Skip to content

Releases

2.4.0, upload paths, iOS 13 floor, IPv6, debuggable and 64-bit ABI checks

Choose a tag to compare

@mjmirza mjmirza released this 15 Nov 20:08
d2156ba

This release closes gaps found by a second survey of comparable open source repositories, a line by line read of the App Store Connect release notes, and an internal audit of the guard. Eight adversarial review rounds attacked the changes before release, and every reproduced bug became a test. The new findings are high or medium except the debuggable Android build, which Play refuses at upload, so nothing that passed 2.3.0 is blocked unless it carries a debuggable release manifest. It is a minor version.

Added

  • The guard now scans uploads made through helm-asc (builds upload and --submit), the asc CLI, and the Codemagic app-store-connect publish command. Before, these reached App Store Connect without a scan.
  • APPLE-MIN-DEPLOYMENT-TARGET-IOS13. App Store Connect refuses uploads that target below iOS 13. The guard reads the project, xcconfig files (including [sdk=...] qualifiers) and the Podfile. Reported as high, because a dependency or test target below 13 does not block an upload on its own.
  • APPLE-2.5.5-IPV4-ONLY-NETWORKING. App Review runs on an IPv6 only network, so inet_addr, inet_aton and gethostbyname are flagged, including in the app's own C, C++ and Objective-C++ files. Vendored native code is skipped.
  • APPLE-5.1.3-HEALTH-DATA-ICLOUD. HealthKit plus CloudKit or iCloud key-value storage is reported for a human to confirm that no health data reaches iCloud.
  • ANDROID-DEBUGGABLE-RELEASE (critical). A release manifest with android:debuggable set to true. Debug-only manifests are ignored.
  • ANDROID-FGS-SPECIAL-USE-UNDECLARED. A specialUse foreground service without its own subtype property, counted per manifest.
  • ANDROID-64BIT-ABI-MISSING. An abiFilters line that ships armeabi-v7a without arm64-v8a, or x86 without x86_64.
  • APPLE-ODR-UPLOAD-BLOCK-2027. From early 2027 App Store Connect will no longer accept On-Demand Resources uploads from new apps or apps that have never used the feature. Tracked from the earliest possible date as a reminder, not as Apple's enforcement date.
  • docs/COMPETITIVE-GAP-ANALYSIS.md records the second survey, what was adopted, what was rejected and why, and the larger work left for later releases (built artifact inspection, runtime flow proof, a Data safety CSV generator).

Questions about each check are answered in Discussions #2321 to #2328.

2.3.0, deadline provenance and the November sweep

Choose a tag to compare

@mjmirza mjmirza released this 15 Nov 09:15
fa1c356

Version 2.3.0 ties every regulatory deadline to the page it came from, adds ten new deadlines read on their primary sources, resolves the September sweep's unverified dates, and publishes the playbook as a searchable website.

What is new

  1. Every deadline names its source and the day it was read. Before, a reader followed two hops from a deadline to a doc section to find where a date came from, and nothing could tell a fresh entry from a stale one. Now each entry in data/regulatory-deadlines.json carries source_url and verified_on, plus source_note when the page was read through a mirror, an archive, a secondary summary, or the EU Publications Office record. 92 sources were fetched and confirmed on 14 November 2026. The two Ofcom pages refuse scripted reads and keep the date they were last read in a browser.
  2. Stale sources can no longer hide. scripts/validate.py fails any deadline missing either field, rejects a source that is not a web address, an empty source note, or a read date in the future, and warns once a source is more than 180 days old. scripts/verify-citations.py already sweeps the new links.
  3. Ten new deadlines, each read on its primary source.
    • Apple. New age rating descriptors for Shared Location and Dating and Relationships (12 January 2027), and the expiry of the original Developer ID Sub-CA (1 February 2027), which affects apps signed outside the App Store.
    • Google Play. Payments authorization for developers selling into or out of Japan (1 December 2026), subscriptions with ongoing eligibility-based discounts (27 January 2027), the AppFunctions API policy (14 April 2027), and the Local Network and Manage Contacts permission declarations (25 August 2027).
    • South Korea. Mandatory PIPC privacy certification for handlers above a decree threshold (1 July 2027), and a new route to reuse lawfully collected personal information for AI development, with a prior risk assessment and PIPC resolution (9 March 2027, Act No. 21910).
    • Indonesia. The PP TUNAS child-protection self-assessment for every product, service, and feature (31 December 2026). Kemkomdigi says a product that misses it can be designated high-risk directly.
    • China. The CAC filing window for platforms with a very large number of minor users or a significant impact on minors, 20 working days from the 19 October 2026 notice, closing around 16 November 2026. That close date is counted by hand and labelled as derived.
  4. September's unverified dates are resolved. The Japan APPI amendment now carries its law number and penal start, the UK subscription regime cites its draft commencement regulations, Android developer verification cites developer.android.com, and the India DPDP rule 4 Consent Manager date (13 November 2026) is carried by the playbook and cites the Gazette copy MeitY hosts.
  5. A readable website. Each doc is now its own page at mjmirza.github.io/app-store-compliance, with a sitemap, search metadata, structured data, and an llms.txt file, so search engines and AI assistants can find and cite it.
  6. Landing page corrections. It used to say every third-party SDK needs a signed privacy manifest, that Stripe can never sell digital goods, that both stores need deletion fully inside the app, and that every AI feature triggers the EU AI Act. Each now states the actual rule and its scope.

Nothing that passed 2.2.0 in the guard is blocked now. The one schema change is for people who add deadlines. A new entry needs source_url and verified_on to validate.

How to update

  1. Pull the latest main.
  2. Run bash scripts/install.sh install.
  3. Restart your coding assistant once so it loads the new data.
  4. Run python3 scripts/deadline-checker.py to see what falls inside your next 90 days.

If you maintain a fork that adds its own deadlines, add source_url and verified_on to each one, then run python3 scripts/validate.py until it reports 0 errors.

How it was verified

  • Each new date was read on the government or platform page that sets it, not on a news summary. Where a page refused a scripted read, the entry says so in source_note.
  • scripts/validate.py reports 105 deadlines, 0 warnings, 0 errors. scripts/test-validate.py adds 9 provenance tests.
  • The full CI workflow passed locally (16 of 16 steps) and on GitHub for #2302, #2303, #2304, #2305, #2306, and #2307.

Still open

Korea's MyData dates and the live eSafety register for the app distribution code are tracked in #659 until they can be read on a primary source.

Full changelog. v2.2.0...v2.3.0

2.2.0, Fleet Sweep and the Ruling Register

Choose a tag to compare

@mjmirza mjmirza released this 05 Nov 19:39
0675e4e

Version 2.2.0 checks a whole fleet of apps in one run and tracks the release rules a code scan cannot see.

What is new

  1. Fleet Sweep (scripts/fleet-sweep.py). One run covers every app in a fleet file and returns one verdict, worst app wins. Exit codes are 0 ready, 3 ready with open actions, 1 not checked, 2 blocked. A partial scan, a declared platform the guard did not find, a timeout, or a report whose counts contradict its findings is NOT_CHECKED, never ready.
  2. The Waiver Ledger. A reviewed false positive can step aside for one exact occurrence (app, finding, severity, file), with a reason, evidence, a reviewer, and an expiry at most 180 days out. Expired, incomplete, duplicated and stale waivers are named and the finding comes back.
  3. The Ruling Register (data/ruling-register.json, scripts/ruling-register.py). Release requirements outside the source code, each with its legal standing, how it is proven, and a quoted source with the date it was read. A rule blocks only when quoted sources cover both its obligation and its start date. 14 first entries across India, the EU, the US, China, Indonesia, Vietnam and Australia. Entries whose source could not be confirmed are marked unverified, so they never block.
  4. A code-scanning export (--scanning-export, SARIF 2.1.0) for GitHub, with waived findings marked as suppressed with their evidence, plus a full JSON result (--json).
  5. A Coverage. line in the guard output, so a blocked run still says whether every app root was scanned.
  6. An eval scorecard (scripts/eval-scorecard.py). 32 gold scenarios check the verdict, the exit code, that the fleet and each app agree, and which rule ids must and must not appear. Any miss fails CI.

Nothing that passed 2.1.0 in the single-app guard is blocked now.

How to update

Run bash scripts/install.sh install from a fresh pull, then restart your coding assistant once. Start with docs/FLEET-SWEEP.md and docs/RULING-REGISTER.md. Templates are in templates/.

How it was verified

  • Two adversarial reviews by a second model reproduced 20 defects. All are fixed and kept as regression cases.
  • validate 117 patterns, 0 errors. check-docs 60 files, 0 problems.
  • fleet-sweep-test 42 of 42, ruling-register-test 74 of 74, eval scorecard 32 of 32 with 0 false passes.
  • CI green on #1009 and #1010.

Full detail is in the changelog.

2.1.0, the October 2026 App Review Guidelines revision

Choose a tag to compare

@mjmirza mjmirza released this 17 Oct 10:06

Apple revised the App Review Guidelines on 16 October 2026. This release follows that revision. It adds three patterns and no new guard rule, so it is a minor version.

Changed

  • Apple revised the App Review Guidelines on 16 October 2026, and the playbook follows the live text, checked on 17 October 2026 against the revision note and Apple's safety post.
    • 1.2 now names social media, user-generated content platforms, and AI content creation apps, lists nonconsensual intimate imagery, and warns of a Developer Code of Conduct violation for missing safeguards or violations left unfixed after notice. APPLE-1.2-UGC-24H-ACTION and BOTH-AI-GENERATED-CONTENT carry the new wording, and the UGC fix no longer promises a 24 hour window the guideline does not set.
    • Hidden, dormant, and undocumented features moved from 2.3.1 to the new 2.5.19. The 2.3.1 row in docs/APPLE.md and the future_functionality mapping now say so, and RN-OTA-UNDECLARED names the bait-and-switch risk.
    • 3.1.1(a) folds the Music Streaming Services Entitlement into the StoreKit External Purchase Link bullet and requires a permitted destination. APPLE-3.1.1-EXTERNAL-LINK-REGION-GATING and its row say so.

Added

  • Three patterns, each checked by hand because the evidence lives outside the project. APPLE-2.3.14-CREATIVE-ASSET-CONTENT (medium), APPLE-2.5.19-HIDDEN-OR-DORMANT-FEATURE (critical), and APPLE-5-TRAFFICKING-CHILD-EXPLOITATION (critical), with matching rows in docs/APPLE.md and two new lines in the pre-submission checklist.

v2.0.0

Choose a tag to compare

@mjmirza mjmirza released this 16 Oct 14:56

This release changes the licence and the release audit's exit codes, so it is a major version. Everything below landed since 1.0.0.

Thanks

  • Omar Albeik (@omaralbeik) found that the sanctions section still listed Syria as an embargoed territory after the OFAC program ended, and opened #967 with the correction. His fix landed in #1001, where he is a co-author. Developers shipping to Syria now get the rules that actually apply instead of a block that no longer exists.

Changed

  • Shell tools are matched against an exact name list, which adds runInTerminal and run_in_terminal. An MCP tool whose name merely contains terminal or shell is no longer scanned.
  • The default branch is now main. The CI trigger, the README badge and SECURITY.md follow it, and the bot branch ruleset excludes whichever branch is the default.
  • The bot PR gate. The workflow from #442 closed every PR whose branch looked bot-generated, which caught the owner's own Jules scheduled tasks. It is replaced by a gate that closes only pull requests from bot accounts not on the allowlist (Jules, Copilot, Dependabot, GitHub Actions), without deleting the branch. Owner-run bot PRs stay open, land on the project board as Backlog through the auto-add workflow, and are reviewed by hand. The eleven Jules PRs closed on 14 September 2026 had their branches deleted by the old workflow and cannot be reopened.

Added

  • A bare Xcode or Android app under a root framework app (Expo, React Native, Flutter, Capacitor), with its own Info.plist or AndroidManifest.xml and no root build file that names it, is scanned as its own app. It counts as an app only with an .xcodeproj that builds an application (or names no product type) or an Android application module, so a library or a local native module stays with the app that uses it, and it splits off only when the root has its own project for that platform, so a managed Expo root keeps its checks. A native host that embeds React Native, Expo, Flutter or Capacitor (Swift, Objective-C, Kotlin or Java code, a Podfile or a Gradle file naming it, Cordova plugins included, in Objective-C or Swift, an RCTAppDelegate, an Expo app delegate, import React with a React surface view, or an Android ReactActivity or ReactNativeHost) is brownfield and stays in the root scan, so its over-the-air updater and other framework findings still count. A native app that does split off, and the root it left, both keep every framework check the pooled tree showed (a Capacitor root's UIWebView check, or a Cordova config.xml inside the split app, for example), exactly as when they were scanned as one. The root and its split native apps are also scanned together as in 1.0.0, and any finding the separate scans did not list at the same severity is added once, so a critical finding is never dropped beside a high one with the same id, so evidence that spans both (a shared source file and the split app) still counts. A project in a folder named exactly demo, example, sample or playground (or the plural) stays in the root scan, so its findings still count (#656).
  • Install for every major AI coding tool. install.sh --tool codex|gemini|opencode|cursor|copilot|grok|qwen|kimi|droid|auggie|windsurf installs the guard, the skill and, where the tool supports it, the /app-store-audit command, and registers the hook in that tool's own config, merged with every other setting and backed up. --tool all installs into every tool it finds. Each install ends with Install verified. only after it feeds a submit command to the hook in that tool's own payload format and sees it blocked. OpenCode gets a plugin, because it has no hook config, written to both plugins/ (current OpenCode) and plugin/ (OpenCode before 1.0, which reads only that folder). Checked live on this machine. The real Codex CLI and the real Qwen Code 0.25.1 each stopped a submit command with the guard's findings. The real Gemini CLI 0.22.5 and OpenCode 0.14.1 loaded the hook and the plugin, and their model call could not run here (no Gemini licence on the account, no free OpenCode model available), so their blocking is proven by feeding each tool's own payload to the installed hook. 203 tests in scripts/install-harness-test.sh, in CI.
  • GitHub Copilot in every place it runs. --tool copilot covers Copilot CLI and Copilot in VS Code with one file that carries both hook formats. --tool copilot-repo --repo <app> writes the hook into the app's own .github/hooks for the Copilot cloud agent and JetBrains, with nothing written to your home folder.
  • Kimi Code gets one marked block appended to its config ($KIMI_CODE_HOME/config.toml when that is set), and the rest of that file, which can hold API keys, is never rewritten or printed. Uninstall restores the file exactly. A config that already defines hooks inline, as a table or as a quoted key, or with an escaped quoted key, or that would not stay valid TOML (on a fresh append and on a repair of its own block), is left unchanged and the block to add by hand is printed, because TOML cannot mix the two forms. Factory Droid keeps using settings.json when its hooks already live there, because Droid ignores them once hooks.json exists.
  • The guard reads the request format of Grok, Factory Droid, Augment Auggie, Windsurf, VS Code, Kiro and Goose, and finds the project folder from each tool's own field.
  • A version you can see. VERSION holds 2.0.0, install.sh --version prints the checkout's version, every install records it next to the skill, and doctor prints the installed one and says when the checkout is newer.
  • On Windows (Git Bash, MSYS, Cygwin) the installer registers the hook for both Bash and PowerShell, and doctor proves a PowerShell submit command is blocked.
  • Guard app root detection, two edges left open by #612. A React Native or Expo app with no app.json or app.config is now its own app when its package.json lists react-native or expo under dependencies and it owns an ios or android folder, so a member whose native folder holds no project file yet no longer pools into its sibling (#655). A dev or peer dependency never counts, and a native-module library with peerDependencies stays a library. A directory linked into the tree from outside it is followed one level when its target carries an app marker, unless the link is named like a vendored folder (vendor, build, node_modules and the other names the scan already skips), and scanned through its target as its own app (#657). A link back into the tree, to an ancestor, or to itself is skipped, and the target is read without following any further link, so a cycle cannot hang the scan. A linked ios or android platform folder is still not followed. How a nested bare Xcode or Android project under a root framework app is scanned is covered by the #656 entry above. A linked app is scanned beside the project, never in place of it, so a clean linked app cannot hide the project's own findings, and an empty project root beside a linked app is not reported as NOT CHECKED, unless a framework app under it still needs its native folders, which keeps NOT CHECKED. A linked Unity, MAUI, Tauri, Cordova or Capacitor app is followed too, and two links to one app scan it once. 13 new test cases, 302 total, green on bash 5 and macOS bash 3.2.
  • README first-run path. A Start here table routes four kinds of reader to one action each, with the time it takes. The two-command check comes first, followed by what each last line of the guard means. The install section lists everything it writes and the undo command before the paste block. New short sections cover troubleshooting and the five words a newcomer has to know. Driven by 40 persona walks (a second model as 24 readers, and two reviewers who ran every README command in a throwaway home folder).
  • scripts/check-docs.py. Fails CI when a Markdown file names a link, heading anchor, image, repo path, script or flag that does not exist. 11 tests prove it fails on each kind.
  • The guard ends every run with a verdict. CLEAR. when no critical risk was found, NOT CHECKED. with exit 1 when the folder holds no iOS or Android project (an empty folder and an Expo app with no native folders used to look like a clean pass), and a Next. line after BLOCKED.. As a hook it still never blocks a project it could not scan.
  • monitor.py --full. A live report over 25 items now opens with a count per track and shows the first 10 (3,503 lines became 194). --json is never cut.
  • The checklist opens with the two checks that stop most rejections and says how long a full pass takes. CONTRIBUTING explains how to fix a wrong citation.
  • README. A second approval proof, a developer's App Store Connect approval mail shared with thanks for the playbook, stored as assets/app-store-approval-thanks.png with the sender's name removed.
  • One command install for Claude Code. scripts/install.sh copies the guard, the skill (now with .citation-allowlist) and the /app-store-audit command, adds one PreToolUse hook entry to settings.json while keeping every other key and writing a backup, then runs the guard against a sample app that must be blocked. doctor checks an existing install, uninstall removes it, --dry-run writes nothing. The check runs the registered command string itself, so an entry on the wrong matcher, behind a comment, or with an unquoted path fails it. A settings.json that is read-only, not valid JSON, or repeats a key is never rewritten. Before this the install was a prose prompt with no settings example and no way to tell whether it worked. 62 tests, in CI.
  • Apple, September 2026. The EU App Tracking Transparency alternative prompt from iOS 27.2 (the only version in Germany, France, Italy, Poland, and Romania), multiseat subscription purchases enabled by default w...
Read more