AI Agent ignores guardrails; Preview stuck in refresh loop; outputs diverge from prompts (iPad & desktop)

Bugs
1

Proposal: a reliability set for the AI Agent

  1. “Safe Mode: additive-only patches” to prohibit destructive changes.
  2. Mandatory pre-flight diff/approval before file writes.
  3. Automatic checkpoints/one-click restore after each run.
  4. Visible session spend/time caps.
  5. Per-run change logs.

These would prevent runaway edits, make failures recoverable, and give users confidence to iterate.

1 Like