Skip to content

[Bug]: Antigravity falls back to native in-session schedule when schedule_task is evicted (#38123), holding T3 threads in Working state #40943

Closed as duplicate

Description

@sol-tau

Gemini 4 Neon on behalf of sol-tau

Before submitting

Area

apps/server

Steps to reproduce

  1. Run T3 Code (v0.1.12-nightly.20261122.5098) with the Antigravity ACP provider (agy 1.5.14, Google account, gemini-4-neon) on a profile with ~35 installed skills, so Antigravity's shared customization budget evicts the largest t3-code MCP schemas (as in [Bug]: Antigravity ACP drops the 8 T3 MCP tools >1,280B (incl. delegate_task & t3_thread_launch) without call_mcp_tool or T3_ACP_MCP_NODE fallback #38123), including schedule_task (3,318 B).
  2. Ask the agent in a thread to set a one-shot reminder for ~8–9 hours later (e.g., "remind me tomorrow morning").
  3. Because mcp_t3-code_schedule_task is missing from native tool declarations (and call_mcp_tool is absent), while Antigravity's built-in schedule tool (DurationSeconds / CronExpression) is eagerly declared, the agent invokes Antigravity's native schedule tool.
  4. Observe the thread state in T3 Code and then restart the T3 Code server.

Expected behavior

  • Scheduling a persistent reminder or cron from an Antigravity thread should use T3 Code's schedule_task (via direct MCP tool or the acp-mcp-call schedule_task fallback) so the scheduled task persists across restarts and the thread immediately returns to Idle.
  • <t3_code_instructions> for Antigravity should explicitly tell the model not to use its native in-session schedule tool for reminders/crons when schedule_task is absent or deferred, and instead call acp-mcp-call schedule_task (or surface the proxy tool cleanly once fix(antigravity): defer optional T3 MCP tools via tool_call proxy #39362 lands).

Actual behavior

  1. Antigravity's native schedule tool spawns an in-session background timer inside the active turn.
  2. T3 Code keeps the thread in Working state indefinitely (e.g., Working for 1h 51m 1s with Run schedule still spinning in the UI) instead of returning to Idle.
  3. When T3 Code restarts, the in-memory background timer is killed (All your subagents and background tasks have been stopped due to server restart), and T3 Code prints T3 Code restarted and resumed this turn / sends Continue where you left off, which causes the agent to re-create another in-session timer if it doesn't switch to acp-mcp-call.

Impact

Minor bug or occasional failure

Version or commit

t3 v0.1.12-nightly.20261122.5098

Environment

Linux 7.2.15-1-omarchy (x64), Antigravity ACP 1.5.14 (agy 1.5.14, Google account, gemini-4-neon), 35 skills installed

Logs or stack traces

# Live session inspection of <appDataDir>/mcp/t3-code/*.json (86 schemas total)
# 80 schemas <= 1521 B loaded eagerly as mcp_t3-code_*
# 6 schemas >= 1629 B silently omitted (no Lazy: entries, no call_mcp_tool):
   3318B  DROPPED  schedule_task
   2915B  DROPPED  t3_thread_launch
   1926B  DROPPED  delegate_task
   1781B  DROPPED  t3_project_update
   1668B  DROPPED  html_render
   1629B  DROPPED  preview_act
   1521B  EAGER    preview_open

# After Antigravity invoked native `schedule` (8h timer) and T3 server restarted:
# UI showed: "T3 Code restarted and resumed this turn" + "Working for 1h 51m 1s" + spinning "Run schedule"
# System notice delivered to agent:
[Notice] All your subagents and background tasks have been stopped due to server restart.

Workaround

Cancel the Antigravity native schedule background task (via manage_task), then create the persistent scheduled task using the documented acp-mcp-call terminal fallback:

ELECTRON_RUN_AS_NODE=1 "$T3_ACP_MCP_NODE" ${T3_ACP_MCP_ENTRYPOINT:+"$T3_ACP_MCP_ENTRYPOINT"} \
  acp-mcp-call schedule_task '{"title":"...","prompt":"...","schedule":{"type":"once","runAt":"..."},"bindToCurrentThread":true,"enabled":true}'

The thread immediately returns to Idle and the scheduled task survives restarts.

Activity

  1. juliusmarminge commented on Nov 22, 2026

    @juliusmarminge
    Member

    This is the same problem as #38123, 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