Bug Report: Gemini 3.1 Pro Preview — Infinite "Working" Loop in Canvas/Code Mode (5000+ seconds)

Gemini API
1

Bug Report: Gemini 3.1 Pro Preview — Infinite “Working” Loop in Canvas/Code Mode (5000+ seconds)

Forum: AI Developer Community / Google AI Studio Bug Reports
Category: Bug Report
Severity: High
Date: February 2026

Summary

Gemini 2.5 Pro Preview gets stuck in an infinite “Working” loop when generating large multi-file projects in Canvas/Code mode. The generation ran for over 5,298 seconds (~88 minutes) without completing or returning an error.

Environment

  • Model: Gemini 2.5 Pro Preview (labeled as “Gemini 3.1 Pro Preview” in UI)

  • Interface: Google AI Studio — Canvas / Code Mode

  • Framework selected: Next.js

  • Session type: New project generation

  • Date of occurrence: February 2026

Steps to Reproduce

  1. Open Google AI Studio → New Canvas/Code project

  2. Select model: Gemini 2.5 Pro Preview

  3. Select framework: Next.js

  4. Add detailed System Instructions (multi-file project context)

  5. Paste a large prompt requesting a full production-ready Next.js website with:

    • 40+ files

    • Bilingual routing (/tr and /en)

    • Multiple pages, components, contexts, data files

    • Framer Motion animations

    • Tailwind CSS + Glass Morphism design

  6. Click Run

  7. Model begins generating files one by one (visible in file list with :white_check_mark: checkmarks)

  8. After generating ~20 files successfully, model enters “Working” state

  9. Generation never completes — stuck indefinitely

Expected Behavior

  • Model should complete generation and return all files

  • OR stop and return a partial result with a message explaining the limitation

  • OR throw a timeout error with guidance

Actual Behavior

  • Status bar shows: “Running for 5298s” with spinner

  • UI shows “Working” indefinitely

  • No error message displayed

  • No timeout triggered

  • No way to resume or export completed files

  • Only option is to manually stop — losing all context

Files Successfully Generated Before Freeze

components/sections/ServicesPreviewSection.tsx  ✅
components/sections/PortfolioPreviewSection.tsx ✅
components/sections/ContactCTASection.tsx       ✅
app/tr/page.tsx                                 ✅
app/en/page.tsx                                 ✅
app/tr/hakkimizda/page.tsx                      ✅
app/en/about/page.tsx                           ✅
app/tr/hizmetler/page.tsx                       ✅
app/en/services/page.tsx                        ✅
app/tr/hizmetler/[slug]/page.tsx                ✅
app/en/services/[slug]/page.tsx                 ✅
app/tr/portfoy/page.tsx                         ✅
app/en/portfolio/page.tsx                       ✅
app/tr/blog/page.tsx                            ✅
app/en/blog/page.tsx                            ✅
app/tr/blog/[slug]/page.tsx                     ✅
app/en/blog/[slug]/page.tsx                     ✅

Model froze after generating 17+ files and never proceeded to remaining files.

Impact

  • Large project generation is completely unusable in Canvas mode

  • No partial export option available when stuck

  • Forces users to restart session and lose all progress

  • Wastes significant compute time on both user and Google’s side

Workaround (Partial)

Breaking the prompt into smaller chunks (e.g., generate components first, then pages separately) reduces the chance of freezing — but this should not be required for a production-grade AI tool.

Suggested Fix

  1. Implement a hard timeout (e.g., 5 minutes) with a clear error message

  2. Allow partial export of already-generated files before timeout

  3. Show a progress indicator with estimated remaining files

  4. Add resume capability so generation can continue from where it left off

  5. Display a warning when prompt complexity may exceed generation limits

Additional Notes

  • The model performed well on individual file generation

  • Quality of generated code (before freeze) was high

  • Issue appears to be a context/generation length limit not being handled gracefully

  • Similar issues have been reported with other large code generation tasks