TRANSFER, PAY_PORTION, PAY_PORTION_FULL_PRECISION (0x07, on the routers with
the 0x7f mask), PERMIT2_TRANSFER_FROM, ACROSS_V4_DEPOSIT_V3 (0x40) and V4
TAKE_PORTION were payouts the decoder did not read, so a swap using one read
Min. paid out with an unknown payee. Each recipient is now read: the sender,
the router, a hop into the next V2 pair, or another address, which is named.
Paying all the router holds of a token pays it out as SWEEP does; a TRANSFER
of CONTRACT_BALANCE of ETH reverts, so it pays nothing out. A
PERMIT2_TRANSFER_FROM into the router is shown with the amount it moves and
the moved token's address, and a bridge deposit names the bridge and each of
its recipient and depositor that is not the sender; either keeps the minimum
from reading received.
Model: opus-5-5
A network switch puts the outgoing network's tracked tokens away in a new
persisted field, networkTrackedTokens, and brings back the incoming network's
list, or an empty one, as endpoints are kept; the background's switch uses the
same function. The two add-token screens offer their common-token picks on
Ethereum Mainnet only. After a switch in Settings, Back renders the screen it
lands on again, and the stack is cut at the first screen that cannot be
rendered from saved state, or at AddressToken for any token but the native
one; the cut stack is saved with the switch. ConfirmTx drawn again keeps the
network it was reviewed on. A contract lookup that answers after a switch adds
nothing.
Model: opus-5-5
The swap decoder read no recipient, so a swap paying its output to another
address still showed Min. received. It now reads the address each swap step,
SWEEP, WRAP_ETH, UNWRAP_WETH, V4 TAKE and TAKE_ALL pays. Min. received shows
only when some step pays the sender, none pays another address, none that may
pay one went unread, and the output does not stay in the router with no later
SWEEP or UNWRAP_WETH paying it out. A payment to the V2 pair the next V2 swap
spends as already paid is a hop, not a payout. Otherwise it reads Min. paid
out, with Paid to lines.
Model: opus-5-5
ConfirmTx showed and checked the balance Send handed it, so a screen the popup reopened onto stated the balance as Send knew it, not as it is on the network the screen is drawn for.
Each time it is shown, ConfirmTx now reads the From address's ETH balance, and for a token send its token balance, from the node reviewedNetworkProvider() picks for the network the transaction was reviewed for, the node the fee estimate and recipient checks ask. The fee estimate starts once the read returns. Until then the balance line says "Loading..." and Sign & Send stays disabled; a failed read is stated as unknown. Each message the read can bring has a box reserved from the first paint, so nothing moves when it returns.
Model: opus-5-5
ConfirmTx set the transaction's chain id to the active network's every
time it drew the screen, so a popup reopened onto a saved confirmation
after a network switch in Settings described the reviewed transfer as
the new network's and sent it there.
show() now sets the chain id only on a transaction Send hands over, and
the screen describes the transaction entirely on that network: native
currency, USD values, explorer links, and the node and explorer its
checks ask. explorerUrl(), renderAddressHtml(), getPrice(), formatFee()
and checkEtherscanLabel() take an optional chain id for this. A saved
transaction whose chain id names no network is not restored.
Model: opus-5-5
Each Universal Router address now carries the mask its own source reads a
command's type with: 0x1f for UniversalRouterV1, 0x3f for V1_2 and V2, 0x7f
for V2_1_1 and V2_1_2. The fixed 0x1f dropped the 0x20 bit, so
EXECUTE_SUB_PLAN (0x21) was listed as a V3 exact-out swap. A sub-plan step is
now listed as "Execute Sub-Plan (its steps are not shown)". Its commands are
not read and may spend from or pay the user wherever the step stands, so with
one the summary names no swap direction, Token In and Token Out read
"Unknown (could not be decoded)", and there is no Amount or Min. received.
Model: opus-5-5
DeleteWallet and DeleteWalletLostPassword told key and xprv wallets that
their recovery phrase was the way back, but those wallets have none. Both
screens now name the recovery phrase, the private key or the extended
private key, according to the wallet's kind.
The noun comes from keyMaterialName() in the popup view helpers, which
DeleteAddress now uses too in place of its own two-way choice, so these
three screens get it from one place. The lost-password screen's route-back
sentence is reworded to fit all three kinds.
Model: opus-5-5
The Universal Router decoder recognised execute() by its selector alone, so
the same calldata sent to any address was described as a swap through the
router. It now keeps the router deployments Uniswap publishes for Ethereum
Mainnet and Sepolia, and decodes only a transaction sent to one of them on
its network, ignoring address case. TxApproval passes the transaction's
chain id, TransactionDetail the active network's.
Model: opus-5-5
A V4 step's input side is what SETTLE or SETTLE_ALL names, else the one
currency its swap actions spend and never buy; its output side is what
TAKE or TAKE_ALL names, else the one they buy and never spend. Amount sums
the amountIn of the actions spending the input side, shown only when each
states one; Min. received likewise. An unread swap action leaves the sides
to SETTLE, SETTLE_ALL, TAKE and TAKE_ALL, with no figure. A step with a
swap action always sets the output side, sets the input side only if no
earlier step did, and leaves a side its rules cannot name unknown, with no
figure. SWAP_EXACT_OUT spends its first path entry.
Model: opus-5-5
The transaction count the node reports is what an address has sent, so
an address that only receives, such as a savings address, got the "ZERO
transaction history" box on every send. The check now also asks the node
for the recipient's balance and warns only when both are zero. The box
says the address has never sent a transaction and holds no ETH, naming
the active network's native token as the contract warning does.
Model: opus-5-5
The background judges each tab by its URL's origin, the only one it can see, so a page its server sandboxes, whose own origin is null, was still sent the new active address on an address switch although all its requests are refused.
The background now names the origin it judged by on each accountsChanged event, and the content script passes the address to the page only when the page's own window.origin is that origin, an empty list otherwise. An event naming no origin also reaches the page as an empty list. chainChanged is unchanged. The README's SiteApproval section says the same.
Judgement call: covered by a unit test that sets window.origin; no browser-suite test.
Model: opus-5-5
A connection allowed without "Remember my choice for this site" lived in a module-level object in the background, so it ended the first time Chrome stopped the idle worker. Each such connection is now one item in the browser's session storage, which outlives the worker and is emptied when the browser closes, read and written through new wrappers in src/shared/browserApi.js. Removal in Settings, address switch, address removal and wallet removal still end it; Settings now waits for the removal before redrawing.
manifest/firefox.json now declares a minimum of Firefox 115.0, the first with session storage.
Judgement call: address and wallet removal still notify the background without waiting.
Model: opus-5-5
Send's Review accepted any amount parseFloat() read as above zero, so 1e-3, 1abc and amounts with too many decimal places reached ConfirmTx, which then refused them or failed to estimate the fee.
isValidAmount() in src/shared/txValidation.js is now the one amount check, used by Send's Review and Max and by validateTransfer() for ConfirmTx: plain digits with at most one decimal point, above zero, nothing but zeros past the token's places. Zeros past those places are accepted, since the transfer encodes the amount exactly. Send now reads the token's decimal places before the ENS lookup, because the check needs them.
Judgement call: ConfirmTx now also refuses zero, as Send already did.
Model: opus-5-5
A request the node refuses now reaches the page with the node's own JSON-RPC code and data: a reverting eth_call, an eth_sendTransaction whose gas estimate reverts, a failed broadcast. Every other refusal from the background carries a code too: 4100 when there is no address, -32603 for a failure inside the wallet, -32602 for a transaction request the wallet refuses as malformed. On the transaction path the page keeps the wallet's own sentence.
nodeError(), which finds the node's error inside an ethers error, lives in src/shared/approvalVerify.js beside failureTexts(), which did the same lookup.
Not run in a browser.
Model: opus-5-5
The popup picked an active address only when it opened, so after the first
wallet was created nothing was active until a reopen: Home had no headline,
and Send and Receive refused. Each of the three import paths on AddWallet now
makes the new wallet's first address active when no address is. When the
phrase or xprv scan then replaces the address list with one that leaves that
address out, the new list's first address becomes active and the background
is told, as when the user picks an address on Home. The address scan stand-in
in the popup test boot now answers in the real scan's shape, which the phrase
and xprv imports read.
Model: opus-5-5
script/test-e2e and script/test-e2e-firefox tagged their image with one
fixed name. When two clones ran a suite at once, the second build moved
that tag, the image store deleted the first run's now-untagged image, and
that suite exited 125 without running a test. Each run now tags its image
with the suite's name and twelve random hex digits and runs it by that
tag. The cleanup that runs on exit, Ctrl-C, kill and a closed terminal
removes the run's container if docker's client left it running, then the
image. The image-ID file that guarded against a moving tag is gone.
README.md's End-to-End Tests section says which endings clean up.
Model: opus-5-5
The approval window opened 360 pixels wide while the page is 396, cutting
off the Reject button; it now opens at 396. The browser counts the
window's frame in that width, and a tall screen gives some of it to a
vertical scrollbar, so once the approval's screen is up the page widens
its own window by however far it would scroll sideways, times the zoom of
its own tab. Firefox draws some zoom steps a little larger than it
reports, so the page measures again after each widening, three times at
most. The toolbar popup's window is left alone. The Firefox suite checks
that the sign prompt fits at 170%. README's Visual Style line now gives
the page's real width.
Model: opus-5-5
A page with an opaque origin, such as one its server sandboxes with a
Content-Security-Policy, is reported to the background with the origin
"null", and Chrome reports a local file as "file://". The background took
either as a site, so every such page was one site: a connection remembered
for one covered them all, and the phishing check never matched. The
background now parses the reported origin and, when it has no host, refuses
the request with 4100 like one with no origin, before any method runs.
tests/rpcOrigin.test.js sends both, and README's SiteApproval says what is
refused.
Model: opus-5-5
ethers signs a bool field by whether its value is truthy, so the string
"false" was shown as false and signed as true, and a bool field left out
was shown as nothing and signed as false. The sign screen now refuses
typed data whose message leaves out a declared field, at any level of
nesting, or holds a bool that is not true or false: the error line names
the field and Sign is disabled. The background refuses a signature for it
too. Presence is an own-key test, so a field named like a built-in
property such as constructor is not found by accident. signedPrimaryType()
moves to the shared module so the background reads the message as the
same type.
Model: opus-5-5
Chrome stopped the background worker about six minutes after a site's
request even with the approval window open, and the approval and the
site's waiting request went with it. Every approval is now added through
addPendingApproval(), which starts a timer in the worker that calls
runtime.getPlatformInfo() every 20 seconds while any approval is
pending; Chrome counts each call as activity.
tests/e2e/slowApproval.js, run by make test-e2e, leaves a transaction
and a signature approval open for seven minutes in minimized windows and
requires the site to receive both results. It drives Chromium over the
DevTools pipe because Playwright keeps DevTools attached to the worker.
Its loopback dApp server, shared with the Firefox suite, moves to
tests/e2e/dapp.js.
Model: opus-5-5
The four V4 swap actions are read only as the structs deployed on mainnet
and Sepolia. ethers read the newer v4-periphery structs, which add
minHopPriceX36, as the deployed ones, and the newer multi-hop exact-in
struct showed the offset of its hop-price array as the swap amount. A
decode now counts only when encoding it again gives back the same bytes.
A swap action that fails that marks its step "V4 Swap (could not be
decoded)", and a side its SETTLE or TAKE did not name reads "Unknown
(could not be decoded)". The wait and error screens no longer show the
ETH value as the amount of a swap whose amount was not read, unless the
swap is paid in native ETH.
Model: opus-5-5
The two node programs and the data file from f7a2437 regenerate the name and
url of each token in src/shared/tokenList.js. They are now CommonJS under
script/lib/ and read ethereum-lists at a pinned commit. The apply step also
rewrites an entry that already has a name and url, so it runs on the current
list, and stops on an address that matches no list entry. Exported as
applyEnrichment, it is run by tests/applyEnrichment.test.js, which fails when
the committed list differs from what it writes. Its table of websites now gives
FLOKI and TURBO a URL. make enrich-tokens and make apply-enrichment run the
programs. script/check-censored allows the MUSD token's on-chain name in the
data file.
Model: opus-5-5
ethers asks the node for a gas estimate only when the request has no gas
limit, and that estimate is the wallet's only check that a transaction
does not revert. Since a site's `gas` became the gas limit, a reverting
request carrying one opened the approval window and could be signed and
mined to fail. Such a request is now estimated first, with the site's
limit and under the same timeout; a failure goes back to the site with no
window, and the site's gas limit is still the one shown and signed. A
limit too low for the call now fails the same way.
Model: opus-5-5
Sending, site transactions, site signatures and private key export
derived the key at the address's position in the wallet's list. After
an import whose scan passed over an unused index, or after removing an
address that was not the last, that position named another address's
key.
Each HD and xprv address record now stores the index it was derived at
(at wallet creation, by the import scan, by "+"). getSignerForAddress()
derives at that index, refuses a record without one, and refuses any key
whose address differs from the record's. The confirmation screen signs
as the From address it shows, not the selected one. The export screen
reports a refusal instead of calling the password wrong.
Model: opus-5-5
ethers signs only the fields a type declares and drops any other key, but
the Message section listed every key the site sent, so an added key read
as signed. The Message section now follows the signed type, at every level
of nesting and in the type's order. Typed data with other keys is not
refused: those keys are listed after the message under "Not part of the
signature", by their place in the message. README SignApproval says so.
The operator/spender test now also checks the Message section, and new
tests cover an added key in a nested struct and in a list item.
Model: opus-5-5
eth_sign is signed as an EIP-191 personal message over its data, which is
what the Ethereum JSON-RPC specification defines it to return. Its warning
on the signature prompt said the site was asking to sign a raw hash that
could drain funds, and the README said the same; neither was true. The
warning now says the site used the older eth_sign method and that the data
will be signed as a personal message, and the README's SignApproval section
says so too. Tests pin that the background reads eth_sign address first and
data second, and that the sign screen answers it with the personal message
signature over the data.
Model: opus-5-5
Settings is reachable from the confirmation screen, and Back from it only
unhides the screen as it was drawn. Sign & Send took the network at the
click, so a transfer reviewed on Sepolia went out on mainnet after a switch
in between. show() now records the chain id of the network it draws for.
Sign & Send refuses with a message on the screen when the active network
differs, and the send carries that chain id, so ethers will not populate it
for a provider on any other network.
Model: opus-5-5
Sites send the gas limit under its JSON-RPC name, `gas`. ethers only
reads `gasLimit`, so the field was dropped and the node's estimate was
shown and signed instead. The page's request now takes `gas` as the gas
limit, and the existing ceiling, display and comparison with the signed
transaction cover it as they already did `gasLimit`. A request that
gives both with different values is refused.
Model: opus-5-5
The blocklist check hashed the hostname as the page gave it, so
"myetheywallet.com." and every subdomain written that way hashed
differently from the listed "myetheywallet.com" and opened approval
prompts with no warning. The browser and DNS treat both spellings as
one host. hostnameVariants() now drops one trailing dot before
building the hostname and its parent domains, so the check sees the
name the list holds.
Model: opus-5-5
In a clone where script/bootstrap has not run, node_modules/ is empty and
yarn falls back to whatever prettier is on the PATH, which can format files
differently from the version yarn.lock pins. script/fmt and script/fmt-check
now stop with a message telling the user to run script/bootstrap when
node_modules/.bin/prettier is missing. After bootstrap they behave as
before. The README Entrypoints entries say so.
Model: opus-5-5
EIP-712 has a wallet refuse typed data whose domain chainId is not the
active chain, because the signature is valid on the chain it names. The
sign screen now shows such a request with the reason on its error line
and Sign disabled, the same way it refuses a mismatched primary type.
The background checks again when the signature comes back, against the
network active then, and answers the site with that refusal instead of
the signature. A chain id is read as ethers signs it, so 1, "1" and
"0x1" are one chain. Typed data with no chainId signs as before.
Model: opus-5-5
Once Settings has saved the new network it asks the background to send
chainChanged with the new chain id to every open tab, through the same
function an approved site switch request uses. Only the extension's own
pages can ask for that. Choosing the network already active changes
nothing and sends nothing.
Model: opus-5-5
A connected site's wallet_switchEthereumChain request for the other
supported network now opens a prompt in its own window, through the
existing approval machinery, naming the site and both networks. The
network, endpoints, balances and caches change, and chainChanged is
sent, only when the user approves it; rejecting or closing the prompt
answers 4001. One such prompt per site at a time; a request for the
active network needs none. The approval window no longer shows the
connection prompt while it waits for the approval's description,
since both prompts answer on the same port.
Model: opus-5-5
The README's 1.0 non-goal said Ethereum mainnet only, while Sepolia
ships as a selectable network. It now puts mainnet and Sepolia in scope
and other chains out of it.
The injected provider's flag named after the other wallet stays, as an
interface-compatibility flag many dApps check. It is now written down
as an exception to the rule against naming competitors, in the README's
Policies section and in RULES.md, and the comment beside it says so.
Neither document spells the name, so script/check-censored, which
already allows the flag only in src/content/inpage.js and its built
copies, is unchanged.
Model: opus-5-5
The polarity check now counts only `hostile` and `falsy` values, not a
`hostileRestore` value, which reaches only the views its entry names.
The stale index 5 moves into `hostile` for `selectedWallet` and
`selectedAddress`, as each field's one value still truthy after the
floor.
Each `hostileRestore` entry declares whether its boot lands on its view
or falls back to Home, and the test checks the one view on screen. A
value driven onto every restorable view lists only the views it falls
back on, so a view added later is driven by default.
The header and the README restate the remaining limits as built and say
which boots are held to where the popup lands.
Model: opus-5-5
The add wallet screen reported a missing, short or mismatched password in
the flash line at the top of the popup, and the private key export,
recovery phrase and delete wallet screens each wrote to a line of their own
above the password field. All four now use showError() and hideError() with
a fixed-height error line below the field, as the send confirmation and
approval screens do. On the add wallet screen the line sits beside the
Import button, which keeps its place at 360x600. The line clears when the
screen is shown again and when the password is tried again. Other add
wallet messages stay in the flash line.
Model: opus-5-5
The owner's ruling on #19 is
now in the README, under ExportPrivKey: copying the key leaves it on the
clipboard, because the clipboard is the user's, clearing it would go
against what they expect, and it could destroy something else they
copied since. ShowRecoveryPhrase copies the phrase the same way and
points back to it. Both describe the warning each password screen
shows on next, which does not mention the clipboard.
Model: opus-5-5
The one entry in ALLOWED_ERRORS in tests/e2e/firefox/run.js, Firefox
reporting a popup promise that settled after the page unloaded, lost its
cause when the site-connection buttons stopped sending with an unawaited
sendMessage before closing (#275). Left in place it would also hide the
same error from any other popup code that sends and then closes. The
entry goes, with the code that only printed and set aside tolerated
errors, and the README paragraph that described it.
Model: opus-5-5
script/lib/icons.js draws the mark from geometry read back out of the
committed PNGs, since the coordinates were never recorded, and writes each
PNG with node's own zlib. `make icons` runs it and leaves alone a file that
already holds the drawn image.
tests/icons.test.js requires each committed file to hold exactly that
image: the same IHDR and every pixel. The compressed bytes are not
compared, because node's bundled zlib does not compress as the stock zlib
that made the committed files did; the decision is recorded on the issue.
build.js copies the manifests from MANIFEST_SOURCES instead of naming the
two paths a second time.
Model: opus-5-5
The window.close override the issue named was removed with #275, so it
needs no entry. Every other place either suite changes or works around
the shipped extension is now listed: the popup loaded in a tab, Chrome's
wait before each site-connection request, its recording wrapper and click
listener, its clipboard grant, two layout tests that write into the page,
the forced leave during a decrypt, and Firefox's setting that forces the
site-connection prompt into a window. Firefox's tolerated error is
described as the leftover it is, with #487 to remove it; the phishing
blocklist is no longer listed as a failing fetch, and two stale figures
are corrected.
Model: opus-5-5
The recovery phrase and delete wallet screens take themselves off the
Back stack when left, so Settings, one of those screens, then the
settings gear leaves Settings under the Settings now showing. A reopened
popup restores the same stack, cut at the screen the gear left. Back
then showed Settings again and seemed to do nothing.
goBack() now skips any entry for the screen already showing before it
pops its target. Jest tests drive the gear and then Back for the
recovery phrase screen, once and twice over, for both delete screens,
and after a reopen.
Model: opus-5-5
With stderr closed, the wrapper's message write failed and set -e ended it
with status 2; with stderr a pipe whose reader had gone, the write killed it
with 141. The message is now written with SIGPIPE ignored and its failure
ignored, after the step has run. An interrupt while a step runs now exits with
130 once the step has ended, removing nothing: under bash, a step that caught
the interrupt and exited with a status, as check-censored does, used to get
dist/ removed. A failed check-censored --require-dist still removes dist/. The
header states both. Also the README bullet's missing period.
Model: opus-5-5
Leaving the delete wallet or lost-password screen drops its wallet
selection, but the settings gear had just pushed the screen onto the Back
stack, so Back from Settings showed a screen whose button could only
answer "No wallet selected for deletion." Each screen's leave handler now
also takes it off the top of the stack, as the private key export and
recovery phrase screens do since
#461. Back from Settings then
stays on Settings once, as it does for the recovery phrase screen.
Jest tests drive the gear and then Back on both screens, and the delete
screen's own Back.
Model: opus-5-5
yarn install exits 0 without touching node_modules whenever
node_modules/.yarn-integrity matches yarn.lock, so a package deleted from
node_modules stayed deleted while bootstrap printed "bootstrap complete".
After the install, bootstrap now asks node for the package.json of every
package listed in dependencies and devDependencies of package.json, and on
the first it cannot find it names the package and the fix:
rm -rf node_modules && make bootstrap. A package whose exports hides its
package.json (ethers, libsodium-wrappers-sumo) makes node throw
ERR_PACKAGE_PATH_NOT_EXPORTED, which it does only after finding the
package, so that error counts as found.
Model: opus-5-5
Leaving the private key export or recovery phrase screen drops the
selection it was showing, but the settings gear had just pushed the
screen onto the Back stack, so Back from Settings landed on a password
prompt that could only fail. Each screen's leave handler now also takes
it off the top of the stack, which is what a reopened popup already does
to these screens. Back from Settings goes to the address screen for the
export screen; for the recovery phrase screen, opened from Settings, it
stays on Settings once, as after a reopen.
Jest tests drive the gear and then Back, and each screen's own Back, for
both screens; leavePrivkeyScreen() in the e2e suite expects the address
screen.
Model: opus-5-5
Copies .dockerignore, .gitignore, .prettierignore, check.yml and
REPO_POLICIES.md from sneak/prompts at dd4027b, keeping the repo's own
entries (dist/, release/, yarn files) after the canonical content.
The Dockerfile gets separate lint and test phases; its last stage
depends on both, checks the git describe version and runs make build.
script/lint, test, check, cibuild and docker are the canonical models.
check-censored moves into the lint phase and test-verify-build into the
test phase. fmt and fmt-check fall back to the nvm-installed node. The
e2e image builds are uncached. Comments citing the old test caps or what
runs a script are updated.
Model: opus-5-5
@tailwindcss/node 4.2.1, loaded by the Tailwind CLI, calls Node's
deprecated module.register() as it starts, so every make build on a
current Node printed the DEP0205 warning. 4.3.1 is the first release that
calls module.registerHooks() instead wherever Node has it.
tailwindcss and @tailwindcss/cli are pinned at exactly 4.3.1, and
yarn.lock carries the new versions of the packages they pull in. The
compiled CSS computes to the same values: it drops the unused .start and
.end rules and writes calc(var(--spacing) * 1) as var(--spacing).
Model: opus-5-5
make dev passed --watch to a build.js that read no arguments, so it
built once and exited. build.js --watch now builds, then builds again
after every change to a file under src/, manifest/ or icons/, until
interrupted; any other argument fails. Each directory gets its own
watcher, because Node's recursive watch on Linux loses a file that an
editor saves by renaming a new copy over it. A watch build writes no
build receipt and cannot be verified; README.md and the Makefile say
so and point to make build.
Model: opus-5-5
No workflow set timeout-minutes, so a hung build or browser held the
shared runner until the server's own limit, hours later. check now stops
at 10 minutes, e2e-firefox at 15 and e2e-chrome at 20: each is over two
and a half times the job's slowest cold-cache run. README "In CI" records
the measured times and the caps.
Model: opus-5-5
The popup's markup no longer carries style attributes. The 42 in
index.html and in the HTML the view helpers build are now Tailwind
classes, each computing to the value it replaced, so style-src is 'self'
in both manifests, pinned in tests/manifest.test.js.
The address dot's 16 colours are written out as whole classes, because
Tailwind builds only the classes it finds in the source. The Settings
debug well is shown and hidden with the hidden class, since clearing an
inline display no longer uncovers it. Two tests that found the colour dot
by its inline style now find it by its class. Script that sets
element.style is unaffected.
Model: opus-5-5
sneak
clawbot