Skip to content

[Bug]: Antigravity native schedule tool trips 15m inactivity watchdog and poisons the session DB into an infinite 15m timeout loop #40646

Closed as duplicate

Description

@pointauction

Area

apps/server (Antigravity ACP adapter)

Summary

In T3 Code 0.1.9, the Antigravity ACP adapter sets promptInactivityTimeout: "15 minutes" (the only provider with this watchdog) and only pauses it while an ACP-visible foreground tool or subagent is active (AcpAdapterV2.promptInactivityPaused).

Antigravity's native schedule tool (step_type = 132) is not surfaced as an active ACP tool call, yet agy_acp_server keeps session/prompt open while waiting for the timer. As a result:

  1. Any native schedule timer >= 15 minutes (or a recurring cron) deterministically hits the 15-minute inactivity watchdog and kills agy_acp_server:

    Antigravity sent nothing for 15 minutes, so T3 Code stopped the turn and restarted the agent. Send a message to continue.

  2. Killing the ACP runtime mid-timer orphans the schedule step in antigravity-acp/conversations/<sessionId>.db at steps.status = 2 (CORTEX_STEP_STATUS_RUNNING) and protobuf field 4 = 0x20 0x02 inside step_payload.
  3. On every subsequent turn after session/load, agy_acp_server never reconciles the orphaned status = 2 row—even if the agent calls manage_task(Action="kill") (which returns "Task ... cancelled." without updating SQLite). Every future turn delivers its full assistant reply (step_type = 15, status = 3), never settles session/prompt, and fails 15 minutes later with the same error in an infinite loop.

Steps to reproduce

  1. Start a T3 Code thread using the Antigravity provider (e.g. gemini-4-neon).
  2. Ask the agent to wait 20 minutes before continuing (or run a workflow where the agent calls native schedule(DurationSeconds=1200, Prompt="...") and ends its tool calls to wait for the reactive wakeup).
  3. Wait 15 minutes.
  4. Observe T3 Code kill and restart the Antigravity runtime with:
    Antigravity sent nothing for 15 minutes, so T3 Code stopped the turn and restarted the agent. Send a message to continue.
  5. Send any short follow-up message in the same thread (e.g., "status?").
  6. Observe that the agent replies immediately, but the turn stays running with no active tools and fails again exactly 15 minutes later with the same error—and repeats on every subsequent turn.

Expected behavior

  • Long waits (CI/CD, remote jobs, user-requested delays) should not crash Antigravity threads at 15 minutes, or native schedule should be disabled in favor of T3 Code's own watch_pull_request / schedule_task MCP tools.
  • Restarting the ACP runtime after a watchdog timeout should never leave the session DB in a permanently wedged state where every future completed reply hangs for 15 minutes and fails.

Actual behavior

  • Observed 6 consecutive watchdog failures across 2 separate Antigravity threads (4 consecutive failures on one thread after a 27m schedule(DurationSeconds=1620) call; 2 consecutive failures on another thread after a 20m schedule(DurationSeconds=1200) call).
  • Inspecting <GEMINI_HOME>/antigravity-acp/conversations/<sessionId>.db on both threads confirmed a single orphaned step_type = 132 (schedule) row frozen at status = 2 (RUNNING), with step_payload prefix 0884012002 (field 1 = 132, field 4 = 2), even after multiple restarts and manage_task(Action="kill").
  • Flipping that row's status column and protobuf field 4 byte from 2 (RUNNING) to 6 (CANCELED, prefix 0884012006) while the DB is idle unsticks the session.
  • (Separately, a 7th 15-minute failure in the same 48-hour window was caused by two silent upstream 502 "The model is currently unreachable." steps step_type = 17, matching Antigravity silent 502s wait out the full 15m inactivity timer and fail to parse in AntigravityTurnOutcome #33583.)

Impact

Major degradation or frequent failure (any long-running CI/CD, scrape, deploy, or delayed-resume workflow on Antigravity/Gemini threads repeatedly fails and permanently wedges the thread until a fresh session is created or the SQLite DB is manually patched).

Version

  • T3 Code Desktop (Beta): 0.1.9
  • Antigravity ACP server (agy_acp_server.par): 1.4.5 (darwin_arm64)

Environment

  • OS: macOS 26.6.2 (arm64)
  • Provider / Model: Antigravity (gemini-4-neon)

Suggested fixes (T3 Code side)

  1. Disable native schedule in antigravitySessionMeta (quickest high-leverage fix): T3 Code already passes _meta.antigravity.disabledTools: ["run_workflow"] on Antigravity 1.4.5+. Adding "schedule" to disabledTools (or surfacing schedule as an active ACP tool that pauses promptInactivityPaused) prevents Antigravity from entering untracked in-prompt timer waits when T3 Code already provides watch_pull_request and schedule_task.
  2. Prevent or recover from orphaned status = 2 steps after watchdog teardown: When retireInactivePrompt tears down agy_acp_server, or before session/load, detect/reconcile orphaned status = 2 (RUNNING) schedule / background steps (or avoid failing an already-delivered step_type = 15, status = 3 reply).
  3. Make promptInactivityTimeout configurable per provider instance (or pause it when Antigravity reports active waiting/thinking), as also requested in [Bug]: Antigravity sent nothing for 15 minutes, so T3 Code stopped the turn and restarted the agent. Send a message to continue. #34160.

Workaround

  1. Configure an Antigravity PreToolUse hook in <GEMINI_HOME>/config/hooks.json with "matcher": "schedule" that blocks CronExpression and DurationSeconds > 600 (10m), steering the agent to use watch_pull_request, schedule_task, or <= 10m chained check-in hops.
  2. For already-wedged threads, once idle, back up <sessionId>.db and update orphaned step_type = 132, status = 2 rows (both steps.status and step_payload prefix 0884012002 -> 0884012006) to 6 (CANCELED), or start a fresh thread.

Related issues

Report prepared with Antigravity (gemini-4-neon).

Activity

  1. juliusmarminge commented on Nov 22, 2026

    @juliusmarminge
    Member

    This is the same problem as #34160, so I'm closing it in favour of that one. If it is different, say how and it will be reopened.

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