harden: keep the test recovery phrase out of release bundles, match committed keys by content (closes #351)
check / check (push) Failing after 1s
e2e / e2e-chrome (push) Failing after 1s
e2e / e2e-firefox (push) Failing after 1s

The 12-word BIP-39 test phrase survived in every release bundle as dead
text: module.exports keeps DEBUG_MNEMONIC live even though wallet.js's only
use of it folds away in a release build, so it could not be tree-shaken.
Putting the value itself behind the __BUILD_DEBUG__ define makes esbuild fold
it to null, so no distributed bundle carries it. script/verify-build now
fails a release build if the phrase appears in any emitted file, so the fold
cannot silently regress; test-verify-build covers both the release failure
and the debug allowance.

tests/extensionId.test.js now scans the content of every tracked file for a
PEM private-key header instead of matching filename extensions alone, so a
key committed under an unexpected name is caught.

Model: opus-4-8
This commit is contained in:
2026-09-21 07:55:39 +00:00
parent 99292b9188
commit 84be04d6e9
6 changed files with 100 additions and 11 deletions
+27
View File
@@ -13,6 +13,12 @@
# fallback branch; the property only exists in the emitted output, so it has to
# be asserted against the emitted output.
#
# The DEBUG half also checks the phrase directly: a release build must not carry
# the test recovery phrase in any emitted file. The phrase is behind the
# __BUILD_DEBUG__ define in src/shared/constants.js and folds away in a release
# build, but the marker only proves DEBUG compiled off, not that the fold
# removed the string; the phrase grep is the assertion that it did.
#
# Which mode to expect is an ARGUMENT (--expect release|debug) and is never
# taken from this script's environment. It used to be read from
# AUTISTMASK_DEBUG here, which meant an operator with AUTISTMASK_DEBUG=1
@@ -67,6 +73,15 @@ TAB=' '
MARKER_ON="autistmask-build-debug=on"
MARKER_OFF="autistmask-build-debug=off"
# The 12-word BIP-39 test recovery phrase from src/shared/constants.js. It is
# behind the __BUILD_DEBUG__ define there, so a release build folds it out of
# every bundle; this is the assertion that it stayed out. The phrase is a
# publicly committed test value rather than a secret, but a BIP-39 phrase in a
# distributed wallet artifact is exactly the string a scanner or auditor has to
# stop and reason about, so a release build must not ship it. A debug build
# ships it on purpose, so this is checked only when release is expected.
TEST_MNEMONIC="cube evolve unfold result inch risk jealous skill hotel bulb night wreck"
RECEIPT_HEADER="autistmask-build-receipt v1"
# Set by the arguments.
@@ -283,6 +298,18 @@ check_entry() {
receipt records $ENTRY_HASH and the file on disk is $SHA. Something wrote
to dist/ after the build, so this artifact is not the one that was built."
# No emitted file of a release build may carry the test recovery phrase.
# Checked on every file, not only the audited bundles, so a copy that
# reached some other emitted file fails here too. A debug build ships the
# phrase deliberately, so this runs only when release was expected.
if [ "$EXPECT" = "$MARKER_OFF" ] && has_marker "$TEST_MNEMONIC" "$ENTRY_PATH"; then
fail "$ENTRY_PATH carries the BIP-39 test recovery phrase, which a
release build must fold out. The __BUILD_DEBUG__ define in build.js is what
drops it from src/shared/constants.js; check that DEBUG_MNEMONIC is still
behind that flag. A recovery phrase in a distributed bundle is exactly the
string an auditor or scanner has to stop on, so this is a hard failure."
fi
if [ "$ENTRY_FLAG" = A ]; then
read_marker "$ENTRY_PATH"
[ "$MARKER" = "$EXPECT" ] ||