Deriving the next HD transparent address read the key count under a
read lock, released it, derived the address, and only then took a
write lock to append it. Between reading the position and appending,
another caller - for instance a scan deriving an address at the same
time - could read the same position, derive the same address, and
append it too, leaving a duplicate transparent key in the wallet.
- The whole read-modify-write now runs under a single write guard, so
the position a caller derives from reflects every key already
added. The slow seed stretch is computed once before the guard is
taken, so holding the lock does not block on it.
- As a further guard, derivation skips a position whose address is
already present and advances to the next, so no path can produce a
duplicate even if positions were ever to overlap.
- The starting index is computed exactly as before (it still counts
imported keys too), so a wallet with no contention derives the same
addresses in the same order.
Uncontended derivation produces the same addresses in the same order,
verified against the differential harness. The wallet file format is
unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Saving the wallet serializes it to a temporary file under the sync
lock, but it took a write guard on the wallet to do so. Serialization
only reads wallet state, so the write guard needlessly excluded other
readers for the whole duration of the write, buffering, and flush.
- The save path now takes a read guard to serialize. A read guard
still excludes writers, so nothing can mutate the wallet mid-write;
it only lets other readers proceed in parallel. The surrounding
sync lock still serializes saves against syncs, and the atomic
temp-file-then-rename sequence is unchanged.
The serialized bytes are identical, verified against the differential
harness. The wallet file format is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Opening a wallet reads each shielded note from disk, and every note
record embeds a 169-byte extended full viewing key that was decoded
from scratch for every note. That decode is an elliptic-curve parse;
a wallet with many notes sharing a handful of keys paid for it once
per note, which dominated open time.
- Note reads now go through a small content-addressed cache keyed on
the exact 169 raw key bytes. The first note with a given key
decodes it; the rest reuse the result. Because the cache is keyed
on the raw bytes and the decoder is deterministic, the cached key
is identical to a fresh parse.
- The cache is thread-local and naturally bounded by the few distinct
keys a wallet holds, so it adds no cross-thread state and cannot
grow without bound.
The bytes on disk are read exactly as before - this only avoids
re-parsing identical bytes - so the loaded wallet is byte-identical,
verified against the differential harness. The wallet file format is
unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Each block range is streamed over one gRPC connection into a channel
that a worker thread drains, encodes, and scans. That channel was
unbounded, so when the server delivered blocks faster than the
scanner consumed them the backlog piled up in memory with no ceiling
- on a fast link against a large range it could grow without limit.
- The channel is now a bounded synchronous channel holding at most
2048 blocks. When it fills, the fetch side waits, applying
backpressure up the stream instead of buffering endlessly, so peak
memory is capped no matter how far ahead the server runs.
- The batch size is untouched: it still streams a whole range over a
single connection, which is what keeps the connection count low
enough to avoid socket-buffer exhaustion on Windows. Only the
in-flight depth between fetch and scan is now bounded.
The same blocks are fetched and scanned in the same order, so the
resulting wallet is byte-identical, verified against the differential
harness. The wallet file format is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When sync hit a reorg it stepped the scan back to just before the
invalidated block, then set the top of the re-scan range to that
point plus the full forward batch size. That window is far larger
than any reorg can be: the batch size is tuned for forward streaming
(tens of thousands of blocks), while a reorg is capped at the maximum
reorg depth and the pass aborts once it is exceeded. So the recovery
pass re-fetched and re-scanned a large span of already-good blocks
before it could resume moving forward.
- The recovery range now reaches only the maximum reorg depth (plus
one) above the rewind point, which is the most a reorg can ever
disturb. Forward syncing still uses the full batch size; only the
pass immediately after a reorg is bounded this way.
- The set of blocks that ultimately get scanned is unchanged - the
next pass picks up from where recovery stops - so the resulting
wallet is identical, just reached with far less redundant fetching
and scanning.
The wallet file format is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On quit the lite controller's destructor never runs, because main()
ends the process with _Exit(). Any state the background refresh had
scanned since the last save was lost: the destructor is where that
final save would normally happen, and it was being skipped on the
one path that always runs.
- shutdown() now calls a new flushOnExit() that persists once on the
way out, bounded to 4 s so the existing 15 s shutdown watchdog
budget still holds. On timeout it abandons the save thread and lets
_Exit follow.
- It is a plain save through the same backend path, which writes a
temp file and renames it into place. That cannot corrupt the wallet
if it is cut off or abandoned, and it does not change the wallet
file format - it only persists sooner on the quit path.
- It reuses the destructor's guards and also skips when a wallet op
is already in progress: shutdown() has already waited for ops to go
idle, and a still-pending op holds the sync lock and issues its own
save, so flushOnExit would only block the budget on a save that
cannot take the lock.
Tests cover a save on the open, synced, idle path (exactly once), a
blocked save returning within the budget, and the retired and
never-opened guards skipping the save.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The sidebar unconfirmed-tx badge was recomputed only when the tx list
changed, keyed on last_tx_update plus the list size. That missed the
one case that matters most: a transaction confirming without the list
count changing. last_tx_update did not move for a same-second content
change, and lite never stamped it at all, so the badge stuck on a
stale 0-conf count until some unrelated refresh happened to shift the
list size.
- The badge now keys on tx_revision, which the state bumps on every
content change - including a confirmation flip at an unchanged tx
count - and which lite stamps too, so the count tracks confirmations
on both the full node and the lite wallet.
- The sentinel starts at 0, which tx_revision never takes (it starts
at 1), so the first frame always computes once.
- The recompute still runs only when the revision actually moved, so
the per-frame work the previous guard avoided is still avoided.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Transparent history is gathered by skip-paging listtransactions. The
underlying list can shift between the per-page RPCs - a new
transaction arriving pushes one page's tail into the next page's head
- so a row could be collected twice and show up as a duplicate in
History.
- Pages are now de-duplicated on the daemon's stable per-output
identity: txid, send/non-send class, and vout. Keying on the class
rather than the exact category means a coinbase that matures
(immature to generate) between two fetches is still recognised as
the same row; keeping vout means two equal-value outputs of one
transaction stay distinct. No real transaction row is dropped or
duplicated.
- Rows with no txid or vout ("move") are never keyed and pass through
(they are dropped downstream anyway), so the de-dup fails open.
- The raw page size still drives the stop condition and the next-run
row hint, independent of how many rows were unique.
Tests cover a shifted page whose head repeats the prior page's last
row (dropped), a sibling output of the same tx (kept), and an
un-keyable row (kept).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
listtransactions was called with the default "" account, which
returns only rows in the unnamed account. Transactions tied to a
labelled address were left out of the History tab entirely.
- All three listtransactions call sites now pass "*", which spans
every account, so labelled-address rows appear alongside the rest.
- With "*" the daemon also emits "move" accounting rows that have no
txid, address or vout. Those are not wallet transactions, so they
are skipped and never become blank History rows. No real
transaction row is dropped by this filter.
Tests cover that a "move" row is skipped while a labelled receive row
is kept, and that the transparent collectors now page with "*".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
parseRpcResult copied the "result" node out of the parsed reply,
because indexing the JSON object yields an lvalue reference. On large
replies - full transaction lists, unspent-note sets - that duplicated
the entire subtree for no reason, and it also meant a secret-bearing
result then existed in two trees at once.
- The result node is now moved out of the parse tree instead of
copied, so the secret lives only in the returned value that the
calling code already wipes. The leftover result slot is moved-from
(null), so the existing secret-scrub of the discarded tree still
runs harmlessly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Loopback HTTP was still subject to the process environment's proxy
settings (http_proxy / all_proxy). That routed the local daemon's
basic-auth credentials and wallet data - and the miner's 127.0.0.1
stats API - through whatever proxy the user's shell happened to
export, and could break the connection outright.
- Local daemon RPC now sets curl's no-proxy to "*" when the host is
loopback, so it always connects directly. This also overrides a
no_proxy list. Remote daemons keep full proxy support unchanged.
- The xmrig stats fetch against the local miner API does the same.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The RPC worker loop only caught std::exception. A task that threw
anything else - a non-std exception propagating out of a posted job -
escaped the loop and tore down the worker thread, so every later RPC
silently went unanswered.
- The worker loop now also has a catch-all that swallows non-std
throws and keeps the thread running. Such a job still loses its
main-thread callback, exactly as the std::exception path already
did; the only change is that the worker survives.
- RPCClient's connected_ and warming_up_ flags are now atomic. They
are read from the UI thread while the worker updates them, so the
plain bools were a data race.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Starting the daemon polls for an already-running node holding the
datadir, sleeping up to three 100 ms intervals to let a shutting-down
node clear out. When a start runs on the UI thread, that is up to 300 ms
of blocking sleep on the render thread.
- The lock-wait poll count is now a parameter. A start invoked on the UI
thread passes 0, which does a single check and never sleeps; every
other caller keeps the three-poll wait. The default is unchanged, so
background starts behave exactly as before.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
At quit the UI thread drained and joined the mining-control thread and
then called the miner's blocking stop with a 3 s budget. Both ran on the
UI thread, so the window could stall for up to that long before the
shutdown screen rendered.
- Quit now only signals the mining-control thread (no join) on the UI
thread. The blocking teardown — join the control thread, then stop the
miner — moves to a background thread, so xmrig still has a single
driver but the UI thread never blocks on it.
- On the daemon-stop path that teardown runs first on the shutdown
thread, so the miner is fully stopped before the daemon stop begins.
On the disconnect-only quit paths it runs on its own thread, which
the final shutdown joins before tearing xmrig down, so there is never
a second concurrent join of the control thread.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The daemon monitor thread drained output and then slept a fixed short
interval on every tick, waking several times a second for the life of
the node even long after startup, when the node emits almost nothing.
Stopping it also meant waiting out the current sleep before it saw the
stop flag.
- The monitor now waits on a condition variable instead of a plain
sleep, so a stop request wakes it immediately rather than after the
interval. The stop flag is set under the monitor's mutex so the wake
cannot be lost.
- The poll interval stays fast for the first 30 s after the thread
starts and for 3 s after any output, and otherwise doubles per idle
tick up to a 500 ms cap. Any new output drops it straight back to
fast, so responsiveness during startup, rescan and shutdown is
unchanged while a quiet node is polled far less often. The interval
choice is a pure function.
Tests cover the interval function across the warmup window, the
recent-data window, the reset on new output, the doubling, and the cap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The miner's stats and pool-API fetches are blocking HTTP calls on the
monitor thread with multi-second timeouts. When the miner was stopped,
stop() still had to wait out an in-flight request before the monitor
thread could join.
- Both fetches now install a curl progress callback that aborts the
transfer as soon as a stop is requested, so a request in flight ends
promptly instead of running down its timeout.
- An abort triggered by stop is treated as a clean cancellation, not a
fetch failure, so it is not logged or counted against the failure
warning.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Stopping the miner closed the stdout pipe before joining the monitor
thread, and the UI thread's stats poll also drained stdout. Either path
could read or close a descriptor the other side was still using, so a
read could land on a closed or already-reused fd.
- stop() now joins the monitor thread first and only then closes the
process handle and the stdout pipe, so nothing is draining the pipe
when it is closed.
- The stdout drain is serialized under its own mutex, and the UI
thread's stats poll no longer drains stdout at all: the monitor thread
is the sole reader. The process handle, pid and fds are atomic so they
can be observed safely from the memory probe.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The wallet pins a -dbcache value on the daemon command line. A
command-line flag overrides the same key in DRAGONX.conf, so a user who
set their own dbcache in the conf never got it: the wallet's pin always
won.
- Before spawning, a normal start now reads DRAGONX.conf once and, if it
sets dbcache, drops the wallet's own -dbcache pin from the arguments
so the user's value takes effect. An isolated start keeps its own
datadir and conf and is left alone.
- The conf is now read a single time for both this check and the
existing rpcthreads check instead of twice.
Tests cover dropping the pin when the conf sets dbcache (in various
forms), and leaving the default arguments untouched when it is
commented out, empty, or sets something else.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When the daemon was restarted, the output-parse offset was reset to 0.
The daemon's output buffer is retained across runs, so the next poll
re-read the whole retained tail of the previous run and fed it to the
busy tracker as if it were fresh. A witness rebuild or rescan line left
over from the old run could then be mistaken for activity in the new
one.
- On restart the offset is now reset to the absolute end of the output
at that moment, captured on the lifecycle thread after the stop and
before the start, so the UI resumes parsing only the new run's output
rather than replaying the old tail.
- The capture is handed to the UI thread through a pending flag and
applied there before the next read, both on the normal poll and when
skipping the backlog, so the offset is only ever moved on the thread
that owns it.
Tests cover reading from the captured watermark (only the new run is
seen) versus the old reset-to-0 behaviour (the previous run replays).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The embedded daemon and the xmrig manager each kept their own copy of
the console output ring: append new bytes, trim the front to a cap at a
newline boundary, and hand a reader everything since its offset. The two
copies had drifted. The daemon tracked an absolute byte total so a
reader's offset stayed valid across a front-trim, while the miner keyed
its reader off the live buffer size, so a trim could silently rewind the
console feed.
- The append/trim and the since-offset logic now live in one small
header, outputWindowSince and appendBounded, included by both. The
trim keeps the newline-boundary cut and the guard that stops a lone
trailing newline from erasing the whole buffer.
- The xmrig manager now carries the same absolute byte total, so its
reader resolves against the window the same way and survives a trim
(and a process restart that clears the buffer but keeps the total).
Tests cover feeding the ring well past the trim point while a reader
polls each chunk (the feed is continuous and never dies at the trim),
the lone-trailing-newline guard, and a cleared buffer with a retained
total.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ui.toml and the default banlist were embedded by a CMake function that
read each file and emitted a C array header one byte at a time in a
configure-time loop. That is slow to configure, and because the result
is a header, editing either resource recompiled every translation unit
that included it.
- Both files are now embedded with INCBIN from a small generated
translation unit, the same pattern already used for the fonts. There
is no generated header, so editing a resource re-assembles one tiny
object file instead of triggering a wide recompile, and configure does
no per-byte work.
- Consumers include a hand-written header declaring the symbols and
always use the embedded data (the __has_include fallback is gone). The
data is not NUL-terminated, so every consumer passes an explicit size.
- The old EmbedResources.cmake script is removed, and the generated unit
is added to the app and test builds.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ingesting a scan's worth of chat messages persisted each new message
with its own append() call, so a batch of fresh messages opened and
committed one SQLite transaction per row.
- A new appendBatch persists many messages in a single transaction,
using the same INSERT OR IGNORE dedup key as append() and preserving
input order so load()'s rowid ordering is unchanged.
- Every message is serialized and encrypted before the transaction
opens, so a message with a non-UTF-8 body is skipped on its own
without leaving a transaction open; the valid siblings still persist.
- Ingest collects the newly stored messages and writes them once after
the loop instead of per message. It is inert (returns 0) while locked
or on empty input.
Tests cover a batch insert, dedup on re-insert, empty input, bad-UTF-8
isolation with a valid sibling and a following append(), insert order,
and the locked no-op.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Opening the transaction-history cache always derived its encryption key
with Argon2id, even for an unencrypted wallet. For those wallets the
passphrase is just a value derivable from the database itself, so the
~50-100 ms Argon2id stall bought no additional secrecy and ran on the
UI thread at open.
- Unencrypted wallets now derive the cache key with a cheap,
domain-separated BLAKE2b KDF over the wallet identity, keyed by the
per-wallet salt. The secret is derivable from the DB, so this is
security-equivalent for the unencrypted case and avoids the stall.
- Passphrase-encrypted wallets are unchanged: they keep the Argon2id
key derivation. The two paths use separate domain labels, so the
fast-path key and the Argon2id key never collide for the same DB and
salt; a cache written under one cannot be read under the other.
Tests cover the fast-path round-trip and reopen, and assert that an
Argon2id unlock of the same DB produces a different key and fails to
read the fast-path snapshot.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Drawing the tiled background noise built a viewport-sized pre-tiled
texture and re-created it whenever the viewport changed, copying the
256x256 noise tile across the whole screen in memory.
- On GL the draw now samples the resident 256x256 tile directly with
GL_REPEAT, computing UVs from the screen rect over the tile size with
the phase anchored at the viewport origin. The result is pixel-
identical to the old pre-tiled texture (same generator, same origin)
with no per-viewport copy.
- DX11 keeps the pre-tiled viewport-sized texture, since its sampler is
CLAMP and cannot repeat.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Theme images were decoded straight to RGBA at their full source size
and uploaded to a GPU texture as-is. A large or over-sized image wasted
VRAM and could exceed the GL maximum texture size, and each code path
also hand-managed the decoded pixel buffer with its own free calls.
- Loading from file and from memory now goes through the shared
size-aware decoder with a 2048-pixel cap on the long side. An image
larger than that is downscaled with its aspect ratio kept; an image
whose area is absurd is rejected; a small image is left untouched.
- The decoder owns the pixel buffer for the life of the upload, so the
manual free calls are gone.
Tests cover downscaling a wide image to a 2048 long side with the aspect
kept, rejecting an over-area image, and leaving an under-cap image as-is,
plus the fit-dims math at the theme-art cap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adding a lite server from the network tab restarted the background
server probe immediately. Starting a probe while one is still running
joins the previous worker thread, and that join happened on the UI
thread, so the frame could stall until the outstanding probe finished.
- Adding a server now sets a dirty flag instead of restarting the probe
inline; the next frame starts a fresh probe once the current pass has
finished, so the UI thread never blocks on a join.
- The probe's libcurl transfers honor the cancel flag through a progress
callback and abort cleanly, so a probe in flight stops promptly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Before verifying a downloaded archive, both updaters slurped it into a
string with an istreambuf_iterator, which grows the buffer repeatedly
and reallocates as it reads a multi-megabyte file. It also placed no
upper bound on the size it would read into memory.
- A shared readFileExact helper sizes the file first and reads it in a
single allocation, failing closed if the file is missing, empty,
larger than a caller-supplied cap, or comes up short. A truncated read
can never reach the SHA-256 / signature verifier.
- Both updaters use it with the existing max-archive cap.
Tests cover a good read, a read exactly at the cap, over the cap, an
empty file, and a missing file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The daemon and miner updaters fetch release metadata with a blocking
HTTP GET that ignored the updater's cancel flag. Closing the dialog set
the flag, but an in-flight metadata request kept running to completion
on the worker thread, so the dialog could not tear down promptly on a
slow or stalled connection.
- The metadata GET now passes the updater's cancel flag through to the
HTTP helper, so closing the dialog aborts the request in progress.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Several settings were read with json's value()/get<T>() helpers that
assume the stored type. A hand-edited settings file, or one re-saved
after an inf/NaN float was written back as JSON null, could therefore
hold a value of the wrong type and throw during load. The whole load is
wrapped so a throw quarantines the entire settings file to a
.corrupt- copy, meaning one bad field reset every unrelated setting.
- ui_opacity is accepted only when it is actually a number, and it is
clamped to the same 0.3..1.0 range the setter enforces; the legacy
reduced_transparency migration likewise runs only when that key is a
boolean.
- A non-string name in a blocked chat conversation, and a non-string
scope in a portfolio entry, now fall back to an empty string instead
of throwing.
The file loads with the offending field defaulted, and the rest of the
settings survive.
Tests cover ui_opacity as a string/null/int/float/bool, a non-string
blocked-conversation name and portfolio scope, and the legacy
reduced_transparency migration, asserting no quarantine file is written.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The single-instance lock lived in /tmp, which is world-writable. Creating
it with mode 0666 there gets masked to 0644, so a second local user
cannot open it O_RDWR and is wrongly blocked from running the wallet at
all. A world-writable directory also lets another user pre-create or
symlink the lock path.
- The lock now lives in a per-user directory: $XDG_RUNTIME_DIR when it is
an absolute directory the user owns, otherwise the per-user, per-variant
config directory that startup has already created. Never /tmp.
- It is created mode 0600 with O_NOFOLLOW, and after opening we fstat it
and refuse anything that is not a regular file we own, rejecting a
planted symlink or a foreign-owned file.
- The lock file is no longer unlinked on release; flock drops on close,
and leaving the file in place avoids an unlink/relock race (a stale
file is harmless).
Tests cover acquiring and conflicting on the lock, the 0600 mode, that
release leaves the file in place, and that a symlinked lock path is
refused.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Both the HEAD size probe and the archive download followed redirects
with no restriction on where they could go. A redirect could therefore
downgrade the transfer to plain http or point at a file:// URL, defeating
the TLS verification the initial https request was relying on.
- Redirects on both requests are now limited to https, so a redirect can
no longer leave TLS or reach a local path.
- The limit is set with the string-based curl option where available and
the protocol-mask option on older libcurl.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Resuming a partial download leaned entirely on the remote size from a
HEAD request. When that size could not be determined the code played it
safe by deleting the partial file and starting over, so a flaky HEAD (or
an offline mirror) threw away gigabytes that were fine to resume. It also
could not tell a real size from an error page, and if the server answered
a ranged request with the whole file (200, not 206) libcurl could append
it to the partial and silently corrupt the archive.
- The HEAD now returns transport status, HTTP code, and size together,
and a size counts as authoritative only on a 2xx reply with a positive
Content-Length.
- A pure helper decides what to do from the existing bytes and that
authoritative flag: start fresh, truncate a stale oversized file, reuse
a complete one, or resume with a ranged GET. When the size is not
authoritative the partial is kept and a ranged GET is attempted rather
than discarded.
- A ranged request that comes back as 200 instead of 206 now restarts
cleanly from scratch instead of appending to the partial.
- The free-space pre-check runs only when the size is authoritative, and
a failed download reports how many bytes were kept for the retry.
Tests cover the resume-decision helper across the fresh, stale, complete,
resuming, and non-authoritative cases.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Any verification failure deleted the downloaded archive and forced a
full multi-GB re-download on the next attempt. But most failures are not
a corrupt archive: a missing or unreadable checksum sidecar, or a fetch
blip while pulling the checksum, would throw away a perfectly good file.
On top of that the checksum parser took the first token of whatever came
back, so an HTTP error page could be mistaken for a digest.
- The archive is now removed only on a genuine digest mismatch. A
missing sidecar, an unusable one, or a transient fetch error keeps the
file so the next attempt can reuse or resume it; the no-checksum case
still fails closed and keeps the archive.
- The checksum parser now accepts a token only if it is exactly the
expected length and all hex, lowercasing it; anything else (an error
page, a truncated or wrong-width digest) is rejected.
- The small-file fetch moved onto the shared capped HTTP helper, and MD5
is consulted only as a fallback when no SHA-256 is published.
Tests cover the checksum parser across good, wrong-length, non-hex, and
error-page inputs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The checksum and MD5 passes measured the file for their progress bar
with fseek/ftell. On Windows a long is 32-bit, so ftell overflows on
any archive larger than 2 GiB and hands back a wrong (often negative)
size, which makes the verify progress percentage meaningless on exactly
the multi-gigabyte downloads this code exists for.
- Both passes now take the size from std::filesystem::file_size, which
is 64-bit on every platform, through a small helper that returns 0 if
the size can't be read.
- No seek is performed just to size the file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
While the seed migration sits on the Confirming gate it polled the
sweep's confirmations and a full wallet balance together on every pass.
That balance call is the expensive one, and it only matters once the
sweep is mined and we do not yet have a post-mine figure, yet it ran
every poll even mid-sync when it cannot move the gate.
- The confirmation count is still read on every poll.
- The balance scan now runs only when it can actually advance the adopt
gate: the sweep is mined, no post-mine balance has been recorded yet,
and no sync is in progress. The manual Refresh button forces it
regardless.
- The decision lives in a small pure helper, and the poller snapshots
the main-thread state it needs rather than reading members off the
worker thread.
- The Confirming screen now claims a leftover remainder (and offers the
re-sweep button) only once the balance has actually been measured; while
it is still unknown it keeps waiting instead of prompting on a sentinel.
Tests cover the gate helper across the forced, syncing, and mined cases.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Several HTTP fetches read a remote body into memory without a tight
bound. The price and exchange-ticker calls shared one large 16 MiB
metadata cap, and the bootstrap checksum download had no size limit at
all, so a hostile or MITM'd server streaming an unbounded (and
Content-Length-free) response could grow memory far past anything these
endpoints legitimately return.
- The HTTP GET helper now takes a per-call byte cap and redirect limit.
The write callback refuses bytes past the cap, which aborts the
transfer, and the same value is set as curl's max file size.
- The price call is capped at 256 KiB and the ticker list at 2 MiB,
each sized to its real payload rather than the shared default.
- The bootstrap checksum fetch caps its body at 64 KiB in both the
write callback and curl's max file size.
Tests cover the sink accept/abort boundary.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Loading chat history did more work than it needed to. Seeding the unread
watermark walked every conversation and copied each one's full message
list out of the store just to read the last timestamp, and the database
reload copied every message into the store one at a time with no
reservation.
- The watermark now comes from one pass over the stored messages that
returns the latest timestamp per conversation, with no per-
conversation copy or sort.
- The store gained a move-append and a reserve, so the reload pre-sizes
its vector and dedup set once and moves each row in instead of copying
it. The dedup key is computed before the move.
Behaviour is unchanged: the watermark values and the blocked-conversation
filter on reload are the same as before.
Tests cover the per-conversation latest-timestamp pass and move-append.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two secure-delete paths overwrote a sensitive file and then unlinked it,
but neither reliably forced the overwrite to stable storage. The
plaintext key export used a buffered stream with only a flush, and the
PIN vault removal opened a second descriptor to fsync only on non-
Windows. On a crash, or on Windows, the zeroing write could stay in the
OS cache and the original bytes could survive on disk.
Both now go through one shared best-effort secure-delete helper:
- It overwrites the file in place in chunks, never truncating first, so
the original blocks are the ones rewritten.
- It flushes the data to the device before unlinking (full fsync on
macOS, write-through on Windows), then removes the file.
- It is safe on a missing file and reports whether the file is gone.
This is still best-effort: SSD wear-levelling and copy-on-write or
journaling filesystems may retain copies.
Tests exercise the overwrite-and-remove helper.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A failed wallet unlock was always treated as a wrong guess: the lockout
counter was bumped and the back-off curve armed no matter why the RPC
failed. A busy or warming-up node, a transport timeout, or a wallet that
is not actually encrypted would all eat an attempt and could push a user
into a lockout without a single mistyped passphrase.
The failed call is now classified before it is acted on:
- Only the daemon's "passphrase incorrect" error counts as a wrong
guess and advances the lockout.
- A warming-up or unreachable node, a reply timeout, or a full queue is
reported as a transient busy-node condition with no attempt charged.
- A "wallet is not encrypted" reply refreshes the local encryption state
instead of penalising the user.
- Any unknown error surfaces its message without touching the counter.
The PIN path got the same treatment: when the vault decrypts correctly
but the unlock RPC then fails for a non-credential reason, the attempt
is no longer counted.
Tests cover the classification of each failure kind.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Alt + mouse wheel adjusts the font scale, but it was only handled in the
poll loop that runs after the app wakes. When the app was idle it blocks
in a single timed wait, and that first wheel event went straight to
ImGui, scrolling the window instead of changing the scale until a later
event arrived.
- The Alt+wheel handling is pulled into a shared helper used by both the
idle wait path and the poll loop, so the first event while idle is
consumed and triggers a redraw.
- Behaviour is otherwise unchanged: the visual scale updates smoothly and
the font atlas is rebuilt once scrolling stops.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The window always asked DWM for a full-window blur region and, on
Windows 11, an acrylic backdrop. When the user runs with an opaque
window (low-spec mode or full opacity) that desktop blur is never seen
but DWM still composites it every frame.
- An early read of the saved settings decides whether the window starts
opaque before the backdrop is configured.
- When it does, the full-window blur region is left as SDL's zero-area
default and the backdrop type is set to none instead of acrylic.
Windows-only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The custom minimize / maximize / restore buttons are drawn from inside
app.render(), while an ImGui frame is open. Calling SDL to change the
window state there pushes resize events synchronously, which re-enters
the resize watcher and can nest a new frame inside the open one.
- A click now records the requested op instead of acting on it, and the
window state change is applied once the frame has closed, after the
pacing block so the follow-up redraw is not dropped.
Windows-only; it rides the borderless caption controls.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The live-resize event watcher redrew for every resize event, and each
resize tore down and recreated the DX11 render target even when the
pixel size had not actually changed. A single drag fires both a RESIZED
and a PIXEL_SIZE_CHANGED event reporting the same final size, so the
work ran twice.
- The watcher keys off the real pixel size and returns early when it has
not moved since the last pass, collapsing the duplicate event pair into
one redraw. It also bails out when called from inside an open ImGui
frame, which the caption buttons can trigger.
- The DX11 resize now checks the swap chain's current dimensions up
front and does nothing when the size already matches and the target is
live, skipping the destroy/recreate churn.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On Windows stdout and stderr were reopened onto the same log path with
two separate freopen() calls. Each stream then had its own file handle
and its own write offset into the one file, so the two writers
overwrote each other and interleaved into a garbled log.
- stdout is reopened (truncating the file per run), stderr is pointed at
the same underlying handle via _dup2 so both share one offset and
append in order.
- stderr is first given a valid descriptor, since a GUI-subsystem
process may start with no fd 2, and is left unbuffered so crash
diagnostics are not lost.
Also corrects a stale comment that named the old log file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Startup initialised SDL's gamepad subsystem and turned on ImGui gamepad
navigation, but the wallet is never driven by a gamepad. Initialising
the subsystem enumerates attached controllers and loads the mapping
database for nothing.
- SDL now starts with the video subsystem only.
- ImGui keeps keyboard navigation; the gamepad navigation flag is
removed from both the DX11 and the OpenGL setup paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On Windows SDL3 raises the global timer resolution with
timeBeginPeriod(1) and holds it for the whole process lifetime, which
keeps the system on a 1 ms tick and costs power for no benefit here.
- A startup hint now asks SDL to leave the system timer alone. Nothing
in the app needs sub-15 ms waits: vsync paces drawing, the idle wait
is 200 ms, and background sleeps are 50 ms or longer.
The hint is Windows-only in effect and is ignored elsewhere; an
environment override that turns it off is harmless.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A secret copied to the clipboard auto-clears after a timeout, but the
deadline was stored as an ImGui::GetTime() value. That clock only
advances while frames are being drawn, so while the app idled (no
redraws) it effectively stood still and the pending clear could be put
off well past the intended window.
- The clear deadline is now a steady_clock time_point, which keeps
counting whether or not the UI is rendering.
- The pump uses the armed hash as its "something is pending" flag rather
than a sentinel deadline value, and compares the real wall time.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The hashrate sparkline guessed the age of its oldest sample from the
sample count, labelling the left edge "5m ago", "1m ago" or "start"
regardless of how fast samples actually arrived. Solo and pool samples
come in at different intervals, and a slow or stalled poll made those
labels wrong.
- Each hashrate sample now carries a wall-clock timestamp, kept in
lockstep with the value as the history fills and rolls over, for both
the solo and pool histories.
- The chart's left-edge label is derived from the oldest sample's actual
timestamp, falling back to "start" only when no stamp is available.
Tests: a mining refresh records a timestamp per sample, and the
timestamp history stays the same length as the value history once the
history is capped.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When following the debug.log of a node this session did not launch, every
poll re-read the last 2 MB of the file and re-scanned it for the process
banner. If that node logs to its console instead, the banner is never
there, so the full re-scan repeated on every poll for the life of the
process.
- A miss is now remembered by pid and file size: if nothing has changed
since the last miss, or the file is growing but the backoff has not
elapsed, the scan is skipped.
- The retry schedule re-scans quickly during a fresh-node grace window
(the banner flushes a beat after the pid appears) and then backs off
from 4 s, doubling to a 60 s cap. The memo is cleared on a successful
tie and when the followed process goes away.
Tests: the retry delay is fast within the grace window and then steps
4, 8, 16 s up to the 60 s cap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hovering the Paste button in the import-key dialog read the OS clipboard
every frame to draw a ghost preview of the key, and the trimmed
plaintext was rebuilt each frame and never wiped. SDL's clipboard read
is a synchronous round-trip, and the pasted value is a secret private or
viewing key.
- The clipboard is now read only on hover entry and then at most twice a
second, and the classified preview is cached rather than recomputed
every frame.
- The cached preview is a dedicated secret buffer: it is wiped the moment
it stops being shown, wiped on secret wipe, and the transient trimmed
copy is zeroed right after it is classified.
- Trimming moves to a shared clipboard helper shared with the paste path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DanSandClaude Opus 4.8