[Sadok Barbouche (cr-sbarbouche)]

Motivation & Context

response.output_item.done is where every other hosted-tool item type (shell_call, computer_call, image_generation_call, mcp_call, web_search_call/file_search_call, custom_tool_call, tool_search_call, function_call_output) gets its final, complete state captured during streaming - a couple of the existing branches even say so explicitly in their comments. code_interpreter_call is missing from that dispatch entirely. The only other place a code interpreter call's outputs (logs, generated images) are read from is response.output_item.added, which fires before the call has run, so a streamed code_interpreter_tool_result.outputs was always empty no matter what the sandbox actually produced. See #9743 for a runnable repro against the real client.

Description & Review Guide

• What are the major changes? Added a code_interpreter_call branch to the response.output_item.done dispatch in OpenAIChatClient, mirroring the existing non-streaming parser's and response.output_item.added's handling of the same item shape (parse outputs into code_interpreter_tool_result, parse code into code_interpreter_tool_call). Both are tagged merge_mode="replace", matching how code_interpreter_call_code.done already tags the code, since a normal stream can send both events for the same call and the merge layer needs to recognize the second one as a repeat of the authoritative final value rather than append onto the first.
• What is the impact of these changes? Streamed code interpreter calls now surface their logs/images the same way a non-streaming get_response() call already does. No change for callers who never inspect code_interpreter_tool_result.outputs.
• What do you want reviewers to focus on? Whether replace is the right mode for outputs specifically. Outputs are never streamed incrementally (no delta event for them), so the done item's list is always the complete, authoritative set; I treated it the same as the already-done code rather than append, but wanted to flag the choice explicitly.

Related Issue

Fixes #9743

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🟡 Changes recommended

The generated-image behavior addressed by this fix still needs completed-item regression coverage under the repository's testing standards.

1 open issue

• Add completed-item image coverage to from_updates tests · New

What changed in this PR

This PR addresses #9743 by capturing completed code interpreter outputs in the Python OpenAI client's streaming parser.

Changes:

• Parse logs, images, and final code from response.output_item.done.
• Use replacement merging for completed calls and results to avoid duplicating final code.
• Add regression tests for streamed log capture and code deduplication.

File	Description
python/​packages/​openai/​tests/​openai/​test_openai_chat_client.py	Adds completed-item and streaming-sequence regression tests.
python/​packages/​openai/​agent_framework_openai/​_chat_client.py	Captures final code interpreter outputs and code during streaming.

_{Review effort: Balanced}