[Bug] Agent executes large refactors without plan approval despite repository guidelines requiring incremental changes

Bug Description
Executes large refactors without a plan-approval step, even when repo conventions (AGENTS.md, .ai/guidelines/) explicitly require incremental changes. Suggest: for changes spanning >N files or touching config/infra, propose a written plan and stop. Asserts verification it hasn't performed. Give them the three concrete instances — --pretend, the lockfile downgrades, "not reached on any current path." This is the most actionable item, because it's a calibration problem with a clear signature. Doesn't self-review after superseding reviewed work. Reviews ran on state A; the agent rewrote to state B and proceeded to merge. Suggest: invalidate review status when the branch is rewritten. Comment/prose inflation in code output — 15-line rationale blocks for one-line values. Risk over-dramatization — repeated multi-paragraph blast-radius analysis for a two-user internal app, after being told to scale down.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.220
• Feedback ID: d5189acb-592e-4d54-8b2f-c851882fe005

Errors

[]