Project status

This is the single place that statement lives; everything else on the site and in this manual links here.

Why this page exists rather than a version number

Drivel is a filesystem that deletes files on your behalf, in both directions, partly on the strength of what it infers happened while it was not running. That is the job, and there is no version of it that does not involve acting on a belief about the past.

Data safety describes the rules that govern those decisions, and they are real: a deletion is only ever inferred from a sync baseline, the first run deletes nothing, a local file that diverged is kept and pushed back rather than removed, and a delete pass whose count looks wrong abandons itself instead of trimming. The design is careful on purpose.

What is missing is not the care. It is the mileage. The rules have been reasoned about far more than they have been run, and every serious bug found so far was found by running the thing on a real machine against a real server — never by reading it, and never by a green test suite. The manifesto commits to saying what has not been done. This is that, at the level of the whole project.

What this means for you

• Keep a backup, and restore from it once. A backup nobody has ever restored is a hypothesis. This is the only item on this list that would save you from everything else on it.
• Do not let a Drivel mount hold the only copy of anything. The backing directory is ordinary files and survives Drivel being uninstalled — but it does not survive Drivel deleting from it, which is a thing Drivel is designed to do.
• The first run against an existing tree is the riskiest moment. That is when the enumeration sweep reconciles two sides that have never been compared. It deletes nothing on a first run by design; it can still push, pull, and create conflict copies in bulk.
• Watch the log the first few times. Refusals, skips and conflict copies are all announced. A mount that is quietly doing nothing is easier to notice early than to diagnose later.
• Prefer a scratch copy for evaluation. Point -data at a duplicate of the tree you care about, not the original, until you have watched it behave.
• Be deliberate about -max-deletes. It defaults to 100 and it is the guard that turns a wrong premise into a refused pass instead of a mass deletion. Raising it is a decision, not tidying.

What has actually been run

Every package has tests and CI runs them under the race detector on every push; that is the floor, not the evidence. Above it:

• Google Drive — exercised continuously in development against a real account, with one behaviour (what a removal does to the remote object) pinned by a live test against the real API in September 2026.
• SFTP — verified through a real mount against OpenSSH in September 2026: an existing tree materialised, edits moved both ways, rename and delete propagated, the diverged-copy guard held. Two bugs were found by that single run that nothing else had surfaced.
• Linux is the platform everything is developed on. FreeBSD has been run on its own hardware. macOS has never been run at all — it compiles, and that is the entire claim. See platform support .

What has not happened is sustained multi-machine use by people who did not write it, over months, on data they care about. That is the missing input, and no amount of further reading replaces it.

Known sharp edges

These are documented rather than fixed, because each one’s available fixes cost something worse. They are not the whole list — they are the ones you could plausibly meet.

• Same-name siblings on Google Drive resolve to one visible file, and a deletion can make the path reappear with an older sibling’s contents. See data safety .
• A case-insensitive backing filesystem (the default on macOS) collapses two distinct remote names into one local path. Nothing handles this yet.
• Lazy mode is unsafe on a backing filesystem without extended attributes, and the mount warns rather than refusing. See lazy mode .
• SFTP polls, because the protocol has no change feed, and its default poll interval is far too long for interactive use. See SFTP .

When this page goes away

Not on a date, and not at a version number picked to look finished. It goes away when sustained real-world use has stopped producing the kind of bug that the last several real-world runs each produced — and the honest way to read this page is that until then, the project would rather have your scepticism than your trust.

If you run it and something goes wrong, that is the most useful thing you can do for it: open an issue , and say what you had mounted and what you expected.