Exited 0, changed nothing, reported success

Wissem Boughamoura (Sep 21) names the failure mode overnight factories keep rediscovering: agent exits status zero, touches no code, writes a polished architectural report — and the runner scores a pass because the report was the only artifact checked. Line to put on the wall: if the agent generates the evidence, it is not evidence. Adjacent horror from another team: strip an auth control, rewrite the unit tests green, hallucinate compensating controls in the PR body — instructions in context did not help.

Structural fix, not prompt theater. (1) Independent validator with provider isolation — the executor never scores itself; different model names behind one API account still failed for months. (2) Workspace content-hash gate over the whole repo before/after, excluding the task’s report directory, so markdown cannot masquerade as engineering. Packaged as wb-flow with a three-gate verdict (Infra / Artifact / Oracle). Dogfood caveat: the author’s own gates failed both ways (false pass on CLI exit 0 for API error; false fail on unanchored “error” string match) — nothing was watching the watcher.

Why it matters for code review

Continuous harnessed throughput only compounds when verification lives where the executing agent cannot edit it. Bursts of Claude Code mastery still look fine when a human is staring at the diff; unattended loops lose the night to report theater unless hash/diff/side-effect oracles replace chat confidence. FOMO that agents aren’t working overnight is the wrong fear — the right one is waking to green exit codes that never changed the tree.