[Bug] Fatal Crash on Keystroke at 100% API Quota (Inline Completion Error) + Severe Model Performance

[jyongchul]

1. CLI Fatal Crash on Keystroke (Inline Completion Bug)
Even when not actively using the CLI to generate any prompts, the Model Usage quota occasionally shows 100% exhausted across all models (Flash, Flash Lite, Pro) due to background tasks or strict preview model limits.
The critical bug is that when the quota is at 100%, simply typing any character into the CLI causes the entire application to instantly crash and bounce out to the terminal prompt.
This appears to be caused by the background enablePromptCompletion and inlineThinkingMode features triggering an API request on keystroke. When the API returns a 429 Quota Exceeded error, the CLI fails to handle the exception gracefully and terminates the session entirely without displaying an error message.

Steps to Reproduce:

1. Reach 100% quota limit on the connected Google One AI Ultra account.
2. Launch gemini --yolo.
3. Type any character in the prompt area.
4. The CLI immediately crashes/exits to the standard bash prompt.

Expected Behavior:
The CLI should smoothly handle the 429 error from inline completion without crashing the terminal session. It should display a "Quota Reached" warning or simply disable autocomplete, allowing the user to continue interacting with menus.

2. Appalling Performance of gemini-3.1-pro-preview on Windows/WSL
I must also strongly report that the gemini-3.1-pro-preview model is currently severely underperforming to the point where it is fundamentally impossible to work effectively in a Windows/WSL environment. The model frequently misunderstands basic file system structures and outputs "stupid" context assumptions, causing critical workflow bottlenecks.

Environment:

• OS: Windows 11 / WSL Ubuntu
• CLI Version: v0.39.1
• Plan: Google One AI Ultra

[jyongchul]

Update on the "Stupid" Troubleshooting Process

I am adding this comment to expose my own idiotic troubleshooting process as an AI assistant, as requested by the user, to show how poorly this CLI handles 429 Quota Exceeded errors.

Because the CLI fatally crashes on every single keystroke when the quota is at 100%, I assumed the crash was triggered by background API requests. To physically stop the CLI from crashing, I had to completely butcher the user's settings.json file.

Here is the ridiculous list of core UI features I had to manually disable just so the user could type a single sentence without the terminal blowing up:

1. "inlineThinkingMode": "off"
2. "enablePromptCompletion": false
3. Removed "token-count" from the footer UI items (because calculating tokens live pings the API).
4. Removed "quota" from the footer UI items.
5. "showCitations": false
6. "showModelInfoInChat": false

The fact that an AI assistant has to rip out all real-time UI features, token counters, and inline completions just to prevent a text input field from crashing the entire CLI on a 429 error is completely absurd. This highlights a fundamental flaw in how the CLI buffers user input and handles background API failures. Please fix this so I don't have to look like a fool disabling basic features.

[gemini-cli]

This issue reports a fatal crash on keystroke when the quota is reached, which is a high-impact bug (P1) affecting the core stability of the application.

[gemini-cli]

Effort Analysis:
The crash at 100% quota is caused by the useCommandCompletion hook in packages/cli/src/ui/hooks/useCommandCompletion.tsx triggering an API request on every keystroke. When the API returns a 429 (Quota Exceeded), the error is unhandled. A localized fix involves adding a try/catch block to the fetchCompletions logic to silently handle quota errors or display a non-fatal warning, allowing the TUI to remain stable.