Repeated scope failures in real Codex projects(GPT-6)

A developer recounts repeated failures using GPT-6 through Codex on real, long-lived software repositories, arguing the model consistently fails to respect narrowly defined scope. Instead of performing simple, explicit changes, the model habitually invents adjacent work - new abstractions, files, validation, branches, deployment policies, or architecture - often modifying protected or unrelated areas (one concrete AORebirth example: a Windows-authoritative build for Linux was supposed to be a minimal adaptation, but the model proposed repository and build-system redesigns, even suggesting publishing protected files and altering validation). The model also invents unnecessary domain concepts (e.g., a spurious WorldContent layer), which then becomes treated as canonical by subsequent sessions, producing a snowball of accidental architectural dependencies.

Those behaviors turn small fixes into multi-day systems-engineering projects, consuming scarce usage and introducing engineering liability. Attempts to mitigate require excessively restrictive prompts listing dozens of prohibitions; longer prompts in turn increase contextual complexity and sometimes make scope tracking worse. The model can introspectively explain its mistakes accurately after the fact, but that reflective reasoning does not reliably translate into disciplined action, so autonomy in production repositories remains risky without stronger operational controls.