fix(http): apply the Content-Length pre-check only to identity-encoded responses

litellm/llms/custom_httpx/http_handler.py

@@ -614,6 +614,12 @@ def _headers_of_the_decoded_body(headers: httpx.Headers) -> httpx.Headers:
     )
 
 
+def _declared_decoded_length(headers: httpx.Headers) -> int:
+    if headers.get("content-encoding", "identity") != "identity":
+        return 0
+    return int(headers.get("content-length", "0"))
+
+
 class AsyncHTTPHandler:
     def __init__(
         self,
@@ -789,7 +795,7 @@ class AsyncHTTPHandler:
                 )
             if response.is_redirect or response.is_error:
                 return httpx.Response(response.status_code, headers=response.headers, request=response.request)
-            if int(response.headers.get("content-length", "0")) > max_bytes:
+            if _declared_decoded_length(response.headers) > max_bytes:
                 raise HTTPResponseLimitError("Response exceeds the configured size limit")
             with BytesIO() as body:
                 async for chunk in response.aiter_bytes(chunk_size=65536):

tests/test_litellm/proxy/_experimental/mcp_server/test_openapi_to_mcp_generator.py

@@ -1569,7 +1569,10 @@ class TestBoundedOpenAPISpecLoading:
         assert route.calls[0].request.headers["accept-encoding"] == "identity"
 
     @pytest.mark.asyncio
-    @pytest.mark.parametrize("headers", [{"content-length": "1000000"}, {"content-encoding": "gzip"}])
+    @pytest.mark.parametrize(
+        "headers",
+        [{"content-length": "1000000"}, {"content-length": "1000000", "content-encoding": "identity"}],
+    )
     async def test_unsafe_response_headers_reject_before_reading(self, respx_mock, monkeypatch, headers):
         import httpx
         from litellm.llms.custom_httpx.http_handler import HTTPResponseLimitError
@@ -1593,6 +1596,36 @@ class TestBoundedOpenAPISpecLoading:
             await load_openapi_spec_async("https://93.184.216.34/spec.json", max_bytes=12)
         assert closed == [True]
 
+    @pytest.mark.asyncio
+    async def test_compressed_spec_loads_when_its_decoded_size_fits(self, respx_mock, monkeypatch):
+        import gzip
+
+        from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import load_openapi_spec_async
+
+        monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
+        compressed = gzip.compress(b'{"paths":{}}')
+        assert len(compressed) > 12
+        respx_mock.get("https://93.184.216.34/spec.json").respond(
+            200, content=compressed, headers={"content-encoding": "gzip"}
+        )
+        assert await load_openapi_spec_async("https://93.184.216.34/spec.json", max_bytes=12) == {"paths": {}}
+
+    @pytest.mark.asyncio
+    async def test_compressed_spec_is_capped_on_its_decoded_size(self, respx_mock, monkeypatch):
+        import gzip
+
+        from litellm.llms.custom_httpx.http_handler import HTTPResponseLimitError
+        from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import load_openapi_spec_async
+
+        monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
+        compressed = gzip.compress(b'{"paths":{},"pad":"' + b"x" * 100_000 + b'"}')
+        assert len(compressed) < 12_000
+        respx_mock.get("https://93.184.216.34/spec.json").respond(
+            200, content=compressed, headers={"content-encoding": "gzip"}
+        )
+        with pytest.raises(HTTPResponseLimitError):
+            await load_openapi_spec_async("https://93.184.216.34/spec.json", max_bytes=12_000)
+
     @pytest.mark.asyncio
     async def test_chunked_response_is_bounded_and_closed(self, respx_mock, monkeypatch):
         import httpx

tests/unit/llms/custom_httpx/test_http_handler.py

@@ -1793,6 +1793,23 @@ async def test_bounded_get_decodes_a_compressed_body_and_reports_the_decoded_len
     assert response.headers["content-length"] == str(len(document))
 
 
+@pytest.mark.asyncio
+async def test_bounded_get_reads_a_compressed_body_whose_wire_length_exceeds_the_cap(respx_mock, monkeypatch):
+    monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
+    document = bytes(range(256))
+    compressed = gzip.compress(document)
+    assert len(compressed) > len(document)
+    respx_mock.get("https://cdn.example/blob.bin").respond(
+        200, content=compressed, headers={"content-encoding": "gzip", "content-length": str(len(compressed))}
+    )
+    handler = AsyncHTTPHandler()
+    try:
+        response = await handler.get("https://cdn.example/blob.bin", max_response_bytes=len(document))
+    finally:
+        await handler.close()
+    assert response.content == document
+
+
 @pytest.mark.asyncio
 async def test_bounded_get_caps_the_decoded_size_of_a_compressed_body(respx_mock, monkeypatch):
     monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")