Antigravity version: 2.3.1 (948663537)
System:
- MacBook Pro, Intel
- 1.4 GHz Quad-Core Intel Core i5
- 8 GB RAM
- Intel Iris Plus Graphics 645
- macOS Sonoma 14.5
I am repeatedly encountering a serious issue with terminal tasks launched by the Antigravity agent.
Typical reproduction steps:
-
Open Antigravity.
-
Open a local Python project.
-
Ask the agent to inspect and complete the repository autonomously.
-
The agent launches terminal commands such as:
python3 -m pytest tests/test_public.py -v
or:
source .venv/bin/activate && python -m pytest tests/test_public.py tests_agent/test_agent.py -v
-
The command appears to complete and may already print the final result, for example:
49 passed in 10.45s
-
However, Antigravity continues to show the corresponding task as Running.
-
Clicking Stop Task or Cancel Task has no visible effect.
-
No success, failure, cancellation message, or diagnostic output is shown.
-
In macOS Activity Monitor, the remaining high-CPU processes are displayed with the process name
log, rather thanpythonorpytest. -
These
logprocesses continue consuming CPU after the associated Python or pytest command appears to have completed. -
If several tasks become stuck, multiple
logprocesses accumulate and divide nearly all remaining CPU capacity. -
I eventually have to force-quit the
logprocesses manually through Activity Monitor.
This issue occurs almost every time the agent launches pytest or similar terminal commands.
The problem is specific to commands launched through the Antigravity agent. I have not observed the same behavior when manually running commands in a normal terminal.
I have reproduced the same behavior with all of the following models:
- Gemini 3.6 Flash
- Gemini 3.1 Pro
- Claude Opus 4.6
Because the issue occurs across three different models, it appears to be related to Antigravity’s terminal task lifecycle, process cleanup, output capture, PTY handling, logging wrapper, or task-state synchronization rather than to one specific model.
The lingering processes are displayed as log in Activity Monitor. Since the original commands are Python or pytest commands, this may indicate that the process remaining active is part of Antigravity’s terminal-output capture or logging layer. However, I cannot determine the exact component from the process name alone.
Expected behavior:
- Once a terminal command exits, the task should stop showing as Running.
- Stop Task or Cancel Task should terminate the task and any related child or wrapper processes.
- The UI should report whether cancellation succeeded or failed.
- Completed tasks should not continue consuming CPU.
- Repeated terminal commands should not accumulate orphaned or stuck processes.
Actual behavior:
- The command output may show that execution has completed.
- The Antigravity task remains indefinitely marked as Running.
- Stop Task and Cancel Task do nothing.
- No cancellation feedback or error is shown.
- High-CPU processes named
logremain active in Activity Monitor. - Multiple stuck
logprocesses can accumulate. - CPU usage can rise to nearly all remaining available capacity.
- Manual force-quit through Activity Monitor is required.
Impact:
This makes long autonomous agent workflows difficult to use safely.
The agent may launch replacement or duplicate commands because previous tasks still appear active. Multiple stuck tasks can saturate CPU usage, interfere with test-result reporting, prevent normal task completion, and make unattended workflows unreliable.
The issue has occurred repeatedly across multiple repositories and multiple models.
I do not currently have screenshots or a screen recording of the stuck state, but the problem is highly reproducible. During the next reproduction I can collect:
- Activity Monitor process samples;
- parent-process information for the
logprocesses; - open files and ports;
- Antigravity logs;
- a screen recording showing completed command output while the task remains Running.
Please let me know which diagnostic files or logs would be most useful.