[Bug/Feedback] Antigravity IDE lacks basic context compaction. Claude Opus 4.6 is virtually UNUSABLE due to constant "prompt is too long" agent terminations.

Google Antigravity
1

image
image1236×224 20.9 KB

I am consistently encountering fatal agent terminations when using Claude Opus 4.6 (Thinking) on Antigravity IDE. Instead of managing context gracefully, the agent simply terminates midway through any moderately complex session.

Here is the exact Debug Info I keep seeing:

  • Error: HTTP 400 Bad Request

  • Detail: prompt is too long: 218849 tokens > 200000 maximum

The Core Architectural Defect (NOT a User Error): Antigravity’s current “smarter context packaging” is fundamentally broken when handling long-context reasoning models. The IDE completely lacks a dynamic auto-compaction or context pruning mechanism.

Instead of intelligently summarizing the trajectory when approaching the 200K threshold, it recklessly dumps the entire accumulated history alongside the massive, unpredictable “Thinking” token overhead generated by Opus 4.6 straight into the API payload. This guarantees an instant token overflow and a fatal crash.

The Industry Standard (Why Antigravity is Falling Behind): This is already a solved problem in the industry.

  • Cursor IDE handles this flawlessly with background context summarization.

  • Anthropic’s own Claude Code utilizes robust native compaction logic (e.g., the /compact function).

Why is a flagship, “agent-first” IDE from Google so far behind, lacking even the most basic auto-compaction feature?

As long-context reasoning models like Opus 4.6 become the new standard, relying on hard-coded token limits without intelligent history summarization is fundamentally unsustainable. Please prioritize implementing a native Server-side Compaction mechanism immediately. We are tired of losing our entire task trajectories to a preventable overflow.

2 Likes
2

Hi @Daniel.C
Welcome to the Google AI Forum!!!

Thank you for your valuable feedback. We sincerely appreciate you taking the time to share your thoughts with the community. Your input is crucial for our continuous improvement and we have shared it directly with the relevant internal team for their review.

3 Likes
3

Antigravity has been basically unusable for about a week now.

I keep getting:

“Agent terminated due to error”
and HTTP 503 – MODEL_CAPACITY_EXHAUSTED (claude-opus-4-6-thinking)

This happens several times per minute. I’m constantly forced to click the retry button just to continue working, which completely breaks the workflow.

At this point it’s extremely frustrating.
Instead of focusing on development, I’m spending my time babysitting retries.

Antigravity is not usable for serious work.

I’ve already submitted literally dozens of bug reports through the “Provide Feedback” form (Bug Report category), describing the issue, steps to reproduce, expected vs actual behavior, and including the error messages - but the problem still persists.

1 Like
4

I am experiencing the same issue. Antigravity has become unusable for me. Funny thing is that this wasn’t always an issue, but it is now almost a certainty that my conversation will crash from lack of autocompaction. My only way around it has been to use Claude Code, and/or a hacked Ralph loop style method of shortening my conversation, creating a summary of work, then passing the context along to a new conversation . It is a terrible workflow and it is mind-blowing that using a Google product I’d have to jump through such hoops. It’s a shame because i love Antigravity’s design.

3 Likes