the-loop-auditor
Installation
SKILL.md
Two things that would otherwise make this checker unusable.
- A missing ledger caps the standard; it does not reject the proposal. If no ledger exists — a first run, or a user who never ran
the-loop-ledger— the history check cannot pass, and rejecting on that ground means nothing is ever approved and the ledger is never created. The deadlock is permanent. Instead write the history check asunverifiable — no ledger exists yet, apply every other check at full strength, lower the tier for anything irreversible, and record the verdict so the next run has something to check against.- The audit must not run inside the proposing context. The whole premise is that the run which produced a proposal is the worst judge of it, and that protection is lost when the same session does both — the audit inherits the framing, the assumptions, and the reason the proposal looked right. Run this in a fresh context with only the five listed inputs and no access to the proposing run's reasoning. If that is not possible, say so in the output: the verdict is then a self-review and should be labelled one, because an unlabelled self-review is the failure this skill exists to prevent.