Aether Desktop 1.2.4

What's new

Upgrade notice: 1.2.4 brings the mobile edition's AI assistant to Windows — including a full chat page whose assistant can apply tuning settings for you — rebuilds the entire settings area as the mobile hub-and-subpage menu, bundles Aether Core 1.9.0, and fixes a setting that never did anything: a pinned address range was handed to the engine and read by nobody. Saved profiles load unchanged and the desktop version is 1.2.4.

New in this release

An AI assistant, ported from Aether Mobile 1.2.9. A ✨ button next to every setting explains what that setting does on this machine, with a Did not understand? Ask the assistant footer that carries the question — and the explanation you just read — into the chat. Alongside it: a connection advisor that reads a redacted log excerpt, and a settings advisor whose output passes a hard allowlist before anything is written.

Chat is its own tab, with everything the mobile edition has. Four ready-made questions on the empty page, your message on screen the moment you press send, copy on any answer, editing a message you already sent, deleting one or several, Try again on a message that never went out, and Stop for an answer in flight. A failed request becomes a retryable bubble with a translated sentence and Google's raw text kept as detail, not a line in an error bar. Nothing is ever deleted without a confirmation that names the count.

The assistant can change settings for you. Ask for a change and the answer arrives with a proposal card: every entry as setting: old → new with the model's own one-line reason, an Apply button, and — once applied — a note that tunnel settings reach the engine at start-up, so a reconnect is needed. Two rules make this safe: the proposal is validated against the allowlist before it is drawn, so a change the app would refuse never appears as a button; and the chat's allowlist is deliberately narrower than the advisor's. The network backend, the upstream proxy, routing rules, split tunnelling, a manual endpoint, LAN sharing, the kill switch and every credential are not writable from a conversation, because those decide which traffic is protected and where it goes. Tuning — protocol, obfuscation strength, MTU, fragmentation, keepalive, DNS, IP version, reconnect behaviour — is.

The settings area is now the mobile menu. A hub of grouped rows — icon, title, subtitle, current value, chevron, and a per-row ✨ — opening subpages, with the old flat Advanced page gone and the reset row as a separate confirmed action. Hub and subpages are rendered from one section definition, so no control exists twice and storage behaviour cannot drift between the two.

Aether Core 1.8.0 → 1.9.0. A real three-way merge — upstream 1.9.0, this repository's patched 1.8.0, and the recorded 1.8.0 baseline — with 20 conflicts resolved and none left; the merged core passes its own suite at 265 tests. Upstream absorbed the 1.2.3 throughput work, so the masque_h2.rs patch was dropped instead of carried; keeping it would have meant maintaining a fork of code upstream now owns. What upstream did not absorb still ships as a patch and now carries AETHER-APP-PATCH markers in the source: CUBIC selection in smoltcp (pinned by the socket-tcp-cubic feature), the packet-queue depth cap, and the split SO_RCVBUF/SO_SNDBUF budgets.

Fixed: endpoint mode Manual range had no effect. The app sent AETHER_SCAN_CIDRS, AETHER_MASQUE_CIDRS and AETHER_WG_CIDRS, and no core version has ever read them, so the scan swept its own built-in ranges instead. Both scanners now honour a pinned range, invalid entries are dropped rather than silently collapsed to a single address, and built-in seed addresses outside the range are no longer probed.

Upgrade note: saved profiles load untouched. The AI layer is inert until you enter a key, and every AI control is additive — no existing setting changed its default, its name, or its meaning.

How the AI layer is contained

Six boundaries, because this is a censorship-circumvention tool and an AI feature that leaks is worse than no AI feature:

• the Gemini key is sealed with DPAPI in secrets.bin, separate from profile.json, so Reset all settings does not take it and a copied file is worthless elsewhere;
• every request is dialled through the tunnel's own local SOCKS5 proxy with the host name sent as ATYP=DOMAIN, so the exit resolves the API host instead of the operator's resolver, and a gate refuses to send anything while the tunnel is down;
• the log excerpt is redacted, not just truncated: secret-shaped strings replaced (the user's own key included), public IPv4 masked to /16 and IPv6 to /32, WARP enrolment ids and bare UUIDs dropped, loopback and private ranges kept, and a hard size cap;
• one model allowlist is applied at all four points where a model id can reach a request — fresh list, cached list, default pick, and the id sent to generateContent;
• the model cannot write security settings: only allowlisted keys are applied, an unknown key is rejected and logged, accessSecret and accessToken are deliberately absent, and every value is type- and range-checked before the profile's own normalize;
• the chat's allowlist is narrower still, and a proposal is nothing until you press Apply — the model proposes, you decide, and the app writes through the single gate that also persists and revises the profile.