clothescast

Privacy Policy

Last updated: 2026-06-21

ClothesCast is a daily weather-insight app for Android. This policy describes what data the app handles, where it goes, and what control you have over it.

TL;DR

• ClothesCast has no login and no personal account. The free online voice signs in anonymously — a random ID with no name, email, or password — purely to count your daily usage; no backend holds personal data about you. The app may send anonymous crash reports and aggregate usage analytics to a third-party reporting service so the developer can fix bugs and decide which features to keep — see “Analytics and crash reporting” below for what those payloads include and (more importantly) the hard limits on what they don’t.
• Your approximate location is sent to Open-Meteo to fetch the weather forecast, and (on Play Services devices) to Google’s geocoding service via Android’s Geocoder API to look up the city name shown next to the date. If you search for a place name (from the Settings location picker), the text you type is also sent to Open-Meteo so it can suggest matching places. That is the only user-content data the app sends off your device by default.
• If you opt in to online text-to-speech, the short spoken sentence (e.g. “50% chance of rain at 3pm — take an umbrella”) is sent to the TTS provider you chose (Google Gemini) so it can return the audio.
• If you opt in to calendar extras, the app reads today’s events on your device. The event title may appear inside the spoken sentence (e.g. “Bring a jacket for your concert tonight”), and is therefore also sent to the TTS provider in that one case — only when both calendar extras and online TTS are enabled. If you also enable the MQTT bridge below, one calendar-derived boolean — whether the window has a located event, with no titles, times, or locations — is published to your own broker.
• If you opt in to the Smart Home / Home Assistant MQTT bridge (Settings → Smart Home, off by default), the rendered forecast sentence — plus a small calendar-derived has_events boolean — is published as a retained MQTT message to a broker you configure — typically the Mosquitto add-on running on your own Home Assistant. The broker host and credentials are entirely yours; nothing ClothesCast sends goes to a service the developer operates. See “Smart Home / Home Assistant bridge” below for what’s in the payload.
• Nothing is sold, no advertising, no ad-targeting profiles, no third-party data sharing beyond the services listed below. (The aggregate usage analytics + crash reports above do count as cross-session telemetry — that’s what the install identifier in the crash payload is for. The “Analytics and crash reporting” section spells out exactly what’s in and out.)

Who we are

ClothesCast is an open-source Android app developed by Mikel Ward. The source code is at https://github.com/mikelward/clothescast.

Data the app handles

Approximate location

• What: Coarse latitude / longitude (city-block level, never precise GPS — the app declares only ACCESS_COARSE_LOCATION).
• Why: To fetch the weather forecast for where you are and to look up a human-readable place name to show in the UI.
• Where it goes:
  • Open-Meteo — receives the coarse coordinate to return the weather forecast, and (separately) the raw text you type into a location search (“Edinburgh”, “新宿”) so Open-Meteo can match it to candidate place names. The search text leaves the device only when you actually type into the Settings location picker.
  • Google’s geocoding service, via Android’s Geocoder API on Play Services devices — receives the coarse coordinate so we can look up the city name (e.g. “London”) and the country code (used to pre-select holidays in your country). On AOSP / non-Play devices Geocoder reports no backend and we skip the lookup; the Today header falls back to a date-only label.
  • Google’s Weather API (Google Maps Platform) — receives the coarse coordinate to return a weather forecast, only if you turn on the “Google” forecaster on the Settings → Forecasters page. That forecaster reuses your own Gemini API key (Google’s keys are project-scoped, so the same key calls the Weather API when your project has it enabled), so the key is sent to Google with the request. Off by default; nothing goes to Google’s Weather API unless you enable the forecaster and have set your own Gemini key.

  None of these services receives an account, device identifier, or any other payload beyond the items listed above.

• Stored on device: Your chosen location is saved in app settings so the daily refresh can run without re-prompting you.
• Retention by us: Until you uninstall the app or change locations.
• Retention by Open-Meteo: See https://open-meteo.com/en/terms.
• Retention by Google: See https://policies.google.com/privacy.

Calendar events (optional, off by default)

• What: Today’s events — title, start/end time, location, all-day flag — read via Android’s CalendarContract only when you grant READ_CALENDAR and enable the calendar extras setting.
• Why: To pair a clothes recommendation with a meeting that overlaps bad weather (e.g. “Bring a jacket for your concert tonight”).
• Where it goes: Calendar reading happens entirely on your device. Two narrow paths can let calendar-derived data leave it: (a) if the extras fires for today’s forecast and you have online TTS enabled, the rendered sentence (which can include the event title) is sent to your chosen TTS provider for vocalization; and (b) if you enable the Smart Home MQTT bridge, a single event-presence boolean (has_events — no titles, times, or locations) is published to your own broker. See “Smart Home / Home Assistant bridge” below.
• Stored on device: The most recent forecast snapshot is cached for up to one day so the app doesn’t refetch on every launch. Event titles and locations are never written to disk; only event timing (start, end, all-day) and a “has a location” flag are persisted, since that’s all the insight renderer needs to re-derive the prose.
• Retention by us: Replaced on the next daily refresh; cleared on uninstall.

Calendar-sourced theming (optional, off by default)

• What: When you enable “Holidays” or “Birthdays” in the Holiday settings screen, ClothesCast reads your synced device calendars to detect holiday and birthday events. Detection runs entirely on-device (event title pattern matching + source-calendar account inspection).
• Why: To theme today’s outfit on holidays your curated catalog doesn’t cover (Diwali, Eid, Lunar New Year, etc.) and on detected birthdays — a celebratory palette + a banner naming the day.
• Where it goes: Nowhere off-device. Detected event titles (including contact names on birthday events) are shown only in the themed banner on your screen — they are never read aloud through online TTS, never sent to Gemini, and never written to Firebase Analytics or Crashlytics.
• Permission: Both toggles require READ_CALENDAR, prompted by the toggle itself. Revoking permission later from system Settings doesn’t auto-flip the toggle off; the reader simply returns no events until permission is restored.

Notifications

• ClothesCast posts a local notification at the time you choose. The notification is generated on your device. Nothing is sent to a push service.

Diagnostic logs and crash reports

• What: ClothesCast keeps a rolling on-device file of the most recent log lines — errors, warnings, info — at cacheDir/diag.log (plus one rotated predecessor cacheDir/diag.log.1, capped at ~200 KB each). Lines may include the rendered insight prose (“Insight delivered for …”) and device-resolved coordinates (“Using device-resolved location at lat, lon”). If the app crashes, recent lines plus the stack trace are also written to cacheDir/last-crash.txt. The local bug-report payload includes the most recent ~300 lines of the diag file plus your current settings (schedule, units, TTS choice, clothes rules) and the most recently rendered insight prose.
• Why: So you can hand a complete diagnostic snapshot to the developer when something goes wrong, and so the developer can be alerted to crashes affecting users in the wild.
• Where it goes:
  • Locally, always: after a crash the home screen surfaces a banner offering to share the report; tapping Share report opens Android’s system share sheet so you can pick a destination (email, Slack, Drive, etc.). The on-device file includes the full payload described above (settings + insight prose + log buffer).
  • Automatically, possibly: the app may also send a trimmed crash report — stack trace, app version, Android version, device model, and a non-resettable install identifier used only to group duplicate crashes — to a third-party crash-reporting service. This automatic payload does not include insight prose, calendar data, location, API keys, or the in-memory log buffer. (Your settings values may travel separately as part of the aggregate analytics described below.) See “Analytics and crash reporting” below for the full limits.
• Stored on device: The diag file (cacheDir/diag.log plus one rotated predecessor) persists across launches and app upgrades until rotated out by newer lines or cleared on uninstall / cache-clear, so bug reports filed after a process restart may include lines from earlier runs. A ---- process start <version> ---- marker line at each launch lets readers see where the current process began. The crash file persists across launches until a fresh crash overwrites it, and is cleared on uninstall.
• Retention by us: The on-device file is yours — whatever you do with it is governed by the destination you share to. Anything sent automatically to the crash-reporting service is governed by that service’s privacy policy.

Smart Home / Home Assistant bridge (optional, off by default)

• What: The rendered forecast sentence — the same one displayed in the notification, e.g. “Today, cool and mild. Wear a sweater. Chance of rain at 3pm.” The bridge publishes this string as a retained MQTT message after each scheduled twice-daily refresh, so Home Assistant (or any MQTT-aware consumer) can read the latest value and speak it on a trigger of your choice (a wardrobe door opening, a humidity spike, a fixed time of day, etc.).
• Why: To let you hear your tuned ClothesCast forecast on a Google Home, Nest Hub, Echo, or any other speaker your home automation can reach — without reimplementing the clothes-rule and insight logic in Home Assistant.
• Where it goes: The MQTT broker you configure in Settings → Smart Home. That’s typically the Mosquitto add-on running inside your own Home Assistant instance on your local network. The broker host, port, optional TLS, optional username, and topic prefix are all values you supply; ClothesCast never sends this payload anywhere you haven’t pointed it at. The broker is not a service the developer operates; it’s your infrastructure under your control.
• What’s in the payload: The rendered sentence string, UTF-8 encoded — the same text you see in the notification — and, on their own sibling topics when available, the outfit-card image (the same 800 × 480 PNG shown on a Nest Hub), the spoken TTS audio clip (the same bytes the phone speaks, Gemini engine only), and a combined MP4 that simply muxes that image and audio together for one-shot casting. The image and audio carry the same insight — the card renders the sentence, the audio speaks it — so the MP4 introduces no new data; it only repackages what the image and audio topics already publish. A small now/timestamp topic carries the epoch-millis time of the most recent publish so a consumer can tell a fresh forecast from a retained one. A has_events topic carries a single boolean — true/false for whether that window has a located calendar event — so an automation can tell a deliberately-quiet evening (when “notify only on event nights” keeps the phone silent) from an outage. It’s the only calendar-derived field here, and it carries no event titles, times, or locations. The payload includes any calendar-event extras clause that fires (e.g. “Bring a jacket for your concert tonight”) when you have the calendar extras enabled, because that clause is part of the rendered sentence. No coordinates, no API keys, no settings values, no device identifiers travel with any of these messages.
• Authentication: If you set a broker username, the bridge sends it on connect; the corresponding password is stored on your device encrypted at rest (same Tink-AEAD pattern as the Gemini API key storage above) and only sent to the broker you configured.
• Retention: Retained MQTT messages persist on your broker until you delete them or publish a new value to the same topic. The next twice-daily refresh overwrites the previous payload. Retention by the broker is governed by your own broker configuration.
• Setup guide: See docs/smart-home.md for step-by-step configuration on both the ClothesCast side and the Home Assistant side.
• Local-network discovery: When you tap “Find broker” on the Smart Home settings page, ClothesCast issues standard mDNS / DNS-SD queries (_home-assistant._tcp. and _mqtt._tcp.) over your local network multicast group so it can offer to pre-fill the broker host. Nothing about you or your forecast is included in the queries — they’re indistinguishable from any other mDNS browser on your network — and discovery stops as soon as you leave the page or pick a result. No discovery happens unless you tap the button.

Online TTS: shared key by default, your own key on request

Online TTS (the Gemini voice) has two paths.

• Default — shared key via the ClothesCast TTS proxy. When you have not set your own Gemini key, online TTS requests go to a small ClothesCast-operated Cloud Function instead of directly to Google. The function holds a Google Gemini API key paid for by the developer, forwards your request to Google, and returns the audio. It does not log the spoken prose or the audio response.
  • First use — an anonymous identifier from Google. Before the first shared-key request, the app signs in anonymously to Firebase Authentication (a Google service). This is not a user account: there is no name, email, password, or sign-in screen. It returns a random, app-generated identifier — just a string of letters and digits, e.g. a1B2c3… — that is not linked to your name, email, phone number, device, or location, and is never used for advertising or tracking across apps. It exists only so the proxy can count how many free clips this app install has used today. On the shared path your data is handled by ClothesCast and Google servers.
  • Sent on each request: the rendered insight sentence (the same text the phone speaks), the anonymous identifier above (carried as a signed Firebase ID token in the Authorization header, which the proxy verifies to read the identifier — the client cannot forge or swap it), a Firebase App Check attestation token (proves the request came from a genuine install — verified in-memory and discarded), and the model name. During the current rollout the app also sends a Firebase Installation ID for backward compatibility (so older proxy versions still recognize the install); it is likewise an anonymous, random identifier with no personal information, and is dropped once the rollout completes. No coordinates, calendar metadata as fields, user account, advertising ID, or hardware / device identifiers are added.
  • Stored: a Firestore document at quota/<your anonymous ID> with a first-use timestamp, a successful-call counter, and a most-recent-use timestamp, capped at 5 successful clips per UTC day. Used to enforce the free daily limit and for capacity planning. Not used for advertising or cross-app tracking. Your anonymous identifier resets if you reinstall the app or use “Clear data” in Android Settings.
  • Cloud Functions infrastructure logs (operated by Google Cloud, not configured by us) record the client IP and HTTP status of each call. We do not associate the IP with your anonymous identifier.
• Your own key (BYOK). Open Settings → Voice → Gemini API key and paste your own key. Keys are stored on your device, encrypted at rest using a key sealed by the Android Keystore. With your own key set, online TTS requests skip the ClothesCast proxy entirely and go straight to Google — no anonymous identifier, no App Check token, no anonymous sign-in, no proxy involvement. Your key is never shared with us or any third party.

Third-party services

ClothesCast talks to these services on your behalf. Their own privacy policies apply to anything they receive:

Service	What we send	When
Open-Meteo	Coarse coordinate (forecast); your typed search text when you use the Settings location picker	Always for forecast; only when you search for a place name
Google’s geocoding service (via Android’s Geocoder on Play Services devices), governed by Google’s Privacy Policy	Coarse coordinate	Always on Play Services devices (city / country lookup for the Today header); skipped on AOSP devices
Google Weather API (Google Maps Platform), governed by Google’s Privacy Policy	Coarse coordinate and your own Gemini API key	Only if you enable the “Google” forecaster on the Forecasters page and have set your own Gemini key
Google Firebase Authentication	An App Check attestation, to mint and refresh an anonymous identifier (no name, email, or password)	First use of the shared-key path, then periodic token refreshes
ClothesCast TTS proxy (Cloud Function, ClothesCast-operated)	The short rendered insight sentence, an anonymous app identifier (as a signed Firebase ID token; plus a Firebase Installation ID during the current rollout), a Firebase App Check token, and the Gemini model name	Online TTS with the shared key (default)
Google Gemini API	The short rendered insight sentence (forwarded by the proxy, or sent directly when you use your own key)	Online TTS, either path
Your self-hosted MQTT broker (e.g. Mosquitto inside Home Assistant)	The short rendered insight sentence; the outfit-card image, TTS audio, and muxed MP4 on sibling topics; an update timestamp; and a calendar-derived has_events boolean (no event titles, times, or locations) — all as retained MQTT messages	Only if you opt in to the Smart Home bridge and configure a broker
Analytics / crash-reporting service (e.g. Firebase Crashlytics + Google Analytics for Firebase)	Aggregate usage events and crash diagnostics — see “Analytics and crash reporting” below for what’s in and out	Possibly always, in all builds

These providers act as service providers fulfilling a single request and returning the result. When the ClothesCast TTS proxy is in the path, it adds the anonymous identifier and App Check token described above; when you use your own Gemini key, the request goes straight to Google and no anonymous identifier is attached.

Note on Gemini API: request inputs are not retained for training by default. See the provider’s policy linked above for the authoritative terms.

What we do not collect

• No user accounts and no personal sign-in — no name, email, password, or login screen. (The free online-TTS path uses an anonymous, randomly-generated identifier to count daily usage, described under “Online TTS” above; it identifies an app install, not you, and carries no personal information.)
• No advertising identifiers, no ad networks, no ad targeting.
• No precise GPS location.
• No contacts, photos, microphone, or files.
• No data is sold or shared for advertising, profiling, or model training by us. (Provider terms govern what they may do; the linked policies above describe their commitments for API access.)

Your controls

• Location: Revoke the Location permission in Android Settings, or change / clear it in ClothesCast’s Settings.
• Calendar: Revoke READ_CALENDAR in Android Settings, or turn off “Use calendar events” in ClothesCast’s Settings.
• Online TTS: Switch the voice engine to “Device” in Settings → Voice to keep all spoken text on-device.
• API keys: Clear them from Settings → API Keys at any time.
• Everything: Uninstalling the app deletes all locally stored data (settings, cached insight, API keys).

Analytics and crash reporting

To spot bugs and decide which features are worth keeping, ClothesCast may send two kinds of payload to a third-party reporting service (e.g. Firebase Crashlytics + Google Analytics for Firebase, or equivalent). This may be present in all builds and run for all users by default; the goal is to inform product decisions, not to identify you.

What’s sent:

• Crash reports: stack trace, app version, Android version, device model, and a non-resettable install identifier used to group duplicate crashes. Sent automatically when a crash occurs.
• Aggregate usage events: the values of your in-app settings — TTS engine, schedule cadence, delivery mode, units, notification time, clothes-rule customisations, and the like — plus basic lifecycle events such as app open and daily refresh, so unused options can be pruned and defaults tuned.
• API-call outcomes: one event per outgoing network request to Open-Meteo, Google’s Weather API (only if you enable the Google forecaster), or Gemini, carrying the endpoint identifier (e.g. open_meteo_forecast, google_weather), the HTTP status code, an outcome bucket (success / http_error / timeout / network_error / other_error), and the request latency in milliseconds. No URL parameters, no request or response bodies — the endpoint identifier is a fixed string from a short enum, not a captured URL. Used to spot rate-limit hits and transient server issues in aggregate. Failures that look like “device offline / airplane mode” are filtered out before the event is sent so the stream reflects real provider issues rather than the user’s local radio state.
• Notification delivery timing: when the daily / tonight alarm actually posts its notification, one event carries the slot (today / tonight) and two delay numbers in milliseconds — alarm fire delay (how late the OS released the alarm after the scheduled trigger, an indicator of Doze deferral) and total delay (how late the notification reached the system, which also folds in WorkManager waits for connectivity). No timestamps, no location, no calendar or insight content. Used to spot regressions in delivery punctuality. Powered-off / airplane-mode misses simply don’t appear in the stream — the event only fires when a notification is actually posted, so unfired alarms aren’t reported.
• Daily refresh outcomes: when the scheduled fetch + insight + notify pipeline reaches a terminal result, one event carries the slot (today / tonight), an outcome bucket (success, forecast_error, no_location, cancelled, or other_error), and the wall-clock latency from worker start to terminal result in milliseconds. No forecast content, no insight prose, no location coordinates, no error messages. Used to spot regressions in delivery reliability (“what percentage of scheduled refreshes actually completed today?”). WorkManager retries between transient failures (no connectivity, 5xx from upstream) aren’t reported individually — only the terminal outcome the user actually sees.
• Settings snapshot: once per app process start, and again each time the user changes a non-voice setting later in the same process, one event carries the resolved values: unit settings (auto vs. explicit choice + the unit in effect), delivery mode for the daily and tonight slots, theme mode, colour palette, the default-bottom fallback choice, the four schedule booleans (tonight enabled, notify-only-on-events, daily-mention-evening-events, use-calendar- events), the day-of-week counts for daily and tonight schedules (1..7), and the schedule times bucketed to the hour (“00”..”23”). No exact local time, no calendar content, no location, no insight prose. The launch-time emission is intentional — it samples the population’s default behaviour without requiring the user to interact, so reports can answer “which features do users actually configure and which defaults serve them best?”
• Clothes-rule customisation snapshot: same cadence as the settings snapshot — once per app process start, then again whenever the user edits their clothes rules in the same process. One event carries: the count of catalog defaults the user has changed; the count of extra rules added beyond the catalog defaults; a sorted comma-joined list of which catalog categories were customised; an all_defaults flag; per-category integer Celsius deltas from the default threshold for the temperature rules (sweater / jacket / coat / gloves / shorts), clamped to ±5°C (so e.g. moving the jacket threshold from 10°C down to 8°C reports -2); and, for the two rain-gear rules (umbrella, rain jacket), a signed percentage-point delta of their rain-probability gate from the default, rounded to the nearest 10 and clamped to ±50 (so raising the umbrella gate from 20% to 60% reports +40). A rain-gear rule whose only change is its weather-code floor (drizzle / rain / off) still shows up in the customised-categories list, but the floor value itself is never sent. No raw thresholds, no weather-code values, no user-added rule items.

What’s not sent — these are hard limits, not “best-effort”:

• No calendar event data. Not titles, not times, not locations, not attendees, not whether you have any events at all. The calendar-sourced holiday/birthday classifications described above stay on-device — the themed banner is rendered locally and never read aloud through online TTS, never sent to Gemini, and never written to Firebase.
• No user names, account identifiers, email addresses, or contact info. This explicitly includes contact names that may appear on birthday-themed banners on the Today screen.
• No location coordinates or geocoded place names.
• No insight prose, notification text, or anything else that could carry free-form user content. Exception: if you opt in to the Smart Home / Home Assistant MQTT bridge in Settings → Smart Home and configure a broker, the rendered insight sentence is sent to that broker — your own infrastructure, not the analytics service.
• No API keys or other credentials.
• No precise GPS or advertising identifiers.

The reporting service receives only what’s described above and is bound by its own privacy policy (linked from the third-party services table above once a specific provider is chosen). ClothesCast is open source, so you can audit exactly what’s instrumented at https://github.com/mikelward/clothescast — or build a copy with the reporting calls stripped out, if you’d rather not participate.

Children

ClothesCast is not directed at children under 13 and does not knowingly collect personal data from them.

Changes

If this policy changes, the updated version is committed to https://github.com/mikelward/clothescast/blob/main/PRIVACY.md and the “Last updated” date at the top reflects the change. A short summary of each material change is added to the Changelog below, and material changes are also noted in the app’s release notes.

The full revision history of this file is viewable on GitHub: View all changes to PRIVACY.md →

Contact

Open an issue at https://github.com/mikelward/clothescast/issues or email the address listed on the Play Store listing.

Changelog

• 2026-06-21 — The Smart Home / Home Assistant MQTT bridge now publishes a has_events boolean (true/false) on each window’s topic (and the now mirror): whether that forecast window has a located calendar event, so a Home Assistant automation can tell a deliberately- quiet evening (when “notify only on event nights” keeps the phone silent) from an outage. This is a new calendar-derived field crossing the device boundary — but only to your own broker, only when you have both the calendar extras and the bridge enabled, and it carries no event titles, times, or locations, just the single boolean. See the updated “Smart Home / Home Assistant bridge” and “Calendar events” sections above.
• 2026-06-17 — Rain gear (umbrella, rain jacket) now also fires on the forecast weather code, catching drizzle the probability gate misses. The clothes_rules_snapshot analytics event gained one aggregate param, rain_jacket_delta_pct: the rain jacket’s rain-probability gate as a signed percentage-point delta from its 50% default, bucketed exactly like umbrella_delta_pct (rounded to the nearest 10, clamped to ±50, MISSING if deleted). No raw thresholds, no weather-code values, no calendar / location / insight content — same hard limits as the rest of the event.
• 2026-06-01 — The umbrella is now a default clothes rule keyed on rain probability. The clothes_rules_snapshot analytics event gained one aggregate param, umbrella_delta_pct: the user’s umbrella rain-probability gate as a signed percentage-point delta from the 30% default, rounded to the nearest 10 and clamped to ±50 (e.g. "+30", "-20", MISSING if the rule was deleted). Same hard limits as the rest of the event — no raw thresholds, no calendar / location / insight content.
• 2026-05-31 — Transitional note for the identifier change below: during the staged rollout the app sends both the new anonymous Firebase Authentication identifier and the older Firebase Installation ID, so older versions of the proxy keep working while clients update. Both are anonymous, random identifiers with no personal information. The Installation ID is dropped once the rollout completes.
• 2026-05-31 — Hardened the shared-key TTS identifier. The free online-TTS path now identifies each install by an anonymous Firebase Authentication identifier, verified on the server, instead of the Firebase Installation ID the app previously sent in a header. This stops a modified client from rotating the identifier to bypass the daily free-clip limit. The identifier is still random and carries no personal information; the daily counter moved from installs/<install ID> to quota/<anonymous ID>. To obtain the identifier the app now signs in anonymously to Firebase Authentication (a Google service) — no name, email, or password, and no user-facing account. See “Online TTS: shared key by default, your own key on request” and the third-party services table above.
• 2026-05-31 — Online TTS now has a shared-key default: when you haven’t supplied your own Gemini API key, the spoken sentence is sent to a small Cloud Function operated by the developer, which forwards it to Google’s Gemini API and returns the audio. The function holds the Gemini key, doesn’t log the spoken prose, and records only a per-install counter (keyed by your Firebase Installation ID) for capacity planning. A Firebase App Check token is sent alongside each request to prove it came from a genuine install; it’s verified and discarded. The BYOK path (your own key in Settings) is unchanged and bypasses the proxy entirely. See “Online TTS: shared key by default, your own key on request” and the updated third-party services table above for the full breakdown.
• 2026-05-29 — The optional usage-analytics settings snapshot now records one analytics event per Settings page (Schedule, Clothes, Format, Region, Display, Calendar) instead of a single combined event. Content-neutral: the same coarsened settings values are reported, just grouped by page; nothing new leaves the device and the same privacy toggle still gates it.
• 2026-05-23 — Internal change to how the daily insight is cached on device: the cache now stores the upstream forecast snapshot plus minimal event timing (start, end, all-day, a “has a location” flag), and re-derives the insight against your current settings on read. Event titles and free-form locations are not written to disk. No material change to what leaves the device. See the updated “Calendar events” section above.
• 2026-05-16 — Added an optional Smart Home / Home Assistant MQTT bridge (Settings → Smart Home; off by default). When enabled, the app publishes the rendered forecast sentence as a retained MQTT message to a broker you configure (typically the Mosquitto add-on inside your own Home Assistant). This relaxes the existing “insight prose never leaves the device” hard limit — but only to your own broker, never to a developer-operated service or analytics endpoint. The broker password, if you set one, is stored on-device encrypted with the same Tink-AEAD pattern as the Gemini API key. See the new “Smart Home / Home Assistant bridge” section above and docs/smart-home.md for the full setup.
• 2026-05-14 — Added three new aggregate analytics events: daily_refresh (slot + outcome bucket + latency, one per terminal result of the scheduled fetch + notify pipeline; WorkManager retries between transient failures aren’t individually reported), settings_snapshot (the resolved values of the non-voice user settings — units, delivery mode, theme / palette / default-bottom, the four schedule booleans, day counts, and schedule times bucketed to the hour), and clothes_rules_snapshot (counts of customised defaults and extra rules, sorted list of customised categories, plus per-category integer Celsius delta from the default threshold, clamped to ±5°C). No raw thresholds, no exact local times, no calendar / location / insight content; same hard limits in “Analytics and crash reporting” apply.
• 2026-05-12 — Added two new aggregate analytics events: per-request API-call outcomes (endpoint identifier, HTTP status, outcome bucket, latency) for Open-Meteo and Gemini, and notification-delivery delay (slot + two millisecond delays) when the daily alarm actually posts. Neither event carries request bodies, URL parameters, location, timestamps, or calendar / insight content; the offline-failure filter on API events drops airplane-mode noise. Same hard limits in “Analytics and crash reporting” apply; the toggle in Settings → Privacy still controls whether anything is sent.
• 2026-05-12 — Diagnostic logs are now persisted to cacheDir/diag.log (with one rotated predecessor, ~200 KB each) instead of living only in a process-memory ring buffer. This lets errors logged by background workers reach a bug report filed after the OS later kills the process. The categories of data the diag log may contain (rendered insight prose, device-resolved coordinates, error / warning text) are unchanged — they were already shown in the bug-report payload from the in-memory buffer; what changes is that they now live on disk between launches, similar to the existing crash file. Nothing new leaves the device; the bug-report share-sheet flow is unchanged.
• 2026-05-06 — Removed OpenAI and ElevenLabs as TTS provider options; Google Gemini is now the only online TTS provider. Removed the corresponding rows from the third-party services table and the note on OpenAI’s 30-day input retention. Updated the API-key storage description accordingly. No change to what data leaves the device or when.
• 2026-05-05 — Wired the previously-anticipated Firebase Crashlytics
  • Google Analytics for Firebase integration into the app. Default-on with a one-time non-blocking notice on the Today screen pointing the user at the new Settings → Privacy toggle to turn it off. The hard limits described in “Analytics and crash reporting” above are unchanged; this entry just records that the SDKs are now actually loaded (provided the developer has supplied app/google-services.json).
• 2026-05-03 — Permitted automatic crash reporting and aggregate usage analytics in all builds and for all users, with hard limits on what those payloads may include (no calendar data, location, insight prose, API keys, precise GPS, or ad identifiers). Broadened analytics to cover in-app settings values. Simplified the API-key storage description.
• 2026-05-01 — Initial publication.

View full revision history on GitHub →