| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
|
|
|
|
|
| |
createproof's default proof omits invreq_payer_note, and passing include
replaces the default disclosure set — so for zap payments request both
invreq_payer_note and invoice_amount, which NIP-177 verification requires.
Verified end to end on v26.06.2: NWC pay over nip44 returns a proof that
decodes valid with note + amount disclosed and preimage matching the
payment hash.
Co-Authored-By: Claude Fable 5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- sendLightning fetches an invoice from lno1 offers (payer_note passthrough
for NIP-177 zap intents) and pays it through the existing bolt12 path
- NWC: new nwc#2 pay/receive methods (BIP-321 URIs, payer proofs stubbed
until CLN 26.06 createproof), nip44_v2 alongside nip04 with per-request
scheme detection, dual 23196/23197 notifications, 13194 refreshed when
its content or encryption tag goes stale
- standing per-user bolt12 offer via GET /offer; incoming bolt12 payments
record invreq_payer_note as memo
- parse/fetchinvoice routes handle offers; zap display parses kind 9736
receipts (amount tag + embedded intent) and drops a call to a sync()
helper that never existed
Co-Authored-By: Claude Fable 5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
POST /preimages now accepts fiat + optional currency (defaults to the
merchant's) as an alternative to a sat price. The fiat amount is stored
per entry and converted to sats at the moment a usePreimage invoice is
created, so the price tracks the exchange rate instead of freezing at
upload time. The invoice also takes on the entry's currency and rate so
receipts display the fiat price exactly as the merchant set it. If no
rate is available for the entry's currency, the preimage is pushed back
onto the queue and creation fails rather than issuing an any-amount
invoice.
Co-Authored-By: Claude Fable 5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Invoice creation is unauthenticated, so buyers could name their own
price for a queued preimage (create the usePreimage invoice with
amount: 1). POST /preimages now takes an optional price (sats) that is
stored per entry and overrides the requested amount when the preimage
is consumed; omitting it keeps the old any-amount behavior. Also gate
the stored-preimage reveal in credit() on the invoice being fully paid,
since internal sends can settle partial amounts. Bare-hex entries
queued before this change still parse (treated as priceless).
Co-Authored-By: Claude Fable 5 <[email protected]>
|
| |\
| |
| |
| |
| | |
# Conflicts:
# lib/payments.ts
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Merchants POST /preimages (auth) to pre-load 32-byte hex preimages into a
FIFO queue; creating an invoice with usePreimage: true consumes the next
one so the invoice's payment hash is sha256(preimage). Paying the invoice
reveals the merchant-chosen secret to the buyer — via their own wallet on
external settlements, and via GET /invoice/:id once settled — enabling
backend-less digital download stores (encrypted files on static hosting,
preimage + public string = decryption key).
- GET /preimages lists the queue with payment hashes so the merchant can
map files to future invoices; POST /preimages/delete clears it
- empty queue fails invoice creation rather than issuing a random
preimage the buyer would pay for but couldn't decrypt with
- preimages cln rejects as already used are discarded and the next tried
- consumed preimages persist at preimage:${payment_hash} (TTL = invoice
expiry) and credit() prefers them as the settlement ref, so internal
coinos-to-coinos payments reveal the real preimage instead of the
sender's uid
Co-Authored-By: Claude Fable 5 <[email protected]>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
apiv2.nobitex.ir became unreachable from our egress, and it was fetched
inline on every binance tick, so it failed and logged on every tick.
Move the Iranian market-rate fetch onto its own 60s timer with a request
timeout, cache the last-good value (falling back to the fixer rate after
an hour), and throttle failure logging. Switch the source to Wallex
(api.wallex.ir BTCTMN), which is reachable.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
On-chain deposits below 300 sats are dropped by the dust filter in the
confirm/credit path, so a specific-amount bitcoin/liquid invoice under
that threshold can never be credited. Reject creating one at generation
time and point the user at lightning. Any-amount (0) addresses are
unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| | |
| |
| |
| |
| |
| |
| |
| | |
The preHandler request logger wrote the full request body, which included
plaintext passwords on auth requests. Redact password/confirm fields
before logging.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
credit() now credits the tip the sender was actually debited (passed in
by the internal-transfer callers) instead of independently re-reading the
invoice. This closes a time-of-check/time-of-use gap where an invoice's
tip could be changed between the sender debit and the receiver credit,
letting a payer be credited a tip they were never charged.
External receipts (lightning/ecash) keep reading the invoice tip, since
there is no paired coinos debit to diverge from.
Also harden POST /invoice/:id: validate the tip is a sane non-negative
integer and reject tip changes on an already-settled invoice.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
The nobitex endpoint host changed (old host now NXDOMAIN) and the fetch
was swallowed by an empty catch, silently blanking Iranian rates. Point
at the current host, and since the API quotes in Rial despite the IRT
pair name, set IRR from lastTradePrice and IRT = Rial/10. Log fetch
failures instead of ignoring them.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
publishInfo sent the kind-13194 info event over the long-lived NWC socket
fire-and-forget. The nostr lib Relay.send is async (await wait_connected; ws.send)
but was called without await, and success was logged immediately. A send that
rejected (e.g. ws in CLOSING state during a strfry reconnect) was silently
dropped while the log claimed success, leaving NWC clients with "no info event"
until a manual republish.
Publish over a fresh short-lived socket and wait for the relay OK, retrying up to
3x with honest logging (warn on unconfirmed, err on give-up). ensureInfos
receive-based detection is unchanged. Verified on prod: deleted both info events,
restarted, both republished within seconds.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
| |
Emit a greppable EVICTED_AUTH line (username + cf-connecting-ip) when an
evicted account's JWT is used, so the real-time auto-banner can block the
source IP at the edge immediately rather than waiting for the hourly scan.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
Add an `evicted` set check in the JWT strategy: an evicted account fails
authentication on every request and every endpoint, regardless of source
IP. The existing blacklist only blocks sends, so a compromised/abusive
JWT could still authenticate and read/probe; eviction kills it outright.
Matches on uid or username so a rename can't evade it.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
BOLT11 is case-insensitive, but an uppercase-pasted invoice was stored as
the payment hash and the pending-set entry verbatim. check()'s
startsWith("ln") test is case-sensitive, so it treated such pending
entries as non-lightning and evicted them without reversing — a failed
send would never be auto-refunded.
- sendLightning: lowercase the invoice up front so the stored hash, the
pending entry, and xpay all use the same form.
- check(): lowercase before the startsWith("ln") test (defensive; also
rescues any already-stored uppercase entries).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
check() skipped pending payments younger than 10s, but xpay retries for
30s (retry_for: 30). During that window listpays can transiently show all
parts failed before the winning part lands, so check() could reverse a
payment sendLightning was still driving; the part then completed and
finalize() ran against an already-removed record.
- check(): raise the skip threshold 10s -> 60s so sendLightning finishes
finalize()/reverse() and removes the payment from `pending` before
check() ever touches it.
- finalize(): if the record is gone, raise a clear error instead of a
cryptic null-deref.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
generate() unconditionally lPush'd the invoice id to `${aid}:invoices`,
but it's also called from the update() handler (tip/webhook changes, and
repeated bolt12 offer creation) re-using the existing id — so each update
re-appended the same id and the user's :invoices accumulated duplicate
entries, rendering the same offer/invoice multiple times. Only push when
the id isn't already in the list (LPOS guard).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The /lnurl/proxy, /decode and /encode handlers fetched an arbitrary
user-supplied URL via got() and returned the body — an unauthenticated
SSRF reaching internal services. Route those fetches through a new
safeGot() that rejects hosts resolving to loopback/private/link-local/
CGNAT/metadata ranges, re-validates on each redirect hop, restricts the
scheme to http(s), and caps timeout + response size. Also validate the
NIP-57 zap request (kind 9734 + signature) before storing it, instead of
trusting the client-supplied event.
These endpoints can't be removed — they back lightning-address/LNURL
parsing and resolution in the wallet UI.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
After broadcasting an on-chain send via bitcoind, also POST the raw tx to
mempool.space and blockstream (fire-and-forget, non-blocking) so those
explorers see it immediately instead of waiting ~8s for p2p propagation.
Skipped on regtest/testnet.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
getPayment() pre-check in the lightning listener/replay isn't atomic, so
concurrent re-credits of the same settlement each ran incrBy. Claim each
lightning/bolt12 settlement with an atomic SET NX on credited:<ref>
(ref = preimage) and bail if already claimed. Also disable the /replay
route that triggered it.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
The blacklist (send freeze) only matched the current username, so a
compromised account could shake the freeze by renaming itself. Match on
either the lowercased username OR the immutable uid; add the uid to the
`blacklist` set to freeze durably.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The zombie-listener watchdog (bd1ece38) treated "waitanyinvoice blocked > 5min
while cl is alive" as a stall — but that's NORMAL during quiet periods, so it
false-fired and needlessly recycled the cl listen socket every few minutes
(~26x in 6h on 2026-06-07). Receives still worked, but it was churn + noise and
would mask a genuine zombie.
Fix: a blocked long-poll is only a zombie if cl actually has a paid invoice the
listener isn't picking up. When blocked past the threshold, probe cl on the MAIN
socket with waitanyinvoice(pay_index, 2s server-side timeout): if it returns an
invoice there's a real backlog the wedged listener missed -> recycle; if it
times out (CLN code 904 = nothing waiting) the listener is just idle and healthy
-> leave it alone. Only recycle on a confirmed backlog.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Three issues reported by an NWC client integrator (refs #80):
(a) pay_invoice sync reply: the handler polled listpays for ~20s then returned
a misleading "Payment timed out" error on slow routes even though the payment
settled — client never got the preimage. sendLightning already runs xpay to
completion and finalize() sets p.ref = preimage, so return that directly; keep
a longer listpays loop only as a fallback, and on still-in-flight return an
explicit "check with lookup_invoice" rather than implying failure.
(b) rate limiting: raised per-pubkey cap 5/min -> 60/min (normal usage — a
balance check + payment + a few lookup_invoice reconciles — trivially exceeded
5), and over-limit requests now get a NIP-47 RATE_LIMITED error reply instead
of a silent drop that looked exactly like a lost reply. Also: every request
path now always sends a 23195 reply (success, error, or handler-returned-
nothing) so a client never hangs waiting.
(c) notifications: payment_received/payment_sent are advertised in the info
event but app.notify defaulted to false unless the client sent notify:"true"
(only ~1% of apps had it on), so they never fired. Default notify ON; only
disable when explicitly "false". (Existing apps still false — backfill TBD.)
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
double-receive
completePayment logs once when an on-chain deposit is first seen (pending) and
again from /confirm when it confirms, both as bare "received" — so a single
liquid/bitcoin deposit appeared twice in the logs and looked like a double
credit. It never was (the /confirm p.confirmed guard and the seen set prevent
double-crediting); purely misleading log noise. Now logs "receiving (pending)"
on first sight and "received" on confirmation. Instant lightning receives are
confirmed=true so they still log a single "received".
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
pay_invoice derived the amount solely from the decoded invoice, so an
amountless bolt11 yielded amount = round(undefined/1000) = NaN and
sendLightning rejected it with "Invalid amount". NIP-47 allows the client
to specify the amount (msat) in the request for exactly this case (AlbyHub
supports it; Alby Go and others hit the error). Fall back to the request's
`amount` field when the invoice carries none; invoice amount still takes
precedence when present, and a still-missing amount returns a clear error
instead of NaN. sendLightning/sendInternal already handle amountless
invoices given a valid amount — the bug was only the handler dropping it.
Fixes #83
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
generate() stores `created` in milliseconds (Date.now()), but the NWC
make_invoice response returned it directly as created_at and computed
expires_at = created_at + expiry. NWC timestamps are Unix SECONDS (NIP-47),
so clients parsed a ~1.78e12 expires_at that overflowed their date types —
e.g. ptcpay's .NET DateTimeOffset, whose max is 253402300799 (issue #84:
"creating lightning invoices via NWC fails ... Valid values are between
-62135596800 and 253402300799"). Convert created to seconds before use.
The other NWC handlers were already correct (lookup_invoice reads seconds
straight from cln; list_transactions divides p.created by 1000). Only
make_invoice leaked milliseconds.
Fixes #84
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
On 2026-06-04 cl restarted (autoheal, after its 15s healthcheck timed out),
and the app's waitanyinvoice long-poll hung on the now-dead listen socket.
listenerActive stayed true forever, so ensureListenerAlive never restarted
it, and incoming lightning payments stopped crediting for ~2h until a manual
app restart. The per-call RPC timeout in ln.ts can't catch this: waitanyinvoice
is intentionally in NO_TIMEOUT (it's meant to block).
Fix: ensureListenerAlive (runs every 120s) now also detects a stalled listener
— active but blocked in waitanyinvoice for >5min. It probes cl via the main
ln socket (getinfo); only if cl is actually alive does it treat the listen
socket as a zombie, call lnListen.reset() (new: drops the proxy's socket so the
next call reconnects), and re-arm. A listenerEpoch guard makes the wedged
invocation bow out when its waitanyinvoice finally rejects, so we never run two
listeners concurrently and never double-credit. Verified with a state-machine
test covering normal receive, quiet-but-healthy (left alone), stall+cl-alive
(recycled), stall+cl-down (left alone), and the no-double-listener race.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
scripts/analytics.ts — streams live db (SCAN, never KEYS) and reports for
a trailing window (default 30d): MAU/WAU, new users, custody total + top
holders, network volume by type (lightning/liquid/bitcoin/internal/bolt12/
fund), most-active and highest-volume users, and bolt12 recipients with a
"regular" flag for the Ocean mining-reward stream.
scripts/analytics-geo.ts — top countries from the request log. Counts only
/login + /register events tied to a username (raw cf-ipcountry volume is
dominated by datacenter/nostr traffic). Surfaces a distinct-IP-per-country
column so proxy/VPN egress is visible, and filters a known credential-test
bot IP. Honest result: US #1, then GB/CA/JP/DE/ES/BR.
lib/register.ts — stamp `created` (ms epoch) and store `country` (CF/
ipregistry code) on new users, so future reports get real signup timing and
geography from the db directly instead of log archaeology. Existing users
predate this; backfill is out of scope.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
NWC clients (Alby Go, Lightning Piggies) fetch the kind 13194 info event
(capability advertisement) before they'll talk to the wallet; without it
they report "no info event" and show an unknown balance. The events went
missing after strfry restarted on 2026-05-29 and stayed gone for days,
because the app only published them once in the relay's on-open handler
and nothing re-asserted them.
13194 is a replaceable event meant to persist, so instead of blindly
republishing, query the relay on connect and publish only the server
pubkeys whose info event is actually missing (ensureInfo). Re-check
hourly so a relay restart that leaves our socket intact also self-heals.
Interval is cleared on close/reconnect alongside the heartbeat to avoid
leaks.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A unix-domain RPC socket has no TCP keepalive, so if cl goes
unresponsive while the socket stays open (zombie/half-open), an awaited
call hangs forever — the "stuck spinner past /pay/bob, no error"
symptom. Nothing threw, so the existing reconnect-on-error path never
fired and only an app restart recovered it.
Wrap short-lived calls in a 30s Promise.race timeout (matching
health.ts). A timeout rejects with ETIMEDOUT, now part of isSocketDied,
so the client is dropped and the next call reconnects — no process
restart. Long-polls (waitanyinvoice et al), in-flight payments
(pay/xpay/sendpay/keysend), and bolt12 round-trips (fetchinvoice/
sendinvoice/offer) are exempt: timing those out client-side would risk
leaked debits or break intentional blocking.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
freezeCheck only populates ${type}:limit for external asset types
(lightning/fund/ecash/bitcoin/liquid/bolt12), so internal:limit stays 0
forever. The atomic limit check added in 609fcebe applied to all types,
so amount > 0 blocked every internal coinos-to-coinos payment with
"serverLimit 0". Internal sends are pure ledger moves that never touch
an external wallet — there's no hot-wallet liquidity to reserve — so
exempt them, matching the existing soft-freeze and withdraw-lock gates.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Each successful lightning send was emitting 5 INFO lines:
"<user> sent lightning <amount>" (debit, KEEP)
"paying lightning invoice <pr> <amt> <fee>" (drop — redundant)
"payment completed <id> <preimage>" (drop — implicit from absence of warns)
"refunding fee <maxfee> <fee> <delta> <ref>" (drop — the warn for non-integer
delta already covers anomalies)
"sendLightning outcome <id> = finalized" (drop for "finalized" only —
keep for the recovery-via-fallback
variants since those are diagnostic)
Net: only the "<user> sent lightning <amount>" line fires per healthy send,
matching the other asset types' style. Anything that needed a fallback path
or threw still surfaces via the existing warn/err lines.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Each xpay with failed_parts > 0 that ultimately settled (multi-path
retry — common, not concerning) was firing 3 WARN lines:
"xpay returned (abnormal)"
"xpay returned failed_parts without throwing"
"xpay failed_parts but listpays complete — finalizing"
These are all the SAME successful payment. The presence of failed_parts
just means CLN tried multiple paths before one succeeded. Adam was
seeing noise.
Now:
- Only warn on actually-missing preimage (xpay silently failed).
- Drop the unconditional "without throwing" warn — the failed_parts
branch handles the recovery silently in the success case.
- Drop the "listpays complete — finalizing" warn — that's the success
path, not a concern.
- KEEP warns for the genuinely interesting cases: listpays says failed
(we're about to reverse), listpays says pending (debit left for
check() to reconcile), or finalize itself throws.
Net effect: a healthy multi-path payment now emits only the standard
"sendLightning outcome <id> = finalized-via-listpays-after-failed_parts"
info line instead of 3 WARN lines.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Triggered by a recurring "finalize threw despite no failed_parts: ERR
value is not an integer or out of range" warning. Root cause: when xpay's
response shape is unexpected (or invoice is zero-amount and decoded
amount_msat is undefined), the computed p.fee becomes NaN. The
subsequent db.incrBy(balance:<uid>, maxfee - NaN) throws because Redis
can't increment by NaN.
Symptom: leak guard fires "ref set in db despite throw" — payment is
actually fine (p.ref persisted before the throw) but finalize never
completes the fee refund. User loses up to maxfee sats (was already
debited maxfee, gets nothing back).
Fix: validate the computed fee is a finite integer; if not, keep the
originally-reserved maxfee (zero refund, no balance corruption) and warn
loudly. Same guard on the refund delta before db.incrBy.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Two cleanups based on a day of production data:
1. Only log the xpay response shape when something is abnormal
(failed_parts > 0 or missing preimage). The success case (200+/day) was
pure noise — answers Adam's "do we still need these xpay warnings".
2. Re-read the payment record from db before the leak guard fires. Several
"LEAKED DEBIT" lines today were finalize() throwing on a late line
(likely db.incrBy for the fee refund) AFTER db.set had already
persisted p.ref. The payment is functionally settled in those cases —
only flag as a real leak if ref is genuinely missing from the db record.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The first deployment surfaced two LEAKED DEBIT lines that were actually
clean finalizes via the new failed_parts listpays-fallback path — exactly
the path that recovers payments that would have been stuck under the old
code (joho33's case). The leak guard's `outcome !== "finalized"` check
was too narrow: any path that finalizes the payment, regardless of how
it got there, is a success.
Bucket outcomes explicitly:
- finalizedOutcomes — clean successes (finalize ran, ref set)
- safeNonFinalized — reverse ran, OR debit left in pending for check()
to reconcile (these aren't leaks)
- anything else — actually stranded (finalize-threw-*, unknown) → err
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Diagnosing the stuck "optimistic" state seen on payment 30976bc9 (joho33,
2026-05-19): confirmed=true with no preimage, not in the pending set, so
check() couldn't reconcile. Most likely path: ln.xpay returned without
throwing but with failed_parts > 0; the previous `if (!r.failed_parts)`
guard skipped finalize, no catch ran, and the debit was left stranded.
Adds:
- Log the full xpay/keysend response shape (status, failed_parts, parts,
amount_sent_msat, has_preimage) so we can see what CLN actually returned.
- Handle the failed_parts branch explicitly — verify with listpays and
finalize, reverse, or leave pending based on actual outcome.
- Track an `outcome` string through every branch and log it before
returning or throwing.
- "LEAKED DEBIT" err line at the function-end if we return without a clean
finalize — that's the exact stuck state we want zero of. grep `LEAKED
DEBIT` in app logs to catch any future occurrences immediately.
Same instrumentation pattern applied to sendKeysend.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Out-of-band emergency stop. The companion nobal.sh creates a sentinel
file on the host (bind-mounted at /locks in the container) when a
wallet balance drops below its failsafe threshold; debit() checks for
the file on every external send and refuses with "External withdrawals
temporarily disabled". /locks/ALL.locked is a global override.
Lives outside the db so a db-write vector (the May 18 /gateway pattern)
can't disable it. Internal sends are exempt — receives and user-to-user
transfers keep working with the site fully online during a lock.
For bitcoin and liquid this is a secondary defense on top of daemon-
level walletpassphrase rotation. For lightning, it's the only available
mechanism since CLN has no walletlock equivalent.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previous calc summed our_amount_msat across every channel regardless of
state, including ONCHAIN (channels that have been closed and are waiting
for the on-chain sweep). Those funds are not actually usable for
lightning sends, but were inflating lnbalance — and therefore inflating
${type}:limit — by the locked-up amount. With several closed channels
pending sweep, the effective threshold floor was tens of megasats lower
than the configured lightning:threshold suggested.
Filter to CHANNELD_NORMAL (and CHANNELD_AWAITING_SPLICE so an in-flight
splice doesn't briefly pin the limit to zero).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
When ln.listpays throws inside the post-failure verification path in
sendLightning / sendKeysend, the outer catch was swallowing it silently.
That meant a CLN outage during a failing payment was indistinguishable
from clean failure-and-reverse in the logs.
We still want to not-reverse in this case (the debit stays, pending stays
in the reconciler set, check() retries when CLN is back), but the
operator deserves to know the verification step failed. Logged via
warnThrottled so a multi-hour outage doesn't flood the log.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The catch blocks for ln.xpay and ln.keysend reversed the user's debit
whenever the RPC call threw, regardless of the actual payment state in
CLN. If the RPC errored transiently while the payment was still pending,
the user's balance got refunded — and then the lightning payment could
still complete, leaving the user with both the refund and the outgoing
sats.
Both call sites now mirror the safe check() logic:
const failed = !pays.length || pays.every((p) => p.status === "failed");
if (completed) finalize ...
else if (failed) reverse ...
// else (pending) leave the debit; check() reconciles when CLN settles.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Two related resilience fixes for when one chain backend (typically lq via
the cs tunnel) is down or unreachable:
- getAddressType: when lq.getAddressInfo throws on a connection error
(ECONNREFUSED / ETIMEDOUT / "Unable to connect"), fail with
"liquid temporarily unavailable" instead of the misleading
"unrecognized address". Users get an actionable message and we can
tell user error from infrastructure problems in the logs.
- catchUp: wrap the per-chain listTransactions call in its own try/catch
so a downed lq doesn't stop the bitcoin deposit-recovery loop (and
vice versa). Logged via warnThrottled so a multi-hour outage doesn't
spam the same error every tick.
Also drops a redundant /confirm double-call comment from catchUp — with
walletnotify reliable again, a single call is sufficient and avoids
piling on the /confirm race the new atomic lock is preventing.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
build() is called from both /payments (the actual send path) and /fee
(the fee-estimation endpoint). Locking UTXOs in build() meant that every
fee quote was reserving inputs that never got broadcast, slowly leaking
UTXO locks until a bitcoind restart cleared them — eventually starving
the wallet of selectable inputs even when balance was healthy.
The lock belongs at the point where we're actually about to broadcast,
which is sendOnchain. Released in the catch path on broadcast failure;
on success, the inputs become spent and the lock is moot.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Follows up the Apr 29 double-credit incident. Two parallel callers of
/confirm (walletnotify, catchUp, bulk sweeps) could each read
p.confirmed=false and each run the multi() that increments balance —
because main-db pointers were sometimes missing while arc still had them,
gf()'s arc fallback returned stale state. Three pieces here:
- routes/payments.ts: SET NX EX 60 lock around the /confirm credit block
keyed by `confirmlock:${txid}:${vout}`. Concurrent callers short-circuit
cleanly. The lock is released on success and on every early-return.
- lib/payments.ts credit() + routes/payments.ts /confirm: mirror the
payment/invoice records and the `payment:<txid>:<vout>` / `payment:<hash>`
pointer keys into arc as soon as they're written. gf() will then find
the prior credit via arc fallback even if the main-db copy is wiped by
a future data-loss event.
- routes/payments.ts hide `revertedDuplicate` payments from the user's
payment list — matches the same filter added to lib/nwc.ts list_transactions
so reverted records never surface in either rendering path.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Three small NWC robustness fixes that surfaced from the Apr 29
double-credit cleanup:
- fail loudly when an NWC request references a pubkey whose user record
no longer exists (migrated/deleted account); previously crashed inside
handle() on `user.id`
- swallow `pubkey not found` and `user not found` in the per-event catch
so deleted-app/migrated-user spam doesn't fill the error log — these
are client-side staleness, not server faults
- use gf() instead of g() when fetching payment records (consistent with
the rest of the codebase: arc fallback for wiped main-db pointers)
- filter `revertedDuplicate` payments out of list_transactions so
clients never see the doubled records that were later reverted
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The previous check at debit() read ${type}:limit, compared amount, and
deferred any decrement to the next freezeCheck tick (every 10s). That made
the check non-atomic: a burst of concurrent withdrawals could each pass
the same stale limit value before any reconciliation happened. Combined
with the /gateway abuse, this allowed the hot wallet to be drained.
withLimitLock() serializes the check + db.decrBy inside a per-type async
mutex, and freezeCheck also takes the lock when overwriting limits so its
10s reconciliation can't race in-flight checks. Single-process only;
distributed lock or TigerBeetle reservation needed if app ever scales out.
Trade-off: decremented limits are not refunded on payment failure — they
get reset by the next freezeCheck within 10s. Acceptable false-deny
window in exchange for correctness against bursts.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
| | |
|
| | |
|
| | |
|
| | |
|
