The login gate called GET /guilds/{guild}/members/{user} — the Guild
resource's Get Guild Member, which wants a Bot token and the application
present in the guild. Handed a user Bearer token it answers 401, which
discordMember reports as an error, so every sign-in rendered "Discord
sign-in is unavailable right now" and nobody could get in.
The endpoint guilds.members.read actually grants is Get Current User Guild
Member, GET /users/@me/guilds/{guild}/member. Same single-guild question,
same privacy property, and it takes the token we hold. #18 flagged this as
verified from Discord's documentation but never from a live flow; it was
wrong.
The stub mirrored the implementation, so the suite could not see it. It now
serves the OAuth path and answers the bot path 401 the way Discord does —
without that, a regression falls through to 404 and reads as an ordinary
"not a member" refusal instead of failing.
Review findings on the previous commit:
- README's config table listed 8 of the backend's 25 environment variables,
omitting the whole DISCORD_* set that the backend refuses to start without.
The canonical reference cannot be missing the vars that gate startup.
- `docker volume ls --filter name=` is a substring match. Both runbooks used
it to find a volume they then mount read-only or delete; a second matching
volume makes the mount fail obscurely and the delete take both. Derive the
name exactly from the compose project instead.
- CUTOVER §6 removes the retired volume 'once the Postgres data has been
trusted for a while', in a shell where §1's $VOL no longer exists.
- REDEPLOY and DEPLOY still pointed at /opt/bookmarkmanager, so CUTOVER §6's
handoff to REDEPLOY §1 sent the operator to a path that does not exist.
Dry-running CUTOVER.md against production surfaced four things that would
have failed mid-cutover:
- The volume is named after the compose project, which is the lowercased
directory name (mangabookmark), not the repo name. Both runbooks hardcoded
bookmarkmanager_bookmarks-data. Derive it from docker instead.
- §1 asserted a clean stop leaves no -wal. compose stop SIGKILLs after 10s,
and a surviving -wal holds writes bookmarks.db alone does not, so the
export would silently lose them. Check rather than assume.
- jq is not installed on the server; §5's read-path check now uses python3,
which the runbook already requires.
- REDEPLOY §1 still listed the pre-split table set, omitting readers and
sessions from both the \dt output and the pg_restore contents.
Also records the git-pull failure the redeploy hits on a checkout whose
remote is the HTTPS clone URL.
sulthan