Skip to content

Switching a rate-limited Codex thread to Antigravity can stay "Working" with nothing happening (not reproduced) #39412

Closed as not planned

Description

@juliusmarminge

What users report

From feedback on the Antigravity integration (post):

I couldn't get it to hand off, I had a task going with codex, and I hit a rate limit, so I switched the model, and it just wouldn't go. Said "working" but nothing was happening

Not reproduced

main @ 6b6950adfe2, Linux x86_64, agy CLI 1.5.12, isolated vp run dev state.

I tried the plain switch without a rate limit: a thread completed a turn on Codex (gpt-6-luna), then I switched the composer model to Antigravity Gemini 3.8 Flash (Low) and sent a follow-up. It worked: a handoff item completed, then the Antigravity turn read a file and answered (about 15 s end to end).

So the plain provider switch works. The difference in the report is that the Codex thread had hit a usage/rate limit mid-task before the switch. I could not trigger a real Codex rate limit here. That state (a run stopped by a usage limit, possibly waiting for its reset or auto-resume, see #24511 / #38498) is the likely trigger.

To investigate

Fixed looks like

After a provider hits its usage limit, switching the thread to another provider and sending a message starts a turn on the new provider promptly, or shows a clear reason why it can't.

Activity

  1. juliusmarminge commented on Nov 20, 2026

    @juliusmarminge
    MemberAuthor

    Tried to reproduce on main @ 29b42d4 (Linux x86_64, native agy 1.5.12). Not reproduced.

    To put a Codex thread in the real usage-limited state without exhausting an account, I ran a dev Codex instance whose provider points at a local endpoint that answers every /v1/responses call with 429 {"error":{"type":"usage_limit_reached","resets_in_seconds":7200,…}}.

    1. New thread on that Codex instance (gpt-6-luna) → run 1 fails at 01:20:33 with Codex's real message: "You've hit your usage limit. Visit https://chatgpt.com/settings/usage …".
    2. In the same thread, switch the composer to Antigravity → Gemini 3.8 Flash and send.
    3. Run 2 starts on the Antigravity instance at 01:20:57 and completes at 01:21:00. Items: handoff completed → user_message → assistant_message ("handed") → checkpoint. Nothing is left in "Working".

    So a provider switch after a Codex usage limit works on current main, including the handoff. The report differs in that the limit hit mid-task during a longer Codex turn, on an unknown T3 version. If it still happens, the next useful data is the T3 version and whether the thread showed "Limited"/auto-resume before the switch. Suggest closing unless the reporter can reproduce on a current build.

  2. juliusmarminge commented on Nov 20, 2026

    @juliusmarminge
    MemberAuthor

    Closing: not reproducible on current main (agy 1.5.12); test details above. Reopen with a repro on a current build if it still happens.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions