The Login tab creating accounts through Google is such an easy one to miss, because with OAuth there's no separate sign up call at all.
On evannn-dev's question about the race: Supabase has a Before User Created auth hook that runs before the row is written to auth.users, and if it returns an error the user is never created. If your invite codes are tied to an email address, the hook can check the incoming email against that list and reject everything else, which would also save you the create-then-delete step for fresh Google accounts. If codes aren't tied to emails it's harder, since the Google signup doesn't carry the code with it, so the check after the callback probably still has to stay.
Either way, putting the hard lock on the Stripe Connect function is the right call. That's the step you can't undo.
That hook is exactly what I was missing, thanks. Our codes aren't tied to an email, so the Google signup can't carry one, which means the check after the callback stays for now. Tying a code to an email at redemption could let the hook do the job though. Adding it to the list.
One test I’d add to the payout boundary: send a valid JWT for user A with user B’s ID in the body and assert 401, then use an authenticated but uninvited account and assert 403. The unauthenticated check proves the outer gate; these cases pin down that a caller can create a payout account only for themselves and only after passing invite access.
Good call. The function already compares the JWT user with the userId in the body (401) and checks invite access before Stripe (403), but neither is pinned by a test yet. Writing both.
Good call pinning both cases. A 401 for a mismatched JWT user and a 403 for missing invite access make the identity check and the access policy easy to distinguish. I’d also assert that Stripe is never called in either failure case.
Done, both are pinned now: A's token with B's id gets a 401, an uninvited owner gets a 403. The gate runs before any Stripe call, so neither case ever reaches Stripe.
Nice, that closes the two important failure paths: identity mismatch returns 401, and an uninvited owner returns 403 before Stripe is touched. Thanks for pinning the boundary with tests.
One layer worth checking alongside the edge-function gates is direct database access. supabase-js can query tables through PostgREST without going through your edge functions, so an invite check in a function doesn't cover that route. For exposed tables, RLS policies need to enforce the same access rule. Supabase documents that enabling RLS blocks access through the API with a publishable key until policies allow it. I'd add a direct table query from a non-invited authenticated account to the regression tests, alongside the OAuth and payout cases.
Fair point, the edge function gate means nothing if PostgREST lets the same account read the table directly. RLS is on for our tables, but I haven't tested them from an uninvited account. Adding that query to the regression set next to the OAuth and payout cases.
Update: I ran that check. You were right, the database had a side door that the edge function gate couldn't see. Fixed the same day, and I wrote up how I tested it without touching production: meridianbuild.dev/blog/my-invite-g... Thanks for asking the question I was avoiding.
Full-Stack Web Developer specializing in React.js and Laravel. I build dashboards, SaaS applications, and management systems for businesses worldwide. Open to freelance projects.
Location
Lahore, Pakistan
Education
BS in Computer Software Engineering, Gomal University (2021–2025)
The detail that stuck with me is the two defaults pointing opposite ways. Client catch block set inviteOnlyMode to false, the server version fails closed. I had the same bug shape once: a feature flag fetch failed on the client and opened a gate that should have stayed shut. Flipping the client default to locked was the cheapest fix that ever stuck. Your _shared/invite-gate.ts is the grown-up version of that lesson: one function, one default, and every caller reads from it.
"Cheapest fix that ever stuck" is the right way to put it. The scary part is that the client default looked harmless in review, it only meant something when the fetch failed.
when the invite_only flag lives only in the DB, what guarantees do you have that the check runs before the user record is persisted, and have you considered a middleware gate to avoid the race?
Honest answer: nothing guaranteed it at the time, that was the bug. A Google sign in from the Login tab created the auth user before any check ran. Today the check runs right after the OAuth callback and removes uninvited accounts, and the real lock sits on the payout function. Elijah's reply below points to a cleaner fix with the Before User Created hook.
We're a place where coders share, stay up-to-date and grow their careers.
The Login tab creating accounts through Google is such an easy one to miss, because with OAuth there's no separate sign up call at all.
On evannn-dev's question about the race: Supabase has a Before User Created auth hook that runs before the row is written to auth.users, and if it returns an error the user is never created. If your invite codes are tied to an email address, the hook can check the incoming email against that list and reject everything else, which would also save you the create-then-delete step for fresh Google accounts. If codes aren't tied to emails it's harder, since the Google signup doesn't carry the code with it, so the check after the callback probably still has to stay.
Either way, putting the hard lock on the Stripe Connect function is the right call. That's the step you can't undo.
That hook is exactly what I was missing, thanks. Our codes aren't tied to an email, so the Google signup can't carry one, which means the check after the callback stays for now. Tying a code to an email at redemption could let the hook do the job though. Adding it to the list.
One test I’d add to the payout boundary: send a valid JWT for user A with user B’s ID in the body and assert 401, then use an authenticated but uninvited account and assert 403. The unauthenticated check proves the outer gate; these cases pin down that a caller can create a payout account only for themselves and only after passing invite access.
Good call. The function already compares the JWT user with the userId in the body (401) and checks invite access before Stripe (403), but neither is pinned by a test yet. Writing both.
Good call pinning both cases. A 401 for a mismatched JWT user and a 403 for missing invite access make the identity check and the access policy easy to distinguish. I’d also assert that Stripe is never called in either failure case.
Done, both are pinned now: A's token with B's id gets a 401, an uninvited owner gets a 403. The gate runs before any Stripe call, so neither case ever reaches Stripe.
Nice, that closes the two important failure paths: identity mismatch returns 401, and an uninvited owner returns 403 before Stripe is touched. Thanks for pinning the boundary with tests.
One layer worth checking alongside the edge-function gates is direct database access. supabase-js can query tables through PostgREST without going through your edge functions, so an invite check in a function doesn't cover that route. For exposed tables, RLS policies need to enforce the same access rule. Supabase documents that enabling RLS blocks access through the API with a publishable key until policies allow it. I'd add a direct table query from a non-invited authenticated account to the regression tests, alongside the OAuth and payout cases.
Fair point, the edge function gate means nothing if PostgREST lets the same account read the table directly. RLS is on for our tables, but I haven't tested them from an uninvited account. Adding that query to the regression set next to the OAuth and payout cases.
Update: I ran that check. You were right, the database had a side door that the edge function gate couldn't see. Fixed the same day, and I wrote up how I tested it without touching production: meridianbuild.dev/blog/my-invite-g... Thanks for asking the question I was avoiding.
The detail that stuck with me is the two defaults pointing opposite ways. Client catch block set inviteOnlyMode to false, the server version fails closed. I had the same bug shape once: a feature flag fetch failed on the client and opened a gate that should have stayed shut. Flipping the client default to locked was the cheapest fix that ever stuck. Your _shared/invite-gate.ts is the grown-up version of that lesson: one function, one default, and every caller reads from it.
"Cheapest fix that ever stuck" is the right way to put it. The scary part is that the client default looked harmless in review, it only meant something when the fetch failed.
when the invite_only flag lives only in the DB, what guarantees do you have that the check runs before the user record is persisted, and have you considered a middleware gate to avoid the race?
Honest answer: nothing guaranteed it at the time, that was the bug. A Google sign in from the Login tab created the auth user before any check ran. Today the check runs right after the OAuth callback and removes uninvited accounts, and the real lock sits on the payout function. Elijah's reply below points to a cleaner fix with the Before User Created hook.