Product status and acceptance

From mnema-talk v3.3.13, rendered on this site.

This page defines product status independently of release numbering. A build, a passing unit test or an isolated cryptographic component cannot qualify an entire communication path.

CategoryMeaning
SupportedNormal maintained product path. Named tests and explicit limits apply.
ExperimentalAvailable to execute or test; outstanding acceptance prevents a production qualification.
BetaUsable and maintained, but outstanding acceptance (for example platform signing) keeps it out of Supported. Expect rough edges.
PlannedRequirements exist, but no qualified usable product path is available.

Status at 3.3

Mnema Talk 1.0 was the first release under the API stability contract: its public surfaces change incompatibly only in a major release. Every release up to 2.9 kept that contract and only added to it (one announced exception in 1.2.0 for automod report entries, see below); 2.0.0 was a major version for the built-in TURN relay and changed no API. 3.0.0 is a major release and contains the changes listed in api-stability.md. What the label does and does not say:

Supported versions. Security fixes go to the latest minor release and, once a newer minor exists, to the previous minor for 90 days (SECURITY.md). At 3.3.x that means 3.3.x and, for 90 days after 3.3.0 (released 2026-11-15), 3.2.x. CI upgrades from 0.8.16, 0.9.0, 1.0.0, 1.1.0, 2.0.0, 2.9.12, 2.9.19, 3.0.13, 3.2.1 and 3.3.5 to prove the upgrade path; that is a test, not a support promise. Per-version notes: upgrade.md.

AreaCategory at 3.3
Server (Docker, linux/amd64 and arm64), browser app, text chat, roles and permissions, webhooks and bots, notifications, search and onboarding, voice, cameras and screen share in Chromium-based browsersSupported
Direct messages (two to ten members), blocking and reports, saved messages, forwarding, thread follow, search filters v2, built-in slash commands, reading a Talk's chat without joiningSupported (surfaces introduced in 1.1)
2.x surfaces: the built-in TURN relay (2.0), the setup assistant and the guided first run with Test my connection (2.2), several attachments per message and the reworked forum with tag colours, rules and solved posts (2.2), the community icon and Talks started from a direct message (2.2), taking over a call from another device (2.2), camera layers with a per-viewer camera cap and an optional voice upload limit (2.5, 2.6), raid protection, the vanity invite link and self-assignable roles (2.7), event roles with a waitlist, event comments and recurring events (2.8), the wiki with pages, revisions, locks, restore, pictures and search (2.9)Supported (new in 2.x, released 2026-10-29 to 2026-11-04; covered by automated backend, web and end-to-end tests and database migrations 0077–0091)
3.x surfaces: bots added from a manifest with per-channel permissions, buttons, selects and forms on bot messages with interactions (3.1), access to every channel in one step for a bot, messages with a lifetime (timed messages), auto-delete in hours, days or weeks for attachments only or everything with a community default and a dry run, a placeholder for a removed file (3.2), member invites and inviting into a channel from its menu, commands restricted by the member's permissions, bots downloading attachments of messages they can see, the vanity link's single Link active switch (3.3). The official bot SDKs live in their own repositories (bot-sdk-js, bot-sdk-go)Supported (new in 3.x, released 2026-11-13 to 2026-11-15; covered by automated backend, web and end-to-end tests and database migrations 0105–0110)
Browser app on WebKit (Safari, iOS, installed iOS web apps) and Firefox: text chat, search, voice, cameras and screen share with sealed media, as tested in BrowsersSupported for text chat, search and administration. Sealed calls pass on macOS WebKit locally; the CI WebKit job cannot run them because Playwright's Linux WebKit build has no encoded-media API (RTCRtpScriptTransform); see the limits there (not yet tested on real iOS or Android hardware). WebKit occasionally refuses to clone the media key message (web/src/lib/media/guard.ts retries once; if the retry fails the line stays keyless until the next key, the call no longer aborts; seen on macOS WebKit CI runners, not verified on real Safari devices).
Operations: Prometheus metrics and alert rules, operator status page, backup and restore, upgrade tested from 0.8.16, 0.9.0, 1.0.0, 1.1.0, 2.0.0, 2.9.12, 2.9.19, 3.0.13, 3.2.1 and 3.3.5, cosign-signed images, self-update that verifies the image's signature and provenance before it restarts the appSupported
Native mobile apps, text end-to-end encryptionNot offered. Text stays readable by the server, by design. The browser app has a phone layout (channel list as a drawer, full-screen thread and member list, 44 px touch targets, safe areas) that is covered by the mobile-* e2e projects in Chromium and WebKit device emulation, not by tests on real phones (no real iOS/Android hardware qualification).
Desktop client on WindowsBeta (unsigned, no Authenticode; SmartScreen warns; sealed media in WebView2 not verified on real Windows hardware)
Desktop client on macOS and LinuxBeta (macOS: Apple silicon, ad-hoc signed, not notarised, Gatekeeper warns; Linux: Electron AppImage and .deb. Neither is qualified on real hardware. No gaming overlay. The Tauri host on Linux is not built)

Internal security reviews. Before 1.0 the code was reviewed in six lanes (authentication and sessions, permissions, content and media, integrations, realtime and voice, supply chain and updates). It found no Critical and no High issue; every Medium finding is fixed, and a short list was deliberately deferred (see the summary in SECURITY.md). The 1.2 additions went through the same kind of internal review (0 Critical, 0 High; see SECURITY.md and the audit package). These reviews were done by the project itself with automated help. They are not an independent or external security audit, and none has been commissioned.

API compatibility note for 1.2.0. 1.2.0 made message_id optional in GET /api/admin/reports, GET /api/admin/reports/{id} and POST /api/admin/reports/{id}/resolve (automod entries have no message) — a breaking change in a minor release, announced via the BREAKING CHANGE: footer; strict third-party clients must treat the field as optional (see upgrade.md and api-stability.md).

What 1.x does not claim: real 4K/60 capture, system audio from physical hardware, an external security audit, the 250-endpoint capacity target, load results on dedicated hardware, Authenticode or notarised desktop signing, and sealed media qualified in WebView2 on real hardware. These stay outstanding under "Remaining acceptance" below.

Supported browser and server path

The Vue browser app and Docker server for Linux amd64/arm64 are the normal product path. CI exercises backend/frontend regressions, browser end-to-end flows, advancing screen-share frames, image smoke tests, backup/restore with message and attachment checks, and an isolated PostgreSQL upgrade drill. These tests have different scopes: synthetic two-browser screen sharing does not establish real 4K/60 capture, system audio or 100-person capacity.

HTTPS protects browser traffic in transit. Chat messages and attachments stay readable by the server. Microphone, camera, screen video and screen audio are encrypted in the browser with MLS-derived keys; an honest SFU forwards those payloads and cannot read them. From 0.8 on the keys rotate every 30 minutes, members silent for 20 minutes are removed and re-admitted, every sealed frame is signed with its sender's leaf key, and the connection details show a safety code to compare. An active or compromised server is not excluded (it serves the code and asserts group membership), and metadata such as who speaks when stays visible. The transport underneath is DTLS-SRTP. See SECURITY.md.

Other capabilities on this path, each with automated tests:

  • Screen-share layers (source up to 4K at 60 FPS plus smaller layers) with a per-viewer choice or Auto; the browser tests switch layers mid-stream and check that other viewers are not disturbed.
  • Roles and permissions with a protected Owner, overwrites and private channels and Talks, enforced in REST, the hub, media serving and the SFU.
  • Incoming and outgoing webhooks with an SSRF guard for outgoing delivery.
  • Operations: Prometheus metrics, the live operator view, alert examples, cosign-signed release images, and the backup, upgrade (release to release) and PostgreSQL upgrade drills in CI.

Not covered by these tests: real 4K/60 capture, system audio from physical hardware, an external security audit and capacity beyond the recorded runs. Internal security reviews preceded 1.0 and 1.2; the 3.0.x patches came out of an internal defect hunt in November 2026 (its open items are listed under known open defects). None is an independent audit.

Browsers

Sealed media needs three things in the browser: a frame transform that reaches encoded media (RTCRtpScriptTransform), Ed25519 in WebCrypto and a module worker. Chrome and Edge 137+, Safari 17+ (including iOS 17+ and installed iOS web apps) and Firefox 129+ have them. A browser without them does not join a call at all: the join is refused with a message, no microphone is requested and nothing is sent. Media is never sent unsealed, in any engine.

What the end-to-end suite showed (Playwright with the engines it ships: Chromium 156, WebKit 27.2, Firefox 157, two members per call with fake devices; run make e2e or see development):

NotesChromium (Chrome, Edge)WebKit (Safari, iOS)Firefox
Sign-in, chat, search, roles, moderation, webhooks, notificationsSupportedSupportedSupported
Microphone and camera, sealed on every line, with key rotation, late join and reloadSupportedPasses on macOS WebKit; in CI only in e2e (webkit-macos-calls) (Playwright's Linux WebKit has no RTCRtpScriptTransform, so the app refuses the join there, see below)Supported
Screen share with viewer-chosen layers (simulcast)SupportedPasses on macOS WebKit; in CI only in e2e (webkit-macos-calls)Supported (VP8)
Screen share frames advance for the viewerSupportedPasses on macOS WebKit; in CI only in e2e (webkit-macos-calls)Supported
Required CI checkYesNo (e2e (webkit 1/2) and e2e (webkit 2/2): every spec that does not join a call, plus the refusal spec)No (e2e (firefox), smoke and sealed-call subset)
Not covered by the suiteReal Safari and iOS hardware; calls on Linux WebKit (no frame transform in that build, they pass on macOS WebKit)Real screen capture

A real-device call (real camera and microphone, real Google Chrome on macOS, sealed both ways) was verified on 2026-10-22.

What it took:

  • Firefox, screen share. Screen lines asked for H.264 first. Firefox lists H.264 but encodes it with a downloadable OpenH264 module; without it the encoder never produced a frame and the share stayed black without an error. Firefox now keeps VP8, which it encodes itself (sealed the same way).
  • All engines, joining. A browser that cannot seal media is refused at the join (before any device is requested) instead of joining with microphone, camera and screen switched off.
  • Linux WebKit in CI cannot take part in a sealed call. This is the engine build, not the app. Playwright's Linux WebKit (26.6 up to Playwright 1.63, 27.2 since 1.64, in both cases the version of its macOS build; the pinned version is in e2e/package.json) reports RTCPeerConnection and getUserMedia, but RTCRtpScriptTransform, RTCRtpSender.prototype.createEncodedStreams and RTCEncodedAudioFrame are all undefined (checked on x64 and arm64 Linux with both versions; the macOS build has RTCRtpScriptTransform and RTCEncodedAudioFrame). With no way to reach encoded media the app does what it promises for such a browser: it refuses the join with a message, requests no microphone and sends nothing (browser-sealing.spec.ts, second spec, runs in the WebKit CI job). The first spec of that file skips itself only on Linux WebKit and only while RTCRtpScriptTransform is missing, so a build that gains it runs the call again. Every spec that joins a call (calls, screen share, Talk chat, roles rights) stays out of e2e (webkit 1/2) and e2e (webkit 2/2) and runs in six short e2e (webkit-macos-calls, <group>) jobs on macOS runners, not a required check.
  • WebKit 27.2, <select> and the style policy. Playwright's WebKit since 1.64 (27.2) reports one style-src-attr violation for every <select> a page builds (empty sample, the script that created the select as source; the browser's own shadow tree sets an inline style). Nothing is refused: a page with selects renders pixel-identically under 'unsafe-inline' and under style-src-attr 'none', and 26.6 does not report it. The policy is left as it is; the e2e fixtures ignore this one signature in WebKit only (the Chromium and Firefox jobs run the same code and still fail on any style-src-attr event). Not yet checked in a release Safari 27.
  • Firefox, navigations the driver loses. On Linux Firefox (Playwright 1.63 and later, pinned in e2e/package.json) page.goto sometimes never resolves although the server answered 200 and document.readyState is complete: the driver's frame stays on about:blank (no navigated event; 3 to 12 of 40 navigations of a bare page on a 2-CPU container). Not specific to this app; the e2e fixtures retry the navigation once the first attempt has timed out (tests/navigation-diagnostics.ts).
  • WebKit media key cloning. WebKit occasionally refuses to clone the media key message (web/src/lib/media/guard.ts retries once; if the retry fails the line stays keyless until the next key, the call no longer aborts; seen on macOS WebKit CI runners, not verified on real Safari devices).
  • WebKit mutes the microphone of all but one page per browser process (as Safari does between tabs). That limits two members in one Safari window, not two devices; the suite therefore runs each WebKit member in its own process.

Stability: Chromium passed every spec on every run. Firefox passed the whole suite apart from the specs that need Chrome's DevTools protocol, and its CI subset (7 files at the time) 10 of 10. WebKit passed the suite apart from those too, but with one or two timing failures per full run on a loaded machine (a sealed call taking longer than the spec waits to re-key after a reload, a simulcast layer not up in time, a voice-chat leave); each passes when run alone. They are retried once in CI, and the WebKit and Firefox e2e jobs use one retry and are not required checks until they have gone green for a while.

The AI noise filter's worker pipeline processes audio on Firefox wherever the browser has an audio output device. CI's Linux runner has none, and Firefox then never starts an AudioContext (it stays suspended; resume() never settles), so nothing routed through the Web Audio graph is rendered. That used to send silence (the voice gate read a dead meter and stayed shut) without any notice. The page now watches the context: if it is not rendering after 3 s the call is rebuilt around it (the capture is sent as it is, the gate stays open outside push-to-talk) and the member is told once; a filter that processes no frame within 3 s of the context running is given up the same way (web/src/talk/lib/audioHealth.ts). The Firefox CI job covers both: audio-stall.spec.ts on the runner as it is, and late-join.spec.ts (filter running) after e2e/ci-audio-sink.sh gave it a null audio sink.

Known limits: Firefox's fake microphone is a steady tone, which the AI noise filter removes, so the specs measure decoded samples rather than loudness. Real-device behaviour of iOS audio sessions, background tabs and installed web apps is not covered by the suite (no real iOS or Android testing) and stays on the "Remaining acceptance" list.

Beta desktop path

The desktop client is Beta on Windows, macOS and Linux. It is unsigned by decision: no paid code-signing certificate will be bought (see installing the unsigned Beta). macOS is ad-hoc signed and not notarised, so Gatekeeper warns. Linux is an Electron AppImage and a .deb. None of the three is qualified on real hardware, and the gaming integration is Windows only. The Tauri host on Linux is not built. Everything below about the signed bundle, the security floor and updates applies to all three clients.

Windows

The desktop app (this section covers its Windows build) lives in the separate talk-desktop repository and is an unsigned Tauri DEV client, including the Gaming integration described in the usage guide. Its local instance selector checks HTTPS (or a loopback address) and the server's web_client_api and release; it then opens the Vue app from a signed bundle that the client verified (built into the installer or downloaded from this repository's releases, Ed25519 key pinned in the client) at the client's own origin. The client forwards /api/ to the instance, keeps the session cookie and relays the WebSocket, so the instance supplies community data and never executable code (design and threat model). The page receives only the WebSocket commands and the gaming integration's session-bound Talk-state sync; local gaming surfaces hold their separate display/control permissions. Status: unit-tested and run on macOS during development; the bundle verifier has a high-water mark, a security floor that rises with releases whose security fixes carry a Security-Floor: raise trailer and a session cookie kept in the OS keychain. Windows CI on the desktop repository (run 39423960431) checks the current web-client host. WebView2 presents the custom protocol as https://mnema.localhost/ for the selector, the instance window and the gaming surfaces. The instance window leaves the microphone on WebView2's own prompt; the check does not click it. The selector and the gaming surfaces are refused. The host passes a synthetic capture device, and each probe window has its own profile because a refusal is remembered for the origin. A human prompt, a physical device, sealed media and an Authenticode signature remain outstanding. The installer is unsigned.

Voice in the desktop client uses the instance's browser path, so microphone, camera and screen payloads are sealed in WebView2 by the same code as in a browser. This is not yet qualified on real Windows hardware.

Desktop builds come from the talk-desktop repository (releases of this repository stopped attaching them after v0.7.3). WebView2 Runtime is required on Windows. Checksums identify package bytes; they are not publisher signatures. The desktop installer of a release updates itself with signature-verified, newer-only updates; no Authenticode signing exists.

The desktop repository's own CI runs a limited Windows flow (instance selection, sign-in, a persisted message, a live message over the hub connection, sign-out) on generated test accounts and an isolated server.

These checks qualify only that limited flow. They do not qualify production-account usage, S3 uploads, physical microphone/camera capture, system audio, device recovery, E2EE, gaming, overlays, exclusive fullscreen or complete browser parity. The earlier feasibility probe and native-crypto checkpoint are research foundations, not alternate supported clients.

Remaining acceptance, in order

The Windows-first client and its Gaming integration have been released. Versions remain managed by the normal release pipeline. Further runtime qualification and the remaining requirements below are still outstanding.

The Gaming integration connects these surfaces to the existing client: left-hand sidepeek with participants, speaking, mute and screen-share state; Alt+M interactive overlay; controls for the actual own Talk; persisted, configurable mute, deafen, push-to-talk and overlay bindings. Both surfaces require an own connected Talk and an active detected game. See the Windows Gaming DEV usage and limits. Topmost windows do not establish exclusive fullscreen or anti-cheat support.

The remaining fixes and acceptance work follow:

  1. Client parity: Windows CI checks the trusted bundle's custom-protocol origin (https://mnema.localhost/), that the instance window leaves the microphone on WebView2's own prompt, and that the selector and the gaming surfaces refuse one. The check uses a synthetic capture device and does not click the prompt. A human prompt, a physical microphone or camera, display capture, sealed media in WebView2 on real Windows hardware and a signed installer remain outstanding (desktop = Beta, unsigned; Windows SmartScreen warns, macOS Gatekeeper warns). Desktop builds for macOS and Linux are published and have no hardware qualification either.
  2. Chat messages and attachments stay readable by the server and are not deleted by age (unless per-channel auto-delete is configured). Sealed media is verified in Chromium, WebKit and Firefox (see Browsers), and in Chromium for calls between current and one-release-older tabs. WebKit occasionally refuses to clone the media key message (web/src/lib/media/guard.ts retries once; if the retry fails the line stays keyless until the next key, the call no longer aborts; seen on macOS WebKit CI runners, not verified on real Safari devices). The e2e WebKit and Firefox jobs use one retry and are not required checks. Sealed media in WebView2 is not verified on real Windows hardware, no real iOS/Android testing exists, and no external security audit has reviewed the app.
  3. Capacity: make load-test is reproducible (smoke, the 100-member full profile, join and reload stall checks, packet loss and jitter on the lightweight viewers, a two-hour long profile). On a single machine make load-test full is not met: it misses its gap and heap thresholds while all 84 viewers stay connected, so what remains is recording results on dedicated hardware with more real decrypting viewers (capacity notes). A viewer's bad reports move only that viewer to a lower screen layer (TestOneViewersTroubleLeavesTheOtherOnTheHighLayer); voice forwarding does not follow screen layers. The additional 250-endpoint target is deferred.
  4. Independent operation. main is protected by a ruleset (pull request required, CI and CodeQL checks must pass, squash merge only). Review approvals are not required. Dependency advisories remain part of release checks; a green build does not mean zero risk.

Desktop requirements and the native research checkpoint describe these boundaries in more detail. Existing coverage and security gates remain in force.

Something wrong on this page? Tell us.