fix(responses): map Bedrock Mantle context overflow to ContextWindowExceededError (#37862)

Mantle reports context overflow as a structured 400 validation_error rather than
the plain-text patterns Bedrock itself uses, so callers such as Claude Code that
key reactive compaction off the phrase "prompt is too long" never see it. Detect
the pattern and normalize the message to that phrase.

litellm/litellm_core_utils/exception_mapping_utils.py

@@ -811,6 +811,24 @@ def _map_openai_like_exception(
             )
 
 
+_BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERN: Final = re.compile(r"prompt tokens \((\d+)\) exceed model maximum \((\d+)\)")
+
+
+def _get_bedrock_mantle_context_window_message(error_str: str) -> str | None:
+    """
+    Mantle reports context overflow as a structured validation error rather than
+    the plain-text patterns Bedrock itself uses, so it needs its own detection and a
+    message clients recognize as context overflow (litellm/litellm#36546).
+    """
+    if "invalid_request_error" not in error_str and "validation_error" not in error_str:
+        return None
+    match = _BEDROCK_MANTLE_CONTEXT_WINDOW_PATTERN.search(error_str)
+    if match is None:
+        return None
+    prompt_tokens, max_tokens = match.groups()
+    return f"prompt is too long: {prompt_tokens} tokens > {max_tokens} maximum"
+
+
 def _map_bedrock_exception(
     *,
     model: str,
@@ -821,6 +839,14 @@ def _map_bedrock_exception(
     exception_provider: str,
     extra_information: str,
 ) -> None:
+    if custom_llm_provider == "bedrock_mantle":
+        mantle_context_window_message = _get_bedrock_mantle_context_window_message(error_str)
+        if mantle_context_window_message is not None:
+            raise ContextWindowExceededError(
+                message=mantle_context_window_message,
+                model=model,
+                llm_provider=custom_llm_provider,
+            )
     if (
         "too many tokens" in error_str
         or "expected maxLength:" in error_str

tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py

@@ -785,3 +785,27 @@ def test_bedrock_mantle_400_maps_to_bad_request():
 
     assert excinfo.value.status_code == 400
     assert "Invalid 'input'" in excinfo.value.message
+    assert type(excinfo.value) is litellm.BadRequestError
+
+
+def test_bedrock_mantle_context_overflow_maps_to_context_window_exceeded():
+    from litellm.llms.base_llm.chat.transformation import BaseLLMException
+
+    original_exception = BaseLLMException(
+        status_code=400,
+        message=(
+            '{"error":{"code":"validation_error",'
+            '"message":"prompt tokens (1055489) exceed model maximum (1050000) for openai.gpt-5.6-sol",'
+            '"param":null,"type":"invalid_request_error"}}'
+        ),
+    )
+
+    with pytest.raises(litellm.ContextWindowExceededError) as excinfo:
+        exception_type(
+            model="openai.gpt-5.6-sol",
+            original_exception=original_exception,
+            custom_llm_provider="bedrock_mantle",
+        )
+
+    assert excinfo.value.status_code == 400
+    assert "prompt is too long: 1055489 tokens > 1050000 maximum" in excinfo.value.message