# CRITICAL BUG REPORT: Client-Side Crash Loop Causing Unfair Token/Quota Drain on Ultra Tier
## 1. Issue Summary
The Antigravity desktop IDE is entering a crash loop where the local Language Server fails to initialize. This triggers a false “Servers are experiencing high traffic” error. Critically, during this local crash loop, the background agent continuously re-reads the prompt context and attempts re-execution, which rapidly and unfairly drains the user’s premium token quota and 5-hour refresh bucket.
## 2. Environment Details
* **Subscription Tier:** Ultra (Priority Access)
* **Version Affected:** April 2026 Build (v1.2x)
* **Models Affected:** Auto-subscribe models (specifically heavily impacting Claude Opus 4.6 and Gemini 3.1 Pro)
* **Operating System:** Windows
* **Frequency:** 30 to 40 times per day during active development.
## 3. Steps to Reproduce (Observed Behavior)
1. User submits a standard prompt during active workflow in the desktop IDE.
2. The local Language Server Client experiences a connection refusal/initialization failure (often local loopback 127.0.0.1 drop).
3. The Antigravity IDE UI displays the error: `“Error Our servers are experiencing high traffic right now, please try again in a minute.”`
4. **The Critical Bug:** Under the hood, the system’s retry mechanism repeatedly attempts to execute the prompt against the auto-subscribed model without waiting for the local client to stabilize.
5. Because models like Opus 4.6 are “heavy,” this rapid retry loop consumes Massive amounts of context tokens for failed generations.
## 4. Impact Assessment (Why this is Critical)
This is not just a stability issue; it is a **billing and quota integrity issue**.
* **Misleading UX:** The error blames cloud server capacity, but the community/logs confirm it is a local client-side software defect. Priority Ultra load balancers are never even reached.
* **Financial/Quota Drain:** Users are losing their paid Ultra quota (and potentially overage AI credits) due to ghost-retries from a software bug, not from actual usage. The 5-hour refresh bucket is being wiped out in minutes due to background prompt looping.
* **Workflow Paralysis:** The user is forced to manually force-close background processes, clear `CachedData` / `Code Cache`, and restart continuously.
## 5. Expected Behavior
* The IDE should correctly differentiate between a `Local Language Server Crash` and `Cloud 429/503 High Traffic`.
* If the local server crashes, the agent **MUST halt all token consumption and API calls immediately**. It should not loop and burn user quota on failed local states.
* Ultra subscribers should have their burnt tokens refunded if consumed during a verified local client crash loop.
## 6. Temporary Workarounds Currently Forced Upon Users
We are currently forced to downgrade to the December 2025 build (v1.19.x) to stop the crashing, or abandon the desktop IDE entirely to use the Web Editor just to preserve our paid tokens. This defeats the purpose of the premium desktop experience.
**Request:** Immediate patch to the retry-logic to prevent token drain during local disconnects, and a fix to the local Language Server stability.
It’s certainly not the Ultra crowd. We’re too busy staring at quota limits and ‘resource unavailable’ screens to actually put any load on the servers.