You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Reviewed clean always fails on System & app caches: live caches invalidate inode-pinned preview #526
macOS with Full Disk Access granted to dev.caezium.Burrow
Reproduced from the GUI Clean pane → System & app caches
Problem
Every attempt to clean System & app caches ends with:
The reviewed files changed
Nothing was cleaned. Rescan before trying again. (The cleanup selection no longer matches the reviewed preview.)
Rescanning and retrying does not help. The other Clean entries (Project build artifacts, Leftover installers) are fine; this is specific to the long, live-cache scan.
What we observe
A System & app caches scan routinely takes several minutes and surfaces thousands of items / tens of GB (local logs show sessions like 4351 items, 36.71GB).
Those paths are actively rewritten by browsers, IDEs, cloud sync, etc. while the preview is open.
Confirming cleanup then fails closed — nothing is removed. ~/Library/Logs/mole/operations.log shows the session filled with SKIPPED / protected entries and no successful removals for these attempts.
From a user perspective this makes the most valuable Clean category effectively unusable on a busy machine: long scan ⇒ live mutation ⇒ guaranteed stale preview.
Why this looks structural (not user error)
In macos/Sources/CleanupAuthorization.swift, a reviewed clean pins filesystem identity (device:inode:uid:mode) and gives the snapshot a fixed lifetime (CleanupSnapshot.lifetime = 300 seconds). Launch validation requires every selected identity to still matchesCurrent(). That is the right fail-closed model for a deliberate review of static paths, but it fights System & app caches:
the scan itself takes a large fraction of (or exceeds) the 5-minute window on a full disk;
password / Touch ID elevation adds more delay before the boundary re-check;
live cache trees change inodes constantly, so any busy Mac will invalidate the plan.
So “files changed mid-review” is the normal state for this category, not an edge case.
Expected behavior
For System & app caches, either:
Soften per-item: at confirm time, drop or re-pin entries whose identity changed, and still clean the rest (report skipped/changed items honestly); or
Category-level engine clean: for this category, run the engine’s own cache selection (the path described as “without a per-item review”) instead of an inode-pinned reviewed plan; or
Honest UX: if the reviewed path cannot work for live caches, say so up front and offer the non-reviewed / engine path, instead of an endless rescan loop.
Happy to test a build that tries any of the above.
Related
This does not appear to be tracked as its own issue. Closely related design work made the reviewed plan stricter (e.g. Safe Release / reviewed-plan enforcement), which is good for safety — the ask is a workable path for the live-cache category specifically, not a rollback of fail-closed deletes elsewhere.
Activity