OpenAIResponsesCompactionSession: compacted items fail replay with "Unknown parameter: 'input[N].created_by'"

[omikader]

Bug Summary

After OpenAIResponsesCompactionSession compacts a session's history, the very next turn fails with a 400 on the Responses API:

openai.BadRequestError: Error code: 400 - {'error': {'message': "Unknown parameter: 'input[1].created_by'.", 'type': 'invalid_request_error', 'param': 'input[1].created_by', 'code': 'unknown_parameter'}}

This happens on the turn immediately following a successful compaction, every time — the compacted history is effectively unusable once produced.

Root cause

responses.compact() returns a ResponseCompactionItem (the output type), which carries a created_by field (openai/types/responses/response_compaction_item.py). But the input type for replaying a compaction item, ResponseCompactionItemParam (openai/types/responses/response_compaction_item_param.py), has no created_by field at all — only encrypted_content, type, id.

OpenAIResponsesCompactionSession._run_compaction_locked stores the raw output items (via _normalize_compaction_output_items, which just does item.model_dump(exclude_unset=True, warnings=False) — created_by survives because it's explicitly set on the constructed object, not merely a default). Those stored items later get replayed as input for the next turn via prepare_input_with_session() in agents/run_internal/session_persistence.py:399:

strip_internal_input_item_metadata(ensure_input_item_format(item)) for item in history

strip_internal_input_item_metadata() (agents/run_internal/items.py:715) only strips TOOL_CALL_SESSION_DESCRIPTION_KEY and TOOL_CALL_SESSION_TITLE_KEY today — created_by isn't in that list, so it rides along into the next request's input array and gets rejected.

Prior art

This is the same class of bug as #1911 (closed as a duplicate of #1882/#1883) — an output-only field leaking unstripped into replayed input, just for a different field (status).

It's also already fixed in the sibling JS SDK: openai/openai-agents-js#1744 ("fix(core): strip created_by from RunItem replay", merged by @seratch, who's also the author of the two most recent Python-side compaction fixes — #… preserve concurrent writes during compaction (0.22.1), normalize compacted Responses user inputs before session reuse (0.14.2)). The JS fix strips created_by from top-level items, providerData, and shell output chunks at the same "replay boundary" concept strip_internal_input_item_metadata() already exists for on the Python side — it just doesn't cover this field yet.

Repro

"""
Reproduces: BadRequestError - Unknown parameter: 'input[1].created_by' (code: unknown_parameter)
on the second turn of a run using OpenAIResponsesCompactionSession.

Requires: pip install openai==3.16.2 openai-agents[redis]==0.22.3
Set OPENAI_API_KEY in environment.
"""
import asyncio
from agents import Agent, OpenAIResponsesModel, RunConfig, Runner, function_tool
from agents.memory.openai_responses_compaction_session import OpenAIResponsesCompactionSession
from agents.memory.sqlite_session import SQLiteSession
from openai import AsyncOpenAI


@function_tool
async def lookup(question: str) -> str:
    return "42"


agent = Agent(name="test", instructions="Answer using the lookup tool.", tools=[lookup])


async def main():
    client = AsyncOpenAI()
    underlying = SQLiteSession(session_id="repro-created-by")

    session = OpenAIResponsesCompactionSession(
        session_id="repro-created-by",
        underlying_session=underlying,
        client=client,
        model="gpt-5",
        compaction_mode="input",
        should_trigger_compaction=lambda ctx: len(ctx["compaction_candidate_items"]) >= 2,
    )

    run_config = RunConfig(model=OpenAIResponsesModel("gpt-5", openai_client=client), tracing_disabled=True)

    # Turn 0 succeeds and auto-triggers compaction (threshold met). The stored
    # compaction item retains `created_by`, copied straight from the output type.
    result = await Runner.run(agent, "How many are there?", session=session, run_config=run_config)
    print("turn 0:", result.final_output)

    # Turn 1 fails: the stored compaction item (with `created_by` still attached)
    # gets replayed as `input`, and the Responses API rejects it.
    result = await Runner.run(agent, "And how many of those are urgent?", session=session, run_config=run_config)
    print("turn 1:", result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

Debug information

• Agents SDK version: 0.22.3 (latest on PyPI at time of filing)
• openai version: 3.16.2 (latest on PyPI at time of filing)
• Model: reproduced via a corporate Azure OpenAI-backed gateway (litellm) fronting the model; the bug itself is client-side (the SDK constructs the invalid input payload before it ever reaches any provider), so it should reproduce identically against the direct OpenAI API — I haven't verified against a bare OPENAI_API_KEY myself, flagging that as unconfirmed rather than assuming.

Expected behavior

A turn following a compaction should replay cleanly, the same way the JS SDK now handles it after #1744.

[leilei3167]

I'd like to take a look at this — planning a small fix in strip_internal_input_item_metadata so compaction replay drops created_by (same idea as the JS SDK strip). Happy to open a PR if that sounds right.

[leilei3167]

Quick follow-up: CONTRIBUTING limits PRs to collaborators, so I'll leave the patch for maintainers.

The leak is created_by on compacted Responses items surviving into the next turn's input. A small created_by pop in strip_internal_input_item_metadata (same replay boundary as the JS SDK) plus a regression on a compaction item dict should cover it. Happy to answer questions about the repro if helpful.