Bug: _chunk_bytes_to_sse raises AttributeError when chunk.get('bytes') is None in Bedrock stream decoder #1975

Activity

1. kimnamu commented on Oct 16, 2026

   

   kimnamu

   on Oct 16, 2026

   More actions

   Thanks for the clear write-up with a one-line repro, and for opening #1976 so quickly.

   As far as I can tell, None can't reach the helper through a real stream on main (bcd55fe, v1.15.0). botocore's JSON parser skips null and missing members, so {}, {"bytes": null} and an empty body all parse to {'chunk': {}}, and if not chunk: return None (_stream_decoder.py L65) drops them first. botocore 1.31.57 (the pyproject.toml floor) parses them the same way.

   # main @ bcd55fe
   {} / {'bytes': null} / empty body / unknown type ['message_start', 'message_stop']
   {'bytes': ''}                                    ['message_start'] then JSONDecodeError
   # #1976 @ bfe041a
   {} / {'bytes': null} / empty body / unknown type ['message_start', 'message_stop']
   {'bytes': ''}                                    ['message_start'] then JSONDecodeError
   # main + call-site falsy check (below)
   {} / {'bytes': null} / empty body / unknown type ['message_start', 'message_stop']
   {'bytes': ''}                                    ['message_start', 'message_stop']
   

   The one input I found that gets past that line is {"bytes": ""}, and the isinstance check in #1976 lets b"" through. If you'd like to keep a guard, a falsy check at the call site mirrors the line above and needs no pyright: ignore:

           raw = chunk.get("bytes")
           if not raw:
               return None
   
           return _chunk_bytes_to_sse(raw)

   With it, uv run pyright src/anthropic/lib/bedrock is clean and the four Bedrock/AWS test files pass (205). Happy to leave it with you.

   Repro script

   import json, base64, struct, binascii
   
   import httpx2
   from anthropic import AnthropicBedrock
   
   
   def frame(event_type: str, body: bytes) -> bytes:  # same framing as tests/lib/test_bedrock.py::_eventstream_frame
       headers = {":message-type": "event", ":event-type": event_type, ":content-type": "application/json"}
       h = b"".join(bytes([len(k)]) + k.encode() + b"\x07" + struct.pack(">H", len(v)) + v.encode() for k, v in headers.items())
       prelude = struct.pack(">II", 16 + len(h) + len(body), len(h))
       msg = prelude + struct.pack(">I", binascii.crc32(prelude)) + h + body
       return msg + struct.pack(">I", binascii.crc32(msg))
   
   
   def chunk(payload: dict) -> bytes:
       return frame("chunk", json.dumps({"bytes": base64.b64encode(json.dumps(payload).encode()).decode()}).encode())
   
   
   start = chunk({"type": "message_start", "message": {"id": "msg_1", "type": "message", "role": "assistant", "content": [],
                  "model": "claude", "stop_reason": None, "stop_sequence": None, "usage": {"input_tokens": 5, "output_tokens": 1}}})
   stop = chunk({"type": "message_stop"})
   cases = {
       "{} / {'bytes': null} / empty body / unknown type": frame("chunk", b"{}") + frame("chunk", b'{"bytes": null}')
       + frame("chunk", b"") + frame("somethingNew", b"{}"),
       "{'bytes': ''}": frame("chunk", b'{"bytes": ""}'),
   }
   for name, middle in cases.items():
       body = start + middle + stop
       transport = httpx2.MockTransport(lambda request, body=body: httpx2.Response(
           200, content=body, headers={"content-type": "application/vnd.amazon.eventstream"}))
       client = AnthropicBedrock(aws_region="us-east-1", aws_access_key="k", aws_secret_key="s",
                                 http_client=httpx2.Client(transport=transport))
       seen = []
       try:
           for event in client.messages.create(model="anthropic.claude-x", max_tokens=8, stream=True,
                                               messages=[{"role": "user", "content": "hi"}]):
               seen.append(event.type)
           print(f"{name:48} {seen}")
       except Exception as e:
           print(f"{name:48} {seen} then {type(e).__name__}")

   # EventStreamJSONParser + bedrock-runtime ResponseStream shape, botocore 1.31.57
   b'{}'              -> {'chunk': {}}
   b'{"bytes": null}' -> {'chunk': {}}
   b''                -> {'chunk': {}}
   b'{"bytes": ""}'   -> {'chunk': {'bytes': b''}}
   
2. chauvuusvn commented on Oct 16, 2026

   

   chauvuusvn

   on Oct 16, 2026

   Author

   More actions

   Thanks for the great feedback and the thorough analysis, @kimnamu!

   I've updated PR #1976 following your suggestion:

   • Guarded raw = chunk.get("bytes") at the call site with if not raw: return None, cleanly dropping both None and empty b"".
   • Restored the original clean signature for _chunk_bytes_to_sse(raw: bytes) without any pyright: ignore comments.
   • Added a regression test (test_stream_skips_empty_and_null_bytes_chunk) simulating frames with both empty and null bytes.

   All Bedrock tests and type checks are passing cleanly. Appreciate the guidance!

   ❤️React with ❤️1Reacted by Lucas Kim
3. chauvuusvn commented on Oct 21, 2026

   

   chauvuusvn

   on Oct 21, 2026

   Author

   More actions

   Thanks @kimnamu for the thorough investigation and the clean call-site approach! I've updated #1976 accordingly with the falsy check (if not raw: return None), removing the pyright ignore and adding the regression tests for empty and null byte frames. All tests pass locally clean.

   ❤️React with ❤️1Reacted by Lucas Kim