Description
What happened
In an Antigravity thread on T3 Code, once a turn is started by a background-task / subagent wake (post-settle continuation), steering stops working for the rest of that turn — even after no subagents are currently running.
Clicking Steer on a queued message cancels the queued entry and immediately bounces the same message back into the queue. Steering remains impossible until the wake turn finishes on its own or the user manually clicks Stop and starts a new turn.
Diagnosis
PR #39165 (fix(server): steering an Antigravity turn no longer stops its subagents) added native prompt steering (AcpAdapterV2.steerWithPrompt) for Antigravity (agy >= 1.5.0), but unconditionally declines steers whenever context.promptSequence === 0 (packages/provider-acp/src/server/adapter.ts:9558-9565 at commit 59651c1761):
// A wake turn has no prompt of its own to steer; its reply ends the run soon.
if (context.promptSequence === 0) {
return yield* new ProviderAdapter.ProviderAdapterSteerDeclinedError({
driver,
providerThreadId: turnInput.providerThread.id,
providerTurnId: turnInput.providerTurnId,
});
}Why this breaks Antigravity threads in practice:
- When a previous run settles and Antigravity later wakes autonomously on a background command or subagent completion,
startTurnUnlockedenters theisContinuationTurn === truepath (packages/provider-acp/src/server/adapter.ts:9400-9460), which drains wake notifications without issuingsession/prompt. As a result,context.promptSequenceremains0for the entire lifetime of that run. - The assumption "its reply ends the run soon" does not hold for Antigravity: once woken, Antigravity frequently continues running a full autonomous multi-step loop (commands, builds, file edits, and further subagents). In the observed thread, continuation Run 333 ran for 28 minutes (
13:35:11Zto14:03:00Z) withpromptSequence === 0. - Every steer attempt during that turn fails with
ProviderAdapterSteerDeclinedError→ProviderTurnControlService.steerwraps it withsteerDeclined: true(apps/server/src/orchestration-v2/ProviderTurnControlService.ts:487-496) →EffectWorkercatchessteerDeclinedand silently re-dispatchescommand:steer-follow-up:...withdispatchMode: { type: "queue_after_active" }(apps/server/src/orchestration-v2/EffectWorker.ts:361-409). - Secondary impact observed in
orchestration_v2_effect_outbox: an automaticdelegated-completionsteer (provider-turn.steer) targeting the same wake turn also hit thesteer-follow-uppath and ended up failing in the outbox withOrchestratorCommandPreviouslyRejectedError. - Once the user manually clicked Stop (
14:03:00Z), the next queued user message started a normal prompt turn (promptSequence = 1), andAcpAdapterV2.steerWithPromptimmediately succeeded on the very next steer (14:03:12Z).
Steps to reproduce
- Start a thread using Antigravity (
agy 1.5.13) on a build containing PR fix(server): steering an Antigravity turn no longer stops its subagents #39165 (0.1.11-nightly.20261120.4979). - Run a turn that launches background work / subagents and completes, then lets Antigravity wake autonomously into a post-settle continuation run (
Background task completed.,creationSource: "provider"). - While Antigravity continues working in that wake turn (executing tools/commands or subagents), send a message (it queues) and click Steer.
- Observe that the queued message is cancelled and immediately re-queued instead of steering the running turn; repeat even after all subagents finish — steering continues to bounce back to the queue until the turn ends or Stop is pressed.
Version
0.1.11-nightly.20261120.4979 (commit 59651c1)
Environment
Server: Linux x64 (7.2.15-1-omarchy), Node v24.22.1, Antigravity CLI (agy) 1.5.13; Client: T3 Code Desktop 0.1.11-nightly.20261120.4979 on Windows connecting to remote Linux server
Evidence
# Thread timeline (UTC, redacted)
13:35:04.338Z Run 332 completed (normal prompt turn, last session/prompt succeeded at 13:32:09.963Z)
13:35:10.670Z Run 333 requested -> provider-turn:323 started at 13:35:11.452Z
userMessage: createdBy=agent, creationSource=provider, text="Background task completed."
notification: { source: { kind: "background_task" }, outcome: "updated" }
No session/prompt sent -> context.promptSequence === 0
# server.trace.ndjson during Run 333 (multiple user clicks on Steer)
2026-11-20T13:58:45.477Z AcpAdapterV2.steerWithPrompt exit=Failure cause=ProviderAdapterSteerDeclinedError
2026-11-20T13:58:45.477Z AcpAdapterV2.steerTurn exit=Failure cause=ProviderAdapterSteerDeclinedError
2026-11-20T13:58:45.433Z ProviderTurnControlService.steer exit=Failure cause=ProviderTurnControlError
-> EffectWorker dispatches command:steer-follow-up:effect:<id>:provider-turn.steer:... (dispatchMode: queue_after_active)
2026-11-20T13:58:49.939Z AcpAdapterV2.steerWithPrompt exit=Failure cause=ProviderAdapterSteerDeclinedError
2026-11-20T14:00:58.785Z AcpAdapterV2.steerWithPrompt exit=Failure cause=ProviderAdapterSteerDeclinedError
# After user pressed Stop on Run 333 at 14:03:00.513Z -> Run 339 started as normal prompt (provider-turn:324)
2026-11-20T14:03:12.473Z AcpAdapterV2.steerWithPrompt exit=Success
2026-11-20T14:03:12.368Z ProviderTurnControlService.steer exit=Success
Related issues
Regression / design gap from PR #39165 (fix(server): steering an Antigravity turn no longer stops its subagents). No existing open issue covers steerWithPrompt declining steers on long-running Antigravity wake/continuation turns (promptSequence === 0).
Fix applied or workaround
No local patch applied. Workaround: manually click Stop to interrupt the running Antigravity wake/continuation turn, then let the next turn start from a normal user prompt (promptSequence >= 1), after which steering works again.
Filed by
Antigravity (Gemini 4 Neon) via T3 Code triage
Activity