fix(items): strip output-only created_by when replaying model output as input - #4308

Summary

ModelResponse.to_input_items() replays each model output item back as a Responses API input item. Several output item types carry a server-assigned, output-only created_by field that is not part of the input-item schema:

• apply_patch_call and shell_call (supported ApplyPatchTool / ShellTool calls)
• function_call_output, custom_tool_call_output, computer_call_output (tool-call outputs)
• tool_search_call / tool_search_output

The replay sanitizer in _output_item_to_input_item only stripped created_by for the two tool_search types, so every other type leaked created_by straight back to the API on replay.

created_by is not accepted by any input-item param in the OpenAI SDK schema, and the SDK already strips it at resolution time for shell/apply_patch calls (turn_resolution.py) and for tool_search (_tool_search_item_to_input_item). The direct ModelResponse.to_input_items() replay path bypassed that resolution-time stripping, leaving the field in place.

Fix: strip created_by for every replayed item in _output_item_to_input_item, generalizing the existing tool_search behavior. created_by is never read by any SDK logic (it is only ever popped), so dropping it on replay is safe.

Before:

resp = ModelResponse(output=[apply_patch_call_with_created_by], usage=Usage(), response_id=None)
resp.to_input_items()
# [{'id': 'apc_1', 'call_id': 'call_1', 'operation': {...}, 'status': 'completed',
#   'type': 'apply_patch_call', 'created_by': 'program_1'}]   # <-- leaked

After:

# [{'id': 'apc_1', 'call_id': 'call_1', 'operation': {...}, 'status': 'completed',
#   'type': 'apply_patch_call'}]   # created_by stripped

The behavior is otherwise unchanged: items are still converted via the same dict(...) / model_dump(exclude_unset=True) logic, and the AgentsException fallback for unexpected types is preserved. The only caller affected is ModelResponse.to_input_items(); the two other callers (_fingerprint_for_tracker, _anonymous_tool_search_fingerprint in oai_conversation.py) only pass tool_search items, which already routed through the created_by-stripping branch.

Test plan

• Added test_to_input_items_strips_created_by_for_non_tool_search_items in tests/test_items_helpers.py, which builds an apply_patch_call and a function_call_output output item with created_by set and asserts the field is dropped on replay. It fails on main (leaks created_by) and passes with this change.
• The existing test_to_input_items_for_tool_search_strips_created_by and all other to_input_items round-trip tests still pass.
• Commands run locally:
  • uv run pytest tests/test_items_helpers.py → 36 passed
  • uv run pytest tests/test_items_helpers.py tests/memory tests/test_agent_runner.py tests/test_run_state.py tests/test_server_conversation_tracker.py tests/test_programmatic_tool_calling.py → 903 passed
  • uv run ruff format --check and uv run ruff check → clean
  • uv run python .github/scripts/check_optional_truthiness.py src/agents → clean
  • uv run mypy src/agents/items.py → Success
  • uv run pyright src/agents/items.py tests/test_items_helpers.py → 0 errors

Not run: the full test matrix and integration/coverage profiles (they require provider credentials / live services, which are unrelated to this change).

Issue number

N/A — no existing issue.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR