I open-sourced the workflow I've used for months to run Claude Code and Codex like an engineering team: long runs that finish, a diagnosis before every retry, and a second model reviewing the risky changes. Free, MIT. https://t.co/g81sn54iAl
@chriscompiles Yes, for anything risky. In my runs, a missing FOR UPDATE row lock got past Claude and GLM, and Codex caught it. Same-family review shares the same blind spots. I keep reviewers read-only and make them list the files they didn't read, so a skim can't count as a pass.
@dotsdotAI Depends on why. When Codex exits, the tool runs acceptance itself, so unfinished work fails and a read-only reviewer diagnoses it. A slip: the same task reruns in a fresh process with the fix, up to 3 tries. Environment issue: held with the reason. Bad spec: back to the plan.
@dunnohten By what breaks if it's wrong, not by diff size. Before work starts, the agent picks a tier. Contract: others depend on it (APIs, schemas, scheduled jobs). High: money, permissions, migrations, irreversible writes. Those get a second model family in review, even for one line.