From 8cfb20122e162830625c0dc2354e21cb60af7da8 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Wed, 28 Oct 2026 20:04:48 -0700
Subject: [PATCH] fix(bedrock_mantle): map gpt-oss context overflow to
 ContextWindowExceededError

Bedrock Mantle's gpt-oss models report a too-long prompt as "Input length (N)
exceeds model's maximum context length (M)", which the Mantle mapping missed, so
callers got a BadRequestError (or a 503 when streaming), context window fallbacks
never fired, and Claude Code never compacted. Match that wording too and return
the existing "prompt is too long: N tokens > M maximum" message.
---
 .../exception_mapping_utils.py                | 16 ++++---
 .../test_exception_mapping_utils.py           | 47 +++++++++++++++++++
 2 files changed, 57 insertions(+), 6 deletions(-)

diff --git a/litellm/litellm_core_utils/exception_mapping_utils.py b/litellm/litellm_core_utils/exception_mapping_utils.py
index 47ec8331c83..c37db350c6d 100644
--- a/litellm/litellm_core_utils/exception_mapping_utils.py
+++ b/litellm/litellm_core_utils/exception_mapping_utils.py
@@ -892,7 +892,10 @@ def _map_openai_like_exception(
             )
 
 
-_BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERN: Final = re.compile(r"prompt tokens \((\d+)\) exceed model maximum \((\d+)\)")
+_BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERNS: Final = (
+    re.compile(r"prompt tokens \((\d+)\) exceed model maximum \((\d+)\)"),
+    re.compile(r"Input length \((\d+)\) exceeds model\\?'s maximum context length \((\d+)\)"),
+)
 _BEDROCK_MANTLE_CONTEXT_LENGTH_EXCEEDED_CODE_PATTERN: Final = re.compile(
     r"""\\?["']code\\?["']\s*:\s*\\?["']context_length_exceeded\\?["']"""
 )
@@ -912,12 +915,13 @@ def _bedrock_mantle_context_window_response(original_exception: _ProviderHTTPExc
 
 def _get_bedrock_mantle_context_window_message(error_str: str) -> str | None:
     """
-    Mantle reports context overflow as a validation_error carrying the token counts, or as
-    OpenAI's context_length_exceeded code, either in a 400 body or in a streamed error event
-    with no error type. Clients such as Claude Code only treat "prompt is too long" as
-    overflow (litellm/litellm#36546).
+    Mantle reports context overflow as a validation_error carrying the token counts (worded
+    differently for gpt-oss), or as OpenAI's context_length_exceeded code, either in a 400 body
+    or in a streamed error event with no error type. Clients such as Claude Code only treat
+    "prompt is too long" as overflow (litellm/litellm#36546).
     """
-    match = _BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERN.search(error_str)
+    matches: Final = (pattern.search(error_str) for pattern in _BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERNS)
+    match: Final = next((found for found in matches if found is not None), None)
     if match is not None:
         prompt_tokens, max_tokens = match.groups()
         return f"prompt is too long: {prompt_tokens} tokens > {max_tokens} maximum"
diff --git a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
index 55e7c47e2c9..1a96a090f3b 100644
--- a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
+++ b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
@@ -1491,6 +1491,53 @@ def test_bedrock_mantle_context_overflow_maps_to_context_window_exceeded():
     assert "prompt is too long: 1055489 tokens > 1050000 maximum" in excinfo.value.message
 
 
+_GPT_OSS_OVERFLOW_EVENT: Final = (
+    'ErrorEvent { error: APIError { type: "BadRequestError", code: Some(400), '
+    "message: \"Input length (360068) exceeds model's maximum context length (131072).\", param: None } }"
+)
+
+
+@pytest.mark.parametrize(
+    "original_exception",
+    [
+        pytest.param(
+            BaseLLMException(
+                status_code=400,
+                message=json.dumps(
+                    {
+                        "error": {
+                            "code": "validation_error",
+                            "message": _GPT_OSS_OVERFLOW_EVENT,
+                            "param": None,
+                            "type": "invalid_request_error",
+                        }
+                    }
+                ),
+            ),
+            id="400-body",
+        ),
+        pytest.param(OpenAIError(status_code=503, message=_GPT_OSS_OVERFLOW_EVENT), id="streamed-event"),
+        pytest.param(
+            BaseLLMException(
+                status_code=400, message=f"Error code: 400 - {repr({'error': {'message': _GPT_OSS_OVERFLOW_EVENT}})}"
+            ),
+            id="python-repr",
+        ),
+    ],
+)
+def test_bedrock_mantle_gpt_oss_input_length_overflow_maps_to_context_window_exceeded(original_exception):
+    with pytest.raises(litellm.ContextWindowExceededError) as excinfo:
+        exception_type(
+            model="openai.gpt-oss-20b",
+            original_exception=original_exception,
+            custom_llm_provider="bedrock_mantle",
+        )
+
+    assert excinfo.value.status_code == 400
+    assert excinfo.value.response.status_code == 400
+    assert "prompt is too long: 360068 tokens > 131072 maximum" in excinfo.value.message
+
+
 def test_bedrock_mantle_openai_envelope_context_overflow_maps_to_context_window_exceeded():
     from litellm.llms.base_llm.chat.transformation import BaseLLMException