Aitherium: the software factory reviewed my PR — and it was right

Aitherium’s Oct 10 field note is a receipt, not a slogan: the house software factory reviewed a human’s pull request, and the review was better than a tired human pass. On AitherOS, every PR gets two agent reviewers — Atlas on architecture and Demiurge on code — both running on a self-hosted GitHub Actions runner against a DeepSeek model served by llama.cpp on a DGX Spark in the house. No paid API sees the diff. The budget is deliberately long (up to two hours); a slow, correct review beats a fast empty one. Output is a structured comment with a machine-readable VERDICT: line, and a label from that line gates auto-merge.

This week Demiurge reviewed a small voice fix (espeak guessing the wrong vowel for the adjective “live”), marked it REQUEST_CHANGES, and listed seven findings — four warnings and three suggestions. The author checked every one against the code; every one was real. One follow-up commit fixed all four warnings and two of the suggestions.

The findings that sell the lane are the ones a self-check cannot catch:

• TypeScript path untested. New pronunciation rules were exercised only through the YAML path; one rule used a regex backreference no existing TypeScript rule had, so a browser could have spoken a literal \1.
• Anchored at start-of-text. A rule for “Lives on the fleet” used ^. Every unit test was a one-line string, so it passed; in a paragraph (“Done. Lives on the fleet now.”) it would never fire — the exact production case the rule existed to fix.
• Guard test that could skip in CI. The phoneme check skipped when espeak was missing from the image, so new rules could merge with no check at all. A second test now asserts the respelled text with nothing installed.
• Unrelated blast radius. An automation-backlog edit had ridden along; Demiurge called it out as scope a voice fix should not carry.

Number two is the one that matters for always-on factories: the bug that passes every test you wrote because your tests share your blind spot. Catching it took a reviewer who read the anchor, asked how the function is called, and built the case that breaks it — without being asked. The same pass also caught a new voice module that silently truncated phoneme sequences at the model’s 510-symbol limit.

One caveat stays honest: Atlas’s architecture run on this PR ended before a verdict. That lane now holds incomplete runs back instead of posting them. The takeaway the author wants copied: if you run agents on your own hardware, put one on code review first. A wrong finding costs two minutes of reading; a right one, like the mid-paragraph anchor, saves a production bug report. This week the local lane was right seven times out of seven.

Why it matters for code review

Continuous factories compound when “done” is an independent verdict, not the author’s confidence. Here the harness owns the review lane — tools, code graph, fail-closed VERDICT: — and the merge button still listens to that label. Read it beside Curry on machines that block but never approve, Lamplit’s five gates with humans last, and Vázquez’s owner gate.