Skip to content

OpenAIConverseConverter restores an empty history when one blob or text payload fails to parse #751

Open

Description

@abhiramshaji-cyber

Hi team, hope you're doing well!

This is a follow up from #743, where @kimnamu spotted it while testing #749. Thank you, @kimnamu!

Describe the bug

With converter=OpenAIConverseConverter, one stored event that the converter can't parse makes AgentCoreMemorySessionManager.list_messages return [], so the restored agent starts with no history, valid messages included. Three shapes trigger it on current main:

  • a dict blob, the shape in create_blob_event's docstring example (client.py L589 to L596)
  • the blob "[3, 4]"
  • deeply nested conversational text

#749 fixes the same kind of gap in AgentCoreMemoryConverter but doesn't touch converters/openai.py, so the branch behaves the same as main here.

To Reproduce

  1. list_messages passes every listed event to self.converter.events_to_messages and catches any exception by logging it and returning [] (session_manager.py L795 to L814).
  2. In OpenAIConverseConverter.events_to_messages (converters/openai.py L159 to L186):
    • a dict blob reaches json.loads(payload_item["blob"]), which raises TypeError, and the blob except only catches (json.JSONDecodeError, ValueError)
    • "[3, 4]" passes the list of length 2 check, then json.loads(blob_data[0]) gets 3 and raises TypeError
    • deeply nested text makes json.loads(conv["content"]["text"]) raise RecursionError, which the conversational except (json.JSONDecodeError, KeyError, ValueError) doesn't catch
  3. The exception escapes events_to_messages, list_messages returns [], and the agent starts with an empty history.

Minimal repro

Mocked MemoryClient, no AWS needed. Tested on main at 8467c9a (1.27.0) with strands-agents 1.56.0:

import json
from unittest.mock import patch

from bedrock_agentcore.memory.integrations.strands.config import AgentCoreMemoryConfig
from bedrock_agentcore.memory.integrations.strands.converters.openai import OpenAIConverseConverter
from bedrock_agentcore.memory.integrations.strands.session_manager import AgentCoreMemorySessionManager


def stored(role, text):
    return {"payload": [{"conversational": {"role": role.upper(), "content": {"text": json.dumps({"role": role, "content": text})}}}]}


config = AgentCoreMemoryConfig(memory_id="mem", session_id="sess", actor_id="actor")
with patch("bedrock_agentcore.memory.integrations.strands.session_manager.MemoryClient") as client, patch("boto3.Session"):
    client.return_value.list_events.return_value = []
    manager = AgentCoreMemorySessionManager(config, converter=OpenAIConverseConverter)

history = [stored("assistant", "hi there"), stored("user", "hello")]
odd = {
    "dict blob": {"payload": [{"blob": {"note": "x"}}]},
    "blob [3, 4]": {"payload": [{"blob": "[3, 4]"}]},
    "deeply nested text": {"payload": [{"conversational": {"role": "USER", "content": {"text": "[" * 100000 + "]" * 100000}}}]},
}

manager.memory_client.list_events.return_value = history
print("valid events only:", len(manager.list_messages("sess", "agent")))

for name, event in odd.items():
    manager.memory_client.list_events.return_value = [event] + history
    print(f"plus one {name}:", len(manager.list_messages("sess", "agent")))

Output:

valid events only: 2
Failed to list messages from AgentCore Memory: the JSON object must be str, bytes or bytearray, not dict
plus one dict blob: 0
Failed to list messages from AgentCore Memory: the JSON object must be str, bytes or bytearray, not int
plus one blob [3, 4]: 0
Failed to list messages from AgentCore Memory: maximum recursion depth exceeded while decoding a JSON array from a unicode string
plus one deeply nested text: 0

Expected behavior

The one payload that can't be parsed gets skipped with a log line, and the rest of the history is restored, the same way the converter already skips a conversational payload that isn't JSON.

Related: blob content in the ERROR log

The blob except logs the whole payload: logger.error("Failed to parse blob payload: %s", payload_item) (converters/openai.py L177). Once TypeError is caught there, dict blobs land on that line and their content goes into the ERROR log. #749 switched the matching line in AgentCoreMemoryConverter to log only the event id, logger.error("Failed to parse blob content in event %s", event.get("eventId")), and the same would work here.

Environment

  • bedrock-agentcore: main at 8467c9a (1.27.0)
  • strands-agents: 1.56.0
  • Python 3.10, macOS

Additional context

A possible fix: catch TypeError and RecursionError in both branches, log only the event id, and continue, mirroring #749. Does that sound right to you, or would you rather handle it differently? I'm happy to send a PR, or fold it into #749 if you'd prefer both converters in one place. Thanks so much!

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions