Code-Enforced Review Gates Beat Prompt Conventions in AI Agent Pipelines

A software project called ZOdyssey has implemented a code-enforced orchestration pipeline where an 8-phase review gate physically prevents AI agents from executing code until a verified plan approval is confirmed. Unlike most agent pipelines that rely on prompt instructions to enforce planning and review discipline, ZOdyssey uses a hook mechanism that reads a state field, making the review gate structurally impossible to bypass. The project's history highlights a recurring vulnerability: a critical Bash write-gate was silently deleted twice across versions, and three separate external audits failed to catch the regression because each only reviewed the immediate diff rather than re-checking earlier invariants. In response, the team embedded a regression test file whose sole purpose is to enforce that the gate exists, with its header explicitly documenting the deletion history as justification. The case illustrates a broader engineering principle — that security and workflow invariants in AI pipelines must be expressed as executable code, not documentation or model instructions.