harnesswright: an agent's done is a claim — the gate returns a receipt
Jul 2026 Falco: zero-runtime CLI that never runs agents; ledger slices + verity claims reconciled to filesystem, git HEAD, and exit codes — exit 0/1/2; dogfooded 18+ green claims that caught its author.
Source: harnesswright — governance-first, verification-first harness generator
pietro-falco/harnesswright (repo Jul 2) states the overnight-factory axiom in one line: an agent’s “done” is a claim, not a fact. The CLI never executes or orchestrates agents. init drops a governance pack into any repo — evidence-gated slice ledger (.harness/ledger.md), config, ADR scaffold, verity claims manifest, pre-commit hook, per-session worktree script, CI gate.yml, plus AGENTS.md / SKILL.md. next is read-only: the unlocked slice and its criteria. gate reconciles declared claims against the literal filesystem, git HEAD, and command exit codes — exit 0 all pass (receipt), 1 a claim fails (full stop for operator review; the agent does not retry its way to green), 2 config error. No LLM-as-judge: a model check is an opinion; a receipt is evidence. Framed as the Truth layer under Spec Kit (intent) and orchestrators (execution).
Credibility argument is dogfood: the repo adopted its own harness and is gated by gate in CI — 18+ claims green at HEAD, three of them because the harness caught its author. Complements, not competitors: write intent upstream, scale execution with whatever agent you already run, let a deterministic gate decide whether “done” is true.
Why it matters for code review
Modest verified momentum overnight only compounds when completion authority sits outside the worker. harnesswright relocates the morning review surface from chat confidence to a machine receipt shared by laptop and CI — the same shape as continuous factories that refuse to merge on agent narration. Peak-session Claude mastery still wins when the question is product judgment; unattended loops win when a failing claim is a hard stop, not another prompt round.