Antigravity: What was happening behind the 'Infinite Loop/Crash Bug'?

Hello! This is Takahashi (*'▽')

Although it has already been resolved,
the 'infinite loop during prompt input' and 'process termination during image loading (Agent execution terminated due to error.)' that I reported in the following article were quite alarming bugs for engineers who use Antigravity on a daily basis.

What is particularly noteworthy is that this phenomenon occurred only in the 'Antigravity IDE' and worked normally in the desktop version (Antigravity 2.0).

In this article, while following the flow of the tumultuous updates over the past few days, I will delve into and examine the technical background and structural causes of why this bug occurred only in the IDE version.

1. Timeline of recent updates (v2.7.1 to v2.8.1)

First, let's look back at the release notes for the last three days to see what updates were made before the bug occurred.

• v2.7.1 (August 11)

  • Addition of side-by-side preview feature for image diffs

  • UI optimization and accessibility improvements

• v2.8.0 (August 12)

  • Added session persistence for sidebar folder open/close state

• v2.8.1 (August 13)

  • Improvements to message responsiveness during interaction with the agent

  • Fix for issue where processing would hang (become unresponsive)

It is clear that feature additions and responsiveness improvements were being deployed daily. The problem is thought to have been caused by the combination of this 'v2.7.1' and 'v2.8.1'.

2. [Discussion] Why did the 'infinite loop' and 'crash' occur only in the IDE version?

In a previous article, I described the behavior where
'it crashes when attaching images or loading them via path specification'
and 'it recovers if you roll back', but I suspect that behind this phenomenon lies
a conflict between two factors.

Hypothesis: Inconsistency between image analysis processing and retry logic

1. Trigger (v2.7.1 feature) In v2.7.1, image-related features such as "Image Diff Preview" were expanded. When attaching an image to a chat, an error occurred in handling the image data for some reason while passing the process to the IDE's internal MCP (Model Context Protocol) server or the image rendering pipeline.

2. Loop occurrence (v2.8.1 improvement) In v2.8.1, improvements were made for "hang prevention and responsiveness enhancement." As a result, it is believed that the logic for the agent to spontaneously retry was strengthened when internal errors or response delays were detected. Consequently, it is thought that the system entered an infinite loop of "image loading error -> automatic retry -> error again -> automatic retry...".

3. Process termination (crash) It appears that this infinite loop rapidly consumed memory and CPU resources, or the error logs exceeded a threshold, causing the system to output "Agent execution terminated due to error." as a safety mechanism and terminate the process (crash it).

3. Why was the "Desktop Version (2.0)" not affected?

Despite using the same Gemini model and agent engine, the reason this bug did not occur in the desktop version lies in the difference in architecture.

• Antigravity IDE Because it is built on the VSCode extension foundation, UI rendering on the editor side, file system access, and communication between MCP servers are strongly dependent on the VSCode process (Electron / Node.js environment).

• Antigravity 2.0 (Desktop version) It runs on an independent platform specialized for parallel agent execution and task management (a multi-agent structure optimized in languages like Go).

In other words, this bug was highly likely not an "AI model-side issue," but rather a
"problem with the interaction between the UI/communication layer of the IDE version (which is a VSCode derivative) and the agent's retry logic".

4. Countermeasures in the field and how to deal with it in the future

In an environment where development tools are updated at high speed like this, the following risk hedges are effective.

1. Immediate "rollback" when a bug occurs By rolling back the chat history to just before the statement that triggered the issue, such as an image attachment, you can prevent context corruption and resume work.

2. Temporarily switch to the desktop version when the IDE version is malfunctioning Even if there is a problem with the communication layer of the IDE version, you may be able to send tasks to the agent without issues using the desktop version (2.0). It is recommended to keep the desktop version open as a "separate execution environment."

3. Wait for "minor patches" immediately after consecutive daily updates Updates released in quick succession, such as v2.8.0 to v2.8.1, are often a sign that major internal behavioral changes have been made. When working with critical code for business, it is a good idea to wait and see how the community reacts before applying them. *Though I'm one to talk 💦

Because we are in a transitional period for AI development environments, understanding the structure behind such bugs will help you deal with them calmly when trouble occurs.

Also, thank you for commenting with valuable verification information in the previous article!

That's all for this time (*'▽')ﾉｼ