Cleanup targets found on disk, Podman host-disk reclaim, and readable container images

[rducom]

Low-Hanging Fruits only ever offered paths written down by hand. This adds a second kind of target, whose paths are found on the disk, plus a host-side reclaim for Podman and a readable container image list.

Figures below come from one developer Mac and are there to show scale, not as general claims.

New: discovered targets

CleanupDiscovery finds targets whose location depends on what is installed. Each one pairs a candidate directory with a fact on disk that explains it, and offers nothing it cannot pair:

Target	Offered when	Found here
bin/, obj/	a .csproj/.sln sits beside it	8,360 dirs, 20.8 GiB
target/	a Cargo.toml sits beside it	—
node_modules	a package.json beside it and no install for 6 months	5 projects, 2.0 GiB
workspace state	its workspace.json names a folder that no longer exists	6 of 193
Chrome/Firefox caches	inside the browser's own cache tree	5 profiles

The marker rule is what makes this safe. A find -name bin sweep over a home directory deletes every Python virtualenv on it; the rule rejected all of them across 8,360 real matches.

Two behaviours are new and worth knowing about:

• removal: .directory — discovered targets delete the directory itself, where caches only ever had their contents emptied. Build tools recreate it.
• regenerable: false — targets whose bytes never come back (Trash, orphaned editor state). Excluded from Select all, badged not a cache, named individually in the confirmation. Replaces the hardcoded id == "trash".

Podman: reclaim host disk

Pruning inside the VM frees guest blocks but leaves the host image allocated until the guest discards them, which CoreOS only does weekly. A new action runs fstrim on demand and reports the change in the image file's st_blocks.

fstrim's own output is not usable here — it printed 39.5 GiB for 2.06 GB actually returned — so the figure is measured before/after, never reported.

Container images

227 rows, 200 of them <none>, most the same size. Two changes:

• Identity. A displaced image keeps its former tag in History; 198 of the 200 resolve. Full ladder: live tag → repository by digest → former tag → OCI labels → <none>. Everything below the first rung is inference, so the row carries a badge saying so. Resolved names group, collapsing the list to ~12.
• Size. An image's size includes layers its neighbours also hold; the rows summed to 100.8 GiB against 17.3 GiB of real images. podman system df -v gives the bytes an image holds alone, shown when it diverges (906 MiB listed, 11.6 KiB alone).

No reclaim figure is offered per group. For 38 rebuilds of one tag, size sums to 42.6 GiB and per-image unique size to 0 B; neither is what removing the set frees, so the group says the layers are shared and defers to system df.

Fixes

• Container prune buttons never opened their confirmation. Three .alert modifiers on one view; SwiftUI presents one. Verified against 8aa8fcd, so this predates the branch — three destructive actions have been unreachable in shipped builds. All dialogs now route through one enum and one modifier.
• The installed skill was never refreshed. plan() treated "directory exists" as "up to date", so corrections shipped in the app never reached the assistant reading it (86 lines installed vs 157 bundled on this machine). Now compared by content, with an Update action in the panel.
• Catalog paths. uv keeps its cache under XDG on macOS, so ~/Library/Caches/uv alone found nothing while ~/.cache/uv held 4.19 GiB. Same for pre-commit/prek, gh, opengrep, opencode. ~/.npm/_npx joins the npm target.
• ToolActivity gap: go mapped to go-build but not go-mod.
• The journal would have written one 8,360-path line per clean; capped at 32 with the true count kept.

Review focus

• CleanupDiscovery.swift — the marker rules, the prune list, and isOldEnough. This decides what gets deleted, and it is the only place in the app that does so from disk contents rather than from a constant.
• CleanupEngine.clean — the .directory branch. It deletes a whole directory, fenced by passesFence, which discovery reuses unchanged.
• ContainerQueries.identify — every rung below the first is a guess presented as an image's name.
• ContainerQueries.parseUniqueSizes — parses a table because podman refuses --format json with --verbose. Read from both ends; CREATED is a human duration of unpredictable width.

Testing

232 tests (+22). The refusals are pinned before the successes: virtualenv bin/ in four spellings, unmarked bin, node_modules not descended into, the depth bound, the fence on discovered paths, live vs orphaned workspaces, unresolvable workspace.json, the buildah hex-stem rule, a three-token CREATED column, and the group that must not invent a total. explain is covered through a real tools/call.

Every UI change was opened against the running app and read on screen; the four container dialogs were opened and cancelled rather than confirmed.

Known gaps

• The "installed skill is outdated" banner has test coverage but its rendering was never seen — syncing the copy on this machine makes the state unreachable.
• denseBurstCoalescesToItsClusterAncestor fails about 1 full-suite run in 5, alternating between two assertions, never in isolation. Pre-existing timing race in the test, untouched here.
• Discovery adds ~10 s to the mode's load on a large home. It runs off the main actor, alongside catalog sizing rather than before it, so rows still stream in.