TL;DR: Codex skills can’t raise a modal to stop and ask you at a fork (it’s behind manual Plan Mode). Claude Code lets a skill do it. Easy fix, already half-built — upvote #11180.
I moved one of my agent workflows onto Codex. The model’s genuinely great. But I keep hitting one wall.
You tell it “get the tests passing.” A test expects an order total of $42; the code now produces $46.20 — broken code or stale test? The agent can’t know (the right number’s in your head, not the repo), so it picks whatever goes green, edits the test, and ships the bug. Green check the whole way.
The fix is obvious: at that fork, stop and ask. A Codex skill can ask — but only in chat, which you scroll past and answer with a vague half-sentence. What a workflow actually needs is a modal: a structured checkpoint it can branch on.
Codex already has it — request_user_input, plus Plan Mode’s reviewable plan. The catch: the question modal is gated to Plan Mode, Plan Mode is user-only (a skill can’t enter it — verified on 0.133.0), and the flag that unlocks it in default mode (default_mode_request_user_input) is off and “under development.”
So the gap isn’t capability — it’s permission. Two ways to close it:
- let a skill enter Plan Mode (an
EnterPlanMode/ExitPlanModetool — GitHub issue #11180) - ship
default_mode_request_user_inputon by default.
Not asking to weaken autonomy — keep it author-placed, sparing, human-approved. The skill only raises the checkpoint; you still decide.
If you’ve hit this too, upvote the GitHub issue #11180 — features get prioritized by issue upvotes.