With proof of work skipped, chaining headers from a checkpoint up to a
transaction's block proves nothing: a server can fabricate a header and
fake headers after a real checkpoint. The header must now hash-chain up to
the nearest checkpoint at or above its height, whose hash is hardcoded.
This also anchors heights below the first checkpoint. The cached anchor
state is renamed AnchoredDownTo and caches without it are re-verified
once. Blocks above the newest checkpoint remain a documented limit.
See ADR 0040.
After connecting, the Send tab offers Fast/Normal/Economy rates from the
server's fee estimates (never below the relay fee, implausible values
ignored). A manual rate below the relay fee is refused before building and
a fee above 1% of the amount is flagged in the preview. See ADR 0039.
The password no longer has to be a command-line argument: it can come from
--password-env, --password-stdin or a no-echo prompt (--password still works
but warns). Plaintext storage is now an explicit opt-in (--no-encryption)
instead of the result of a forgotten flag. See ADR 0038.
Adding a contact now checks the address against the wallet's network and
rejects duplicates; whitespace in a pasted address is removed. Before, any
string was stored and the error only surfaced at send time.
Address rows and the address detail sheet no longer carry the WIF of
every listed address for the whole session. The key is derived after the
password gate and dropped when hidden, when the sheet closes or when the
wallet closes. See ADR 0037.
After the handshake the wallet checks server.features genesis_hash and,
when the profile has checkpoints, the header at the highest checkpoint
not above the server's tip (mainnet shares Bitcoin's genesis, so the
checkpoint is what tells the chains apart). A server on another chain is
refused with WrongChainException and the next candidate is tried. Used by
the GUI, peer discovery, the CLI and the server probe. See ADR 0036.
Property tests saturate the pool while network tests wait on continuations,
so a loopback handshake could exceed its timeout and fail the run
intermittently. A higher floor removes the starvation.
A hostile server could stream an endless line with no newline and exhaust
the client's memory. A single reply is now capped at 16 MiB (configurable
via maxLineBytes in ConnectAsync); on excess the connection fails with
InvalidDataException. See ADR 0035.
A server that stops answering without closing the socket left a running
sync waiting forever. Requests now fail with TimeoutException after 30 s
(10 s for the handshake, both configurable in ConnectAsync) and the
connection is marked dead so the app reconnects. See ADR 0034.
On Italian/German systems "1.5" was read as 15 sat/vB. The Send form,
the Donate form and the CLI now share CoinAmount.TryParseFeeRate
(invariant culture, '.' or ',' as decimal separator, range (0, 1000]).
Donate and the CLI report an invalid rate instead of silently using 1.
Coverage settings exclude the fuzz harness, server probe and test assembly;
CI fails if Core drops below 95% lines or 88% branches. The gitignore no
longer swallows tests/coverage.runsettings.
ViewModels run on Avalonia.Headless: wizard flows, password and file-lock
handling, seed and private-key reveal gating, contacts, settings, sync and
send against the fake server. The CLI entry point is invoked in-process
end to end. Localization checks cover the 7 languages, placeholders and
keys used in sources. Internal seams (UpdateCheck, ProfileOverride) keep
the tests offline.
Empty header range during checkpoint anchoring, send-all limits, SyncCache
roundtrip, corrupted config, derivation edge cases, spendability
boundaries, unit formatting, loader errors, in-flight cap, disposal under
load and wallet-lock contention. Also raises the synchronizer partial-result
timeouts so they hold under load.
After a protocol error (e.g. a malformed line) the read loop ended but
the client still reported IsConnected and later requests waited forever
for a response that could never arrive. The read loop now marks the
connection closed before failing pending requests, and RequestAsync
throws IOException on a closed connection. Adds regression tests and a
raw-line helper to the fake server.
docs/docs-as-code.md defines the workflow; docs/adr holds 31 ADRs
(retroactive, reconstructed from code and changelog) with a template and
index. docs/check-docs.sh validates ADR structure, index, cited code
paths, relative links and CLAUDE.md/AGENTS.md sync.
ServerProbe is a read-only conformance/health checker for Electrum
protocol servers (Fulcrum, ElectrumX): handshake/TLS, features and
genesis, header linkage, Merkle proof vs header root, fees, latency.
LiveServerTests runs it from dotnet test when PALLADIUM_TEST_SERVERS is
set. Also ignore .claude/, apk/aab and local.properties.
Fills in the CHANGELOG.md entry for the watch-only address-only import
mode, progressive Merkle verification, transaction size-limit fix, and
Android sync-reconnect fixes since 1.0.0, and bumps <Version>/versionCode
across the App and Android head csproj files ahead of the tag.
The tab requires an open wallet to send from, so showing it before
one is loaded led to a dead end. Gate it behind IsWalletOpen, same as
the other wallet-only UI.
Width-capped containers (wizard steps, private-key prompt, wallet info
overlay) used HorizontalAlignment="Center" with MaxWidth, which sizes
the panel from its content's DesiredSize. Since PlaceholderText only
contributes to that measurement while a TextBox is empty and
unfocused, clicking into an empty field shrank the whole panel.
Switch to HorizontalAlignment="Stretch": Avalonia's ArrangeCore sizes
Stretch from the available arrange rect (clamped by MaxWidth) instead
of DesiredSize, and centers the result exactly like Center alignment
does when MaxWidth caps it — so the box stays a fixed width regardless
of focus/placeholder state, on both desktop and Android.
Three issues surfaced by testing the previous lock/unlock reconnect fix
on Android with a large wallet:
- The keep-alive ping had no timeout, so a "half-open" TCP connection
(remote end gone with no FIN/RST ever delivered, the common outcome
of Doze/mobile-radio suspend after a longer lock) left it awaiting a
response that never arrives — the failure path that tears the dead
client down was never reached. Bound the ping to 8s and added a
re-entrancy guard so overlapping 20s ticks can't pile up concurrent
reconnect attempts while one is stuck.
- CheckConnectionOnResumeAsync bailed out whenever a sync was already
in progress, so a lock/unlock during an active sync (which has no
per-request timeout of its own) left it hung indefinitely instead of
recovering. It now cancels the stuck sync and tears down the dead
client, marking the interruption as self-inflicted (_resumeRecovering)
so the UI shows "reconnecting" and restarts immediately instead of a
transient error message.
- WalletSynchronizer.ExportCaches persisted raw transaction bytes only
for already-verified transactions, discarding anything downloaded but
not yet through Merkle-proof verification when a sync was interrupted
mid-way — forcing a full re-download of a large wallet's transactions
on every resume instead of resuming straight into proof verification.
Track confirmed txids at download time (_confirmedTxids), independent
of verification status, and export/preload against that instead.
Also unified the three sync progress messages onto one template
reporting against the sync's total transaction/proof counts rather than
this session's download count, so a resumed sync immediately shows
"transactions n/n" instead of a misleading "0/0" before jumping into
proof verification.
TcpClient.Connected only reflects the last known socket state, so a
connection killed silently while the phone was locked (Doze, mobile
radio suspend, NAT timeout) still reported IsConnected == true. The
keep-alive ping's failure was swallowed by an empty catch, so the
stale "connected" state never cleared and sync kept retrying on a
dead socket instead of reconnecting.
Treat a failed keep-alive ping as a disconnect (tear down the client
and reconnect), and add OnPause/OnResume to the Android activity to
force an immediate health check on resume instead of waiting for the
20s timer, which may itself be suspended during Doze.
README.md was out of date after several recent features: it listed 6
UI languages (missing Chinese Simplified), described watch-only as
xpub-only (address-only import also exists), overclaimed multisig as a
working PSBT flow (derivation for it actually throws — not
implemented), and its CLI quick-reference omitted restore-address and
servers. Also links the new USERGUIDE.md, previously unreferenced from
README.
SECURITY.md's "known limitations" list didn't mention multisig is
unsupported, worth stating explicitly since it's fund-safety adjacent.
Explains why initial sync scales with transaction count rather than
wallet age (one Merkle-proof round trip per confirmed transaction, no
batching in the Electrum-style protocol) and why it's slower on mobile,
plus why later syncs are fast (cache persists proofs/headers/anchoring
state). Advises against using this wallet as a mining payout address:
the many small transactions typical of payouts make sync noticeably
slower (observed ~2 minutes past 5,000 transactions).
Adds a pure address-only watch-only wallet mode (Core + CLI + App wizard
+ Send UI), on top of the existing xpub/xprv/WIF import flows which are
unchanged. TransactionFactory already refuses to sign for any IsWatchOnly
account, so the new address-only accounts inherit that guarantee for free.
Keeps the key-management section of the threat model accurate for the
new restore-address / pure address import path (SECURITY.md must stay
in sync with any change to key/seed handling per project convention).
Send already refused to enable Confirm for watch-only accounts (Signed
false), but the App had no way to get the resulting PSBT out to sign
elsewhere — only the CLI printed it. Adds a base64 PSBT box + copy button
in the Send summary card, and a visible warning banner explaining the
wallet cannot sign, on both desktop and mobile layouts.
New wizard flow mirroring the existing xpub/WIF import steps: paste one
or more plain addresses (no key material at all) and go straight to the
password step, since there is no derivation and no script-type ambiguity
to resolve for a fixed address list.
Parallels restore-xpub but for one or more plain addresses with no
extended key and no private key — output is explicitly labelled
"cannot sign" since the resulting wallet can never produce a signed tx.
Adds a WalletDocument.WatchAddresses field and WalletLoader.NewFromAddresses
to build an ImportedKeyAccount with no private key at all (unlike xpub/WIF
imports, which can still derive/hold key material). ScriptKind is inferred
from the address itself via a new DerivationPaths.KindFor helper. Signing
already refuses to run for any IsWatchOnly account (TransactionFactory), so
this only had to wire up construction/reload of the new account shape.
MaxInFlight was a hardcoded constant (32); different indexing servers
tolerate different levels of request concurrency before throttling, so
expose it as an optional ConnectAsync parameter instead of requiring a
recompile to tune initial-sync throughput against a given server.
_anchoredUpTo tracked which heights were already proven to hash-chain
back to a checkpoint, but only in memory: every app restart re-walked
and re-verified the whole header chain from the checkpoint even when
the header bytes themselves were already cached on disk, dominating
reconnect time on large wallets. Round-trip it through SyncCache like
the other sync caches (RawTxHex/VerifiedAt/BlockHeaders).
Translate every key in Loc.Strings and register "zh" in Loc.Languages
/LanguageNames. Wire it into the Settings language picker, which is a
hand-written RadioButton list (not generated from Loc.Languages), so
add IsLangZh to MainWindowViewModel and the matching button in
MainView.axaml. Update CLAUDE.md/AGENTS.md language count and note the
non-dynamic picker to avoid the same gap next time.
Sending a large amount from a wallet with many small UTXOs could
produce a transaction over the standard relay size limit, surfaced
only as a cryptic "Invalid transaction: Transaction's size is too
high" from builder.Verify() after everything else had already
succeeded — with no way for the user to recover other than manually
picking fewer coins.
Build() now orders spendable UTXOs largest-first and binary-searches
the smallest prefix of them that produces a valid transaction, so it
naturally prefers big coins over dust and avoids the limit whenever a
smaller input set can cover the amount. NBitcoin's own coin selector
already refuses to assemble a combination over the size cap and
reports it via NotEnoughFundsException with a distinctive message
rather than ever handing back an oversized transaction, so that case
is now recognized and translated into a clear, actionable error
instead of being read as "insufficient funds".
Balance/history now render as soon as tx downloads finish instead of
blocking on every historical Merkle proof, critical for mobile where
proof-checking can take much longer than the download itself. Proofs
continue to be checked in the background and each tx's Verified flag
catches up progressively; header ranges are now fetched in batches
(blockchain.block.headers) instead of one call per header to keep this
fast over high-latency links.
Coin selection (UtxoSpendability.IsSpendable) refuses to spend a UTXO
until its Merkle proof is actually checked, regardless of confirmation
count, so a server that fabricates a confirmed balance can get it
displayed early but never spent before the forgery is caught. The disk
cache only ever persists the fully-verified end state of a sync.
UI surfaces the new PendingVerificationSats/SpendableSats split with a
"verifying..." badge, and the sync save now runs off the UI thread to
avoid freezing on slower hardware.
Fills in the CHANGELOG.md entry for the checkpoint-anchoring security fix,
fuzzing infrastructure and its findings, OP_RETURN/coinbase-tag decoding,
and localization fixes since 0.9.1 (see previous commits), and bumps
<Version>/versionCode across the App and Android head csproj files ahead
of the tag.
Transaction detail view silently dropped OP_RETURN messages and coinbase
scriptSig text (e.g. pool tags), since TransactionInspector discarded
scriptPubKey/scriptSig bytes once no destination address could be derived.
Each output/input is decoded independently, so multiple OP_RETURN outputs
in one tx are all shown.
Adds a "User guide" button next to "Report a bug" in the Help > Info
tab, linking to USERGUIDE.md on GitHub via the same Launcher pattern
used for bug reports and release pages.
Convert dense prose paragraphs to bullet lists (Commands, Architecture,
GUI conventions, Working rules) and trim redundant wording; no content
removed. Both files kept in sync per the header sync rule.
Sync AGENTS.md/CLAUDE.md with the fuzzing workflow, add a README section
pointing to tests/PalladiumWallet.Fuzz, and record in SECURITY.md the two
guarantees the fuzz suite enforces: a bounded PBKDF2 iteration count on read
and a typed InvalidDataException for any malformed encrypted container.
New tests/PalladiumWallet.Fuzz project: one target per parser that
consumes untrusted input (block headers, Merkle proofs, peer lists,
wallet files, user-pasted mnemonics/keys/addresses/amounts), each
encoding the parser's documented error contract - any exception beyond
that contract is a finding, reported as a crash.
Three ways to run: the seed corpus (with regression inputs for every
crash found so far, including the two fixed in prior commits) replays
automatically inside dotnet test via FuzzCorpusTests, so a fixed
contract violation can't silently return; a built-in random-mutation
mode needs no external tooling (dotnet run -- <target> --random N);
and fuzz.sh drives real coverage-guided campaigns via afl++ +
SharpFuzz instrumentation for pre-release or post-parser-change runs.
A tampered or corrupted wallet file could reach Decrypt with broken
JSON, missing fields, invalid base64, or a wrong nonce/tag size, all of
which previously escaped as raw JsonException/FormatException/
ArgumentNullException instead of the documented WrongPasswordException/
InvalidDataException contract. Worse, the iteration count is read
straight from the (attacker-controlled) container with no upper bound:
a tampered file demanding e.g. 2^31 PBKDF2 iterations would hang the
wallet at open - a DoS via a file the user merely tries to open. Now
every malformed shape maps to InvalidDataException and the iteration
count is clamped to a sane maximum (10,000,000, well above the current
600,000 default).
server.peers.subscribe is attacker-controlled data: ParsePeers assumed
the standard [ip, hostname, [features...]] shape and threw on anything
else (non-array response, wrong element types), and GetString() on a
JSON string containing invalid UTF-8 bytes threw InvalidOperationException
instead of failing to parse — both found by fuzzing. Any unexpected
shape or unparseable string now degrades to "no usable data" instead of
propagating an exception into the sync/discovery path.
Wordlist.AutoDetect lands on NBitcoin's internal "Unknown" language for
text resembling no wordlist and throws NotSupportedException instead
of failing gracefully — found by fuzzing the restore-mnemonic input.
TryParse now treats that (and the same failure mode from the Mnemonic
constructor) as "not a valid mnemonic", consistent with every other
rejection path.
davide