Managed Codex Workspace diff boundary workflow: keeping coding agents inside the approved change set

A coding agent can respect scope only if scope is concrete enough to respect.

The common failure mode in agent-assisted coding is not that the agent writes syntactically invalid code. It is that the agent writes plausible code across files nobody expected to change, silently edits a neighboring system to make tests pass, or pulls a broad cleanup into a narrow fix because the repo made that path convenient. When that happens, reviewers lose time reconstructing intent instead of evaluating the requested work. A Managed Codex Workspace diff boundary workflow gives the agent and the reviewer the same starting contract: what may change, what must not change, what tests matter, and when a spillover concern needs a separately approved expansion. The useful role for AI is scope framing, neighboring-risk detection, and diff explanation. It is not granting itself permission to widen the change because the wider edit looks tidy.