Break Out of Antigravity's Refactoring Loop

[1bcMax]

The Problem

Antigravity enters a "recursive optimization loop" — it keeps refactoring its own refactored code. Each fix introduces a new error. The undo history gets shredded by dozens of micro-edits.

You're stuck watching your codebase get worse with every iteration.

The Fix: Call an External Model

When Gemini is stuck in a loop, a fresh perspective from a different model can break the cycle.

# Get GPT to review what Gemini did:
"blockrun GPT review what Gemini just did and suggest a different approach"

# Or ask Claude for a second opinion:
"blockrun claude analyze this refactor and identify the root issue"

Why This Works

• Different training, different perspective: GPT and Claude may approach the problem differently than Gemini
• No shared context: The external model starts fresh, without Gemini's accumulated confusion
• Quick escape hatch: One call to break the loop, then continue with the suggested fix

Example

from blockrun_llm import setup_agent_wallet

client = setup_agent_wallet()

# Send the problematic code to GPT for review
code = open("problematic_file.py").read()
response = client.chat(
    "openai/gpt-5.2",
    f"Gemini has been refactoring this code in circles. What's the root issue and how should we fix it?\n\n{code}"
)
print(response)

Cost

A single GPT-5.2 call for code review: ~$0.02

Cheaper than losing an hour to a refactoring loop.

---

Has anyone else hit these loops? What's your escape strategy?