A 100k-line codebase now contains thousands of lines of AI-generated code, but the reasoning behind it is lost. Reviews can't ask why a change was made, only if it's broken. This erodes long-term control and maintenance as AI authorship becomes the norm.

The solution isn't better review tools, but a methodology shift. Plan Stack introduces a formal plan as a first-class artifact, reviewed and committed with each pull request. This lets humans judge intent and trade-offs before code is generated, keeping human oversight scalable.

Plans become a permanent decision history, preserving context for future changes. Unlike distant ADRs, they are tied directly to code changes. This externalizes human reasoning, making AI-generated systems maintainable and preventing them from becoming untouchable legacy code.