A stored record a newer build wrote opens the popup on the recovery
screen. Export Saved Data puts that record, exactly as stored, in the
text box; a near-miss confirmation phrase erases nothing; the exact
phrase erases it and reloads into Welcome. Chrome and Firefox run the
same four cases, each under its shipped CSP.
They run before any wallet exists: with no wallet nothing saves on a
timer, so no save can write a good record over the unreadable one, and
the erase leaves the popup on Welcome for wallet creation. If any of
them fails, the last one removes the record so later tests still start
from Welcome.
Model: opus-5-5
AddressDetail and AddressToken defined their own isoDate() and
timeAgo(), hiding the shared pair in helpers.js, so a fix there would
not have reached them. The copies were identical and are deleted;
blockieHtml() and tokenLabel(), each defined twice, move to helpers.js.
A new test shows the history rows and the transaction detail view
write the time with the shared pair.
Deleted as never called: explorerLink(), ETHEREUM_SEPOLIA_CHAIN_ID,
getWalletValue() and getTotalValue() with their tests. Home's "Total:"
is the active address's total, as README.md already says.
addressColor() and etherscanAddressUrl() are no longer exported.
Model: opus-5-5
The background centred each approval window on the last focused window,
which could be an earlier approval window still open; headless Chrome
reports one as 1280x720, the browser refused the resulting position, and
the request failed with no window. It now centres only on a browser
window, and when the browser refuses a position it asks again without one.
In the Chrome suite a test could raise its prompt while the previous
test's window was still closing. After a passed test the runner now gives
approval windows five seconds to close and fails the test if one is still
open; after a failed test it closes them.
Model: opus-5-5
Max fills in a token's balance, cut down to the 18 decimal places the
confirmation screen accepts, or for ETH the exact balance minus the fee
reserve the confirmation screen's balance check gates on. An ETH fee estimate
that finishes after the Send screen was left, or its address, holding,
recipient or amount changed, fills nothing in. The confirmation screen works a
max ETH amount out again from its own fee estimate and signs it with that
estimate's fee fields, so a fee that rose before signing cannot push amount
plus fee above the balance. validateTransfer() still gates every send, the
check that ETH covers a token send's fee included. Where there is nothing to
fill in, a flash message says why.
Model: opus-5-5
The Settings round trip switched the theme and the network and closed
the popup at once. A close before the change handler's save lands loses
the switch, and the suite then ran on Sepolia.
tests/e2e/run.js now has one helper that polls a field of the stored
record until it holds the expected value, in place of the wait that
only read viewStack. Each Settings switch and spam-filter toggle waits
for its save, the recovery-phrase reopen waits for its saved view, and
reopenPopup() waits until the view it expects to reopen on is the saved
one. README.md no longer lists #446 among the open reports of the
Chrome suite failing under load.
Model: opus-5-5
When a site-connection prompt was decided before the toolbar popup raised
for it had loaded, that popup was torn down, chrome.action.openPopup()
rejected, and the background opened its fallback window for the answered
approval and only then removed it. In the Chrome end-to-end suite the next
test could take that window for its own prompt and lose it under its wait.
openApprovalWindow() now returns before creating a window when the approval
is no longer pending.
The blocklist test clicked its self-closing Reject with a plain click; it
now clicks it as the other site Reject does, with the click witnessed.
README.md and the e2e workflow comment no longer name this issue as what
keeps e2e-chrome from being a required check.
Model: opus-5-5
A wallet named only with spaces compared equal to an empty field, so
typing nothing would have deleted it, and a zero-width space in a name
made the name impossible to type back.
An empty typed confirmation is now refused whatever the name is. The
characters src/shared/symbolSpoof.js already defines as painting nothing
are removed from both sides before comparing. A name that shows nothing
at all is shown on the delete screens as "Wallet N", so it can still be
typed back.
Model: opus-5-5
parseHoldersCount used parseInt, which reads "1,000" as 1, "0x10" as 0 and
"1e3" as 1: a reported low count, which hides the token in the transaction
history and the send-screen token selector. It now accepts only a whole
number of zero or more, or a string of digits alone, no larger than
Number.MAX_SAFE_INTEGER, and returns null for anything else. The balance
list's holders !== null check did nothing, since null >= 1000 is already
false, and is dropped. README.md and docs/README.md say how each filter
treats an unknown count and that the token screen then leaves out its
Holders row; README.md lists src/shared/holders.js.
Model: opus-5-5
networks.js gives each network a nativeCurrency (ETH, SepoliaETH) and nothing
read it: every screen wrote ETH. The wallet's balances and the Send and
confirmation screens now use the active network's. A transaction's figures use
the network its chain id names, through nativeCurrencyByChainId(): the
approval value and fee, the wait, success and error screens, history entries,
the detail screen and the fee-limit refusal, so a site switching networks
cannot make one read as another network's coin. The "ETH" that selectedToken
and txInfo.token hold is the native token's id and is unchanged. A token
reporting any network's nativeCurrency is a spoof, and the detail screen calls
an entry a token transfer when it has a token contract.
Model: opus-5-5
A popup already open when the stored profile became unreadable stayed on the
last good profile until reopened. Every save already runs the check loadState()
runs at open; a save refused by it now stops the ten-second refresh, runs the
leave cleanup of the current screen, and raises the recovery screen. From then
on showView() shows nothing else in that popup, so a transaction wait or a later
save cannot take the user off it or clear an export or a typed confirmation.
That is held in memory, never as the saved current view, so a popup opened
after the record is erased elsewhere opens normally. Any other failed save
keeps the "NOT SAVED" banner. The popup test harness now honours
clearInterval().
Model: opus-5-5
decode() rendered the Deadline line with toISOString(), which throws on a
date past 275760-09-13, the last a JavaScript date can hold. A later
deadline, such as the uint256 maximum, therefore left the whole swap
undecoded, with nothing saying why. That line now reads
"After 275760-09-13 00:00:00 (no deadline in practice)".
Model: opus-5-5
A V2 exact-in amountIn of zero is the router's ALREADY_PAID marker: an
earlier step sent the tokens to the pair and the swap spends all of them.
Amount showed 0.0000 for it; it now reads "Whatever an earlier step sent
to the pair (V2 already paid)", in the style of the V4 open delta line.
A BALANCE_CHECK_ERC20 passes whenever the balance is at least minBalance,
so a zero one guarantees nothing. It now sets the output side only when
that side holds no minimum at the point the check is reached; a nonzero
one sets the output side as before.
README's Display Consistency text and TODO.md are updated to match.
Model: opus-5-5
The signature screen showed only the text a personal message decodes
to, with bidirectional, right-to-left and zero-width characters acting
on it, so a site could make the message read differently from the
bytes that are signed, and a message that was not hex was decoded into
NUL characters. The screen now shows the hex as "Raw data" alongside
the text, lays the text out left to right in byte order, and shows each
control character, line and paragraph separator, and character that
paints nothing (the set src/shared/symbolSpoof.js already strips) as a
U+XXXX mark. A message is hex when getBytes, which signing uses, reads
it; one that is not cannot be signed, so it is shown as plain text with
"Sign" disabled.
Model: opus-5-5
toDecimals() accepted any uint8 scale, but formatUnits() and parseUnits()
refuse more than 80 decimal places. A token reporting 81 to 255 made the
formatter throw, and the catch in the swap decoder and in the ERC-20 decoder
turned that into an undecoded approval screen with nothing saying why.
MAX_DECIMALS is now 80, the formatter's own limit, so such a scale is
treated exactly like an unknown one: both approval paths show the base-unit
amount with the scale stated as unknown. The balance list, the history list
and the Send screen use the same check.
Model: opus-5-5
With debug mode on, debugFetch logged every request's full URL and body,
so an RPC endpoint with an API key in its path or query string printed
that key to the console on every request. It now logs the HTTP method,
the URL's origin and, for a JSON-RPC body, the method name. The balance
refresh and token lookup log the RPC endpoint by its origin too. Failed
RPC calls print ethers' short message, since its full message for an
HTTP error carries the request URL. A failed endpoint check in settings
prints the endpoint's origin, since fetch's error for a URL with a user
name and password carries the whole URL. The README's DEBUG Mode Policy
says what debug mode logs.
Model: opus-5-5
Each eth_requestAccounts or personal_sign call opened another approval
window, so a page calling in a loop could cover the screen with identical
prompts. While a site's connection or signature prompt is unanswered, a
further request of that kind from the same site is now refused with
EIP-1193 -32002 and opens no window; all signing methods count as one
kind. A connection prompt whose toolbar popup closed before it connected,
and which the toolbar popup no longer opens, is shown again by the site's
next request instead of refusing the site until the address changes.
Model: opus-5-5
A site could fix the nonce of the transaction the user was asked to
sign: the same nonce as a pending transaction, at a higher fee,
replaces it, and a nonce above the account's next one leaves the new
transaction stuck behind a gap. `nonce` is no longer one of the fields
taken from the request, so the transaction always gets the account's
next nonce from the network, and that is the nonce the approval screen
shows and the popup signs.
Model: opus-5-5
allowedSites and deniedSites held the bare hostname, so a grant to
https://dapp.example also authorised http://dapp.example and every port
on that host, and the connection, transaction and signature prompts
named only the hostname. Both lists now store and match the full origin
(scheme://host[:port]), the key the connections approved without
Remember already used. The prompts, the Settings site lists and
AUTISTMASK_REMOVE_SITE use the origin too. Entries saved by hostname
are not migrated (pre-1.0): they match no site.
Model: opus-5-5
When two addresses' explorer reports disagree on a token's decimals, the
Send screen showed the stored figure while the confirmation screen said
the balance was unknown. One function in send.js now gives both the
balance and scale, so both read `unknown (SYMBOL)`.
The confirmation screen's fee-unknown message names its cause: for an
unknown scale it says the wallet does not know the token's decimal
places and the transaction cannot be sent, instead of asking for a
retry that cannot help. Other causes keep the old sentence.
Model: opus-5-5
decode() had no arm for Universal Router command 0x09, so the approval
screen showed no token or amount for it. It now takes the path's first
token and amountInMax as the input side, the last token and amountOut as
the output side. With such a step, the Amount figure reads "Up to
<amount>" whichever step set it, on the approval, wait, success and error
screens, unless it reads "Unlimited", as an unbounded PERMIT2_PERMIT does,
or "All available (V4 open delta)". In every swap, UNWRAP_WETH makes
Token Out ETH only when the output side is WETH, on mainnet or Sepolia, or
no step set it; otherwise the output keeps its own token and figure.
decodeV2SwapExactOut() loses its eslint-disable comment.
Model: opus-5-5
A transaction with no `to` showed a blank address, with a colour dot whose
colour was `undefined`, on the wait, success and error screens, the
transaction detail view and the history rows on Home, AddressDetail and
AddressToken. The approval screen showed "(contract creation)".
All of them now say "This transaction creates a new contract. It has no
recipient." The three history lists draw a row's counterparty lines through
one helper in helpers.js. A transaction with a real `to` is unchanged. The
new test drives each screen and list both ways.
Model: opus-5-5
The stored ETH and token balances and the send-confirm screen's fee were each
cut to six decimal places, and a token holding cut to zero was dropped, so a
value below 0.000001 read as zero. Balances are now stored exactly, whatever
decimals a token declares, and every nonzero token holding is kept; the balance
check reads a token balance to its first 18 places. The balance lists, the
send-screen token selector, the address total and the remove-address warning
leave out a holding below 0.000001 themselves, through isBelowOneMillionth().
The send and send-confirm screens' balances, reserve and insufficient-balance
messages go through truncateAmountNeverZero(). The send-confirm and approval
screens both render the fee through formatFee(), which prices the exact fee in
USD.
Model: opus-5-5
A site allowed without "Remember" lives only in the background's
in-memory connectedSites map. Settings never listed it, and
AUTISTMASK_REMOVE_SITE, sent on every remove, did nothing, so the user
could not end such a connection.
Settings now asks the background for those sites and lists them under
Connected Sites. Removing a site from Allowed Sites or Connected Sites
drops its remembered entry under every address and sends
AUTISTMASK_REMOVE_SITE with the hostname; the background deletes every
matching connectedSites entry and sends accountsChanged with an empty
list to the site's tabs. Only the extension's own pages may send either
message.
Model: opus-5-5
A site connected without "Remember" lives only in the background's
in-memory connectedSites map. Removing an address or deleting a wallet
dropped the remembered permissions but never told the background; the
entry went only as a side effect of the accountsChanged broadcast, which
empties the whole map when the active address changes.
dropSitePermissions(), shared by both removal paths, now sends
AUTISTMASK_ADDRESSES_REMOVED with the removed addresses, and the
background deletes their entries. Only the extension's own pages may
send it.
Model: opus-5-5
The README said a genuine zero always renders `0.0000`, beside a
`Min. received` example, though a zero minimum there now reads
`None (no minimum guaranteed)`. It now says the amount rule renders a
zero as `0.0000` and that two swap lines say a zero in words instead:
`Min. received` for a zero minimum, and `Amount` when it shows a V4
exact-in `amountIn` of zero.
A new list says what the decoded ERC-20 and swap amount lines on the
transaction approval screen can read, including that a zero
`BALANCE_CHECK_ERC20` `minBalance` now reads the no-minimum wording.
The token permission warning on the signature screen is left to the
SignApproval section.
Model: opus-5-5
The typed-data screen listed a Permit or Permit2 signature as plain key/value lines, like a sign-in message. It now shows a red warning naming the spender and each token and amount, read only from the fields the signed type declares (a Permit's token is the domain's verifyingContract); anything those fields do not give reads Unknown.
The screen printed the page's primaryType, but ethers signs the type it derives from types. It now shows that type and refuses typed data whose stated type is missing or differs: Sign disabled, checked again where signing starts.
Deviation: the warning names no deadline or expiry; see the issue.
Judgement call: DAI's older permit and Permit2's batch and witness transfer types are recognised too.
Model: opus-5-5
Add docs/RELEASE.md, linked from README.md's Release Artifacts section, giving
the release procedure as a numbered list: confirm main is green in CI, confirm
the one version in package.json and the two manifests matches the intended tag,
make package from a clean checkout, verify SHA256SUMS, create the annotated tag
vX.Y.Z, then distribute per browser. Each step names who performs it, marks the
owner-only ones, and states the check that it worked. Every repo command cited
(make setup, make check, make package) exists on next; tagging and verification
use standard git and coreutils, and the CRX pack line is README's own.
The per-browser distribution step is written as pending the owner's choice on
issue 386, with the Firefox and Chrome options named but none presented as
settled. Docs only: no code or test changes.
Model: opus-4-8
A stored allowedSites whose value was not a list rendered a working popup and then made every subsequent save fail silently, so the user operated a wallet that persisted nothing -- worse than a blank popup, which is at least visibly broken. fraudContracts and selectedToken had the same shape: a container floored by truthiness or not at all, while its entries were dereferenced. Entries are now floored as well as containers, following the idiom #311 established, and a failed save raises a persistent banner instead of vanishing into a swallowed rejection.
The per-field justifications that used to live in a hand-written header are replaced by a contract test that drives each field's hostile and falsy values through a real popup boot, so a claim about a field answers to the code rather than to prose. Its guarantee is stated narrowly and deliberately: no structural dereference on the code paths a wholly-corrupted profile takes, which is not every path a stored record takes. The paths it does not drive are named where the claim is made, and are tracked in #379.
parseInt(decimals || "18") ran before writing stored tokenBalances[].decimals, so an explorer reporting no decimals produced a fabricated 18 indistinguishable from a real one at read time. That defeated the resolve-or-refuse guarantees of #306 and #340: their refusal paths were intact but never fired, because the guess was laundered upstream of them.
An absent scale is now stored as unknown, and a holding whose scale nothing knows carries a null balance -- unknown, never zero -- with six reader sites saying so rather than printing 0.0000. The Send screen resolves the display scale rather than reading the stored one, so a bundled token whose explorer row omits decimals still sends; when the scale cannot be resolved the stored quantity is withdrawn too, so the user is told the balance is unknown rather than only that the fee failed.
Existing fabricated 18s cannot be told apart retroactively and are replaced wholesale on the next balance refresh. An explorer-sourced scale stays trusted -- only fabrication is removed; the reasoning is recorded on the issue.
Stored state had no version and no structural validation, so a corrupt blob produced a completely blank popup with no message and no recovery control, and made every dApp RPC call from every page answer a generic -32603. There was no reset or wipe control anywhere in the UI.
saveState() now stamps a schema version and loadState() validates the shape. A version it does not understand, or a wallets array it cannot parse, lands on a recovery screen that names the problem, offers the stored record verbatim for export, and offers a destructive reset behind a typed confirmation. Unversioned but valid state -- which every existing install has -- migrates in place and keeps working; it is never shown a wipe prompt. A dApp call against unusable state answers -32007, which EIP-1474 leaves unassigned, rather than -32603. networkById() refuses an unknown id loudly instead of returning mainnet, and networkId is validated so a corrupt value cannot be used as an object key.
Fields the gate does not refuse are floored by type, container and entries both: a malformed trackedTokens or tokenBalances entry is dropped rather than dereferenced. Verified by an independent sweep of 1152 corrupt blobs producing no blank popup, with the same harness showing 9 blanks against the previous revision.
tokenInfo() returned decimals 18 for any token absent from the bundled list, so the swap approval line rendered a real 1000.00 of a 6-decimal token as 0.000000000001. The scale is now resolved from what the wallet already holds (bundled list, tracked tokens, explorer-reported decimals) or refused outright, matching the rule set for the ERC-20 path in #306. A refusal reuses unknownDecimalsAmount(), so it reads as "base units (decimals unknown)" with no decimal point and no symbol, and the same string propagates to rawValue so no downstream screen can render a figure the approval screen refused. No new network call on the approval path. Verified green on all three CI contexts: check, e2e-chrome, e2e-firefox.
manifest/chrome.json now carries a fixed public key, so the extension id and the chrome.storage.local partition holding the wallet stay stable across checkout moves and re-clones instead of being derived from the absolute path. A release entrypoint produces a self-contained versioned artifact per browser, including the files that sit at dist/ root outside both browser directories. One version source of truth, enforced: the build fails naming the culprit when the two manifests and package.json disagree, and BUILD_COMMIT now marks a dirty tree as dirty. Firefox ships an unsigned XPI; the README states that release Firefox and ESR refuse it, that Developer Edition or Unbranded is required, and that Remove is irreversible except from the recovery phrase, which is asserted by test.
An amount below the 4-decimal display floor now extends to its first significant digit on the approval and confirmation screens, instead of stating a real transfer, allowance or swap Min. received as 0.0000. The rule had been implemented three times; all three now share src/shared/amountDisplay.js, which holds the plain truncation and the floored variant side by side. History and balance lists keep the unfloored rule, pinned by test.
A failed release build no longer leaves a complete, loadable debug bundle in dist/ whose every wallet uses the publicly committed test recovery phrase. Each step of the release build runs through script/discard-dist-on-failure, which removes dist/ on failure, says on stderr that it did and why, and returns the step's own status. build-debug is deliberately unwrapped. script/verify-build is untouched.
README, the script synopsis, its header paragraph and the check_dist_tree comment now all say the same thing: regular files and symlinks under dist/ are covered; fifos, sockets, device nodes and empty directories are not, and why. No behaviour change — the walk is untouched.
A user who forgot their password but held their recovery phrase was permanently
locked out: deletion was password-gated and re-importing the phrase was refused
as a duplicate. Their only escape was destroying extension storage through
browser internals, taking every other wallet with it.
DeleteWallet gains an "I have lost my password" route that destroys the stored
secret after the wallet's name is typed back. No password gate was added:
requiring one to discard a secret protects nothing, since an attacker who wants
destruction can uninstall the extension, and the only person it stops is the
legitimate user who lost it. The screen is excluded from RESTORABLE_VIEWS and
registers an onViewLeave cleanup.
Deletion was chosen over re-import because a key wallet is duplicate-checked by
address rather than xpub, so an xpub-only relaxation would leave that user
still wedged; because re-import makes the user retype their recovery phrase
into a live popup merely to change a password; and because it reaches no end
state that delete-then-import plus scanForAddresses() does not. The attacker
argument did not decide it — re-import clears the "no worse than the phrase
alone" bar.
All three AddWallet password hints now state the password cannot be recovered
or reset and name that mode's only backup, the xprv mode correctly claiming no
recovery phrase. deleteAddress.js no longer tells the user that deleting a
wallet asks for a password, which this change made false.
The typed confirmation collapses internal whitespace on both sides: a wallet
renamed with two spaces displays with one, so the string a user could see and
type could never match, making the confirmation untypable on the one screen
whose purpose is un-wedging a stuck user.
Measured, not reasoned, after review found the first reserve twice too large
and pushing the Import button below the fold: #btn-add-wallet-confirm bottom
628.13 -> 580.13 at 360x600, scrollHeight 636 -> 600, hint box 48px identical
across all three tabs and on re-entry. make check 40 suites / 828 tests,
test-e2e 55/55, test-e2e-firefox 8/8.
verify-build read its expectation from AUTISTMASK_DEBUG in its own environment
and the Makefile invoked it bare, so an operator with that variable exported
who ran the release target got an INSECURE debug build — every wallet it
creates uses the publicly committed test phrase — verified green, exit 0. It
also had no provenance: a 26-byte file containing the right marker string
passed, the content script and manifest.json were never inspected, and an
entire hand-written dist/ passed.
--expect release|debug and --receipt PATH are now both required, with no
defaults and nothing read from the environment. build.js records every file it
emits with its sha256 and writes the receipt; the Makefile mktemps it outside
the repo per invocation with a trap, and build.js refuses a receipt path inside
dist/. Verification runs three passes in a load-bearing order — receipt shape,
full dist/ walk, then per-file bytes — so an unwalkable subtree cannot make
files look absent. dist/constants-bundles.txt, which was an unsigned trust root
living inside the tree it vouched for, is gone.
What this proves is bounded and stated as such: dist/ is byte-for-byte the
output of the build.js run that just finished, within one make build
invocation. It proves nothing about the honesty of the source tree or build.js,
and nothing to anyone handed a dist/ from elsewhere — that is signing, #310.
The standalone make verify-build target is removed because its only input would
be dist/ itself, i.e. the artifact vouching for itself.
Verified: make check green, test-verify-build 39 cases (was 18), test-e2e 55/55
and test-e2e-firefox 8/8 with make build running uncached inside both images.
All four original bypasses now exit 1. Mutations: digests disabled fails
exactly 4 cases, dropping the dist/ walk fails exactly 8, restoring the ambient
fallback fails exactly 1.
A hostile ERC-20's symbol() reached an innerHTML string unescaped, and neither
manifest declared default-src, so an attacker deploying a token with 1,000+
holders and airdropping one unit could render a full-viewport cross-origin
iframe over the wallet's own UI, on screens where the user types their
password.
escapeHtml is now a pure string replace over & < > " ' — the old version
round-tripped through textContent, which escapes neither quote, while already
being used inside data-copy="...". All 19 files in src/popup/views/ were
audited: beyond the reported symbol site, the explorer-supplied directionLabel
in all three transaction lists, wallet.name, addr.ensName, the blockie data:
URIs and two ad-hoc quote-only escapes were also unescaped. Explorer URLs now
go through one helper that percent-encodes the path segment.
Both manifests add default-src 'self', frame-src 'none', form-action 'none' and
base-uri 'none'. Three loosenings are pinned in tests/manifest.test.js and
justified in README.md: style-src 'unsafe-inline' (39 static style attributes;
Firefox implements neither style-src-attr nor 'unsafe-hashes'), img-src data:
(blockies), connect-src https: http: (user-configurable RPC).
Note frame-src 'none' blocks a frame loading, not the element existing, so the
zero-iframe assertion is a claim about the escaping alone; the test asserts the
element count and the literal rendered text separately, taking the count before
any click an overlay could intercept.
Verified: make check 39 suites / 811 tests, test-e2e 55/55 including the
WebAssembly-under-CSP assertion, test-e2e-firefox 8/8, zero CSP violations
asserted rather than merely unobserved. Reverting only balanceLine's
interpolation reproduces the attack as 2 iframes on the address screen.
The send screen was built from the indexer's decimals while the transfer was
encoded from the contract's decimals() read at signing time, with nothing
comparing them. A token whose scales disagree moved 10^12 times the approved
amount.
The displayed scale is now carried on pendingTx from the same tokenBalances
entry the amount, balance and symbol were rendered from, and both encode sites
use it. transferAmount.js refuses rather than falling back when the two scales
disagree or either is unusable.
Adds the first end-to-end coverage of the popup's own Send -> ConfirmTx ->
Sign & Send path; #btn-confirm-send had never been clicked by any test.
The EIP-1193 error code now survives the last hop into the page, so the third standing limit in the End-to-End Tests section no longer describes the code. Verified against src/content/inpage.js (ProviderRpcError carries code and data) and tests/e2e/run.js (the page-side error is asserted, not only the wire).
The dApp signing path was the largest unverified surface in the milestone: the
only place where the content script, the inpage provider, the background worker
and the popup all have to work together, with unit tests covering each side in
isolation and none covering the seam.
A page served by the harness speaks EIP-1193 to the real provider -- asserted by
EIP-6963 object identity, not by shape -- and eth_requestAccounts, personal_sign,
eth_signTypedData_v4 and eth_sendTransaction are each driven through to approval
and to rejection.
Every signature is recovered and compared to the approved address; the broadcast
transaction is parsed from the bytes captured at eth_sendRawTransaction and
checked for signer, recipient, value, calldata and chain. A signature that
merely came back would pass against a wrong key, a wrong message or a wrong
chain, so each assertion was demonstrated failing against a variant that is
wrong in exactly one of those ways.
The password is asserted absent from every message crossing the extension
boundary, which gives #157's fix a permanent floor rather than a one-time
review.
Two defects this surfaced are tracked separately: EIP-1193 error codes never
reach the page (#274), and approving a site connection races the popup teardown
(#275). Neither is asserted as correct here. A real dApp with real funds against
mainnet remains an uncovered human pass and is documented as such.
clawbot