Repository navigation
Description
Hi, i am keraliss from walletscrutiny. i was trying to check the reproducability of this wallet , the android version. but there seems to be a version mismatch. the release version in this repo is 2.0.1, same as the app store version. but in play store, the version is 2.0.0, and i can't find corresponding release in the repo.
so is the 2.0.1 and 2.0.0 same? if yes, then why different names? and if no, then are you planning to release the updated version on the play store, or publish the code for version 2.0.0?
waiting to hear from you guys!
Product triage
| Field | Assessment |
|---|---|
| Severity | S3 (provisional until reproduced) |
| Implementation and verification confidence | 4/5; estimate, not a probability |
| Effort estimate | 1 points |
| Priority score | 8.0 |
Initial planning assessment; reproduction and acceptance evidence are required before implementation or closure.
Score uses availability × severity weight × confidence ÷ effort. A zero score does not mean low severity or that the issue is fixed.
Activity
ben-kaufman commentedon Feb 21, 2025
Hi, the 2.0.0 and 2.0.1 are the same version for Android, we just had a small issue firing the iOS release process due to which we needed to increment the version.
keraliss commentedon Feb 22, 2025
ah, thanks for the clarification, i tested the app with that assumption, and upon testing we found a bunch of diff between the official apk and the build version, here is the full diff -
Would be great if you can look into why so many diff, that would be great!
ben-kaufman commentedon Feb 24, 2025
Thanks for sharing this!
Could you please share the process you used for building the app and comparing the diffs?
We will make sure to check it internally.
keraliss commentedon Feb 25, 2025
https://gitlab.com/walletscrutiny/walletScrutinyCom/-/merge_requests/1002/diffs#a35857d69dc3c6a25118f9dfb38cb7fa1248b066
you can find the script and dockerfile here. basically i created a dockerfile that built the apk, grabbed the apk from play store and compared the diff.
ben-kaufman commentedon Feb 27, 2025
Thanks! We will make sure to inverstigate it.
ben-kaufman commentedon Apr 8, 2025
Hi @keraliss , thanks for raising the issue. The team was reviewing the differences, and it seems to be due to two reasons: first, the optimization techniques used by the Google Play Store, which modified the files, and second, there were some residual files in the codebase that are no longer referenced. We have worked to remove the latter from the codebase.
keraliss commentedon Apr 12, 2025
Sounds great! will test the new version and share my findings here!
keraliss commentedon Apr 12, 2025
hey, ran the test again with the new version, got this -
i can see the diff reduced a bit, but there's still quite a lot. would be great if you guys can look into it once more!
ben-kaufman commentedon Apr 15, 2025
Thanks @keraliss, sure, we will take another look.
15 remaining items
xrviv commentedon Aug 25, 2025
Latest verification by @keraliss
keraliss commentedon Aug 25, 2025
tried testing the latest version of keeper, got this -
Giszmo commentedon Feb 14, 2026
Hi, following up ... @keraliss recently tested v2.5.8 (Feb 2026) and it's still not reproducible. Same pattern: dex files, baseline.prof, AndroidManifest, and resources differ.
Has there been any progress on build reproducibility since our last report? Happy to help identify specific blockers if you share your CI/build config.
antuz123 commentedon Sep 28, 2026
Thank you @keraliss, @xrviv and @Giszmo for the repeated independent verification and the build scripts/diffs. We owe you a substantive response, and I am sorry the February follow-up was left unanswered.
We are taking the next investigation on the Keeper side: pinning the source and toolchain for a released artifact, reproducing the AAB/split-APK comparison, and examining the remaining differences using the evidence already supplied. Version alignment, a successful build and valid signatures do not establish reproducibility; we are not claiming this is resolved or attributing all differences to Google Play without comparison evidence.
Our next useful update should contain a reproducible build recipe and measured results, or the specific blocker we encounter. No additional test run is requested from you now, and we are not committing to a release/date before that work is validated. Thank you for continuing to hold us to this standard.
antuz123 commentedon Oct 7, 2026
@keraliss @Giszmo — we completed a measured v2.6.3 investigation. Thank you for the earlier diffs and for continuing to check this. We are not claiming reproducibility yet.
v2.6.3, commit872cbc4b16ddcb0a8f3b5b1d7ebcde28e06ed096. The direct GitHub APK isBitcoin_Keeper_v2.6.3.apk, SHA-256a20d934ebd80ece779d3c171c7906bb4aff010337989ec1b826cab55c50eba46, packageio.hexawallet.bitcoinkeeper, version 2.6.3 (626). Its release checksum, detached signature, and APK signature verify.:app:assembleProductionRelease. It used a placeholder environment containing onlyENVIRONMENT=PRODUCTIONand a debug signing key because the exact production environment is not public. Output APK SHA-256:cc4102251a11cbbfa104808adc4906a3a0e3b2efd9a14ab0785d66f5c9d5143b. Comparing uncompressed APK entries while keeping code/resources comparison-critical found 1,424 matching entries and 192 differences: manifest (1), baseline profile (1), Hermes bundle (1), DEX (5), native libraries (56), resources (128). This mismatch under placeholder inputs is not proof that the published binary came from different source.rn-nodeifyand to a Metro asset path affected by thenode_modulesdirectory layout. A separate bundle-only experiment with Node 25.9.0, mirrored directory layout, and a temporary Metro resolver change matched every published Hermes byte outside two 20-byte digest regions. That experiment is not an exact tagged build. Of the 139 native libraries, 83 match, 48 differ only in GNU build IDs, and 8 have executable/data differences (Realm and Screens across four ABIs); those eight still need investigation.The concrete blocker is that
.env.productionis ignored and unavailable in the public tag, whilereact-native-configembeds its values in native output. The historical release lane also requires fresh store-baseline state, and there is no pinned Docker/Podman/Nix third-party recipe. We need a safe, deterministic public production configuration, a pinned build including dependency layout and native inputs, and a rebuilt AAB versus complete Play split-set comparison before a WalletScrutiny verdict. This is now active in our backlog. We are leaving this issue open and are not asking you to repeat a test.antuz123 commentedon Oct 7, 2026
A concrete follow-up to the v2.6.3 measurements above:
We completed one Play-delivered device-profile set from the official generated-APK metadata: SDK 33, arm64-v8a, xxhdpi, English (base plus three configuration APKs). All four verify with the Play signing certificate. The six DEX files, Hermes bundle, and 35 arm64 libraries — 42/42 corresponding code entries — are byte-identical to the direct GitHub APK. The 844 ordinary resource files also match by content after accounting for Play's renamed paths. This strengthens the published-channel comparison; it still does not compare Play output to an independently rebuilt AAB.
I opened draft PR #7035 with the APK payload comparator, measured release evidence, a Gradle distribution checksum, and a change that stops the App Center script logging generated environment values. The PR deliberately does not claim a WalletScrutiny verdict or provide a build.sh while the exact production configuration remains unavailable publicly. We are keeping #6097 open. The next engineering dependency is a security-reviewed path to public, nonsecret production build inputs, then a pinned third-party APK/AAB rebuild and split comparison.
xrviv commentedon Oct 12, 2026
Hi @antuz123, thanks for the detailed v2.6.3 write-up and for PR #7035. Measured results like these are exactly what we need.
Quick update from our side: @keraliss is no longer with WalletScrutiny, so I'll be your contact for Keeper from now on.
On the
.env.productionblocker:react-native-configwrites every env value into the APK as a string resource, so the production config of v2.6.3 can already be read from the publishedBitcoin_Keeper_v2.6.3.apk(for example withaapt2 dump resources). Keeping the file out of git doesn't hide those values. Committing the nonsecret ones for each release tag, or a documented step to extract them, would unblock third-party rebuilds today.We'll rebuild tag
v2.6.3(872cbc4b) with values taken from the published APK and your pinned toolchain (Node 22.23.3, JDK 17, NDK 27.1,:app:assembleProductionRelease), compare against the GitHub APK first, and post the measured diff here. We won't post the values themselves.Thanks again for picking this up.
xrviv commentedon Oct 12, 2026
Update: we rebuilt v2.6.3 (
872cbc4b) as an App Bundle and compared it with the Google Play split set installed on a phone (arm64-v8a / xxhdpi / en). Result: not reproducible — 0 of 4 splits match.classes*.dexfiles andbaseline.prof. The 14 differing native libraries have identical code; only compiler identifiers differ..envwas restored from the app's ownBuildConfig(values not published); Node v25.9.0 and the build paths were mirrored from what the APK records.__keeperrelease2515_(e.g.__keeperrelease2515_node_modules_reactnavigation_elements_src_assets_backicon.png), ours do not. The Play AAB was built with a different asset-path layout than the public tree.Full report, script and recording: https://walletscrutiny.com/mobile/io.hexawallet.bitcoinkeeper/#verificationId=29016727eddaa66bfee7975f64eb8a54b1cbb83d560dd8a349ea2e2c0c220b86
Could you tell us the directory layout and Node version used for the 2.6.3 Play build? With those we can rerun and post the diff.
xrviv commentedon Oct 12, 2026
Follow-up on the 2.6.3 comparison above. We traced the main difference and have a few concrete requests.
What the release build did differently. In the build that produced both the Play AAB and the GitHub APK of 2.6.3,
mobile/node_modulesappears to have been a symlink to../../keeper-release-2.5.15/node_modules, with that directory added as a Metro watch folder in a localmetro.config.js(the committed one has none). Rebuilding tagv2.6.3with that layout reproduces the official__keeperrelease2515_…asset names and the/assets/../../keeper-release-2.5.15/…strings in the Hermes bundle. Both artifacts also embed the same macOS paths (/Users/devaccount/Workspace/keeper-channel-fix/mobile/…,/Users/devaccount/.gradle/…).Native libraries. 14
.sofiles differ only in their compiler identification (.comment) and GNU build-ID; their code is identical. The build-ID is a hash that includes debug information with build-machine paths, so it changes with the machine and directory.Requests that would make releases reproducible by anyone:
node_modulesand no uncommittedmetro.config.jschanges.-ffile-prefix-map=<checkout>=.for C/C++,--remap-path-prefixfor Rust) so build-machine paths and build-IDs do not depend on where the build runs..nvmrcorengines): 2.6.3 was built with Node v25.9.0, not the 22.23.3 used in your test..env.productionvalues per tag (they are already readable from the APK).We'll rerun 2.6.3 with the layout above and post the result here.
xrviv commentedon Oct 12, 2026
Update: we reran 2.6.3 against the same Play split set, this time recreating the release machine's layout (
node_modulessymlinked to../../keeper-release-2.5.15/node_modulesplus that directory as a Metro watch folder).base.apknow reproduces — the JavaScript bundle, all dex files,baseline.profand resources match; only Play's own additions (SourceStamp, Play manifest entries) differ.split_config.arm64_v8a.apk: 14 native libraries differ only in compiler identification and build-ID (code identical), from the build-machine paths noted above.META-INF/directory that exists only in the official config splits; we still need to confirm it holds only Play's signature files.So the source matches what Play ships, once the undocumented local layout is reproduced. Building releases from a clean checkout in a pinned container, with path mapping for native code, would let anyone reproduce them without that guesswork.