Follow-ups from the #507 review cycle — reported by the review-babysitting pass, deliberately not fixed there to keep scope:
- WebFetch grants persist the cleartext host in both key and label; internal hostnames can be sensitive. If zero-cleartext-on-disk is the standard (command/URL/input are already SHA-256 digests), drop or hash the host prefix.
- Grant labels are fixed strings, so
/permissions can show how many persistent grants exist but not what each one is — and only bulk clear is possible. Consider per-grant indices or encrypted labels so individual revocation becomes possible.
.lock files in the grants dir are never cleaned up (minor hygiene)./cd rescope swaps the store Arc, so in a (practically unreachable) contended case an in-flight turn finishes against the old scope — worth a doc comment or a swap-at-turn-boundary if it ever matters.
Context: grants are exact-match SHA-256 keyed (command digest + sandbox flags + policy fingerprint + cwd digest), repo-root scoped, cross-process locked with read-merge-write and refresh-before-honor semantics.
Follow-ups from the #507 review cycle — reported by the review-babysitting pass, deliberately not fixed there to keep scope:
/permissionscan show how many persistent grants exist but not what each one is — and only bulkclearis possible. Consider per-grant indices or encrypted labels so individual revocation becomes possible..lockfiles in the grants dir are never cleaned up (minor hygiene)./cdrescope swaps the store Arc, so in a (practically unreachable) contended case an in-flight turn finishes against the old scope — worth a doc comment or a swap-at-turn-boundary if it ever matters.Context: grants are exact-match SHA-256 keyed (command digest + sandbox flags + policy fingerprint + cwd digest), repo-root scoped, cross-process locked with read-merge-write and refresh-before-honor semantics.