Small enough that one person's bug report gets fixed the same day. Today that happened to Quill twice, both times thanks to you.
Big networks have more people. Small ones still have conversations.
Login to reply
Replies (12)
That speed cuts both ways. A one-person client fixes your report in a day and goes quiet for a week when that person takes a holiday. Fewer users also means fewer people tripping over the same bug, so it can sit unnoticed for months before anyone files the report at all. The same smallness that makes the fast fix possible is what makes the client fragile. Nostr at least splits that risk: relays, clients and wallets fail separately, so a burnt-out maintainer costs you one tool rather than your identity.
Fair, and that's the real argument for building clients this way: if I disappear for a month, you lose a tool, not your account. Quill holds nothing of yours: your key stays in your signer, your notes on your relays, and the code is MIT, so anyone can fork it or host their own copy. A small client should be cheap to walk away from.
Agreed on the principle, and the client is the easy half of it. State is the hard half.
Forking Quill takes an afternoon. Getting your notes back out takes an audit, because the copies are not where you assume they are. I sent my last note to eight relays and counted them just now: six have it. One of the two missing is a relay I run myself. Nothing told me either; the write simply never landed. Client portability is a solved problem. Portability of the state is not, and that is the gap a user finds only when they try to leave.
Deletion is worse than migration. My first attempt at removing a long-form post sent the event id alone; every relay kept its copy. It disappeared only after I also named the replaceable slot by address. Same lesson, harder: walking away means taking the thing with you, not leaving a copy on somebody else's disk.
A test I would trust is blunt. Export everything, delete the account, import it somewhere else, and see whether the thread still renders. If a client passes that, "cheap to walk away from" is a property of the software instead of a line in the README.
Agreed, state is the hard half, and your count matches what I saw. That's why Quill has a check for exactly that audit: it asks each of your relays which of your notes it holds, re-sends the missing ones as they are (already signed, nothing re-signed), and saves a file with everything it found: notes, articles, profile and lists.
Your full test it doesn't pass yet: nothing reads that file back in. Good next step, thanks.
Right, and that is where I'd expect it to bite: import is not the mirror image of export. Export reads, import writes, and writes are the policed half. My own relay takes writes only from a whitelist, and plenty of public ones quietly drop a kind they don't care about or shut the door after a burst. So a restore that reports "done" is worth nothing. It has to report per relay what was accepted and what was refused, and hold the refused ones in a queue instead of pretending they landed.
Two things I'd put in the file before you build the reader.
Tombstones. An export of what currently exists will, on a fresh relay, bring back the notes I deliberately deleted, because deleting is a separate act and nobody remembers to repeat it. My long-form removal only took effect once I sent a kind 5 that named the address of the replaceable slot, not just the event id. If the file doesn't carry those, then import is also an undelete.
The event ids. Replies to me point at the ids of my notes. An import that re-signs anything doesn't restore a thread, it forks one, and the old replies end up hanging under notes nobody will open again. Signed bytes or nothing.
And I'll take the blunt test on myself: when there's an import, I'll run it against my own key, delete, import, then check thread by thread whether anything still renders, per relay, refusals included. If your client passes that, it's the first time I've seen a Nostr app prove the README instead of writing it.
It exists since this morning, so the test is on. Where it stands against your list:
- Per relay: each relay's accepted and refused count, with the relay's own reason. No queue yet; running it again is safe, already-held events count as accepted.
- Tombstones: the file keeps your kind 5s and leaves out what they delete, by id and by address.
- Ids: signed bytes only, nothing re-signed.
Settings → Your notes, safe. Tell me what breaks.
Good, then the interesting part starts. Three places I'd expect it to break, in the order I'd bet.
Accepted is not held. A relay has answered my write with a success frame and then never served the event; nos.lol did exactly that, and nothing in the response said so. If the number comes from the relays' answers rather than from a second question about what they hold, a green report can still be wrong. The count I'd trust asks each relay afterwards which of the ids it actually has. It costs another round trip and it is the only one that measures their disk instead of their mood.
Order of the tombstones. If a kind 5 reaches a relay before the note it deletes, there is nothing there to delete, and the note arriving afterwards simply stays. Sending deletions last costs nothing and closes it. Mine needed the address form next to the id, as you have it, but the pair only works if it lands after the thing it kills.
Replaceable kinds are a snapshot, not a state. Profile, follow list, relay list: the file carries the version that existed at export. Import it into a relay that already holds a newer one and yours loses on created_at, which is harmless but gets reported as refused. Restore onto a fresh relay and the reverse happens quietly, with a stale profile winning because it is the only one there. Printing created_at per replaceable in the report would make the direction visible instead of guessed.
No queue is fine by me, provided the refused list comes out as a file I can feed back in. Then the rerun is the queue, and I prefer that to one that lives inside the app and dies with the tab.
On the test I promised: I will run the delete and restore on my own key, but in a planned window rather than on a whim, because there is a zap address and a profile in there. The half that can be done without risk is also the half that catches most of the above: export, then ask every relay which of the file's ids it really holds, and check that the kind 5s carry both forms. If that passes, the restore is worth the risk.
All four are in now. Deletions go last. After sending, each relay is asked by id what it really holds; "said ok but not served" shows in red per relay. What didn't land (refused, never tried, or not served) downloads as a file you feed back in, so the rerun is the queue. And a restore lists each replaceable it sent with its date.
Your safe half works as you described: download the copy, then "Check my notes" asks every relay what it holds. Thanks, this was a better spec than mine.
Then one hole in the check as written: asking by id only works for kinds that keep every version. For a profile, a follow list, a relay list or a long-form slot, a relay keeps the newest and drops the older ones, so an exported kind 0 that had already been superseded comes back as "not served" in red while the relay is doing exactly what the spec says. Ask for replaceable kinds by author plus kind (and the d tag for 30023) and compare created_at against the file; keep the by-id question for kind 1 and 7, where every version is meant to persist.
And the red list gets more useful if it names which question failed: never sent, refused at write, or held but not served. Those are three different bugs and one column hides two of them.
On my side the safe half is running as soon as you say it is live. The delete-and-restore on my own key stays in a planned window, but I'll report back per relay either way, refusals included.
Both fixed, and it's live. Replaceables are now checked by author, kind and d tag: a newer version on the relay counts as held and shows as "newer version there", not red, and stays out of the not-landed file. Notes, comments and deletions are still checked by id. Each relay's line keeps the three apart: refused (with the relay's reason), never sent, and said ok but not served.
Ready for your safe half.
The line that earns everything in this thread is 'said ok but not served', in red. A relay's ack is a handshake, not a delivery, and most tools quietly treat it as one. Keeping refused / never sent / not served apart, and making the not-landed file the queue for the rerun, is the same move twice - the report is the work order. Thanks for shipping it in public, one fix at a time.
Thanks. Credit where it's due: Roger wrote most of that spec in this thread, I just kept saying yes and shipping it. "The report is the work order" is a better summary than anything I put in the README.