Five defects traced to one fact: src/background/index.js read and wrote the
module-level `state` singleton in src/shared/state.js, which the MV3 service
worker never populates and which answered an unpopulated read out of
DEFAULT_STATE in silence. Every previous fix added a loadState() before the
access, and that is what produced the fifth: a load detaches the objects an
in-flight handler is holding.
So the reachability goes rather than a sixth call site.
The background now has its own storage layer, src/background/state.js:
getState() is a detached, normalized per-call read, and updateState() is a
queued read-modify-write whose read is one storage round trip ahead of its
write. Nothing in the background holds an in-memory copy of the profile. The
write is the whole record, and updateState()'s header now names what that costs:
a popup write landing inside that one-round-trip window is reverted.
- Every handler takes one snapshot and answers from it, including the address
it names: activeAddressOf(s) replaced a second, later storage read that
could disagree with the first.
- wallet_switchEthereumChain applies applyChainSwitchFields() (split out of
chainSwitch.js, which keeps the singleton path for the popup) inside
updateState() instead of calling onChainSwitch() on the singleton.
- The remembered site decision is a read-modify-write, not a load-mutate-save
around a prompt the user takes seconds to answer.
- backgroundRefresh() refreshes a private copy of the wallets and applies the
balances that came back by address, so it never publishes an object other
in-flight work holds, and a wallet added or deleted during the round trip
survives its write.
- The transaction attempt takes its chain id and its endpoint from the same
snapshot. They used to come from different moments, so a chain switch
committed in between moved the endpoint under an artifact already verified
against the old chain.
getProvider(rpcUrl, networkId) now REQUIRES the network id and validates it
against networks.js. That closes the cold-worker wrong-chain send at its shape
rather than at one call site: the hint used to default to currentNetwork() off
the unpopulated singleton, so the endpoint was the user's chain and ethers
fixed chainId at 0x1, and the wallet's own verifySignedTx then refused every
non-mainnet dApp send. refreshBalances(), lookupTokenInfo(), scanForAddresses()
and resolveEnsName() carry the id through; balances.js no longer requires
state.js at all.
The prohibition is enforced mechanically, not by review, and it is enforced by
the bundler rather than by a guess at what the bundler does. The table of
modules an entry point's bundle may not contain lives in
script/lib/forbiddenBundleInputs.js — one copy, read by both layers that act on
it — and build.js's assertNoForbiddenInputs() fails the build when esbuild's
metafile reports src/shared/state.js as an input of a background bundle, naming
the import chain from the metafile's own graph. That is the resolution the
shipped bundle was built from, so no specifier syntax, no hop and no resolution
rule can slip past it; Dockerfile:42 runs make build, so it holds in CI.
A background entry point the table does not name fails the build as well. The
five defects were accidents, and so is adding a second worker entry point
without knowing that a table elsewhere needs a line for it: entry points under
src/background/ are prohibited by default and must be listed, rather than
protected only when someone remembers. That prefix is the build's only notion of
"the background", and eslint.config.js scopes the lint rule from the same
constant so the two layers cannot disagree about it.
Every way the table can rot is a failure rather than a quiet pass: a key no
bundled entry point matched, a listed module this build bundled nowhere, and an
entry that lists no modules. The second is what makes a rename of
src/shared/state.js loud instead of silently disarming the check, and it is
stronger than an existsSync() because it also fails when the module is still
there but has dropped out of every bundle. The third is refused at require time,
where the table is defined, because an empty list also empties the lint rule's
forbidden set — one character, and a plain require of the singleton in the
worker was green in make test, make lint and make build alike.
An entry is recorded as checked only once its bundle's inputs are in hand. It
used to be recorded before the output lookup that produces them, so an early
return past that point left both halves of the guarantee satisfied by a bundle
nothing had examined.
What the assertion does NOT cover is a COPY of the singleton at another path: it
is keyed by path, so a copy builds and lints clean. That is stated where the
table lives, with what the residual actually is — a copy carries the singleton's
own guard, so an unloaded read is a loud StateNotLoadedError and defects 1-3
cannot recur silently, but a copy carries loadState() too, so defects 4 and 5
(a stale read several awaits after a load, a load detaching objects an in-flight
handler is mutating) would recur over it in silence.
make check does not run make build, so the assertion is unit tested against
synthetic metafiles in tests/buildForbiddenInputs.test.js: build.js runs its
build() only as a program now and exports the checks. Executing a check in CI
is not testing it — without that file, inverting the condition leaves every
check in this repo green with the singleton back in the worker. Each vacuous
pass above has a case, including the output lookup that finds nothing, the empty
list, the unlisted second entry point, and recordBundledInputs() itself, which
every other case used to hand-seed.
A custom ESLint rule walks the CommonJS require graph from every src/background/
file and reports the same thing in the editor, before a full bundle. It reads
the same table, and it matches specifiers textually, so it is best-effort fast
feedback and not the guarantee — two earlier revisions of it shipped holes (a
template literal, a dynamic import(), a comment inside the call, a directory
resolved through package.json main). Those are covered now and pinned by
tests/backgroundStateLintRule.test.js. Two shapes it does not report are pinned
there as asserted non-reports, so the header's list of its bounds is measured
rather than claimed: a computed specifier (require("../shared/" + "state"),
which esbuild constant-folds into the bundle) and a symlink to the module
(esbuild reports the real path). Each is make lint exit 0 and make build exit 2.
Reading a persisted field of the singleton before any load now throws
StateNotLoadedError instead of serving DEFAULT_STATE.
Test stubs: chrome.storage.local is a serialization boundary, and eight files
stubbed it with an aliasing get, so the object a module held and the object
"storage" held were one object — an assertion could pass on a build that never
wrote anything. Every test that drives real persistence now goes through
tests/support/storageStub.js, which structured-clones in both directions.
closes#320
Verification compared the signed artifact against the dApp's request object.
For every field the dApp omitted -- normally nonce, gas limit and all the fee
fields, since the popup filled them in -- the number the user actually read on
screen was verified by nothing, and only absolute ceilings stood behind it.
The transaction is now populated in the background before the approval window
opens, and that populated object is both what the popup displays and what the
signed artifact is verified against. Every consequential field becomes an
equality comparison; the ceilings remain as a backstop. Population failing
means no approval and no window, and the error goes to the requesting page --
earlier than before, where the same estimate failed after the password had been
typed.
The account is pinned too: `from` is compared against the address named at
approval time rather than whichever address is active at signing, so switching
accounts mid-flow refuses instead of signing from an account the approval did
not name. The message-signing path had the same defect and gets the same fix.
Nonce selection moves earlier as a consequence; the concurrent-approval case
that follows from it is tracked at #271.
approvalVerify now compares every field of the signed artifact against the
approval, not a subset. Transaction types are allowlisted to 0/1/2 and any
field the module does not check is refused outright, so a future transaction
type cannot smuggle consequential fields past verification -- an EIP-7702
type-4 artifact that delegates the signer's own EOA while matching every
displayed field was accepted before this change. The serialized bytes handed
to broadcastTransaction are compared against the parsed artifact, so the
guarantee covers the bytes that actually go to the node.
Signing failures in the popup are retryable again. To make that safe, an
approval is claimed synchronously before the first await and every path that
resolves or removes one goes through a single chokepoint that refuses a claimed
approval. Without it, closing the approval window, switching the active address
or a late reject would report "User rejected the request." to the dApp while
the broadcast completed -- the user then redoes the transfer at a fresh nonce
and it sends twice.
Failure copy distinguishes the stage reached, so a user is never told to start
again from the site when the first attempt may already have reached the network.
