machinist

Agents are good at producing code and bad at judging their own output. machinist splits those concerns: agents propose — code, review findings, verdicts, prose — while a deterministic orchestrator disposes, deciding every stage transition through a pure, exhaustively tested state machine. The result is a pipeline autonomous enough to ship a pull request unattended, and accountable enough that every decision it made is readable in the issue thread afterwards.

• An interactive /spec flow turns a one-paragraph request into a validated spec — testable acceptance criteria, non-goals, a self-review checklist — filed as a GitHub issue with priority and effort labels.
• A human applies the approval label; that single label event (verified to come from a write-access user) triggers a GitHub Actions run that drives every remaining stage inside a non-root Docker sandbox.
• An implementer agent builds the change on a branch and opens a draft PR; four read-only specialist reviewers — correctness, security, tests, style — then critique the diff in parallel.
• An adversarial challenger tries to refute every finding they raise. Only confirmed blocker-level findings send the code back to a fixer; the loop escalates to a human after three strikes.
• The orchestrator itself runs the test suite — the gate is exit-code truth, never an agent's claim — and a scribe writes the final PR description before the draft flips to ready.
• Every control-flow decision lives in a pure, exhaustively unit-tested state machine over the issue's labels. Agents produce artifacts; they never decide what happens next.

Issue #12 — a tagging feature for the demo API — ran unattended on GitHub Actions for just under nine minutes. The implementer built the feature; the review panel’s test specialist caught that the tag-filter test couldn’t distinguish exact matching from substring or case-insensitive matching — precisely the behaviors the spec ruled out — and the challenger confirmed it. Strike one. The fixer added near-miss fixtures to pin the behavior down. In round two, a security finding about unbounded tag counts was struck down by the challenger as out of spec. The test gate passed, and the scribe flipped PR #13 to ready with the whole honest history in its description — one real defect caught, one false positive killed, in the same run.

Multi-agent orchestration where the orchestration itself is boring, testable code; adversarial review as a first-class mechanism rather than a prompt suggestion; and the discipline of deterministic gates around probabilistic workers. The pipeline was also debugged in production against the real world — headless-CLI deadlocks, Windows argv mangling, GitHub’s bot-identity restrictions — and each incident is documented in the repo’s history with a regression test attached.