Module clean 

The disk janitor: finished runs get their worktrees folded, worktrees whose run record is already gone get reclaimed too, and the shared build cache is pruned to its cap.

A run’s state being written by an older schema is not the same thing as it being unreadable, and this module used to conflate the two: fold_due treated any run.json its version check rejected exactly like one that failed to parse at all, so a single schema bump silently stopped every automatic fold in the fleet the moment it shipped, and did so with no counter and no log line to say so. A record magi genuinely cannot parse — missing fields, broken JSON, a schema newer than this build has ever heard of — is still left alone here, still counted in Housekeeping::unreadable, and still only ever removed by an explicit operator action (magi fold, or the equivalent phone route). One written by a schema this build merely disagrees with the meaning of is not that: as long as it still parses, folding proceeds regardless of the number in its schema field.

Everything policy-shaped — which statuses are foldable, how long a finished run is left alone, whether the cache is over its limit — is a pure function injected with numbers, so nothing here has to ask the operating system to be testable. The only I/O is the removal itself.

Structs

Housekeeping

What one janitor pass did, for the caller’s log line.

Functions

abandon_settled_questions

Abandon every open question whose run has already settled into a status nothing comes back from, worded with what the run became - the same cleanup graph::Runner::settle_questions runs the moment status lands there, for questions that missed it.

cache_report

The cache’s path, size and cap, for magi cache show and the health view. None when the config declares no CARGO_TARGET_DIR to aggregate.

cache_size

Size in bytes of the shared build cache.

clear_abandoned_active

magi fold’s recovery path for a run whose worktrees are already gone — so crate::graph::fold_run removed nothing — but whose run.json still lists active seats nobody is left to answer for: no live daemon claims the run, and every one of those seats has overrun its own timeout (see RunState::active_all_overrun). Clearing them and failing the run is what lets it be deleted afterward — RunState::ensure_can_delete only ever checks whether a live daemon is working on the run and whether its candidates are folded, not status, but a run stuck implementing forever with an empty worktree still reads as unresolved everywhere else (magi show, the deck, the phone) until this runs.

due

Is updated old enough, measured against now, that the run may fold?

fold_due

Fold every run that is finished, older than the grace period, and not being worked on; return (folded, unreadable).

fold_orphaned_worktrees

Reclaim worktrees under worktrees_root that no run record in runs claims anymore, and return how many were removed.

fold_unreadable

Remove a run that cannot be read: its state directory under runs and its worktree directory under worktrees_root.

housekeep

Run the janitor: fold due runs, reclaim orphaned worktrees, prune stale worktree registrations, then prune the cache if it is over its cap.

prune_cache

Delete files from the shared build cache until it fits its cap.