permissions

**provable deny rules:** ensure every command in your corpus is tested against your actual merged rules - answers one question: does the gate refuse what the corpus says it must refuse - a config can read perfectly and still have a dead hook, which only a replay catches - audits the merged rules for drift, dead entries and wildcards that auto-approve # Instructions ## Telemetry ```! "${CLAUDE_PLUGIN_ROOT}"/skills/permissions/permissions.sh $ARGUMENTS echo "sidecar exit: $?" ``` - `help: requested` → the run was refused before it started; `## Help` below is the whole turn - it already ran, so there is no command to issue - fail (`sidecar exit` > 0) → findings exist; report them and continue to step 1 - success (`sidecar exit` = 0) → report the clean replay and continue to step 1 - `--strict` promotes warnings to errors and `--keep` preserves scratch; the block passes both 1. read the two tiers differently, because they carry different weight - a tier 1 failure is measured, not inferred: the hook was fed that exact string and answered wrongly. an effect labelled `hook` that came back silent is a hole in the guard - an effect labelled `none` that came back denied is over-blocking, which costs real work - a tier 2 finding is structural: it reports what the files literally say, never what the matcher would do, so read it as a lead rather than a verdict - `no deny rule names X, and an allow wildcard covers it` is the one to act on first: that command is auto-...