What’s new
What changed for you in each release, newest first.
0.1.12, 2026-11-11
- No change to what shotfleet does. The README is now a short page for visitors, and the technical reference moved to
docs/guide.md; the examples page has a fifth real run, Quickspan on Android in five languages, with what it does not cover.
0.1.11, 2026-11-11
- shotfleet is now free and open source under the MIT license. Every command and every language works, with no key, no account and no license server:
runcaptures every language andexportneeds nothing.shotfleet activate,deactivateandSHOTFLEET_KEYare gone. If you bought a key, thank you; versions up to 0.1.10 keep working with it. - The code is on GitHub, and you can support it through GitHub Sponsors.
checkno longer calls a screenshot that a language skips on purpose missing.- Android:
pt-BRis captured when the app only hasvalues-pt, output folders are named after the language you asked for (id,he,pt-BR), and tablet screenshots no longer show the taskbar or the Android 17 app handle: tablets use an API 33 to 35 image when one is installed (shotfleet warns if not) and the launcher is disabled for the run (hide_system_chrome = falseturns that off).
0.1.10, 2026-10-30
- Your own Android screenshot sizes: an entry of
[android] devicescan be a table with asize, adensityand anorientation(landscapestarts the emulator wider than tall), exported to the Play folder of thebasedevice it names. A changed size creates a new emulator. - For apps whose demo data is built around “today”:
[android] clock = "2026-10-27 10:00"sets the emulator’s date and the status bar time, and[android] root = truerunsadb rootbefore yourshellcommands.[ios] clocksets the status bar time. Both need an emulator image that allows root; on a Play Store image the run stops and says so. - A new flow step,
- waitForExpected, waits in every language for the texts of the next screenshot’s[expect]table, for screens that fill in late with text that differs per language. runFlowtakeswhen: { defined: <input> }, so a block of steps runs only in the languages that have a value for that input and is left out, and counted as skipped, in the others.runwarns, once per run, when a text selector such asWeekis the text of several unrelated strings, and suggests anid:orindex:selector. A failing[android] shellcommand says when root is the reason.initandrunsay when your APK is a Jetpack Compose app and none of its screens exposed a resource id, and name the fix (testTagsAsResourceId). New devicesfold,fold-bookandfold-tabletopcapture a foldable emulator in its unfolded, book and tabletop postures.- dismissIfVisible: ["Not now", "Skip"]taps those texts while they are on screen and does nothing otherwise.initfalls back to a non-debug Gradle variant (benchmarkRelease, staging, release) when there is no debug APK. The docs show how to skip onboarding, set a different app state per screenshot and screenshot the notification shade.- Fewer dead ends on a first run: a flow Maestro refuses to read is named instead of “flow never ran”, an empty
launchAppis written the way Maestro accepts, a failed language still gets a failure screenshot,run --devices phone,tabletnarrows the devices without editing the config, and a relativeapkor flow path is looked for next to the config file. - After the release, clock and root, your own sizes with landscape, waitForExpected and runFlow with when: defined were run on an Android emulator (2026-10-30). One thing came out of it: check calls a screenshot that a language skips on purpose missing; that is not fixed yet. The foldables, the Compose hint and the notification recipes were tested with fakes and unit tests only, not run on a device.
0.1.9, 2026-10-11
check,runandinitnow stop with exit code 2 on a Flutter, React Native, MAUI or Unity build that has no strings, and say what to pass:--stringsforcheck, astringsline inshotfleet.tomlfor the other two. These builds only hold the libraries’ own text, socheckon the APK of a React Native app then reported every language as untranslated andrunwent on and compared your flow with text that was never yours.initstill writes the config, with a commentedstringsline for you to fill in, and no longer lists the libraries’ languages as your app’s. It exits 2 until the line is set, unless it finds translation files in your project.
0.1.8, 2026-10-08
scrollUntilVisiblewith a plain text (element: "Settings") is now translated, like the object form. Before, the plain string stayed in English, so the step failed in every other language.- Before each
takeScreenshot,runwaits up to one second for the screen to stop moving, so a sliding sheet or a splash on its way out is less likely to end up in the picture.settle_msin the config changes the time, and0turns it off. An app that holds one splash image for several seconds still needs a wait for a text on its first real screen. - Ask for
zh-Hanson an Android app that shipszh-CNandrunnow useszh-CNand says so, instead of telling you the app has none of your languages.zh-Hantandzh-TWwork the same way, and so doptandpt-BR. When two codes would fit, it doesn’t guess: it names the ones the app has. - A text selector that is a regex, or that comes from a format string like
%@ on %@, can’t be translated.runnow says so by name before it starts and suggests anid:or a label the app translates. Before, such a flow passed in English and failed in every other language with nothing said. - When a language fails, the failure screenshot is kept in
out/failures/and named in the message, also after a failed retry. “iOS driver not ready in time” on a simulator’s first boot counts as a device failure now, so it is retried once. - Two causes that used to fail with no hint now get a name: a splash screen that never hides (on Android) and a second Maestro job using port 7001. I tested these messages with fakes, not yet on a device.
check --appon a Flutter build tells you to pass--strings 'lib/l10n/*.arb', and--stringstogether with--appnow works. On the parent folder of a multi-device run, a missing iPad set shows asmissingand the exit code agrees with the problem count.- You can add
key = "benchmarks_title"to an[expect.<screenshot>]table.checkthen needs the app’s own text for that string key, in the screenshot’s language, to be on screen, which catches a green run that captured the wrong screen. Nothing changes unless you add it. doctorsays which iOS runtimerunwill use, and why when a newer one is skipped. On Android with a Maestro older than 2.7.0,runsays thatcheckwill read the text by OCR and suggests Maestro 2.11.0.
0.1.7, 2026-10-08
checknow ends with a store coverage table: one row per language, one column per device class (App Store: iPhone 6.9", iPhone 6.5", iPad 13"; Google Play: phone, 7" tablet, 10" tablet), each cellok,missing,too manyorwrong size. It is also inindex.htmlandcheck.json, and it changes no exit code.check --compare <earlier folder>lists, per language and screen, what is new, removed or changed since an earlier shotfleet or fastlane folder. It is a report, and it doesn’t read any text.export --print-uploadprints thefastlane deliverorfastlane supplycommand that uploads what you exported. It only prints it: shotfleet still never uploads. The docs now show theasc screenshots uploadroute too.runskips iOS 27.0 simulators unless you ask for them: they froze at boot in some of my runs. With noruntimein the config it uses the newest other one you have installed and says which, and why, in one line. A runtime you set is always used, anddoctorwarns if 27.0 is the only one installed.- When the Mac is very busy (swap almost full, critical memory pressure, or a load far above the number of cores),
runsays so once, in plain words, and suggests fewer parallel devices. Listing, creating and deleting simulators no longer wait forever: after a time limit they say the Mac is probably too busy. run --retry-failedreruns only the languages that failed in the last run of that config. The ones that passed stay as they were. Andrunnow retries a language by itself once, only when the simulator or emulator failed or the screenshot came out blank (the app painted its first frame late), not when your flow or app failed, because that would fail the same way again; it tells you which languages it didn’t retry.- Smaller: export hooks get
SHOTFLEET_PLATFORMon iOS too; over MCP therunsummary says when the free tier cut the languages; the Linux build no longer warns about a missing release date; the guides have an Atom feed.
0.1.6, 2026-10-06
- New in 0.1.6, on iOS:
runcan run your own XCUITest screenshot tests (XCTAttachmentor fastlane’sSnapshotHelper) instead of a Maestro flow, on several simulators at once, one language each, and then checks what they saved. For that you need neither Maestro nor Java. I tried the idea by hand first, and the finished feature then passed its device test on iOS 26.4 simulators. The docs show how to set it up. - Over MCP,
runnow refuses a config with a[capture]build,[ios] simctlor[android] shellcommands unless you passhooks: true, the same opt-in as for[hooks].checkover MCP no longer takesocr_command; use--ocr-commandin your terminal. checkunderstands fastlane supply’s file names, like1_de-DE.png. Before, it said every language was missing every other language’s screenshots. On a real Google Play set in 15 languages, 15 false “missing” lines are gone.- A Russian screen with English job titles in it is no longer called Kazakh when the app’s own Russian strings match it. A Ukrainian screen in the Russian folder is still caught.
- When a language fails,
runnow says where it stopped: the step, the path to the failure screenshot, and up to 12 texts that were on that screen. - iPad in landscape: when Maestro taps at the portrait position and misses,
runsays so in one line and tells you what to do instead. - The “this Mac was short of memory” note only shows up when a failure really looks like that, not on every failed flow.
- From 0.1.6, a key can be activated 5 times (0.1.5 and earlier stop at 3): use it on up to 3 of your Macs, and the spare activations cover reinstalls. If you run out, email hello@shotfleet.com for a reset.
- In CI, a
SHOTFLEET_KEYthat is refused or can’t be checked now stopsrunwith exit 2, instead of quietly capturing 2 languages. With Gumroad unreachable, a key only counts on a machine where it passed a check in the last 30 days; on hosted runners, cache~/Library/Application Support/shotfleet/ci-checked.jsonbetween jobs. - “A year of updates” now means exactly that: your key covers every shotfleet version released within a year of your purchase. A version that comes out later runs as the free tier and says which versions your key covers; the versions you have keep working. If you bought in October 2026, every version released until October 2027 is covered.
0.1.5, 2026-10-06
Works with what you already use
- Already use fastlane snapshot or screengrab?
checknow works on those screenshots, no Maestro needed. It ignores the timestamp screengrab adds to every file name, so it no longer says every language is missing screens. The docs have a ready-made fastlane lane that stops before the upload whencheckfinds a problem. - Paparazzi, Roborazzi or Compose preview renders: copy each language into its own folder, then run
checkon them. The docs show a short loop that does the copying. I haven’t tried it on a real Paparazzi project yet. - Homebrew:
brew install noloman/tap/shotfleetinstalls the same signed build. Homebrew doesn’t install Maestro or Java, butshotfleet doctor --fixdoes. When a new version is out,doctortells you to runbrew upgrade.
More reliable on a busy Mac
- You never need to restart your Mac when macOS text recognition is busy.
checkwaits for it up to 8 minutes and then finishes anyway: it still checks sizes and missing screens, and it marks the screens it couldn’t read as “not verified”. Run it again later, or install Tesseract. Add--strictif unread screens should fail the check. - No system banners over your screenshots. On iOS, shotfleet tells each simulator right after boot that the “Ready for Apple Intelligence” banner was already shown. On Android it turns off pop-up notification banners during the run. I couldn’t make that iOS banner show up in a test yet, so I can’t prove this part works. As a backup,
checknow flags the banner at the top of a screenshot when it reads the pixels (Android, or screenshots you bring yourself).
Clearer results
checkpicks the right store by itself, from fastlane’s folder layout, your app build or the screenshot sizes. A plain folder of Android screenshots no longer gets judged by App Store sizes. Without--app, it also finds your app build in your project and tells you which one it used.- The HTML report works better with screen readers. Every screenshot has alt text with its language, screen and verdict, and problems are marked in words, not only with a red border.
- The memory warning now says how many devices your run really uses. Before, it could say 4 when shotfleet had already cut the run to 2.
0.1.4, 2026-10-06
- For AI tools:
shotfleet skill installnow puts the three shotfleet skills where Claude Code, Cursor, Codex, Gemini CLI and GitHub Copilot look for them. - The skills are also in a public repo, noloman/shotfleet-skills. In Claude Code you add them as a plugin, and the plugin also starts the shotfleet MCP server.
- The MCP server gives short answers: the counts, the first 100 problems and the path to the full report. Before, a big report could be 67,000 characters. Long runs send progress, and
runskips your config’s hooks unless you ask for them. - iOS: a simulator that stops answering now costs 60 seconds instead of 300, and the next run replaces it with a fresh one. Only shotfleet’s own simulators are ever deleted.
- When no simulator answers, the message gives you the exact command to shut down shotfleet’s own simulators. It no longer tells you to restart your Mac.
- Text recognition copes better with a busy Mac. It tries again when the macOS text reader quits without an answer, and it waits when macOS starts preparing the reader in the middle of a check.
- All nine demo clips are on See it work: from
doctorandinitto catching the wrong language, store rules, export and CI.
0.1.3, 2026-10-05
- On Android with Maestro 2.7.0 or newer, shotfleet reads each screenshot’s text straight from the app instead of from the pixels, so it’s exact in every script.
- The first time macOS prepares its text reader,
checkwaits for it, up to 6 minutes, instead of switching to the slower Tesseract reader after 90 seconds. If macOS still doesn’t answer, it tells you what to do. runwarns you when your Mac is short of memory or already swapping hard.capacity = "off"in the config turns the warning off.- Every release now stays at its own download address, like
https://shotfleet.com/download/0.1.3/shotfleet-macos-arm64.tar.gz, and those files never change. - A new page, See it work, with the first clips and my dated list of the real runs behind what this site says.