Codex harness files: plans are not evidence
Aug 3 Tony Alaniz (OpenAI Cookbook): phased harness/build files, human gates, and an append-only build-log that records observed verification — so long-running agent work cannot mark a phase complete on narration alone.
Tony Alaniz’s Aug 3 OpenAI Cookbook recipe is a concrete file layout for continuous agent work that refuses to confuse plan with proof. Optional conventions around Codex — AGENTS.md, GOALS.md, PLANS.md, PROMPTS.md, plus harness/build/phase-*.md, harness/context/, and harness/build-log.md — turn a long-running coding agent from a chat transcript into a gated factory with an audit trail.
Ownership boundaries are the reusable part. Build files describe one bounded phase: objective, non-goals, approval gate, red/green/refactor/verification plan, acceptance criteria, and the evidence required before “done.” They are plans, not claims that work finished. Phase context files capture only material decisions and constraints (not narration). The build log is the source of truth for observed progress: exact commands run, pass/fail results, blockers, limitations. Statuses are explicit (Not started / In progress / Blocked / Complete), and completion is forbidden without evidence that tests and verification actually ran — never inferred from “code exists” or a prior conversation. Activity entries are append-only; corrections add a new entry rather than rewriting history.
The workflow is deliberately human-gated between phases. Prompts require read-only preflight, an approval summary before writes, and a stop after the selected phase — no silently starting the next one. Commits, pushes, deployments, and credentials stay behind separate explicit approval. When the manual process gets heavy, the recipe shows skill-ifying it (harness-author) with a deterministic validator that rejects planning-only runs that sneak in application scaffolding, absolute paths, or unapproved artifacts. A blameless retrospective loop then promotes friction back into instructions, validators, or regression tests — the self-improving factory as file discipline, not vibes.
Why it matters for code review
Always-on agents without an evidence ledger produce morning inventories of confident narration. This cookbook relocates review into the harness: phase contracts reviewers can sample, and a build-log that cannot promote planned checks into passing evidence. Continuous verified momentum looks like append-only receipts across overnight phases — not a longer peak session that marks itself complete.