Skip to content

Latest commit

 

History

History
230 lines (207 loc) · 17.3 KB

electron-privacy.md

File metadata and controls

230 lines (207 loc) · 17.3 KB

Electron privacy and retention

Historical context: this page was written while the Electron and Tauri apps coexisted. Mentions of Tauri helpers, pages, tests and regression results describe that migration period; the Tauri shell has since been removed and the shared code now lives in electron/src/shared/. Existing Tauri installs: see the migration guide.

Settings > Privacy exposes the existing invisible-watermark setting, analytics consent, automatic error reports and generation-history retention.

Diagnostic bundles and prefilled bug reports scrub known credential query parameters, including share PINs, API keys and passwords. The backend and Electron report builder also recognize percent-encoded parameter names, preserve unrelated query parameters, and do not treat a substring such as spin as pin. Quoted credential values, including log-escaped quotes, are removed in full. Query scanning uses a finite known-name bound so malformed log lines cannot trigger repeated suffix scans. Enclosing URL bounds are verified from structured JSON or the complete URI token; ambiguous quoted tails are redacted conservatively.

Manual bug reports prioritize the existing scrubbed, bounded faulting-thread excerpt for the current native backend exit ahead of verbose status tails. This keeps useful crash context within the issue URL's size budget; it does not fix or infer the native crash's cause. Ordinary reports retain their section order. The report still opens for review before submission, with no automatic raw-log upload or additional network request.

Watermark controls appear when the backend reports AudioSeal available and affect new audio through the existing marking path. No audio producer bypasses mark_synthetic. Analytics consent uses the existing backend opt-in endpoint and remains unchanged until the user explicitly switches it. After consent, the renderer records fixed screen labels and workflow action labels (clone, design, dubbing, transcription); the backend records lifecycle/error events. Privacy-safe pageviews use fixed app paths under app.voicestudio.sh, giving aggregate screen counts without sending renderer URLs or dynamic route segments. The renderer never contacts the analytics service itself: it hands allowlisted events to the local backend, which forwards them only while consent is on. Uncaught renderer errors are reduced to a coarse class and stage; no message, stack, credential, home path, text, audio, filename, or user-named content is sent. Raw exception autocapture, DOM autocapture, automatic pageviews, and session recordings remain disabled. Unavailable features do not display inert toggles, and failed saves preserve confirmed state.

The analytics destination is the public Supermetry site id in config/analytics.json (host https://app.supermetry.com); every build uses it, and SUPERMETRY_SITE_ID / SUPERMETRY_HOST override it for developers. Nothing is baked in at build time.

Optional VoiceStudio Cloud sign-in validates the discovery issuer and binds its token and revocation endpoints to that issuer's origin and HTTP(S) protocol. Missing endpoints, different hosts or ports, and embedded-origin schemes such as blob: are refused before credentials are sent. HTTPS is required outside the explicit local development configuration; redirects remain disabled. Account features stay off until enabled, independently of analytics consent.

Automatic error reports

Automatic reports have their own explicit consent, separate from usage analytics and off by default. Existing analytics consent does not enable them. A fresh desktop install asks before runtime setup, so an opted-in installation can report a setup failure even when Python has not installed or cannot start. Settings > Privacy changes this choice later. Declining leaves the app usable.

Reports cover observed app/backend/renderer crashes, failed or stalled setup and downloads, incompatibilities, model loading, engine execution, speech generation and other failed operations. Normal cancellation, validation prompts, waiting for a network or scheduled retry, and an intentionally paused download are not failures. An internal generation deadline remains a reportable timeout, distinct from a user pressing Cancel. A no-progress observation does not cancel a valid running operation. Reporting is best effort: an abruptly terminated or offline process cannot guarantee delivery; eligible crash summaries can be recovered on a later launch.

Packaged builds configured with an app ID and write-only ingest key send failure reports to Supermetry's app intake (/api/public/v1/app-events). The key identifies the application, never the user, and cannot read reports. It ships in the app's private build metadata; it is never committed, logged by the app, or included in stored reports. Because it is write-only and ships in every build, it is not treated as a secret: when read from a repository variable, CI build logs may show it. The default host remains https://app.supermetry.com from config/analytics.json. A separately configured SUPERMETRY_APP_HOST selects an operator's instance. The owner-configured packaged destination is https://app.voicestudio.sh/api/public/v1/app-events, owner-set 2026-11-16; any other production host needs owner approval. Release workflows read the repository variables SUPERMETRY_APP_ID, SUPERMETRY_INGEST_KEY and SUPERMETRY_APP_HOST, falling back to repository secrets of the same names; the ingest key is write-only and ships in the app. The app intake stores installation and optional session IDs only as keyed hashes. Without valid app credentials, the existing public events request (/api/public/v1/events, event diagnostic_report) and site remain unchanged. Usage analytics keeps its existing destination and request format. config/diagnostics.json is the shared closed metadata policy. Reports contain technical categories/codes, app/build/platform labels, known first-party source locations, bounded counters and a separately generated random installation ID. They never contain audio, text/transcripts, user names, filenames or home paths, credentials, URLs, logs, exception messages, source lines, local variables, memory dumps, screenshots or manual report bodies. Fingerprints use sanitized technical fields only. No cookies, user credentials or referrer accompany requests; only app intake uses its write-only app key. Redirects are refused. Diagnostic destinations require HTTPS; HTTP is allowed only for literal localhost, 127.0.0.1 or ::1 test collectors.

Electron main owns consent and transmission independently of Python. Its private diagnostics directory contains consent and sanitized queued records. Managed Python only writes an epoch-stamped local inbox; it makes no diagnostic uploads. Source and Docker backends own their own consent and sender through the administrator-gated diagnostics settings API. A browser connected to a managed desktop changes consent in that desktop. Uncaught/fatal handlers never wait for network delivery or flush a crash dump.

The durable queue is limited to 64 records and 512 KiB per installation, expires after seven days, coalesces repeated failures and uses bounded retries with stable report IDs. Desktop allocates that total between main and the Python inbox. Recovered crashes retain the original accepted fault or heartbeat timestamp; restarting does not extend their retention. Turning reports off stops delivery immediately, cancels active transport, invalidates the consent epoch and purges queued reports. A new opt-in creates a new identity; old queues and pre-consent crash history do not become eligible. Locked retired files are retried every minute. Collection under a new opt-in waits for that cleanup, so a delayed purge cannot delete new reports. Completed-delivery bookkeeping stays local and prevents repeated uploads while producer files are locked. If completion cannot be saved before an abrupt exit, a restart may retry the same stable report ID; exactly-once delivery is not promised. Within a run, rereading the same grant after a storage error preserves unsaved reports and completed-delivery bookkeeping. App-intake batches contain at most 20 events and 16 KiB of encoded JSON. Acknowledged reports and permanent too_old, invalid, or duplicate drops are retired; server throttling and transient failures retain stable IDs for bounded retries. An unauthorized app key pauses sending for the rest of that app run and retains the queue. Delivery results contain only fixed local status codes and counts. Temporarily unread consent stops collection and delivery while saved evidence waits for a bounded retry. Phone Settings and iOS crash/upload handlers follow the recovered current grant; an early Retry never presents unread consent as a saved OFF choice. Automatic read retries wait one minute, while a user-clicked Retry may reread sooner. Saving a choice never replaces unread shared preferences; a hard-disabled launch still invalidates an unread old grant before it can resume. Reopening a saved failed job only displays its old error; it does not create a new report under a later consent choice. Resumed operations report only failures proved fresh under the same grant. The existing job-events API supplies an indexed event_watermark (true retained-tail sequence and server time) from the same read snapshot as its bounded ascending page. Verification starts after that watermark; it never scans all history. Older servers without the watermark still show job progress, but cannot authorize resumed-history diagnostics. Fresh tasks submitted under the current grant retain their independent submission proof. Changing consent in another window invalidates delayed captures. Uncaught startup errors whose reporting permission was unknown remain local; a later opt-in cannot authorize them retrospectively. Requests, optional installs, native screen actions, uploads and player/picker callbacks retain the permission observed before their work starts. Restored callbacks without that observation stay local. Python exceptions retain their original permission through interpreter/thread hooks after cleanup; delayed desktop setup/start attempts and child-exit recovery follow the same rule. Runtime/file IPC work observes permission at entry. Its asynchronous descendants retain that observation through Node's fatal-error and unhandled-rejection hooks. Uncaught faults with unknown provenance remain in the local journal rather than borrowing permission at delivery time or after restart. A consent-stamped process marker can still recover a coarse unclean exit, but raw journal timestamps do not authorize fault details or frames. Confirmation actions retain the choice observed at the click; an unrelated or rewrapped promise cannot acquire a newer grant. Seeking an existing phone playback does not renew its grant. Permission dialogs, voice-edit waits, gallery loading and stalled transfers do not renew an operation's permission. A cached phone download failure from an earlier transfer remains local, including after Pause and Resume; a genuinely new transfer retains the permission observed when that transfer starts. Temporary queue-read failures preserve current in-memory reports and completed delivery decisions. Crash recovery retains its report until the prior session marker can be replaced; orderly close removes only a marker owned by that process. A failed consent write is shown with Retry and leaves the current process off. On iOS, unread diagnostics consent also keeps reporting off and offers Retry. After the initial privacy check finishes, that read failure does not block opening shared files, shared text, or workspace shortcuts. Where storage permits, withdrawal replaces the previous grant or writes a private revocation marker so restart also remains off. If every write or restoration fails, durable withdrawal cannot be guaranteed; the save fails instead of acknowledging a persisted choice. After a choice is saved successfully, a fresh status read can observe another client's new opt-in. Failed or pending writes remain locally closed; an older operation cannot acquire that fresh grant.

In source/Docker builds, switching off closes cached permission and cancels the sender before waiting on storage. The Settings request completes only after saving that choice, or returns a retryable save error. Slow storage can keep the save pending without allowing collection or delivery to resume. Permission-denied storage errors are not confused with desktop-managed consent. Desktop crash-session recovery uses private local IDs. Separate app-intake session pings are implemented behind diagnostics.sessionPings, false by default in generated build metadata. They additionally require diagnostics consent and are never enabled by usage-analytics consent. An explicit build setting SUPERMETRY_DIAGNOSTICS_SESSION_PINGS=1 enables this owner-decision flag. When enabled, a run sends one random session ID in a session event and its new failure reports; queued reports from earlier runs never acquire the new run's ID. Turning consent off retires that correlation for the rest of the run. With the flag off, there are no session pings or session IDs in failure requests; Supermetry can calculate crash-free users, without crash-free sessions. The consent copy is unchanged.

OMNIVOICE_DIAGNOSTICS_DISABLED or OMNIVOICE_ANALYTICS_DISABLED set to 1, true, yes or on disables automatic reports. Loading either hard-disable switch also revokes a previous diagnostics grant and purges its queue; removing the switch requires a fresh Settings opt-in. The usage-analytics consent choice is not changed. SUPERMETRY_APP_ID, SUPERMETRY_INGEST_KEY, and SUPERMETRY_APP_HOST override the diagnostics-only generated configuration at runtime, including explicitly empty or invalid values. Invalid app credentials select the unchanged fallback. Builds generate git-ignored config/supermetry-app.generated.json and bundle it with the backend; Electron main reads that packaged metadata, falling back to embedded build metadata when the file is absent, and native builds generate an equivalent private resource. These values are never exposed through renderer configuration or the diagnostics status API. Fork and unconfigured local builds need no app credentials. Product analytics does not read them. SUPERMETRY_SITE_ID, SUPERMETRY_HOST and, for the desktop/native fallback sender, SUPERMETRY_APP_URL support developer destinations; an empty site disables reporting. Tests use isolated storage and local/fake transports, never the production site.

Both native phone apps expose the same separate consent and Privacy control. Their application-owned reporter covers phone and paired-host operation failures; OS-provided crash/hang information is reduced to the same metadata policy and recovered only for its original enabled consent epoch. iOS MetricKit windows must begin after opt-in; Android historical exits must carry the app's own matching consent summary. OS scheduling and availability limit when these summaries arrive. Result Report sheets are a separate, explicitly submitted content-report feature and are never included in automatic diagnostics.

Translation privacy classification is shared with Tauri. Unknown or unavailable backend data does not claim offline operation. The existing libretranslate identifier is an alias for the backend's local Argos branch, not an assumption about an external LibreTranslate service. The Translation link opens engine settings.

Retention supports 0 (unlimited), validates whole-number limits up to 100000 and explains that cleanup removes old unstarred takes and their audio after generation. A more restrictive cap requires inline confirmation. Raising the cap or disabling cleanup saves directly. No records or files are deleted by simply opening settings or editing the input.

electron/tests/privacy-settings-smoke.mjs covers consent, watermark changes, failed writes, availability, retention confirmation/cancel, unlimited retention and reload with mocked endpoints. Tauri privacy regressions pass after sharing classification. Live read-only checks verified the watermark, analytics and retention schemas without changing the user's privacy preferences or deleting data.

AudioSeal receives 16 kHz audio for embedding and detection, including with AudioSeal 0.2, which no longer resamples internally. Processing remains bounded by chunk size. Only the watermark residual is resampled back and added to the original signal, retaining its sample rate, length, and high-frequency content.

Pending analytics initialization is invalidated when consent is withdrawn, including a consent lookup or SDK import already in flight. Renderer reports retain only validated stack frames; multiline error messages are discarded.

Backend lifecycle analytics runs in a background worker, so analytics file I/O does not delay listening or backend readiness. Crash detection still records its run sentinel before initialization, and version bookkeeping and uninstall-info cleanup still run with consent off. A later consent choice invalidates queued startup events and metadata writes prepared under the previous choice.

If the AudioSeal generator fails to load, generation continues without an invisible watermark. Further load attempts wait ten minutes, avoiding repeated network timeouts on each render; a later render retries automatically. This cooldown does not change startup's default cache-only policy; explicitly setting OMNIVOICE_PRELOAD_WATERMARK=1 allows startup to download an uncached checkpoint.