Cursor's 'fix code' command is inconsistent. Sometimes it makes things worse.

I've been using Cursor's "Fix Code" command heavily for a few months. The inconsistency is maddening. It's not a "sometimes" thing—it's a core problem with how the feature is implemented.

Here's my head-to-head observation on a simple task: fixing a broken Terraform `for_each`.

**Original buggy code:**
```hcl
resource "aws_instance" "example" {
for_each = toset(var.environments)

ami = "ami-12345678"
instance_type = "t3.micro"
tags = {
Name = "server-${each.key}"
Environment = each.value
}
}
```
The error: `each.value` is invalid; it should be `each.key` for the Environment tag.

**Good Cursor Fix:**
- Correctly changes `each.value` to `each.key`.
- Leaves a concise comment: `# Using each.key as the value for the environment tag`.

**Bad Cursor Fix (happens ~40% of the time):**
- Rewrites the entire block unnecessarily.
- Changes `toset(var.environments)` to a complex `{ for env in var.environments : env => env }` map.
- Introduces a syntax error by misplacing the closing brace.
- Makes the code harder to read for zero benefit.

**Root cause analysis:**
The command seems to randomly choose between:
1. A minimal, correct diff.
2. A complete, over-engineered rewrite that often introduces new bugs.

This isn't a model version issue—it's a prompt engineering or context window failure. For a tool that markets itself as "the AI code editor," this is a fundamental flaw. The "fix" should be the smallest possible change that addresses the error. Anything else is noise.

Has anyone found a reliable pattern or workaround? Or is this just a "wait for v2" situation?