PR 4525 --experimental-json dramatically reduces the amount of information compared to --json

What feature would you like to see?

I would like to see information from the deprecated --json returned such as tool call arguments and tool call results. This pr removed very useful information from 'msg' https://github.com/openai/codex/pull/4525

I leave you with the codex change to my code which parsed the original json and its own feelings about the change. Note the sad face emoji from codex itself!

● I've rewritten the decode_event function to handle the new Codex JSON format from PR #4525. Here are the key changes:

  What Changed

  1. New event structure: Events now use type field directly (not nested in msg)
    - thread.started, turn.started, turn.completed
    - item.started, item.completed
  2. Item-based architecture: Different content types are now items with item_type:
    - assistant_message → text response to user
    - error → error messages
    - reasoning → internal reasoning (displayed with 💭)
    - mcp_tool_call → MCP tool invocations (no args/results in new format 😢)
    - command_execution, file_change, web_search, todo_list → placeholders for future data
  3. Token usage: Now comes from turn.completed event's usage field
  4. Loss of information: The new format sadly doesn't include tool arguments or results, so we can only report that a tool was called with server/tool
  name.

  The function now properly handles all the item types you mentioned and maps them to the appropriate types.

Additional information

_No response_