The channel used DropWrite, under which TryWrite reports success and
discards the email, so a full queue still counted the request and never
sent the code. Wait makes TryWrite return false when the queue is full,
without blocking, so the request is released and the user can ask again.
An admin revoke overturns a human decision, so admins may revoke only
admin-approved uplifts. The work-order revoke path let an admin revoke an
auto-approved uplift they had requested themselves. Both revoke endpoints
now refuse it; dispatchers keep revoking their own auto-approved uplifts.
A work order whose uplifts were all cancelled, withdrawn, expired or revoked no
longer matches the advanced-search Has uplift filter, and the board Uplift
column no longer reports it as having an uplift or shows the dead one as its
primary status. One shared live-status list backs both queries. Explicit
cancelled/revoked sub-filter values still find those work orders.
SendCodeAsync validates the invite, then starts the code with a conditional
update that also requires the invite to still be open. When an admin revoked
the link (or it was used) between those two reads, the refusal was reported as
resend_too_soon with a Retry-After, although the link was already dead.
On a refused start the invite is now read again: a closed invite gets the same
generic invalid_invite response as any other dead link, and a cooldown or send
limit refusal is computed from the fresh row.
A standalone severity patch on an Overdue WO still stores the value,
because the board patches severity before type when correcting Overdue
to Emergency/Reactive and that write must not be dropped or rejected.
Project the severity as null for Overdue in the board row mapping, which
also feeds the PATCH response, search results and the detail view, so a
stored value never surfaces on a type that carries no severity.
- Forgot Password is limited to 3 codes an hour and 10 a day per email, and
an account gets 10 failed code checks a day across every code it is sent,
so new client addresses and new codes no longer buy more guesses. Refused
requests answer exactly like accepted ones.
- The reset email is queued to a background sender, and unregistered
addresses store a row no code can match, so both paths do the same work
and return without waiting on the mail provider. Each request also clears
expired codes.
- Code hashes are HMAC-SHA256 under a key derived with HKDF from the JWT
signing secret; rows in the previous unkeyed format stop matching.
- Email and code are read only from the JSON body.
Two concurrent first requests can leave two pending codes for one email.
Exhausting or using one now deletes all of them, so a sibling code cannot
become live afterwards.
Forgot Password answers every address the same way and emails a code only
to an active account. Codes are stored as salted SHA-256 hashes, expire 15
minutes after issue, are replaced by a newer request, and are checked only
against the email they were issued to. Five failed checks delete the code;
attempts are reserved with one conditional UPDATE so concurrent guesses
cannot exceed the budget. VerificationCode requires the email, and email and
code are accepted in the JSON body so they stay out of URLs.
The three anonymous endpoints are rate limited to 10 requests per 15
minutes per client IP. Forwarded headers are trusted only through loopback
and private hops, since the API sits behind the EB load balancer and nginx.
The migration adds hash, salt, expiry and attempt columns and deletes the
old plaintext rows.
Review asked whether rows written by the old DateTime.Now stamps need a
backfill. They do not: the API has only run on Linux Elastic Beanstalk
hosts at their UTC default, and nothing in Terraform, .ebextensions or
.platform sets a time zone, so DateTime.Now already equalled UTC there.
Record that next to the hosting table so the decision is findable.
