Skip to content

Antigravity quota hit after a background-task wake marks the run completed and never auto-resumes #40229

Closed as not planned

Description

@zakius

What happened

Two Antigravity threads (model gemini-4-neon) hit the individual quota while Antigravity was working on its own after a background task finished. Each thread showed the quota notice as a normal assistant message:

⚠ Individual quota reached. Please upgrade your subscription to increase your limits. Resets in 4h1m47s.

T3 Code recorded both runs as completed. There was no usage-limit banner, and Resume after usage limits (autoResumeLimitedThreads: true) never continued either thread when the quota reset at 02:41Z. One thread sat idle for about 6 hours until the user typed "continue" by hand.

Diagnosis

Two separate gaps. Either one alone stops auto-resume.

1. The quota check never runs for a turn that continues after its prompt settled. watchPromptOutcome (watchAntigravityPromptOutcome in apps/server/src/provider/acp/AntigravityTurnOutcome.ts) is called only around session/prompt and steer prompts in packages/provider-acp/src/server/adapter.ts. With enablePostSettleContinuation / deferFinalizeForBackgroundWork, a finished background command wakes the agent with no prompt open. If the quota runs out then, nothing reads the type-17 error step, and the deferred finalize marks the run completed.

2. Antigravity CLI 1.5.13's error step doesn't parse as a usage limit with a reset time. It writes RESOURCE_EXHAUSTED (code 429): … plus JSON ("reason": "QUOTA_EXHAUSTED", "quotaResetTimeStamp": "2026-11-21T02:41:38Z", "retryDelay": "14507.97s"), not the Go map[...] form. At f0a2aa10f516 no regex matches, so it would be a provider_error. On main, STATUS_ERROR makes it usage_limit, but resetAtFrom still finds no reset time, so limitRecoveryCommand never arms.

Steps to reproduce

  1. Turn on Resume after usage limits.
  2. In an Antigravity thread, give a long task that runs a background command (for example a test suite) and waits for it.
  3. Let the first session/prompt settle while the background command runs.
  4. Let the account's individual quota run out after the background task wakes the agent.
  5. The thread shows the ⚠ Individual quota reached … Resets in … notice as an assistant message. The run is marked completed. No banner appears, and nothing resumes the thread after the reset.

Version

0.1.12-nightly.20261120.5009 (server, running as the background service). The source checked is f0a2aa10f516, compared against main at the time of triage. The triage was started from T3 Code Desktop 0.1.11-nightly.20261119.4928.

Environment

  • Fedora Linux 45, x64 (kernel 7.2.14)
  • T3 Code server as a systemd service; the user messages were sent from the mobile app
  • Antigravity CLI (agy) 1.5.13, model gemini-4-neon
  • autoResumeLimitedThreads: true, snoozeLimitedThreads: true

Evidence

- Both runs have `status: completed`. Their last assistant message is the quota notice (22:39:23Z and 22:39:28/22:39:50Z, 2026-11-20).
- First thread's provider event log: `session/prompt` succeeded at 22:19:55Z, then `session/close` at 23:11:50Z. No prompt was open when the quota was hit.
- The native session log has two type-17 error steps (idx 226 and 228) containing `RESOURCE_EXHAUSTED (code 429)`, `"reason": "QUOTA_EXHAUSTED"` and `"quotaResetTimeStamp": "2026-11-21T02:41:38Z"`.
- The next user message ("continue") came at 04:48Z, about 2 hours after the reset. No `limit-resume` message was ever sent.

Related issues

Fix applied or workaround

No local fix. Workaround: after the reset time, send "continue" by hand.

Filed by

Claude Opus 5.5 (claude-opus-5-5) in Claude Code, from a T3 Code triage session.

Activity

  1. juliusmarminge commented on Nov 21, 2026

    @juliusmarminge
    Member

    I tried to reproduce this on the current build and could not.

    Commit origin/main baf7e5b. No live run: it needs a Google-signed-in Antigravity account to run out of its individual quota after a background-task wake, and the VM's API-key hub cannot produce that. Gap 2, unit test: added the reporter's Antigravity 1.5.13 type-17 error step (RESOURCE_EXHAUSTED (code 429): Individual quota reached... followed by Google's JSON body with "reason": "QUOTA_EXHAUSTED", "quotaResetTimeStamp": "2026-11-21T02:41:38Z", "retryDelay": "14507.97s") as a turn in the existing 'fails only turns whose own steps end in an upstream error' case of apps/server/src/provider/acp/AntigravityTurnOutcome.test.ts. vp test run src/provider/acp/AntigravityTurnOutcome.test.ts (from apps/server): failure.class is usage_limit (through STATUS_ERROR), but the assertion that failure.resetAt equals 2026-11-21T02:41:38.000Z fails with actual undefined. Gap 1, code reading only: scheduleDeferredFinalize in packages/provider-acp/src/server/adapter.ts (about line 8730) waits out the debounce and calls finalizeTurn(context, context.promptSettledStatus ?? 'completed') without calling the flavor's watchPromptOutcome check again. That check runs only around session/prompt and steer prompts, so an error step written after a background wake is never read. Open PR #40232 includes a regression test for this that it reports failing on main ('expected completed to equal failed').

    Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.

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

    via-triageFiled through npx t3 triage

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions