Give DodoSSH a phone, and a shared shell for both heads to drive

The Android head from docs/android-port.md, taken as far as its step 6.

Step 3, the spike, is answered and its throwaway screen is gone: libsodium.so and
libe_sqlite3.so are both in the arm64 APK, so NSec resolves its native half on Android
despite shipping no Android build, and the local cache opens. Two findings the audit
could not have had: Avalonia.Controls.WebView only ships net10.0-android36.0, which
settles the open "which Android versions" question at targetSdk 36; and Android has
blocked cleartext HTTP since API 28, so the terminal renderer needs a network security
config scoped to 127.0.0.1 or the WebView loads nothing.

DodoSSH.Client.Shell is new and is why the phone can exist: the view models, the terminal
renderer files and the palette moved there so both heads drive one state machine and draw
from one set of tokens. The desktop head is otherwise untouched and its 144 tests still
pass.

The platform pieces behind interfaces that already existed: the profile directory from
filesDir, a device key wrapped by a StrongBox-backed key that a fingerprint releases, and
a foreground service so a shell outliving a vault lock stays true on a platform that
stops backgrounded processes.

Sign-in is deliberately absent rather than approximated. It needs an app link, because
reusing the desktop loopback listener is the attack RFC 8252 section 8.3 names.
This commit is contained in:
2026-07-31 20:58:48 +02:00
parent 03e902a2d2
commit fe9d7fc289
65 changed files with 3034 additions and 103 deletions
@@ -0,0 +1,330 @@
'use strict';
/*
The renderer half of the terminal data plane.
Two things here are load-bearing and easy to get wrong:
1. Output is acknowledged from term.write's completion callback, never on receipt. The
acknowledgement returns flow-control credit to the host, so acknowledging early would tell
the host the screen has caught up when it has not — and the whole point of the credit
window is that it reflects what has actually been rendered. Acknowledge on receipt and a
remote running `yes` grows this page's memory until the tab dies.
2. Output is written as a Uint8Array, not a string. xterm decodes UTF-8 itself and carries
partial sequences across writes. Decoding here would corrupt any multi-byte character that
happened to straddle a frame boundary, which shows up as occasional mojibake in exactly the
conditions that are hardest to reproduce.
*/
const SERVER_OUTPUT = 1;
const SERVER_SESSION_OPENED = 2;
const SERVER_SESSION_CLOSED = 3;
const SERVER_SESSION_ACTIVATED = 4;
const SERVER_SESSION_REMOVED = 5;
const CLIENT_INPUT = 1;
const CLIENT_ACKNOWLEDGE = 2;
const CLIENT_RESIZE = 3;
const HEADER_LENGTH = 5;
const SCROLLBACK_LINES = 5000;
/*
The way out of the terminal, for someone using only a keyboard.
It has to be handled here rather than by the host: once this page's window owns Win32 focus, the
host's Avalonia window receives no key events at all, so nothing on that side could hear a shortcut.
Ctrl+Shift+F6 rather than Escape. F6 is the Windows convention for moving to the next pane, but a
bare F6 is a real terminal key that TUIs bind — as is Escape, which vim alone rules out. Ctrl+Shift
is the range terminal emulators conventionally keep for themselves and never forward to the remote,
so qualifying F6 with it keeps the convention without taking a key away from the remote shell.
*/
const RELEASE_FOCUS_MESSAGE = 'dodossh.release-focus';
const root = document.getElementById('root');
const statusBanner = document.getElementById('status');
/** @type {Map<number, {term: object, fit: object, pane: HTMLElement}>} */
const sessions = new Map();
/** @type {WebSocket | null} */
let socket = null;
function setStatus(text) {
statusBanner.textContent = text ?? '';
}
/** Builds a frame: opcode, big-endian session id, then payload. */
function frame(opcode, sessionId, payload) {
const body = payload ?? new Uint8Array(0);
const buffer = new ArrayBuffer(HEADER_LENGTH + body.length);
const view = new DataView(buffer);
view.setUint8(0, opcode);
view.setUint32(1, sessionId);
new Uint8Array(buffer, HEADER_LENGTH).set(body);
return buffer;
}
function send(opcode, sessionId, payload) {
if (socket && socket.readyState === WebSocket.OPEN) {
socket.send(frame(opcode, sessionId, payload));
}
}
function sendAcknowledgement(sessionId, byteCount) {
const payload = new Uint8Array(4);
new DataView(payload.buffer).setUint32(0, byteCount);
send(CLIENT_ACKNOWLEDGE, sessionId, payload);
}
function sendResize(sessionId, term, pane) {
const payload = new Uint8Array(8);
const view = new DataView(payload.buffer);
// Columns before rows, matching the SSH request. Swapping them produces a terminal that is 24
// columns by 80 rows, which reads as a rendering bug rather than a protocol one.
view.setUint16(0, term.cols);
view.setUint16(2, term.rows);
view.setUint16(4, pane.clientWidth);
view.setUint16(6, pane.clientHeight);
send(CLIENT_RESIZE, sessionId, payload);
}
/**
* Asks the host to take keyboard focus back.
*
* Optional by design: the bridge only exists under a real embedded WebView, and this page is also
* openable in a plain browser for debugging, where there is no host to ask.
*/
function releaseFocusToHost() {
window.chrome?.webview?.postMessage(RELEASE_FOCUS_MESSAGE);
}
/**
* Swallows the release-focus shortcut so it never reaches the remote.
*
* Returning false stops xterm processing the event, which is what keeps the chord from being encoded
* and written to the pty.
*/
function handleKey(event) {
if (event.type === 'keydown' && event.ctrlKey && event.shiftKey && event.key === 'F6') {
releaseFocusToHost();
return false;
}
return true;
}
function createSession(sessionId) {
const pane = document.createElement('div');
pane.className = 'pane';
pane.dataset.sessionId = String(sessionId);
root.appendChild(pane);
const term = new Terminal({
allowProposedApi: true,
convertEol: false,
cursorBlink: true,
fontSize: 13,
scrollback: SCROLLBACK_LINES,
// Matches terminal.css, so the canvas and the page agree on the background.
theme: { background: '#10131a', foreground: '#d5d8de' },
});
const fit = new FitAddon.FitAddon();
term.loadAddon(fit);
term.open(pane);
// WebGL where it is available. Falling back rather than failing matters because a software
// renderer is slow but usable, whereas a blank pane is not — and remote desktops and VMs
// routinely have no usable GPU context.
try {
term.loadAddon(new WebglAddon.WebglAddon());
} catch (error) {
console.warn('WebGL renderer unavailable; falling back to canvas.', error);
}
term.attachCustomKeyEventHandler(handleKey);
term.onData((data) => {
send(CLIENT_INPUT, sessionId, new TextEncoder().encode(data));
});
term.onResize(() => sendResize(sessionId, term, pane));
const session = { term, fit, pane };
sessions.set(sessionId, session);
activate(sessionId);
resize(session, sessionId);
return session;
}
function activate(sessionId) {
for (const [id, session] of sessions) {
session.pane.dataset.active = String(id === sessionId);
}
const active = sessions.get(sessionId);
if (active) {
active.term.focus();
}
}
// Below this, a pane is not being looked at — it is minimised or dragged to nothing. Fitting anyway would
// be actively harmful rather than merely useless: the fit addon floors its proposal at 2 columns by 1 row,
// so a degenerate viewport reflows the *remote* pty to 2x1 through window-change, and the wrapped
// scrollback that produces cannot be recovered when the pane comes back. A guard rather than a fix for one
// caller, because more than one path reaches here: a minimised window, and a splitter dragged to the edge
// once splits land.
//
// It is *not* what protects the vault's lock screen, which an earlier version of this comment claimed.
// Collapsing the host's WebView hides a native child window without resizing it, so this page's viewport
// does not change, no observer fires and this function is never called — measured with a live shell, and
// confirmed by removing the guard and finding the lock cycle equally clean. See docs/platform-flags.md.
const MINIMUM_FITTABLE_PIXELS = 40;
function resize(session, sessionId) {
const pane = session.pane;
if (pane.clientWidth < MINIMUM_FITTABLE_PIXELS || pane.clientHeight < MINIMUM_FITTABLE_PIXELS) {
return;
}
// fit() throws if the pane has no layout yet, which happens on the very first frame.
try {
session.fit.fit();
sendResize(sessionId, session.term, session.pane);
} catch (error) {
console.warn('Could not fit the terminal yet.', error);
}
}
function handleFrame(buffer) {
if (buffer.byteLength < HEADER_LENGTH) {
return;
}
const view = new DataView(buffer);
const opcode = view.getUint8(0);
const sessionId = view.getUint32(1);
const payload = new Uint8Array(buffer, HEADER_LENGTH);
switch (opcode) {
case SERVER_SESSION_OPENED:
createSession(sessionId);
setStatus('');
break;
case SERVER_OUTPUT: {
const session = sessions.get(sessionId) ?? createSession(sessionId);
// The callback is the acknowledgement, and payload.length is the byte count the host
// reserved credit for. Both must be the raw byte length, not a decoded character count.
session.term.write(payload, () => sendAcknowledgement(sessionId, payload.length));
break;
}
case SERVER_SESSION_ACTIVATED: {
const session = sessions.get(sessionId);
// Ignored for a pane that does not exist. The host sends this when a tab is selected, and a tab
// whose session ended still has its pane — but one the host knows about and this page has not
// created yet cannot be shown, and inventing an empty terminal for it would be worse than waiting
// for the SessionOpened frame that is already on its way.
if (!session) {
break;
}
activate(sessionId);
// Refitted on activation, not only on resize. A hidden pane has no layout, so every resize while
// it was hidden was skipped by the guard in resize() — meaning it comes back holding whatever
// geometry it had when it was last visible, and the remote pty is still sized to match.
resize(session, sessionId);
break;
}
case SERVER_SESSION_REMOVED: {
const session = sessions.get(sessionId);
if (!session) {
break;
}
/*
The tab is gone, so the pane goes with it — and this is the only place that is true. A shell that
ended on its own keeps its pane, because the last thing the remote said is usually why it ended;
a tab the user closed has nothing left to read.
term.dispose() is what actually matters. It releases the WebGL context, and a browser hands out
about sixteen of those: without this, a day of opening and closing terminals ends with panes that
cannot get a renderer, and nothing outside this page would ever say why.
*/
session.term.dispose();
session.pane.remove();
sessions.delete(sessionId);
setStatus('');
break;
}
case SERVER_SESSION_CLOSED: {
const session = sessions.get(sessionId);
const reason = new TextDecoder().decode(payload);
if (session) {
// The pane and its scrollback stay. The user was probably reading the last thing the
// remote said, and that is usually why the session ended.
session.term.write(`\r\n\x1b[38;5;244m── ${reason} ──\x1b[0m\r\n`);
session.term.options.cursorBlink = false;
}
setStatus(reason);
break;
}
default:
// A newer host than this page. Ignored rather than fatal.
break;
}
}
function connect() {
const token = root.dataset.token;
const url = root.dataset.socket;
// The token travels as a subprotocol rather than a query parameter, which keeps it out of
// anything that logs URLs.
socket = new WebSocket(url, ['dodossh.terminal.v1', `token.${token}`]);
socket.binaryType = 'arraybuffer';
socket.addEventListener('open', () => setStatus(''));
socket.addEventListener('message', (event) => handleFrame(event.data));
socket.addEventListener('close', () => {
setStatus('Disconnected from DodoSSH.');
});
socket.addEventListener('error', () => {
setStatus('The terminal connection failed.');
});
}
// One observer for the whole root rather than one per pane: resizes arrive in bursts while a
// window is being dragged, and a single callback coalesces them for free.
new ResizeObserver(() => {
for (const [sessionId, session] of sessions) {
resize(session, sessionId);
}
}).observe(root);
window.addEventListener('beforeunload', () => socket?.close());
connect();