Changelog

All notable changes to FleetML are recorded here. The format follows Keep a Changelog, and versions follow Semantic Versioning (before 1.0, minor versions may change APIs).

0.10.7 - 2026-10-29

What changed for readers

Fixes every finding of the developer review of 0.10.6: one regression of 0.10.6 in the iOS guide, two 0.10.6 docs fixes that stopped one step short, seven product bugs and the docs gaps each platform's reviewer hit. Upgrading from 0.10.6 needs nothing. Requests 0.10.6 accepted that 0.10.7 refuses (400 naming the reason): a deployment whose canary_stages, rolling_config or health_gate does not match its strategy (0.10.6 dropped them, so such a deployment went to every device at once), and a variant pushed into an ARCHIVED version. A version created as v2.0.0 is now stored as 2.0.0 (versions stored with a v before stay as they are).

Fixed

• Regression of 0.10.6 (docs): the iOS guide said a registered device that starts with another project's key is refused; it moves to that project, as the guide says elsewhere.
• Server: canary_stages without "strategy": "CANARY", rolling_config without "ROLLING", and a health_gate on an ALL_AT_ONCE deployment were dropped silently; they are refused with 400 naming the strategy they need. The dashboard sends a gate only for a staged rollout.
• Server, CLI and Python client: v2.0.0 and 2.0.0 became two versions of one model, and fleetml push --version v2.0.0 printed vv2.0.0; a leading v is dropped, so they are one version.
• Server: a push into an ARCHIVED version, which is final, uploaded the file first; it is refused before anything is uploaded.
• CLI and Python client: a push whose upload or check failed left the model and an empty DRAFT version behind; the push removes what it created and says so.
• CLI and Python client: an .onnx file pushed without --framework (target_runtime) became a TFLITE variant; the file's extension picks the runtime (.onnx, .mlmodel/.mlpackage, .gguf, TFLITE otherwise), and a --framework the extension contradicts is warned about.
• CLI: config set-key saved input that is not a key (UTF-16 without a byte order mark, control characters), and every later call said the server was unavailable; it refuses such input, and a saved key that cannot be sent is named as the problem.
• Dashboard: an ARCHIVED version offered Deploy and was listed in the new deployment's picker (the server refused it); neither any more.

Docs

• Incomplete fixes of 0.10.6, now whole: the "never shows model active" entry in a new terminal goes into the clone and sources ~/.fleetml.env before step 4; the Kubernetes second shell is one complete block that starts in the clone.
• Android: step 5 stops at a failed build or install, and Troubleshooting names INSTALL_FAILED_UPDATE_INCOMPATIBLE and a deployment that targets other labels; a second run of "Publish and deploy".
• iOS: 6a prints the model code when the model name is not known yet; the own-app step 3 says when no server runs; the first build's time.
• Linux agent: running the Connect block again after a later deploy moves the machines back to 1.0.0.
• Windows: the clean-up names the pulled images and the build cache; wsl without --exec also breaks the server line; after a port move, step 4's export FLEETML_SERVER line again.
• Helm: after the tour on Kubernetes alone, remove the ~/.fleetml.env it created.
• Docker Engine's port-conflict message; an exported COMPOSE_PROJECT_NAME is not saved by step 2.

0.10.6 - 2026-10-28

What changed for readers

A short, careful release: the four regressions 0.10.5 brought into the docs and the CLI's messages are fixed, five small product bugs from the review of 0.10.5 are fixed, and the README was checked line by line, every block run and every link and claim checked. Upgrading from 0.10.5 needs nothing. A request 0.10.5 accepted that 0.10.6 refuses: a deployment of an ARCHIVED version (400 naming the status), a version string that is not semver 2.0.0 (01.0.0, +1.0.0), and ?status= on the version list with an unknown status (400 with the allowed values).

Fixed

• Regression of 0.10.5 (docs): CONFIG.md said the version list filters by status, which it ignored; the version list now filters by status (any case) and refuses an unknown one with the allowed values.
• Regression of 0.10.5 (docs): the iOS guide's own-app step 3 kept a FLEETML_API_KEY exported for a stack Get started had removed, and found no CLI in a new terminal; the block now works in any terminal, and Get started's clean-up unsets the key.
• Regression of 0.10.5 (docs): the Helm chart's rollback sentence said the server stays Not Ready; it now says the old revision's pod restarts while the current one keeps serving.
• Regression of 0.10.5 (CLI message): the no-key hint gave a bash line, also in PowerShell; it now names fleetml config set-key and FLEETML_API_KEY.
• Server: an ARCHIVED version could be deployed; it is refused with 400 naming the status (DRAFT versions deploy as before: a new version is DRAFT).
• CLI: config set-key saved a UTF-16 key file or pipe (what PowerShell 5.1 writes with >) as garbage; it decodes UTF-16 with a byte order mark.
• CLI and Python client: a push with a quantization, runtime, version or --metadata the server would refuse created the model and an empty DRAFT version first; it is refused before anything is created.
• Server: versions follow semver 2.0.0: 1.0.0-rc.1 and 1.0.0+build are accepted, leading zeros (01.0.0) and a sign (+1.0.0) refused, and the message names the input.
• Server: a health gate's on_failure is accepted in any case, like strategy.

Docs

• Android: the launch lines use adb shell am start -n <id>/.MainActivity (the monkey form launches nothing on API 35 emulators); the In the background: finishing the sync through WorkManager line explained; step 5 run again, a restart's log, one record per FleetML app.
• iOS: The specified device was not found, this device is already registered, the own app's FleetML cannot register:, the masked key field.
• Linux agent: a deleted project's registration failed lines, the window while a new models_dir cannot be written, the Upgrade block, a models_dir that is a mount point.
• Windows: the stack's images go with docker image rm; run fleetml config set server again after moving the REST port.
• Python: what a version string may be; the status transitions listed.
• Helm: a second shell for the tour sources ~/.fleetml.env; fleetml-secrets.env is in .gitignore and removed at clean-up; the broken pipe form of a port-forward ending; StatefulSet names for another release name.
• README and every Get started: step 1 expects Docker Compose v2 or later (current Compose prints v5); the step-4 PATH line in Troubleshooting.

0.10.5 - 2026-10-28

What changed for readers

Fixes a regression of 0.10.4 in the Android SDK and every finding of the developer review of 0.10.4. An experiment can only be concluded with one of its two versions, and a deployment request with a key the server does not know is refused instead of ignored. Upgrading from 0.10.4 needs nothing; a request that 0.10.4 accepted with a misspelled key (stratgy, a canary stage's duration) now answers 400 naming the key.

Fixed

• Android SDK (regression of 0.10.4): a cold start in the foreground with a stored registration took the background path, logging In the background: finishing the sync through WorkManager and scheduling a job; it now starts the timers as before 0.10.4, and a start that really happens in the background still finishes its first sync.
• Android SDK: a model loaded from the cache at start logs Model <name> <version> is active, so Get started's step 5 passes when run again.
• CLI: config set-key from a pipe or a file drops a UTF-8 BOM and CR/LF (a PowerShell pipe saved an unusable key).
• Server: conclude accepted a winner that was neither arm; it must be the control or the treatment version, and a malformed winner_version_id says so instead of "required".
• Server: unknown keys in a deployment body or a canary stage were ignored; they are refused with 400 naming the key.
• iOS sample app: with -fleetmlPort and FLEETML_PORT both given, the config screen showed the argument's port while the SDK used the environment's; it shows the port in use.

Changed

• The CLI's missing-key error names ~/.fleetml.env; server unavailable names FLEETML_SERVER.
• An experiment's split_ratio: 0 is refused with the value given; an invalid semver names the input as given.
• The Linux agent's unit points at a drop-in for ReadWritePaths= instead of a line to uncomment; install.sh replaces the 0.10.4 unit as shipped.

Docs

• Android: the two-device clean-up names ANDROID_SERIAL's device and the line for the other.
• iOS: own-app step 3 pastes as written; 6c says when the screen shows iris and what "bundled iris (no assignment yet)" means; step 8 runs from the clone.
• Linux agent: the download_host_override block can run again and enables the service; Remove also takes the drop-ins, the .new unit and a moved models_dir; the read-only file system error quoted; docker pause timing and wording.
• Windows: config set server in the second "for every window" place, Docker Desktop's port message, the clean-up of the Path entry and the clone.
• Python: push(promote=True) of the same file again is 409; a call after close(); unknown list filters are ignored (CONFIG too).
• Helm: an own-Secret release upgrades without -f fleetml-secrets.yaml; a rollback to a 0.10.2 revision writes that revision's secrets first; helm lint does not check server.env ports; the CLI on PATH in a new terminal; the kubectl apply warning and the port-forward error lines.
• Backend: the tour's second shell goes into the clone; step 2 again recreates the server and dashboard; a program on all interfaces on macOS goes unnoticed too; a second stack's home folder can be anywhere.

0.10.4 - 2026-10-27

What changed for readers

An experiment now belongs to its project the way a deployment does, so one project's devices can no longer be given another project's model. The Linux agent's upgrade keeps a unit you edited, the Helm upgrade note is the first step of Upgrading, and a clean-up whose compose command fails no longer deletes the settings of a running stack. Upgrading from 0.10.3 needs nothing; an experiment request with split_ratio outside 1-99 or with a version of another project, which 0.10.3 accepted, is now refused.

Fixed

• Server: an experiment accepted model versions of another project (an operator of one project could run another project's model on its devices), answered 500 for an unknown project and stored a split_ratio outside 1-99 as 50; arms of another project are refused with 400, unknown projects and versions are 404, and split_ratio must be 1-99.
• Server: a deployment of an unknown version answers 404 instead of "not deployable"; experiments?project_id= that is not a UUID answers 400.
• Linux agent: install.sh replaced an edited unit on every run, dropping a ReadWritePaths= line, so models outside the state folder failed after an upgrade; an edited unit is kept and the release's goes next to it.
• Android SDK: an app sent to the background before its first screen registered and then did nothing until opened again; the first sync now finishes in the background.
• Get started's clean-up in every guide removes ~/.fleetml.env only after docker compose ... down succeeded.

Changed

• fleetml enrollment-key revoke <key ID>.
• The CLI's 401 names the server that refused; config set-key and config set save only what was set (not the default server).
• A STAGED model version can be archived.
• The Windows run's WSL job warns when the runner has no WSL 2 instead of passing silently.

Docs

• Android: the emulator runbook builds in a subshell; install lines pass -PfleetmlDebugPort; Wi-Fi after a second step 4 times out once, and how to finish it.
• iOS: the offline test's closed port through SIMCTL_CHILD_FLEETML_PORT; own-app step 2 waits for the server.
• Linux agent: mkdir -p in the Install block, models_dir on another disk through a systemd drop-in, sudo journalctl, revoking a key.
• Windows: fleetml config set server with the stack's port; the clean-up removes the saved key.
• Python: the README snippet runs as written, twice.
• Helm: the upgrade note first (it also finishes a failed upgrade), git fetch --tags, helm rollback, changing an own Secret, the dashboard port-forward in NOTES.txt and after an upgrade.
• Smaller lines in ADVANCED, QUICKSTART and CONFIG.

0.10.3 - 2026-10-27

What changed for readers

A first start whose port is taken now stops in seconds with a pointer to Troubleshooting instead of waiting three minutes. The Linux agent puts models on any file system, the Helm chart upgrades across chart versions without a manual step (once more for releases installed with chart 0.10.2 or older: see the Helm upgrade note below), and a misspelled value is refused instead of ignored, in the chart's values and in a deployment's health gate. Upgrading the compose stack from 0.10.2 needs nothing.

Helm upgrade note

Releases installed with chart 0.10.2 or older need one step before their first upgrade to 0.10.3: kubectl -n fleetml delete statefulset fleetml-postgres fleetml-nats fleetml-s3 --cascade=orphan (pods and data stay; the upgrade takes them over). Those charts put the chart version in the bundled StatefulSets' volume templates, which Kubernetes does not let an upgrade change. From 0.10.3 on the labels stay the same.

Fixed

• Get started step 2 skips its wait when docker compose up fails and says "Start failed: see Troubleshooting", and prints the project list and the dashboard line only when the server answers.
• Linux agent: a models_dir on another file system than the state folder never activated a model (invalid cross-device link on every heartbeat); the agent now copies the download into place, and reports a model it cannot put into models_dir once as MODEL_LOAD_FAILED. After a change of models_dir it downloads its models again instead of listing a path that does not exist.
• Helm: every upgrade to a new chart version failed on the bundled StatefulSets (spec.volumeClaimTemplates: Invalid value); their volume templates now carry only labels that do not change.
• Server: a deployment's health gate that sets only some keys keeps the defaults of the others ({"min_events": 30} allowed no error at all); an unknown key in health_gate or rolling_config is refused with 400 naming it instead of being ignored.
• Server: the all-zeros UUID as a project or version ID of a deployment or experiment is an unknown ID (404 or 422) instead of "are required"; a malformed experiment body says "invalid JSON body: ...".
• iOS sample app: a refused enrollment key shows "Not registered · key refused" instead of "Server unreachable".

Changed

• fleetml config set-key without the key reads it at a hidden prompt or from a pipe, so it stays out of the command line and the shell history.
• Helm: values.schema.json checks the values (helm lint, install, upgrade); server.env.REST_PORT and GRPC_PORT are refused, since the pod listens on fixed ports.
• Python client: a server URL without http:// or https:// raises a ValueError naming it.

Docs

• README step 7 is QUICKSTART's clean-up, which also stops the tour's simulator and Grafana; the tour and Windows lines list Get started's project.
• Android: the Wi-Fi Troubleshooting covers this device is already registered, a block puts the download address back to localhost, the phone keeps its device ID over Wi-Fi, the Emulator block starts the app.
• iOS: 6c's message in ADVANCED, a device keeps its project whatever the key once registered (by design), the offline test's real log lines.
• Linux agent: the models folder's owner, the override block restarts the agent, network is unreachable as a TLS-mismatch form, what -label does.
• Helm: the upgrade note, the dashboard port-forward after an upgrade, the values schema.
• Python: the server URL's form, the README snippet's project, empty lists for an unknown project; ROLLOUTS: partial health gates; the Grafana port.

0.10.2 - 2026-10-26

What changed for readers

On an iPhone you trust the developer once per phone: Get started's clean-up now leaves the sample on the phone, because iOS forgets the trust when the last app from your Mac is removed, and every later install from that Mac needs no tap. Every step that waits gives up with a pointer to Troubleshooting, and Troubleshooting covers the reason it gave up. Upgrading from 0.10.1 needs nothing; Android and iOS apps register once more after the update (with their token, so nothing to do).

Fixed

• Get started step 2 waited forever when the stack could not start (a port another program holds); it gives up after 3 minutes and points to the Troubleshooting entry, which shows the taken-port case and the fix.
• Android and iOS SDKs: an app built with another enrollment key or server registered only when its labels or SDK version changed, so it stayed in its old project; a new key, host or port now registers it again, and the server moves it to the new key's project.
• fleetml simulate: a device whose token the server refuses (after a JWT secret rotation, for example) registers again with the enrollment key and says so, where it retried the refused token forever without a word.
• Server: a deployment of a model version from another project is refused with 400 (deleting that other project then failed with 500), a deployment in a project that does not exist answers 404 instead of 500, and a deployment body that is not valid JSON says "invalid JSON body: ..." instead of naming the required fields.
• scripts/demo.sh and scripts/demo-agent.sh set their server's download address themselves (in a terminal that had sourced ~/.fleetml.env every upload went to that stack's port) and print a failed push instead of exiting without a word.
• iOS Get started step 6c succeeds when run again on a phone that already registered (it waits for the phone to check in, not for a new device), and it recognises the untrusted and expired-profile launch failures itself.
• Python client: devices.list("<project ID>") raises a TypeError that names project= instead of an AttributeError.

Changed

• iOS: trust the developer once per phone. Get started's clean-up no longer uninstalls the sample from the phone; a separate line does, for when you are done with FleetML.

Docs

• ADVANCED's port check prints the ports in effect; the tour's second-shell pickup puts the CLI on the PATH; the second-stack recipe keeps the Android SDK, adb's key, Gradle's cache and the keychain.
• Android: this device is already registered with the way out, and what pm clear costs.
• Helm: a Secret of your own step by step, port-forwards after an upgrade or uninstall, the dashboard's local port in "Try it".
• Linux agent: the Connect and verify block runs again, the heartbeat errors of a restarted server, the device ID in "Remove the agent".
• Python: unknown filter values and query parameters give an empty list or are ignored.

0.10.1 - 2026-10-26

What changed for readers

The phone steps of Get started now work as written on a phone seen for the first time, and every step that waits says so when it gives up. The iOS guide's iPhone step picks the phone (or the one you name), builds for it so Xcode adds it to the signing profile, and tells you each one-time tap before it is due; the Android guide's phone step says which device it took and stops after three minutes with a pointer to Troubleshooting. ADVANCED's "Android: a phone over Wi-Fi" is a complete path, and "Other ports" is right about the settings file and has a recipe for a second stack. Step 2 can run again from inside the clone. Upgrading from 0.10.0 needs nothing.

Fixed

• Android SDK: going to the background in the middle of a sync (a phone locking between download and validation) is no longer logged as Heartbeat failed and does not mark the device offline; the background job finishes the interrupted sync, or a first sync that had not happened yet (a phone locking right after it registers), at once rather than an hour later.
• Android sample app: a build with another -PfleetmlDebugHost or -PfleetmlDebugPort and the same enrollment key now connects to the new address; it kept the one it had saved, so a phone moved from USB to Wi-Fi went on dialling localhost.
• iOS guide: Get started's iPhone step built for any iPhone, so a phone not yet in the signing profile was refused at install (This provisioning profile cannot be installed on this device); it needed the simulator step first, took the first paired iPhone when there were two, waited for a second device that only the simulator gave, and printed an ERROR line from devicectl every time.
• Docs: the port check in ADVANCED and the iOS guide missed FLEETML_S3_PUBLIC_PORT; "Other ports" said an exported port moves a stack that has started once (the saved settings file wins); the Wi-Fi section used a $EK that Get started never sets; the Python tour's simulator line ignored a moved gRPC port; fleetml-prod-server-1 was hard-coded where the stack can have another name.

Changed

• Docs: every guide creates and reads the enrollment key as FLEETML_ENROLLMENT_KEY. The dashboard step puts the admin key on the clipboard in WSL and on Linux with xclip too, and prints the dashboard's address. The waits for a device print a line when they give up. Troubleshooting covers the Android phone, the iPhone, the simulator and a dashboard that does not open. The Linux agent guide gains the config lines in Connect and verify and a Remove the agent section; the Helm chart's Get started has a local port for the dashboard and stops only its own background jobs at clean-up.

0.10.0 - 2026-10-26

What changed for readers

Every guide now opens with Get started: a numbered list of copy-paste blocks with nothing to choose, from a fresh clone to one device with a model on it, the dashboard and a clean up, each block followed by the line it prints. The README's device is the Linux agent in a container on the same machine; the Android guide puts the sample app on a phone over USB, the iOS guide on the simulator and then an iPhone, the Python guide and the Helm chart have their own. Everything that is a choice (other ports, a phone over Wi-Fi, Android Studio on Windows, the Helm variants, upgrades, the settings file) moved, unchanged, to the new docs/ADVANCED.md, with a pointer where it was. Two CI jobs keep it that way: Docs golden path runs the Get started blocks of the README, QUICKSTART, LINUX_AGENT and PYTHON_CLIENT against each change, and Docs lint fails on a port number in a command, a key on a command line, a version that disagrees with this file, a malformed entry here, or a golden block that differs between two guides. Upgrading from 0.9.1 needs nothing for the server, the CLI or the SDKs; Python callers see the two client changes under Changed.

Changed

• Python client: TelemetryEvent.confidence_score is Optional[float], None for an event that reported no confidence (a MODEL_LOAD, for example), where it was 0.0; push(promote=True) raises when the server refuses a step (from ARCHIVED), where it returned the version unchanged.
• Docs: the guides open with Get started, and their options moved to docs/ADVANCED.md (above). Every command reads the stack's saved ports (${FLEETML_REST_PORT:-18080}, ${FLEETML_GRPC_PORT:-19090}, ${FLEETML_S3_PUBLIC_PORT:-19100}), so a stack started on other ports needs no edits; prose names the defaults in words.
• Docs: no command puts an enrollment key on a command line: Android's Gradle lines take ORG_GRADLE_PROJECT_fleetmlEnrollmentKey, the iOS sample app SIMCTL_CHILD_/DEVICECTL_CHILD_ variables, the Linux container -e FLEETML_ENROLLMENT_KEY and a machine's config read -rs and printf | tee.
• install.sh without systemd says the state folder "is ready" and prints the restart hint only when it replaced a binary.

Added

• fleetml-agent -download-host-override, the config's download_host_override as a flag; FLEETML_ENROLLMENT_KEY also replaces config.example.yaml's placeholder key, and -h names it.
• The Android sample app connects by itself on the first start of a build that carries an enrollment key, again whenever a build brings a new one, and takes its gRPC port from -PfleetmlDebugPort; the iOS sample app takes FLEETML_HOST, FLEETML_PORT and FLEETML_ENROLLMENT_KEY from its environment and logs line by line.
• fleetml push --promote prints what it promoted (v1.0.0 promoted to STAGED → ACTIVE).
• The prod compose file publishes Grafana on FLEETML_GRAFANA_PORT.
• scripts/docs-golden-path.sh and scripts/docs-lint.sh, and the CI jobs Docs golden path and Docs lint.
• The Helm chart's NOTES.txt names the reader's own port variables and prints the object store's port-forward only with the bundled SeaweedFS.

Fixed

The findings of the seventh developer review (review 7, on 0.9.1):

• Python client: push found an existing model or version only on the server's first page of 20 and then failed with 409 CONFLICT; it now pages 100 at a time, as the CLI does. versions.promote(status="active") skipped the STAGED step; the status is now taken in any case.
• fleetml simulate --help's example put the key on the command line; it uses FLEETML_ENROLLMENT_KEY now.
• iOS: step 2's network check and the never-downloads diagnosis named 19100 on a stack with another model files' port (the review's iOS 4); step 6 now says Ctrl-C on the --console line ends the app.
• Android: the adb reverse lines, the emulator section and From Windows named the default ports; the backup rules are in the skeleton's manifest and the Production checklist.
• Linux agent: the agent-side download_host_override lines and config.example.yaml named 19100; -retry-failed without systemd uses setpriv.
• QUICKSTART: the tour's simulate lines follow a moved gRPC port (--grpc); README: the port-80 override paragraph, which QUICKSTART had replaced, is gone, and Upgrading says what to do with a 0.9.0 override file.
• Helm chart README: its example ports collided with the quickstart's, and on macOS a port-forward takes localhost from Docker Desktop without an error; Get started uses 38080, 39090 and 39100, and ADVANCED says how to check other ports. Each restart past server.dependencyTimeout waits another full budget, as the README now says.
• The 0.9.1 entry's paging sentence, which did not parse.

0.9.1 - 2026-10-25

The fixes from the sixth developer review (seven reviewers on the released 0.9.0, each rating it 5): one Android crash, one Helm restart, and the rest of what they found. A client that sends valid requests needs no change. Two refused requests answer differently (under Changed), and DEPENDENCY_TIMEOUT now counts the whole start-up wait. The upgrade from 0.9.0 is the Upgrading block.

Changed

• REST: a limit outside 1 to 100 (1 to 1000 for GET /telemetry/events) or a negative offset answers 400 VALIDATION_ERROR naming the range, for example limit must be between 1 and 100, got 500, instead of being clamped without a word (GET /devices had no cap at all). A client that asked for more than 100 must page with offset, as the CLI and the dashboard already do (review 6; the Python client's push pages since 0.10.0).
• DEPENDENCY_TIMEOUT is the total time start-up waits for Postgres, the object store and NATS together, not the time for each; the Helm chart's 180s and the compose default 60s stay. Meanwhile /livez answers 200 and /readyz 503 with {"status":"starting","waiting_for":"<dependency>"} (review 6).
• Docs: every fleetml simulate line (the tour, the chart README, the line the Python tour prints, scripts/demo.sh) passes the enrollment key in FLEETML_ENROLLMENT_KEY instead of --enrollment-key (review 6).

Added

• Compose: FLEETML_REST_PORT, FLEETML_GRPC_PORT and FLEETML_DASHBOARD_PORT (defaults 18080, 19090, 80) move the published ports, next to FLEETML_S3_PUBLIC_PORT; the Run block saves all four in ~/.fleetml.env. Moving the dashboard off a taken port 80 is one line in that file instead of an override file.
• Android SDK: backup rules for the registration (@xml/fleetml_backup_rules, @xml/fleetml_data_extraction_rules) for an app to point at or copy; the SDK does not set them itself, which would break the build of every app with its own rules.

Fixed

• Android SDK: an app whose stored registration cannot be decrypted (restored from a backup, moved to a new phone, or copied from another install) no longer crashes at every launch with AEADBadTagException. The SDK logs Stored registration unreadable (...); deleted it, registering again with the enrollment key, deletes the file and registers again with the same device ID (review 6).
• Helm: a server that waits for its dependencies on a cold node is no longer restarted by the 200 s startup probe; /livez answers before the waits (review 6).
• CLI: fleetml push --promote of a further variant into an ACTIVE version succeeds (vX was already ACTIVE) instead of failing with cannot transition from ACTIVE to STAGED; it takes only the steps a version has not passed. push finds an existing model or version past the first 20, and project delete's summary counts past 1000 models.
• REST: an unknown version status ("LIVE") answers 400 VALIDATION_ERROR listing DRAFT, STAGED, ACTIVE and ARCHIVED instead of 409 INVALID_TRANSITION (review 6).
• CLI: fleetml telemetry events leaves the confidence blank for an event without one instead of printing 0.000 (review 6).
• Linux agent: no heartbeat failed or sending telemetry failed warning for a call its own shutdown cancelled (review 6).
• Docs: the device guides' S3_PUBLIC_ENDPOINT lines keep a moved FLEETML_S3_PUBLIC_PORT instead of writing 19100; the Upgrading block finds the server container through compose, so it works with COMPOSE_PROJECT_NAME (review 6).
• scripts/demo-agent.sh runs its clean-up on macOS's bash 3.2 when it fails early, and shows why compose did not start (review 6).
• Dashboard: the Drift chart's time axis shows seconds for a short history instead of repeating a minute; the login page names ~/.fleetml.env.
• CI: on windows-11-arm the Windows run's WSL row reads skipped (WSL 2 not available on this runner) instead of pass.
• CHANGELOG 0.9.0: the version label is on the release images; an image built by docker compose is labelled dev unless FLEETML_VERSION is set (the server's own version comes from the clone's tag either way).

0.9.0 - 2026-10-25

The fixes from the fifth developer review (seven reviewers on the released 0.8.0), and dependency updates. Two changes are visible to API clients: a few refused requests answer with other error codes, and three Python client arguments are keyword-only; both are under Changed with what to do. The upgrade from 0.8.0 is the Upgrading block, which now also keeps a 0.7.0 setup's S3_PUBLIC_ENDPOINT.

Changed

• REST: an experiment or a model version asked for a status change its status does not allow answers 409 INVALID_TRANSITION, as deployments always did, instead of 400 VALIDATION_ERROR (the message text is the same); POST /deployments/{id}/progress on a deployment that is not ROLLING_OUT answers 409 INVALID_TRANSITION instead of returning it unchanged; a limit or offset that is not an integer answers 400 VALIDATION_ERROR naming the parameter instead of being ignored; can_progress in a deployment's health is true only when progress would move it now, and reason names the time a stage still holds (min_duration, batch_interval). A client that branched on VALIDATION_ERROR for those status changes checks INVALID_TRANSITION (review 5).
• Python client: project, limit and offset are keyword-only on the list methods (deployments.list, devices.list after labels, experiments.list, projects.list, models.list and versions.list after their parent, telemetry.events). A call that passed them by position, fm.deployments.list(project_id), raised nothing and listed the whole server; write fm.deployments.list(project=project_id) (review 5).
• Dependencies: github.com/jackc/pgx/v5 5.11.0, lucide-react 1.61.0 in the dashboard, org.json 20260814 in the Android SDK, actions/setup-python 7 in CI. The Android toolchain (AGP, Kotlin 2, Gradle 9, protobuf 4, gRPC 1.85) and the libraries that need it (kotlinx-coroutines 1.11, mockk 1.14, robolectric 4.17) move in an SDK release of their own.

Added

• The Android SDK 0.9.0 is on Maven Central and the Python client 0.9.0 on PyPI.
• The server reports its version: "version" in /readyz and /livez, and in the start log line. A server built by docker compose in a clone takes it from the clone's tag; both images carry the org.opencontainers.image.version label (review 5).
• Android SDK: the background job's log line tells a refused request from an unreachable server: Background sync: server reached, request refused (UNAUTHENTICATED: ...) with the server's gRPC status and message, where it said server unreachable before (review 5).
• ANDROID_SDK.md: the command-line project has its manifest (AppCompat theme, launcher activity) and a MainActivity; built from the guide and run on an emulator (review 5).
• CI's "Android sample app on Windows" job checks the JDK is 17 and builds with a key and a dotted -PfleetmlDebugHost, quoted as the guide writes it, in PowerShell 7 and in Windows PowerShell 5.1 (review 5).
• DEPENDENCY_TIMEOUT (default 60s): how long the server waits at start-up for each of Postgres, the object store and NATS; the Helm chart sets 180s (server.dependencyTimeout), inside its 200 s startup probe (review 5).
• CLI: fleetml telemetry events has Time and Version columns (each version looked up once); fleetml simulate reads the enrollment key from FLEETML_ENROLLMENT_KEY when --enrollment-key is not given, which keeps it out of ps (review 5).

Fixed

• CLI: fleetml push --promote reports a refused promotion (uploaded ... but promoting it to STAGED failed: ..., non-zero exit) instead of printing ✓ Pushed with the version left at DRAFT; fleetml project delete --help says it asks for no confirmation and cannot be undone (review 5).
• Docs: fleetml project delete $P in QUICKSTART "Clean up" and the Android guide, for a server you keep; ANDROID_SDK "From Windows": x64 only, a new window after winget, a time-stamped project name so the block runs twice, logcat runs until Ctrl+C, Get-NetConnectionProfile; the deploy get counts before the app registers (review 5).
• IOS_SDK.md: step 1 lists what needs a hand at the phone once (Developer Mode, trusting the Mac, Connect via network, Auto-Lock Never, the first launch's two prompts), so the terminal steps can run unattended; step 2 has a Safari check of /readyz and the object store from the phone; step 5 quotes Xcode 26's CoreDeviceService was unable to locate a device ... next to Connection interrupted, and says a locked phone takes the install but not the launch; log collect takes the phone's name; step 9 says it needs Xcode and someone at the phone (review 5).
• Linux agent: the last telemetry flush on SIGTERM is bounded at 1.5 s (was 5 s), so a stop with the server unreachable takes under 2 s (systemctl stop: 1.64 s, 5.11 s before); install.sh prints exec setpriv --reuid=fleetml-agent --regid=fleetml-agent --init-groups ... for a supervisor, where su stayed as the agent's parent and SIGKILLed it 2 s after a SIGTERM (BusyBox keeps the su form, which works there). LINUX_AGENT.md: when the offline warning comes, the demo's server ports and its down, the enrollment key in docker inspect (review 5).
• Helm: a cold install no longer restarts the server once while the Postgres image is still being pulled (the 60 s dependency wait ran out first); a secrets.adminKey shorter than 32 characters is refused at install time instead of crash-looping the server; the README's install block names its local ports in one line, used everywhere, and says what existingSecret does not roll and how to remove the namespace; fleetml-secrets*.yaml is git- and helm-ignored (review 5).
• ANDROID_SDK.md "From Windows" uses JDK 17 (Temurin, through winget): Android Studio 2026.1 bundles Java 25, on which Gradle 8.7 stops with What went wrong: 25.0.2; the -P arguments are quoted, since PowerShell cuts an unquoted -PfleetmlDebugHost=<LAN IP> at the first dot (review 5).
• README: tensorflow-lite:2.17.0, as the guide and the sample app.
• Compose: FLEETML_S3_PUBLIC_PORT also moves S3_PUBLIC_ENDPOINT's default (localhost:<port>), and the Run block saves the port and the address together; before, a moved port left every upload and download URL on 19100 (review 5).
• QUICKSTART "Upgrading" keeps S3_PUBLIC_ENDPOINT for a ~/.fleetml.env written by 0.7.0 or earlier (which set it on the command line): the block copies the running server's value into the file first, so phones and agents keep receiving models after the upgrade (review 5).

Security

• pgx 5.11.0 includes 5.10's hardening of its decoders and SCRAM against a malicious or compromised database server.
• fleetml simulate reads its enrollment key from FLEETML_ENROLLMENT_KEY, which keeps it out of the process list; LINUX_AGENT.md says the container one-liner's key is visible to docker inspect and how to keep it in a read-only config.yaml instead.
• Helm: fleetml-secrets*.yaml is git- and helm-ignored, and a short adminKey is refused at install time.

0.8.0 - 2026-10-24

The fixes from the fourth developer review (seven reviewers, a real Android phone and iPhone among them), with its bug numbers, and dependency and toolchain updates. No breaking API changes; see Upgrading.

Added

• The Android SDK 0.8.0 is on Maven Central and the Python client 0.8.0 on PyPI (pip install fleetml).
• Python client: project= on devices.list, deployments.list and experiments.list; projects.delete, models.get, versions.promote (through STAGED from DRAFT), deployments.pause, resume, cancel, progress and health (DeploymentHealth), devices.get, telemetry.events (TelemetryEvent), experiments.get, list, pause, resume, conclude and cancel, and drift.status, each one REST call as the CLI makes it. The live test calls every one of them and deletes its projects (review 4, bug 8).
• CI builds the Android sample app with gradlew.bat in PowerShell on windows-latest (the "Android sample app on Windows" job).
• CI has a Shell scripts job: shellcheck on every shell script and actionlint on the workflows.
• iOS SDK: a unit test asserts FleetML.sdkVersion, as the Python client's test asserts its __version__.

Changed

• Deleting a project also deletes its devices; before, a project without telemetry was deleted and its devices stayed, enrolled in no project (with telemetry, the delete failed, see Fixed).
• The server accepts gRPC keepalive pings every 20 seconds or more, also on a connection without a call in progress (gRPC's default answers a client that pings more often than every 5 minutes with too_many_pings).
• scripts/e2e.sh and scripts/python-e2e.sh name their compose project (COMPOSE_PROJECT_NAME, default fleetml) and stop before creating a database unless that project publishes Postgres on the port the server uses, so a run meant for a second stack cannot touch the dev stack. scripts/wait-migrate.sh names the project the same way.
• Building from source needs Go 1.26 or later (golang.org/x/net 0.60, nats.go 1.54 and minio-go 7.4 require it); the server image builds with Go 1.26.
• Server dependencies: gin 1.12, nats.go 1.54, minio-go 7.4, pgx 5.9, golang-migrate 4.20, the Prometheus client 1.25 and gRPC 1.84.
• CI and the release workflow build with Go 1.26.10 (toolchain go1.26.10 in go.mod, which actions/setup-go reads); building from source still needs Go 1.26 or later.

Fixed

• fleetml project delete (DELETE /api/v1/projects/:id) works for a project whose devices sent telemetry; it answered 500. The project goes with its models, versions, deployments, experiments, enrollment keys and devices, and its versions' telemetry, drift results and feature histograms, in one transaction. Deleting a version with telemetry deletes that telemetry (migration 000019). The CLI prints what it removed (fourth developer review, bug 1).
• fleetml telemetry summary (and the dashboard's telemetry chart) take latency and confidence over INFERENCE events only; a MODEL_LOAD or ERROR event no longer counts as a 0 ms inference, and an event without a confidence no longer counts as 0. The event counts are unchanged (review 4, bug 11).
• The telemetry and drift routes for one model version answer 404 NOT_FOUND ("model version not found") for an unknown version instead of an empty result (review 4, bug 12).
• API error messages no longer carry internal function names: project not found instead of project.GetByID: not found, project already exists: slug "tour" instead of project.Create: already exists: slug "tour". The error codes are unchanged (review 4, bug 12).
• Linux agent: a connection that dies without a reset (a Wi-Fi roam, a NAT or router restart) is found within 40 seconds (gRPC keepalive: a ping after 30 s idle, 10 s to answer), every call to the server has a 30 s deadline, and reconnect attempts back off to at most 15 s, so the first heartbeat after the server is back succeeds. Before, the agent hung without a log line until TCP gave up, about 16 minutes. The first failed heartbeat is logged at WARN once, the recovery at INFO (server reachable again). A model download that receives nothing for 60 s is cut off and resumed on the next attempt (review 4, bug 3).
• Linux agent: SIGTERM stops it within 5 seconds while the server is unreachable; the last telemetry flush is bounded and the rest stays queued (review 4, bug 3).
• install.sh prints the supervisor command with both execs (exec su -s /bin/sh fleetml-agent -c 'exec /usr/local/bin/fleetml-agent ...'): without the inner one a supervisor's SIGTERM stopped only su's shell and left the agent running (review 4, bug 10). Run again on a box where the service runs, install.sh restarts it, so an upgrade takes effect (review 4, bug 12).
• Dashboard: the Overview tiles show the server's totals, and Devices, Deployments and Experiments page through the server (Previous and Next, 21–26 of 26), where they stopped at the first 20 records. The project, model and version pickers (Models, Telemetry, Drift, Alerts, Audit, new deployments and experiments) load every record. Deployments has a Project column (review 4, bug 4).
• Android sample app: the status reads online, assigned model failing and the model line names the assigned version and the reason while its download or validation fails, instead of online and no assignment yet. The SDK logs a download under 1 KB in bytes ((603 bytes)), not (0KB) (review 4, cosmetic).
• Helm: s3.publicUseSSL (the server's S3_PUBLIC_USE_SSL, which defaults to S3_USE_SSL) sets whether the download URLs are https, apart from s3.useSSL, the server's own connection to the store. The chart README's first block (s3.useSSL=true with the bundled SeaweedFS) never became ready, and false signed http://...:443 URLs (review 4, bug 2).
• Helm: SeaweedFS restarts when its S3 credentials change (it read them only at start, so helm upgrade with a new s3SecretKey crash-looped the server); the server has a startup probe that covers its 60-second wait for each dependency, where the liveness probe could restart it on a cold install (review 4, bug 9).
• Helm: a server.env name the chart sets itself (S3_BUCKET, say) is set once, to the server.env value, instead of twice; a jwtSecret or signingMasterKey shorter than 32 characters, or the two equal, is refused at install; NOTES.txt names the release's namespace (review 4, bug 12).
• android/gradlew.bat exists: the Android guide's Windows steps ran .\gradlew.bat, which was never in the repository. The wrapper scripts and jar are the ones Gradle 8.7 generates (the jar's sha256 is Gradle's published one), and .gitattributes keeps *.bat CRLF (review 4, bug 6).
• Android sample app: -PfleetmlDebugHost=<host> (or fleetml.debugHost in android/local.properties) sets the address the debug build may reach over plain HTTP and prefills it on the config screen; the debug network security config is generated from it. A phone over Wi-Fi no longer needs a source edit, and -PfleetmlDebugHost=localhost gives a phone over USB the right host instead of the emulator's 10.0.2.2.
• A failed fleetml push upload says upload failed: once, and fleetml telemetry events prints confidence with 3 decimals instead of float32 noise such as 0.9700000286102295 (review 4, cosmetic).
• The Windows run's quickstart script cloned into ~/fleetml after rm -rf fleetml; on a case-insensitive file system that is a checkout at ~/fleetML. It now clones into a directory it creates with mktemp and removes nothing.
• Every shell script passes shellcheck 0.9 and 0.11; docs/screenshots/demo/record.sh no longer leaves its temporary directory behind when it stops early.

Security

• Dependencies: google.golang.org/grpc 1.83.2 (GO-2026-4762, GO-2026-6061, GO-2026-6348, GO-2026-6443), golang.org/x/text 0.41.0 (GO-2026-5970, GO-2026-6629), golang.org/x/crypto 0.55.0 and github.com/klauspost/compress 1.18.7.
• Dependencies: golang.org/x/net 0.60.0 (GO-2026-6603, GO-2026-6610, GO-2026-6611, GO-2026-6612, GO-2026-6617) and github.com/jackc/pgx/v5 5.9.2 (GO-2026-5004).
• Release binaries are built with Go 1.26.10 (0.7.0's were built with Go 1.24.0); with it, govulncheck finds no reachable vulnerability in the server, CLI or agent, where Go 1.26.2 leaves 29 in the standard library (review 4, bug 5).

Docs

• Quickstart: "Next: a real device" before Clean up (Android, iOS, the Linux agent, Python); an "Upgrading" section for the stack (run as written from a 0.7.0 stack with data and devices: data, the admin key and the devices kept, migrations applied), the CLI, the SDKs, the agent and the chart; Clean up says what -v deletes and removes the env file, and no longer claims a Grafana volume; versions' DRAFT status and push --promote are explained once; a Kubernetes pointer; the tour's curl lines use $FLEETML_SERVER (review 4, docs gaps 3, 4 and 14).
• README: scripts/demo.sh leaves the dev stack running and how to stop it; links to Upgrading and the Linux agent. README, the quickstart, the guides and the Python README: Contact names GitHub Issues for bugs; the stale "pip install fleetml from 0.6.0" sentence is gone (review 4, docs gaps 6 and 13).
• The device guides save S3_PUBLIC_ENDPOINT in ~/.fleetml.env instead of setting it for one command, and say what it changes for USB, the emulator and the sample app (review 4, docs gap 1).
• Android guide: keep the phone unlocked with the app in front; Publish and deploy comes before On your phone; a new project from the command line (AGP 8.5.0, Kotlin 1.9.24, Gradle 8.7; built as written against the SDK on Maven Central); the first-run logcat matches the code (the background line comes when the app goes to the background); the stale-settings note and its error for a phone; lifecycle-runtime-ktx 2.8.7 with an AndroidX activity; TensorFlow Lite 2.17.0 throughout; the lock screen in the production checklist (review 4, docs gaps 2 and 9).
• iOS guide: the team ID from a terminal (the certificate's OU, not the ID find-identity prints); keep the phone unlocked for steps 6 to 8; the closed-port offline test on the phone; fleetml telemetry events --device as the telemetry check; log collect needs root, Console.app does not; Secrets.xcconfig written 0600; UIRequiresFullScreen in project.yml (the device build's orientation warning is gone); the trust error after a free-team rebuild; an unavailable wireless phone; a simulator of your own (review 4, docs gaps 2 and 8).
• Linux agent guide: apt-get install -y curl ca-certificates for images without curl; iris in the telemetry example; an Upgrade section (install.sh again keeps the config and state and restarts the agent; a 0.8.0 agent works with a 0.7.0 server); "Try it locally" names what it needs, the compose project it starts and the variables that move it; the archive's LICENSE and NOTICE (review 4, docs gap 10).
• Python client: the tour creates a project of its own name, prints the simulator line with a seed of its own, counts the project's devices (project=), and ends with a Clean up step that deletes the project; the page says where the admin key comes from, that the package is typed, that only server calls raise FleetMLError, and lists INVALID_TRANSITION, the FAILED deployment status, drift's baseline and every new method. The stale "pip install fleetml from 0.6.0" sentence is gone from the page and from the package README (review 4, Python docs gaps).
• Android guide, "From Windows": one linear path in PowerShell (Git for Windows, a clone, publish and deploy with ConvertFrom-Json so the enrollment key stays in one window, the three adb reverse lines, gradlew.bat with the key and -PfleetmlDebugHost=localhost, the logcat lines, the stale-key reset), then Wi-Fi with a firewall rule for private networks only and its removal line, then the emulator. The quickstart's "On Windows" links to it and names Docker Desktop's licence terms and the Docker Engine route the Windows run uses; "Get the CLI" names Smart App Control and $ProgressPreference; the Python client's page has a Windows line (review 4, Windows docs gaps).
• Quickstart and README "Run FleetML": the block writes the three secrets and S3_PUBLIC_ENDPOINT to ~/.fleetml.env (0600, only if the file does not exist yet) and sources it before anything starts, and it prints no secret (the admin key was echoed before the "save them" step, and §6 said echo $FLEETML_ADMIN_KEY). The observability step and the port-80 override source the file first, so a saved S3_PUBLIC_ENDPOINT survives the up -d (review 4, docs gaps 1 and 5).
• The prod compose file's errors for a missing JWT_SECRET or SIGNING_MASTER_KEY say to source ~/.fleetml.env and never to generate new values for an existing stack; they suggested openssl rand, which on an existing stack breaks publishing and every device token (review 4, bug 7).
• The Helm chart's README is rewritten and was run as written on kind: the published images (no build or kind load), the secrets generated once into a 0600 file with what may and may not change, the install with the bundled store and s3.publicEndpoint explained, a try-it block with fleetml simulate --grpc, Upgrading (helm upgrade --reuse-values, what restarts), Production (TLS in front, s3.publicUseSSL, TimescaleDB for an external Postgres, what the chart does not set, backups), every value, and Uninstall (review 4, docs gap 7).
• Linux agent: "Identity and offline behaviour" gives the measured bound (a 3-minute server pause: the warning after 32 s, the first heartbeat 42 s after the server was back) and the two log lines; the container one-liner keeps the state folder in a volume, so a recreated container is the same device; the supervisor and -retry-failed commands use exec.
• Quickstart: the arm64 Windows binary is now run on windows-11-arm by the Windows run workflow, next to the amd64 one on windows-latest.
• RELEASING.md describes the public repository: its code security, Actions and package settings, the branch rule on main, and both version tests among the files a release bumps.
• A Contact section (hello@fleetml.dev, security@fleetml.dev, GitHub Discussions) ends the README, the guides and the Python client's README, and SUPPORT.md says where questions and bugs go.

0.7.0 - 2026-10-23

Added

• The Android SDK 0.7.0 is on Maven Central and the Python client 0.7.0 on PyPI (pip install fleetml).
• scripts/agent-install-test.sh has two more cases: an image with the systemd package but without systemd running, and an Alpine (BusyBox) container without bash or useradd. CI runs the Android sample app's unit tests next to the SDK's.

Changed

• Linux agent: install.sh is POSIX sh and creates its user with BusyBox's addgroup and adduser where useradd is missing, so it runs on BusyBox images. It decides on the service unit by whether systemd is running (/run/systemd/system), not by whether systemctl is installed; without a running systemd it prints No running systemd on this machine: ....
• Linux agent: the example config's post_activate touches /var/lib/fleetml-agent/model-switched, which the shipped unit can write, instead of a file under /run, which it cannot.

Fixed

• Android SDK: FleetML.syncNow() called at start-up (as in the guide's snippet) ran next to the first automatic heartbeat, so the first run downloaded, validated and activated the model twice, onModelReady fired twice and an offline start logged Heartbeat failed twice; two installs of one model at once could also break each other's download and leave the device without the model for an hour. A sync while a heartbeat is in flight now joins it, and the ModelManager skips a variant that is already being installed, so each version is downloaded and activated once and onModelReady fires once.
• iOS SDK: an app that lists com.fleetml.sdk.refresh in BGTaskSchedulerPermittedIdentifiers without calling FleetML.registerBackgroundTask() crashed the first time it went to the background (No launch handler registered for task with identifier); the SDK now skips the request and logs an info line naming the missing call.
• iOS SDK: no more Swift 6 warnings from FleetML.swift ('bgTask' mutated after capture by sendable closure and the main-actor shared access) in apps that build it.
• Linux agent: install.sh on an image with the systemd package but without systemd running copied the unit and exited 1 at systemctl daemon-reload.
• Android sample app: its debug network security config has a line for the Mac's LAN address (mac-lan-ip.invalid to replace), without which a phone over Wi-Fi registered and then failed every download with CLEARTEXT communication ... not permitted.

Docs

• The Android, iOS and Linux agent guides lead with the real device and the real deployment: a production-shaped configuration, the phone (USB, then Wi-Fi) or the server on another machine, publish and deploy with the key an app or edge developer is given, offline and update tests, and a production checklist; the emulator, the simulator and the local try-out follow as shorter variants, then Windows, the sample apps and the contributor sections.
• Android: the host, port and TLS from Gradle properties, with debug builds on localhost, 10.0.2.2 or the LAN address; AGP 8's buildConfig switch; the API's signatures; which key does what (checked with project admin and operator keys); the offline lines and reconnect times as measured; the README snippet gains the TFLite line.
• iOS: the app's settings from an .xcconfig through Info.plist, an AppModel that prints its callbacks, a complete project.yml for a phone, devicectl over USB and Wi-Fi, log collect, the -1009 and -1004 cases, the package: identity, from: "0.7.0" throughout.
• Linux agent: "Server on another machine" as its own section with a sed line for download_host_override; a docker run --init example and the entrypoint exec line; -o json | jq -r .key; the validator by hand with su too; removing the rejected copy; reloading the inference process with a path unit, and why sudo cannot work under the unit's NoNewPrivileges=true.
• Quickstart: section 3 always shows the canary cohort, D2, EXP and one SIM line are saved for a second shell; the CLI installs without sudo; "On Windows" opens with how the two shells and two clones fit together, and adds the PowerShell version check, docker compose version, .wslconfig and Unblock-File.

0.6.0 - 2026-10-21

Added

• The Android SDK 0.6.0 is on Maven Central and the Python client 0.6.0 on PyPI (pip install fleetml), the first version of each published since 0.4.0 and the first on PyPI.
• GET /api/v1/deployments/{id} (and fleetml deploy get) has devices: {targeted, loaded}: the project's devices the deployment targets and how many report its version loaded, since COMPLETE only means assigned.
• GET /api/v1/devices/{id} (and fleetml device get) has loaded_models, what the device's last heartbeat reported, and last_model_event, its newest model load, load failure or signature failure with the error.
• The Linux agent reports a failed model download as MODEL_LOAD_FAILED with the error, once per version, so the server shows it.
• fleetml deploy list --project, and GET /api/v1/deployments takes project_id (400 for a value that is not a UUID).
• scripts/agent-install-test.sh runs the agent's install.sh in Ubuntu containers with and without systemd.
• Linux agent: download_host_override (host:port) sends every model download, full or delta, to another address than the server's S3_PUBLIC_ENDPOINT, keeping the signed URL's path, query and Host header, for an edge team that cannot restart the server.
• Linux agent: -retry-failed retries the versions that failed validation at once instead of after the hour's wait, and the rejected file is kept in <state_dir>/rejected/<model>/<version>-<variant>/model.bin (the last one per model) for running the validator by hand.
• Linux agent: POST /v1/events answers with rejected_reasons, one per rejected event (event 0: unknown model "foo" (active models: iris)), and logs the first one.
• Android SDK: logs Model <name> <version> passed validation, as the iOS SDK does.

Changed

• The Linux agent's GET /v1/models answers with model, version, version_id and path, the same snake_case as POST /v1/events, instead of Name, Version, VersionID and Path, and with [] instead of null when no model is active.
• The iOS SDK logs each download (model, version, bytes) and validation result at info level and every heartbeat at debug level.

Fixed

• Linux agent: a fully downloaded model was readable only by the fleetml-agent user (mode 0600) while a model built from a delta was 0644; every active model is now 0644 in 0755 folders whatever the umask, and models written by an earlier agent are fixed at startup.
• Linux agent: install.sh exited 127 at systemctl on a machine without systemd after installing everything, and nothing created /var/lib/fleetml-agent there; it now skips the unit, creates the folder for the fleetml-agent user and prints how to run the agent.
• A deployment's device lists (rolling batch sizes, max_unavailable, the health gate's cohort) counted devices of other projects with matching labels; they stay in the deployment's project now.
• iOS: every launch logged Heartbeat failed ... GRPCStatus error 1 and reported connectionChanged(false) before true, because the heartbeat loop was started twice and the second start cancelled the first heartbeat; it now starts once, and a cancelled heartbeat is not a failure.
• iOS: when the device token cannot be stored in the Keychain (an app built without code signing, on the simulator), the SDK now logs a warning instead of silently registering again at every launch.
• fleetml push showed a model under a kilobyte as 0KB; small sizes are now in bytes.
• Android SDK: a failed heartbeat, token refresh, registration or telemetry flush logged a gRPC stack trace of about 30 lines at warning level; it now logs one line with the status and cause, and the trace at debug level.
• iOS SDK: an app without the SDK's background task logged Background refresh not scheduled ... BGTaskSchedulerErrorDomain error 3 as an error at every launch; it is now an info line that says how to enable the task, and only a permitted task that fails to schedule is an error.

Docs

• Quickstart: save the three secrets right after starting the stack (the admin key is honoured on the first start only, and compose needs the two secrets for every command), the port-80 override, how the canary cohort is chosen, what experiment conclude --winner does to devices, what COMPLETE means, bounded waiting loops; On Windows: administrator PowerShell, restart and firmware virtualization for wsl --install, Docker Desktop running, the tour only in Ubuntu, the clone's models under \\wsl.localhost, the unsigned exe, the admin key in PowerShell.
• Android: the com.fleetml.sdk package and imports, 10.0.2.2 in the init snippet, an ActiveModel table, syncNow() as a suspend function, an own-app CLI block, ANDROID_SERIAL with a phone attached, and "From Windows".
• iOS: full delegate signatures, import CoreML, log stream --level info, and a "Your own app" checklist.
• Linux agent: S3_PUBLIC_ENDPOINT as a prerequisite before the connect steps, a self-contained publish and deploy block, "Without systemd", file modes, the delta and restart log lines, and the agent in WSL.
• The next heartbeat is 60 seconds with the quickstart's server; the SDKs' and agent's 30 s applies only until the server answers.
• docs/ROLLOUTS.md "Did the model arrive?" describes the new fields.
• Second developer review: Android's emulator port forward for model files next to the init snippet, the enrollment key from a Gradle property, the first-run logcat lines, retrying without waiting for a heartbeat, the permissions the SDK brings, airplane mode for an own app, PowerShell lines in "From Windows"; iOS's key from the Info.plist, an own-app runbook with an XcodeGen project, connectionFailed being registration-only; the Linux agent's version lookup by number, sed instead of sudoedit, the agent-side download_host_override, the rejected file and -retry-failed, the server-side telemetry check, containers' PID 1; the tour's variables saved to ~/.fleetml.env with a resume from another shell, every printed line and the real waits; Windows: the shell on each block, the admin key from ~/.fleetml.env in one line, fleetml config set-key, reserved port ranges, docker-users, and what the Windows run does and does not verify; loaded turns one heartbeat after the device has the model and is counted each time you ask; RELEASING: tarballs built by hand on a Mac.
• Android "From a clone": the two Gradle plugins and the sdkVersion property the SDK module needs from the app's build, without which Gradle stops with Plugin [id: 'com.google.protobuf'] was not found.

0.5.0 - 2026-10-19

Added

• A Python client, pip install fleetml (Python 3.9 or later, httpx only): projects, models, versions, push, enrollment keys, deployments (with wait_for_deployment), devices, telemetry, experiments and drift, with the same words as the CLI and a FleetMLError for every refused request; typed (py.typed). CI runs its lint, type and unit checks on Linux, macOS and Windows and a live test of the tour's sections 1 to 3; the hand-run Python publish workflow publishes it to PyPI with Trusted Publishing. See docs/PYTHON_CLIENT.md.
• Release binaries: every vX.Y.Z tag gets the CLI for Linux, macOS and Windows (amd64 and arm64; .tar.gz, and .zip with fleetml.exe for Windows) and the Linux agent with its installer, plus SHA256SUMS, attached to its GitHub release by the new Release workflow, which can also back-fill an existing tag. make release-binaries VERSION=X.Y.Z builds the same files into dist/.
• A "CLI on Windows" CI job builds fleetml.exe, runs it, and runs the CLI's unit tests and a smoke test of the quickstart's CLI calls (project create, push, deploy create and start, device list) against a fake of the REST API; scripts/e2e.sh runs the same smoke test against a real server.
• A hand-run "Windows run" workflow on GitHub's Windows runners (windows-latest, windows-11-arm): the "Get the CLI" PowerShell block with a release's assets in Windows PowerShell 5.1, the released fleetml.exe against the CLI's fake of the REST API (clitest-server), and the quickstart with the tour's sections 1 to 3 in WSL 2, reached from Windows at localhost.
• fleetml enrollment-key create --project <id> [--name <name>] prints a new enrollment key once (-o json gives it as .key), and fleetml enrollment-key list --project <id> lists a project's keys without them; every device path needs one, and it took a curl call before.
• CLI flags the tour needed: deploy create --canary-stages and --rolling-config, experiment create --target and --min-sample-size, push --metadata for a new version's metadata (such as the requirements override a Core ML FP32 model needs on an iPhone), and device list --project.
• Container images for the Helm chart: the Release workflow's new images job builds the server and dashboard for linux/amd64 and linux/arm64 and pushes ghcr.io/fleetml/fleetml-server and ghcr.io/fleetml/fleetml-dashboard tagged with the version and latest.
• iOS SDK: FleetMLDelegate gains fleetML(connectionFailed:) (with an empty default), called with the error on the first registration failure after start and whenever the error changes, and the SDK's log messages are no longer redacted as <private> in Console.app.

Changed

• fleetml version and fleetml-agent -version report the version set at build time from the release tag (-ldflags -X), and dev for a build from source outside a clean checkout of a tag; the release commit no longer bumps them by hand.
• fleetml push draws its progress bar only when stderr is a terminal.

Fixed

• The Postgres health check in both compose files and the Helm chart now connects over TCP (pg_isready -h 127.0.0.1). The socket check reported ready during the image's first-start initialisation, about a second before the real server listened, so migrate could start too early and exit with "connection refused", and docker compose up then failed with "service migrate didn't complete successfully". Seen on GitHub's Windows runner under WSL 2; possible on any first start.
• scripts/demo.sh prints the Docker Compose output when the dev stack fails to start, instead of exiting without a message; scripts/android-e2e.sh and scripts/signing-e2e.sh use sha256sum where shasum is missing.
• Shell scripts and the Makefile keep LF line endings in Windows checkouts (.gitattributes).
• deploy/agent/config.example.yaml set tls: true, so an agent configured as the Linux agent guide said failed against the quickstart's plain-gRPC server with "transport: authentication handshake failed"; it is now false, the documented default, with a comment on when to set it, and install.sh names tls among the settings to edit.
• A push whose upload failed (an unreachable S3_PUBLIC_ENDPOINT, say) can be run again: the server replaces a variant that was never verified instead of refusing the retry with "variant.Create: already exists"; a READY variant is still refused.
• GET /api/v1/devices honours project_id (it returned every device, even for a bogus project), refuses one that is not a UUID, and keeps its label filter within the caller's projects.
• fleetml experiment conclude no longer panics on a winner ID shorter than 12 characters.

Docs

• Windows: the README's requirements and the quickstart's new "On Windows" section run FleetML in a WSL 2 shell with Docker Desktop's WSL 2 backend, and reach the dashboard and the API from Windows at localhost.
• "Get the CLI" in the README and the quickstart: the release assets, the checksum check and the commands for macOS and Linux (curl) and Windows (PowerShell); the Linux agent installs from its release archive.
• Windows: Docker Desktop's requirements and its early-access status on Arm, a TLS 1.2 note for the PowerShell download, and what the Windows run workflow exercises (the amd64 binary; arm64 not yet run on Windows on Arm).
• First-run guides rewritten against the production compose stack from the quickstart, with the CLI instead of curl, for Android, iOS, the Linux agent and Windows.

0.4.0 - 2026-10-15

Added

• The Android SDK is published to Maven Central as dev.fleetml:fleetml-sdk (with sources and javadoc jars, every file signed) by the hand-run Android publish workflow; the version comes from sdkVersion in android/gradle.properties, and protobuf-lite is now an api dependency because its types appear in the SDK's public API.
• The iOS SDK installs by the repository URL: Package.swift is now at the repository root (the code stays in ios/FleetMLSDK), so Xcode's Add Package Dependencies with https://github.com/fleetml/fleetml and .package(url: "https://github.com/fleetml/fleetml", from: "0.4.0") resolve it. Breaking for apps that added the package by the local path ios/FleetMLSDK: the local path is now the repository root.

Changed

• The README opens with what FleetML does and why it exists, describes each feature in one sentence, lists the quickstart's requirements, shows how to move the dashboard off port 80 with a compose override file, and has a CI badge.
• SECURITY.md adds GitHub's private vulnerability reporting as a second way to report a vulnerability.
• Docs: first-run fixes from a stranger test on macOS and a physical Android phone (zsh comments, port override, ANDROID_HOME and JDK 17, physical device runbook, offline and metered testing, p-value note).
• Docs: a physical device runbook for the iOS SDK (Developer Mode, the Mac's LAN address and S3_PUBLIC_ENDPOINT, personal team signing, ENABLE_PREVIEWS=NO for command-line builds), and both SDKs are stated as verified on a physical phone (an iPhone on iOS 26.6.1, an Android 15 phone).
• Docs: the Android install section leads with the Maven Central dependency and keeps including the module from a clone for unreleased changes; the iOS install section leads with the repository URL and keeps the local clone as a second option; the README shows both, and RELEASING.md has a release checklist and a step that publishes the Android SDK.

0.3.0 - 2026-10-08

Added

• Artifact signing: every project gets an ed25519 key (private keys encrypted with the new required SIGNING_MASTER_KEY), each variant is signed when it is finalized, and Register and heartbeat responses carry the keys and each assignment its signature. The Linux agent and the Android and iOS SDKs verify it before downloading; a failure is reported as the new SIGNATURE_INVALID event, which counts against the health gate. POST /projects/{id}/signing-key and .../rotate (with a grace period). scripts/signing-e2e.sh checks a client against the dev stack.
• Feature-level drift: PUT /models/{id}/features sets each feature's bin edges, devices get them in DeviceConfig.feature_bins, and the SDKs (reportInference(..., features)) and the agent's local endpoint ("features") send per-feature histograms. The drift check computes PSI per feature (metric feature:<name>) with the usual deduplication and alerts. Every check's PSI is kept (GET /drift/{id}/points) and the Drift page draws it over time.
• Helm chart (deploy/helm/fleetml): the server, the dashboard and optional bundled Postgres (TimescaleDB), NATS (JetStream) and SeaweedFS, secrets from values or an existing Secret, and a helm test hook. CI lints and renders it; scripts/helm-e2e.sh installs it into kind and runs the end-to-end suite against it (weekly in the E2E workflow).
• Users and roles: every API key belongs to a user; users are global admins or have a viewer, operator or admin role per project (POST /users, POST /users/{id}/api-keys, PUT /projects/{id}/members). A permission matrix covers every route and refuses unlisted ones; lists across projects show only the caller's projects. Existing API keys belong to the new bootstrap user, a global admin, so nothing changes for them.
• Audit log of every change and every refused request (user, key, route, project, status, client address): GET /audit with filters, GET /audit/export as CSV, and an Audit page in the dashboard.
• Delta updates: when a variant becomes READY the server builds a binary delta (bsdiff with raw DEFLATE sections, internal/delta) from the previous READY variant of the same model, runtime and quantization; assignments carry delta_url, delta_size_bytes, delta_sha256 and delta_base_sha256, and the Linux agent and the Android and iOS SDKs download and apply the delta when they have the base version, falling back to the full model. DELTA_MAX_SIZE_MB (128) bounds the models that get one.
• At-least-once telemetry: accepted batches go into a NATS JetStream stream (FLEETML_TELEMETRY, file storage) before the device is answered, and a durable consumer acknowledges them only after the database commit. Events carry an event_id (set by the SDKs and the agent) and are stored once per ID. New metrics fleetml_telemetry_consumer_pending, fleetml_telemetry_redelivered_total and fleetml_telemetry_duplicates_total; a Grafana dashboard and Prometheus in the compose profile observability.

Changed

• NATS must run with JetStream; both compose files enable it, pin nats:2.12.1 and keep its data in a volume. scripts/e2e.sh and the demos start their own NATS, so they no longer share the dev stack's.
• The telemetry consumer drops only events the database can never accept (invalid values, unknown model versions); other database errors are retried.
• The Linux agent applies a telemetry flush interval from the server at once instead of after its first 30-second wait.
• The CLI, the Linux agent, the dashboard and the Android and iOS SDKs are 0.3.0.
• SIGNING_MASTER_KEY is required; make env-file adds one to an existing .env.
• The quickstart's dashboard hint mentions the login (it still pointed to Settings).
• The dashboard's login asks for "your API key": any user's key logs in, not only an admin's.
• The 20 MB delta timing test no longer runs under the race detector, where it took minutes.
• Both compose files and CI pin timescale/timescaledb:2.30.2-pg16, the tag the Helm chart uses, instead of latest-pg16.
• The iOS SDK's tests, live test and signing check also pass on the iOS 16.4 simulator (before, only iOS 26.5 had been tested); the deployment target stays iOS 15.

Fixed

• DELETE /devices/{id} failed with a 500 for any device that had reported telemetry. Deleting a device now keeps its telemetry, with its model version and experiment arm, and drops only the link to the device (migration 000018); experiment results, drift and telemetry summaries do not change.
• Android and iOS SDKs: the start-up registration and a syncNow() right after init() could both register; each Register issues a token that supersedes the previous one, so the device could keep the superseded token and be refused on its next start. Registrations now run one at a time, and a caller that waited does not register again.
• make dev, make test-integration, make e2e, make cover, scripts/demo.sh, scripts/demo-agent.sh and the E2E workflow waited for migrations with docker compose wait migrate, which fails with "no containers" when the migrations have already finished. They now use scripts/wait-migrate.sh, which waits for the container and returns its exit code.
• iOS SDK: flushNow() called while another flush was running (for example the one a reconnect starts) returned at once without sending; it now waits for that flush and sends what is left, as on Android.
• The Android live tests failed when run again against the same server: LiveServerTest cleared the stored token after each run while the emulator's fingerprint stayed registered in that run's project, so the next run (a new project) was refused; and it took the first model reported, which could be one SignatureLiveTest had left in storage. Each run now registers with a fingerprint of its own (as the iOS live tests do) and waits for the model it deployed. scripts/android-e2e.sh runs only LiveServerTest; SignatureLiveTest is skipped unless scripts/signing-e2e.sh runs it.

0.2.0 - 2026-10-08

Added

• Linux device agent (cmd/fleetml-agent, make build-agent): a static binary for linux/amd64 and linux/arm64 with a systemd unit and install script (deploy/agent). It registers with an enrollment key and a machine-ID-based fingerprint, downloads with Range resume and sha256 checks, validates with a configurable command, switches a current symlink per model, keeps the last-known-good version, runs a post-activate hook, queues telemetry on disk, offers a local HTTP endpoint for inference events, and honours the download policy and urgent flag. docs/LINUX_AGENT.md.
• scripts/demo-agent.sh: four Linux agents take a canary of a model that fails at inference; the health gate rolls it back. Its recording, with two dashboard screenshots, is the README's GIF (docs/screenshots/demo.gif), regenerated by docs/screenshots/demo/record.sh (vhs, Playwright, ffmpeg).
• Documentation site: make docs-site builds the guides, README and changelog with MkDocs Material (pinned, in Docker, strict) into site/ and checks every internal link and anchor offline; CI builds it on every pull request.
• Alert sinks per project (webhook with HMAC signing, Slack Block Kit, SMTP email), managed over REST (/projects/{id}/alert-sinks, /alert-sinks/{id}) and on the dashboard's Alerts page, with a test-send button and a delivery log of the last 50 attempts. Alerts now carry project_id and the model name and version. New settings: DASHBOARD_URL, SMTP_HOST, SMTP_PORT, SMTP_USERNAME, SMTP_PASSWORD, SMTP_FROM.
• Dashboard login: POST /api/v1/auth/login exchanges an admin API key for an httpOnly, SameSite=Strict session cookie backed by a server-side session (SESSION_TTL, 12 hours by default). Writes with a session need its CSRF token (X-CSRF-Token). GET /auth/session, POST /auth/logout, GET /auth/sessions and POST /auth/sessions/{id}/revoke; the dashboard has a login page, redirects to it when the session ends, and lists and revokes sessions in Settings. The REST API still accepts Authorization: Bearer <key>.

Changed

• Linux devices accept CUSTOM model variants, and a linux/arm64 chipset is profiled as arm64.
• The dashboard no longer stores the API key in the browser's local storage, and no longer supports VITE_API_URL: it calls the API on its own origin (FLEETML_API sets the dev and preview proxy target).
• The revert deployment of an unnamed deployment is named revert-<id prefix> instead of revert-, and health gate reasons say "1 cohort device".
• Alert deliveries make up to five attempts (after 1s, 2s, 4s and 8s) instead of four; the last failure is recorded as a dead letter.

0.1.0 - 2026-10-08

The first release.

Server

• Model registry: projects, models, versions and variants (runtime and quantization). Files are uploaded straight to an S3-compatible store with presigned URLs; finalizing streams the object and checks its sha256 and size, and only READY variants are ever assigned to devices.
• Variant selection per device by OS runtime, RAM and capability score, with per-version requirement overrides (fleetml_requirements metadata).
• Deployments: all at once, canary stages and rolling batches on a consistent-hash ring, targeted by device labels; one version per model per device (the newest deployment that reaches it).
• Health gate on by default for canary and rolling rollouts: the exposed cohort's error rate and, optionally, its p95 latency against the previous version; a failing gate rolls back or pauses.
• Rollback (also after completion) to the previous version with a reason and an alert, cancel, pause and resume, with a validated state machine.
• Urgent deployments: their assignments download on any network, ignoring the device download policy; rollbacks are always urgent.
• Device settings from the server (DEVICE_* variables), sent at registration and with every heartbeat: foreground and background heartbeat intervals, telemetry batch size and flush interval, and the download policy.
• A/B experiments: sticky cohorts with stable arm IDs, both arms' models delivered and kept on the device, telemetry attributed per arm, a fixed-horizon Welch t-test, maximum duration and conclusion with a winner.
• Telemetry over gRPC through NATS into a TimescaleDB hypertable (30-day retention, hourly continuous aggregate), with summaries and time series.
• Drift detection: PSI on inference latency and confidence, last 24 hours against the 7 days before, with history; a change into CRITICAL raises an alert.
• Alerts for rollbacks and drift on NATS (fleetml.alerts.*) and to an HMAC-SHA256-signed webhook with retries.
• Security: admin API keys on the REST API, per-project enrollment keys, revocable and refreshable device tokens, devices confined to their project's deployments, experiments and telemetry (a device moves to another project only with its own token and that project's enrollment key), locked-down CORS.
• Operations: /livez, /readyz, Prometheus /metrics, grpc.health.v1, gRPC reflection behind GRPC_REFLECTION, start-up that waits for its dependencies, and configuration checked at start-up with every invalid variable named.
• Device protocol v1 (proto/fleetml/v1): Register, Heartbeat, RefreshToken and ReportBatch.

SDKs and samples

• Android SDK (Kotlin, minSdk 26): enrollment, token refresh, resumable sha256-verified downloads, validation with TensorFlow Lite or ONNX Runtime before activation, last-known-good, offline start, a persistent telemetry queue, and server-set intervals. A sample app runs the Iris TensorFlow Lite model.
• iOS SDK (Swift package, iOS 15): the same lifecycle with Core ML validation and Keychain credentials, and a SwiftUI sample app running the Iris Core ML model.
• Battery and data policy in both SDKs: heartbeats and telemetry flushes on a timer only in the foreground; in the background one check (WorkManager on Android, BGAppRefreshTask on iOS) at the server's interval, hourly by default; telemetry always batched; model downloads on an unmetered network or while charging unless the deployment is urgent. syncNow() runs a check on demand.
• Sample Iris classifiers in TensorFlow Lite and Core ML (models/iris).

Tools

• Dashboard: models with versions and variants, deployment and experiment creation and control (rollback and cancel with confirmation), drift history, telemetry charts; works on phones.
• CLI: fleetml push, deploy, experiment, drift, device, telemetry, and simulate for fleets of simulated devices.
• scripts/demo.sh, the 60-second story against a simulated fleet.
• Docker Compose files for development and for running the server and dashboard in containers.
• CI for the server, dashboard and both SDKs, and a weekly end-to-end run.

Known limitations

• No TLS termination: put the REST, gRPC and object store ports behind a TLS proxy before exposing them.
• The dashboard keeps the admin API key in the browser's local storage.
• One admin role; no organisations or per-project permissions yet.
• The SDKs are used from source; they are not yet published to Maven Central or a Swift package registry.
• The iOS SDK targets iOS 15 and was tested on the iOS 26.5 simulator. The Simulator cannot run background tasks, so the iOS background refresh has been checked on the Simulator only up to scheduling.