < context window should be used as-is."""
+    mock_get_model_info.return_value = {
+        "max_output_tokens": 8192,
+        "max_input_tokens": 200000,
+    }
+
+    llm = LLM(
+        model="anthropic/claude-3-5-sonnet-latest",
+        api_key=SecretStr("test-key"),
+        usage_id="test-llm",
+    )
+
+    assert llm.max_output_tokens == 8192
+
+
 # LLM Registry Tests