codex - 💡(How to fix) Fix Add agent/subagent metadata to SDK streaming ThreadEvent items [1 comments, 2 participants]

For clients that render live activity timelines, grouped logs, or debugging views, it is important to know which agent produced a message, reasoning item, tool call, todo update, or error. Today, the TypeScript SDK exposes item types such as agent_message, reasoning, command_execution, mcp_tool_call, todo_list, and error, but those streamed items do not appear to include stable identity metadata for the active agent/subagent.

TL;DR

The @openai/codex-sdk should be updated to include explicit metadata for the agent or subagent responsible for each ThreadEvent item in streaming events.

Guidance

• Review the existing ThreadEvent item types (e.g., agent_message, reasoning, command_execution) to determine if any current fields can be leveraged to identify the responsible agent.
• Consider adding the suggested fields (agent_id, agent_name, agent_kind, parent_agent_id, parent_item_id, or parent_thread_id) to the ThreadEvent items to provide stable identity metadata.
• Evaluate the relationship between this request and the related issue #20675, which covers similar root/subagent metadata for hook payloads.
• Investigate the feasibility of implementing a workaround that utilizes event ordering heuristics or transcript parsing to attribute streamed activity to the responsible agent.

Example

No code snippet is provided as the issue is focused on feature enhancement rather than a specific code-related problem.

Notes

The implementation of this feature may require significant changes to the @openai/codex-sdk and its underlying infrastructure. The suggested fields may need to be standardized across different agent types and subagents.

Recommendation

Apply workaround: While waiting for the @openai/codex-sdk to be updated, client applications can attempt to attribute streamed activity using event ordering heuristics or transcript parsing as a temporary solution.