Field note

AIFactory: verify against the live endpoint, not the diff

May 20 AIFactory PARR loop: plan→act→deploy→reflect→review — worktree isolation, live HTTPS verify, bounded auto-feedback, human-stop defaults; auto-merge only after reviewer gate.

Source: AIFactory — the PARR pipeline and its guards (plan → build → deploy → verify → merge)

olafkfreund/AIFactory (May 20 2026; platform + PARR guards guide) frames autonomous shipping as a guarded closed loop — Plan → Act → Reflect → Review — where most stages past build are opt-in and every uncertain step degrades to a human stop rather than guessing. Spec-first planning (with optional HMAC-trusted upstream plans) precedes code. Act runs in an isolated worktree; parallel waves merge sequentially with a parent-owned plan and a repeated-failure detector that escalates instead of burning tokens. A mutation ledger checkpoints claimed-vs-actual edits for rollback.

The distinctive verification tax sits in Reflect: tests run against the deployed HTTPS URL, not a local process, so "it builds" is not mistaken for "it runs." Deploy (opt-in) is deterministic Terraform to App Runner with teardown always paired, cost tags, and a sweeper — infrastructure the factory generated, not an LLM improvisation. Review adds a configurable pre-merge reviewer gate, bounded auto-feedback (≤2 fix/re-review cycles, then human), and a merge-conflict stop. Adoption is staged: auto-PR first, then automated reviewer, then auto-merge, then auto-deploy — so continuous factories earn the right to close the loop rather than skipping straight to lights-out merge.

Why it matters for code review

Harnessed overnight momentum usually verifies in the worktree; AIFactory productively complicates that by moving the hard gate to the live endpoint and keeping merge behind a bounded, human-default review endgame. Continuous throughput still compounds only when each stage can refuse — promote green live verify plus reviewer approval, reopen exhausted auto-fix loops, and refuse to treat a local green as production-ready. Peak Claude sessions still win when you are steering the change; the overnight board is for specs whose acceptance criteria survive deploy.