Antigravity IDE 2.0: agents cannot run terminal commands (Git & NPM) – executor has not processed the previous input yet / UNKNOWN service agentSessions

Google Antigravity
1

Hi,

I am facing a blocking issue with Antigravity IDE 2.0 on Windows where agents cannot run any terminal commands (Git, NPM, etc.).
When I ask an agent to execute commands in the integrated terminal, nothing actually runs and the UI ends up stuck.

The errors I see in the logs are:

  • executor has not processed the previous input yet: ConnectError: [unknown] executor has not processed the previous input yet

  • Error [createInstance] aae depends on UNKNOWN service agentSessions

  • Then: An unknown error occurred. Please consult the log for more details.

Log excerpt (renderer / workbench):

text

[Window] [unknown] executor has not processed the previous input yet: ConnectError: [unknown] executor has not processed the previous input yet
...
[Window] [createInstance] aae depends on UNKNOWN service agentSessions.: Error: [createInstance] aae depends on UNKNOWN service agentSessions.
at Jcr._throwIfStrict (...)
at Jcr._createInstance (...)
at Jcr.createInstance (...)
...

How to reproduce:

  1. Launch Antigravity IDE 2.0.

  2. Open an existing project (Git repository).

  3. Create an agent and ask it to run any terminal command (for example: git status or git diff).

  4. Observe that the command does not actually run and that the errors above appear in the logs.

Similar behavior with NPM:

This is not limited to Git commands.
When the agent tries to run NPM commands such as npm install, npm run dev, etc., these commands also do not execute from the agent.
The same NPM commands work fine in a normal terminal outside of Antigravity, but from the agent they either stay in a “running” state or do nothing visible, with the same log errors (executor has not processed the previous input yet / UNKNOWN service agentSessions).

**Expected behavior:**The agent should be able to run terminal commands (Git, NPM, etc.) successfully without getting stuck.

**Actual behavior:**The agent appears to be “stuck” on the previous command and cannot execute new ones, showing executor has not processed the previous input yet, and the logs indicate a dependency on an agentSessions service that seems to be unknown or not registered.

Environment:

  • OS: Windows 11 (64‑bit)

  • Antigravity IDE version: 2.0.x (please fill in your exact build number if needed)

  • Node.js / NPM: (for example: Node 20.x, npm 10.x)

  • The issue only occurs when commands are executed by an agent inside Antigravity.
    Running the same commands directly in a separate system terminal works as expected.

Additional context:

This issue started immediately after installing Antigravity 2.0.
Before upgrading to 2.0, I did not have this problem with agents running commands.
I have also seen similar reports from other users in the community, so it does not seem to be strictly tied to my local machine or a one‑off misconfiguration.

Things I have already tried:

  • Full IDE restart and machine reboot.

  • Opening different projects, including a very simple test folder, to check whether the issue is project‑specific.

  • Changing the default terminal shell (e.g. PowerShell vs. Command Prompt / Git Bash).

  • (If applicable) Starting Antigravity with extensions disabled / using a fresh profile.

  • (If applicable) Deleting Antigravity config/cache folders to simulate a clean first‑run state and reinstalling the latest version.

This issue looks very similar to other reports mentioning UNKNOWN service agentSessions breaking agent features and to threads where the IDE does not realize that commands in the terminal never actually ran.

Could you please confirm whether this is a known bug in the current Antigravity 2.0 release, and whether there is a recommended fix, patched build, or official workaround (for example: specific configuration changes, launch flags, or known‑good versions)?

Thanks in advance for your help.

2

Noone interested by my issue it seems. Too bad.

3

Hi @pstmartin,

We haven’t been able to reproduce this unexpected behavior based on your description. To help us investigate this further, could you please provide a bit more information?

  • What are your Terminal settings under Agent Security Mode (located in Settings > Permissions)?
  • What is your active integrated shell configured inside the app (e.g., PowerShell, native cmd.exe, Git Bash, or WSL2)?
  • Does this issue happen on all of your projects, or is it specific to certain ones?
  • Does the issue happen in a completely fresh, empty directory with a basic git init, or does it only occur in large, existing multi-folder projects?
4

Unfortunately, I’m experiencing the same thing. I’ve deleted all the possible cache available in antigravity, but the issue still persists. I’m forced to use VM to be able to use antigravity.

image
image673×218 3.21 KB

5

@Siddharth_Naik, thanks for the follow-up. Here are the requested details:

1. Agent Security Mode (Settings > Permissions): I am using the default settings — the agent asks for permission before running each command, and I approve them. The commands are approved but never actually execute; they stay in a “running” state indefinitely until I cancel them or the agent times out.

2. Active integrated shell: Windows PowerShell 5.1 (the default). I do have a PowerShell profile that loads conda and PSReadLine, but I have confirmed this is not the cause: I tested the agent after wrapping the profile in a $Host.Name -eq 'ConsoleHost' guard (effectively disabling it for non-interactive shells), and the issue persisted identically. I also confirmed that powershell -NoProfile -Command "echo test" works perfectly in a normal terminal.

3. Does this happen on all projects or specific ones? It happens on every project I have tested, regardless of size or content. Even a single echo "test" command fails — it is not related to Git or NPM specifically. The agent’s terminal is completely non-functional.

4. Fresh empty directory with git init? The issue is not project-specific. Even the simplest possible command (echo "test") in any directory fails from the agent. The commands work perfectly when I run them myself in a normal terminal (both the integrated terminal and an external one). Only the agent’s command execution is broken.

Additional important details:

  • My main workspace path contains spaces and accented characters: d:\Mon ordinateur\Ordi 2011\Programmation\Antigravity\Panier courses. This could potentially be relevant.

  • The exact error from the agent logs: executor has not processed the previous input yet: ConnectError: [unknown] executor has not processed the previous input yet

  • And: [createInstance] aae depends on UNKNOWN service agentSessions

  • This started immediately after upgrading to Antigravity IDE 2.0. The previous version worked fine with the exact same machine, projects, and configuration.

Full environment:

  • OS: Windows 11 64-bit

  • Node.js: v18.15.0 (C:\Program Files\nodejs\node.exe)

  • NPM: C:\Program Files\nodejs\npm.cmd

  • Git: C:\Program Files\Git\cmd\git.exe

  • Shell: Windows PowerShell 5.1

  • All tools work correctly from a normal terminal — only the agent’s terminal execution is broken.

The UNKNOWN service agentSessions error in the logs strongly suggests that a required internal service is not being registered or initialized properly during IDE startup. This looks like an internal dependency injection issue in the 2.0 release, not a user configuration problem.

Looking forward to a fix or a workaround. Thank you.

6

i am using linux CLI and getting the same error -

image
image883×307 9.43 KB

7

Hi @Siddharth_Naik , no reply from your part. Are my answers not sufficiently accurate ?

8

Any breakthrough on this. I am getting the same error since last few days

9

Same issue here: on CLI for windows (powershell or bash), on Antigravity desktop (windows)
Every external command launched kicks u out of the cli, on the desktop app it just restarts it and u can continue the chat but it breaks long prompts/jobs, so basically Antigravity is unusable
best to bring back gemini CLI :wink:

here is a bug report:
Bug Report: Agent Server Restarts After Every run_command Completion

Date: 2026-06-19
Environment: Windows (amd64), PowerShell 5.1.14393.3053
Antigravity App Data: C:\Users\tb\.gemini\antigravity
Conversation ID: 6a5e5ec2-ab63-4ac2-918f-1c644a88ed..

Summary

Every time a run_command tool call completes — regardless of the command, its duration, or its exit code — the Antigravity desktop runner immediately performs a full server restart. The command’s output is returned correctly, but the step is always marked as canceled and a system notice is emitted:

[Notice] All your subagents and background tasks have been stopped due to server restart.

Steps to Reproduce

  1. Open any Antigravity conversation on Windows.

  2. Have the agent invoke any run_command tool call, for example:

    echo "Hello"

  3. Observe that:

    • The command output is returned correctly (Hello).

    • The step status is marked canceled.

    • A system message is immediately received: All your subagents and background tasks have been stopped due to server restart.

  4. Repeat with any command — the behavior is 100% reproducible.

Commands Tested (all reproduce the issue)

Command Duration Result Restart?
echo "Hello" ~0ms :white_check_mark: Output received :white_check_mark: Always
powershell -NoProfile -NonInteractive -ExecutionPolicy Bypass -File <script>.ps1 ~1s :white_check_mark: Output received :white_check_mark: Always
cmd /c powershell -NoProfile ... -Command "<cmd>" ~1s :white_check_mark: Output received :white_check_mark: Always
go version ~0ms :white_check_mark: Output received :white_check_mark: Always
go test -count=1 ./... ~1.5s :white_check_mark: All tests passed :white_check_mark: Always

Observed Behavior

  • Command output is always fully returned before the restart occurs.

  • The restart happens immediately after the child process exits — even for sub-millisecond commands.

  • There are no Windows Application Event Log entries (Event ID 1000 / crash) for the Antigravity runner process at the time of restart.

  • There are no .log files or logs/ directories in C:\Users\tb\.gemini\antigravity that would indicate a crash dump.

  • The runner process appears to exit cleanly with code 0, which is why no crash is registered.

  • Background tasks (launched with WaitMsBeforeAsync less than the command duration) are also killed on restart.

Expected Behavior

  • run_command tool calls should complete with status DONE, not canceled.

  • The server should not restart between agent turns.

  • Subagents and background tasks should not be killed as a side effect of a shell command completing.

Suspected Root Cause

The most likely cause is an IPC/stdio pipe propagation bug in the local runner service:

  1. The runner spawns a child process (powershell.exe / cmd.exe) to execute the command.

  2. On child process exit, the runner either:

    • Accidentally propagates the child’s exit signal to the parent runner process, or

    • Closes shared stdin/stdout pipes that the agent host relies on for MCP/RPC communication.

  3. The agent host detects the closed pipe as a clean disconnection and exits gracefully (no crash, no error log).

  4. The watchdog/UI layer detects the exit and starts a new session, broadcasting the “server restart” notice.

This would explain:

  • Why no crash logs exist (exit is clean/voluntary).

  • Why no Windows Error Reports are generated.

  • Why it affects every command without exception.

  • Why output is always received (it is written to the pipe before the pipe is closed).

Impact

  • Functional impact: Low — command side-effects (file writes, test results) are preserved on disk. The workspace is persistent.

  • Workflow impact: High — every command invocation resets the agent context, kills all subagents and background tasks, and forces a new server session. This prevents reliable use of long-running background services (dev servers, watchers, etc.) and creates noise in every conversation.

  • Reproducibility: 100% — affects every run_command invocation on this machine.

Environment Details

OS:           Windows (amd64)

PowerShell: 5.1.14393.3053

Go: 1.26.1 windows/amd64

App Data: C:\Users\tb\.gemini\antigravity

No crash logs, no .log files, no Event Viewer entries at time of restart.

10

18 days without a solution even a workaround !!! And Antigravity persists in launching commands even though you’ve asked it not to do so. And you have to reload it. :+1:
A little bit frustrating to have answered your question with precision and not to have any solution proposed.

And you know what ? This made me try Codex which works fine with commands… Obliged to use Open AI…

11

Hey @Siddharth_Naik are you still there ???

12

I’ve never seen that. You pay, it doesn’t work and when you ask for help… Will you recharge my account for not being able to use your tool ?

13

I’m seriously starting to get fed up with wasting all my credits having to restart this tool every time it freezes when I run a command. It’s unacceptable to pay for a tool that doesn’t work.

14

To my great surprise, Codex actually lets you run commands, which is already huge, but what’s more, unlike Anti Gravity, it doesn’t overwrite files, it doesn’t destroy them, and it doesn’t accidentally overwrite files. Because that’s the mishap that just happened to me. Not only do you have to slog through all the commands in the console, but you also have to be extremely careful before asking anti gravity for a modification. So I strongly advise against this tool, and I’m going to stop wasting my time on this forum where no answers ever come.

I subscribed to Google One to be able to use Anti-Gravity, but I think I’m going to unsubscribe.

.Thanks for your help.