The registration throttle called record_failure on every attempt, successful
ones included. Five legitimate signups from one shared or NAT address locked the
sixth real user out for up to 600s, doubling from there — while an attacker
sidestepped the limiter entirely through B-54. Failure backoff is the wrong
instrument here: nothing about creating an account is a failed guess at a
secret, so the only people it reliably punished were the honest ones.
RollingQuota says what was actually meant: 5 accounts per IP per hour, in a
rolling window. The caller over it waits exactly until the oldest of the five
ages out — an accurate Retry-After, and waiting never makes the next wait
longer. It is recorded only once an account exists, so attempts that create
nothing (a taken username, a validation error) leave the quota untouched, and
checked before the Argon2 hash, so an IP out of quota costs nothing to refuse.
Bounded like the failure limiter (B-56): the keys are caller-chosen, so the dict
gets both a sweep and a hard cap, evicting keys with room left in their quota
before full ones.
Also fixes the inline comment that cited B-31 (the resubscribe finding) where it
meant B-33.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The login throttle keyed on body.username.lower() while the lookup matched
User.username exactly, so "Bob" and "bob" were two accounts sharing one
rate-limit bucket — each able to lock the other out — and registration happily
accepted near-duplicate names, which on a custodial system is an impersonation
vector.
Uniqueness is now the database's job: a unique index on lower(username), with
register and login both matching through func.lower(). The name is still stored
exactly as typed, since that's what /admin and the audit log display, and the
username pattern is ASCII-only so lower() is the whole of the normalization.
The migration refuses to run if two existing accounts differ only by case. It
can't merge or rename one automatically: both are custodial accounts that may
hold funds, so that would be the migration silently deciding who owns what. It
names the collisions and leaves them to the operator — the container runs
`alembic upgrade head` at startup, so it surfaces as a refusal to start rather
than a half-applied schema. Verified both directions against a scratch DB, plus
`alembic check` (clean) and the collision guard actually firing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
_buckets is keyed by strings the caller chooses — any username, and (before
B-54) any IP — and only ever grew: decay_seconds aged a bucket's counter but
never removed the entry, so hammering login with random usernames was an
unbounded memory leak.
A bucket is "spent" once its lockout has expired *and* its failure count would
decay to zero on the next failure anyway — at which point keeping it and
dropping it are indistinguishable, which is what makes eviction safe. Those are
swept on record_failure (at most once every 60s) and on the read path, so a key
that's merely being probed never leaves an entry behind. That alone holds the
dict at the size of the genuinely active attack surface.
_MAX_BUCKETS = 50_000 is the backstop for a burst faster than the sweep
interval, when nothing has had time to expire. Over it, the entries closest to
expiry go first: what an attacker gets from a successful flood is the loss of
the shallowest, nearly-over lockouts, never the deep ones actually holding an
attack back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Argon2 is deliberately expensive — tens of milliseconds of CPU per call. Called
inline from the async handlers for register, login, change-password and the
admin reset, that cost froze the entire process for its duration: every other
request, plus all six background tasks (scheduler, confirmation poller, RBF
bumper, listener, both reconcilers). A burst of unauthenticated login attempts
was therefore not just slow logins, it delayed draws and confirmations.
hash_password_async/verify_password_async wrap the existing pair in
run_in_threadpool, and every async caller now uses them. The synchronous
functions stay: they're what the wrappers call, and what tests and scripts (no
running loop) use directly.
The regression test runs a heartbeat task alongside the hashing and counts how
often the loop got to run it — 1 tick with the old inline call, many with the
threadpooled one.
Also drops the running "already fixed and removed" list from BUGS.md: the file
tracks open findings, and `git log --all --grep 'B-nn'` is the record of how a
closed one was closed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
GET /rounds/stream capped concurrent subscribers with one global
counter (MAX_SUBSCRIBERS=500): anyone opening 500 connections degraded
every other user to polling. The comment called it a defensive cap;
it was actually the vector, since nothing stopped a single source from
exhausting it alone.
RoundEventBroadcaster now also tracks subscribers per client IP,
capped much lower (MAX_SUBSCRIBERS_PER_IP=5). Past that cap, opening
one more stream evicts that same IP's own oldest connection (woken via
a new EVICTED sentinel so the SSE generator closes it promptly) rather
than refusing the new one or letting one abusive IP crowd out unrelated
clients under the old global-only cap. The global cap stays as a
backstop against overall resource exhaustion regardless of source.
Extracted client_ip() (X-Forwarded-For, since Caddy reverse-proxies
every request) out of auth/routes.py into app/api/client_ip.py so the
login/registration throttles (B-33) and this new per-IP cap share one
definition instead of two that could drift apart.
Not implemented: enforcing the connection cap at the Caddy layer
itself, which the proposed fix also suggested - the standard Caddy
image this project uses has no such directive without a third-party
module, and building a custom image felt like a bigger, separate change
than this fix warranted.
Suite grows from 201 to 211 tests. BUGS.md moves B-38 to Previously
fixed.
Neither self-service password change nor the admin reset invalidated
already-issued JWTs — a 24h-lifetime token stayed valid regardless, so
a stolen token (or an attacker who already had the old password) kept
working past a password change meant to lock them out. The admin reset
exists precisely for the "account compromised" case and didn't evict
the attacker at all.
Add User.token_version (migration 943dbd74d983), embedded in every JWT
as a "tv" claim and checked against the DB on every request in
get_current_user/get_optional_user; a mismatch reads as session_expired.
Both change-password and the admin reset bump it. change-password hands
back a freshly minted token so the caller's own session keeps working
instead of being logged out by its own request; the admin reset does
not, since that session isn't the one making the call.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
POST /auth/login had no rate limiting, no lockout, no delay — a patient
distributed attack could brute-force a password against an enumerable
username list on a custodial wallet, where a guessed password means
withdrawing someone's funds.
Add per-username and per-IP throttling with exponential backoff
(app/auth/rate_limit.py), keyed on app.state like UserLocks rather than
a module global so each app instance gets isolated throttle state.
Unknown-user and wrong-password already shared one response path, so no
enumeration oracle there. Registration is throttled per-IP too, which
also bounds how many accounts one IP can spin up (B-31).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
fee_address was the dangerous one (B-05). PUT /admin/config assigned whatever it
was given, and a well-formed address from another chain (bc1...) parses fine as a
witness program — so every round's 30% commission would be signed and broadcast
to a script nobody holds the key for. A malformed one instead wedged the payout
with an unhandled EmbitError. It now has to pass is_valid_plm_address, the same
check user withdrawals already had. Numeric fields got bounds too:
fee_rate_sat_vb=0 produces transactions no node relays, which stalls bets,
payouts and withdrawals alike, and round_duration_seconds=0 expires a round the
instant it opens.
Config changes are audit-logged (B-10). /pause and /resume were logged but a
config edit wasn't, so the most sensitive setting in the system could be changed
without leaving any trace — contradicting CLAUDE.md, which says audit_log records
what changed. The entry carries a before/after diff per field, computed before
assignment, and no-op updates write nothing. `paused` was removed from
_CONFIG_FIELDS so the maintenance switch has exactly one audited path; it stays
in the response model.
Admin token comparison is constant-time (B-14), with the empty-token check kept
*ahead* of it: compare_digest("", "") returns True, so the obvious ordering would
have opened the panel on any instance without an ADMIN_TOKEN.
Registration input (B-12). It accepted an empty username and a one-character
password while /users/me/change-password demanded 8 — an odd place to be lenient
on a custodial system holding real funds. MIN_PASSWORD_LENGTH moved to
auth/security.py so both share it, and the username is constrained to 3-32 chars
of [A-Za-z0-9_.-]. The IntegrityError handler also distinguishes a username
collision (answers username_taken) from a derivation-index one (retries): a
concurrent duplicate username used to be retried five times and then reported as
derivation_index_conflict, which told the user the wrong thing.
verify_password (B-13) catches VerificationError and InvalidHashError, not just
VerifyMismatchError, so an unparseable stored hash reads as "wrong password"
instead of a 500 — logged as an error, since that one is a data problem.
guida-admin.md gains a table of the audit events worth watching, including
payout_failed, which needs manual intervention.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The dashboard now speaks seven languages but every failure path still showed
the API's raw English text ("insufficient balance", "current password is
incorrect"), which is the most frequent and least forgiving part of the UI to
leave untranslated.
Rather than teach the API about locales, it keeps answering in one language
and hands the client something to translate: `detail` becomes
{code, message, params}, where message stays English for non-dashboard
consumers (curl, tests) and code maps onto `error.<code>` in i18n.js. An
unknown code falls back to message, so a client older or newer than the server
degrades to English instead of a blank toast.
Domain exceptions (BetError, WithdrawalError) subclass the new ApiError and
carry the code from where the failure actually happens; str(exc) is still the
English message, so existing tests keep matching on it. Interpolated values
travel in params rather than baked into the English sentence — amounts as
*_sats, from which the frontend derives a *_plm sibling, so each language can
place them wherever its grammar wants.
admin.js reads detail.message defensively: the admin endpoints still return a
bare string, but the shared auth dependencies now return the structured form.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Balance display: place_bet/request_withdrawal spend whole UTXOs and mark
them spent at broadcast time, well before confirmation, so the confirmed-only
balance could drop by far more than the amount actually moving. Add
compute_pending_balance() (app/wallet/balance.py) to fold the unconfirmed
change from in-flight bet/withdrawal PendingTransactions back in; GET
/users/me now returns pending_balance_sats + has_pending, and the frontend
shows it colored green (settled) or amber (still pending) instead of the
confirmed-only figure.
Round outcome display: the win/lose reveal and the "pagamento al vincitore
in corso" status were fighting over the same UI slot, and the reveal broke
across a page refresh. Now:
- The round-status box (generic phase progress) and the personal win/lose
box are independent and can both be visible at once.
- The win/lose box only renders for users who actually played in that round
(new user_played field on GET /rounds/current, via a new optional-auth
dependency so the endpoint stays usable logged-out).
- The reveal delay is anchored to the round's server-provided closes_at
instead of a client-side "first seen" timestamp, so repeated reloads can't
reset it, and the revealed result is persisted in localStorage so it
survives a refresh even after the round has fully closed.
- GET /users/me/last-round-result is a durable DB-backed backstop for
players who miss the live window entirely (backgrounded tab, offline).
Also hardens the frontend polling loop: call() now times out instead of
hanging forever, and a session-epoch counter stops an in-flight request from
a previous login from resurrecting a duplicate poll loop after logout.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Argon2 password hashing, JWT session issuing/verification
(auth/security.py), register/login routes, the bearer-token
get_current_user dependency, and GET /users/me for address + balance.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
davideandClaude Opus 5