e00c3bd413e0da2ddac539d9f67adc41a47f36ea
2
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 lockbitchat
|
e00c3bd413 |
feat(groups): group chats, and a mesh rather than a star; release v6.1.1
A group is an orchestration layer over the pairwise sessions the app already holds. It owns no transport and no shared key: every frame leaves over a chat that is already SAS-verified and already ratcheted, so a removed member simply stops being sent anything. Membership is a roster the admin signs, ordered by epoch, and the safety code is a commit-then-reveal round over every member's fingerprint and nonce. Delivery was the part that did not match its own description. The admin held a link to everyone and nobody else held a link to anybody, so the relay path — the documented fallback — was in fact the entire topology, and the admin going away partitioned the group. Now, once the code is confirmed, each pair without a link dials one over that relay path. The descriptors are compact enough to ride a group frame and are signed with the sender's group identity key, so the relaying member can drop a dial but cannot substitute one. The member with the smaller fingerprint dials, which is the whole glare protocol. Mesh links are released without a human comparing digits. Twenty-eight codes for a group of eight is not a check anyone performs; the guarantee moves rather than disappears, since the descriptor was signed by a key the signed roster names and the group code covers. markGroupLinkVerified refuses any session whose in-band exchange has not completed and whose peer has not proved possession of that key. An existing 1:1 chat between two members is adopted instead of re-dialled, via a probe bound to that session's own key fingerprint so it cannot be replayed onto another chat to impersonate its author. Security fix: g_hello was accepted on any session from anyone who knew the group id, so any member could publish an identity the admin never invited and have the admin sign and broadcast a roster containing it. It is now accepted only on a session an invitation went out on, which also confines it to a direct link. Mesh connections are kept out of the chat registry and muted from the document events the header listens to, so a routing detail cannot tear down the display of a conversation the user actually opened. |
||
|
lockbitchat
|
2a7142c722 |
feat(webrtc): recover a dropped connection without a signalling server; release v5.6.0
A chat no longer dies when the network moves under it. A NAT rebind, a lift, a Wi-Fi radio parking itself, a phone that dozed: the session repairs its own network path in place, and the messages typed meanwhile go out when it returns. Recovery is an ICE restart, which renegotiates only the transport path — the DTLS handshake, the session keys and the SCTP association carrying the data channel all sit above ICE and survive it. The renegotiation SDP therefore travels over the existing end-to-end encrypted, SAS-verified channel: no signalling service enters the design, and an attacker who cannot already decrypt the session cannot inject a reconnection. A restart is refused outright unless the DTLS fingerprint in the incoming SDP matches the live session's, so recovery can never re-point a conversation at a different peer. When the path is gone for good the session is ended and its data wiped rather than left half-alive: with no server there is nothing to re-signal through, and a conversation whose transport is gone should not leave its plaintext in an open tab. The two cases where that is already certain are recognised in seconds instead of being retried for two minutes — a channel that has delivered nothing at all since the drop cannot carry a renegotiation, and an ICE agent left bound to a network that no longer exists reports zero candidate pairs on every restart. Judging liveness was the hard part. Silence is not evidence of death: browsers freeze backgrounded tabs outright, and a frozen peer answers nothing while being perfectly healthy. What survives that freeze is ICE consent, which the browser runs in its network stack rather than on the page's thread — so a connected ICE state means a silent peer is asleep, and only a degraded one turns an unanswered probe into a teardown. The grace window before a restart is sized to the browser's own timings: 'disconnected' arrives after ~5s of missed consent responses and is held ~25s before 'failed', and that window exists for self-healing, so restarting at the start of it broke connections that were about to recover. Several long-standing bugs surfaced along the way and are fixed here: - handleHeartbeat() was dispatched to but never defined, so every inbound heartbeat threw a TypeError and peer liveness was never observed at all. - Heartbeats were folded into the 5-minute maintenance cycle instead of running on their own timer, far too coarse to notice a dead path. - ondatachannel can hand over a channel that is already open, so the answering side's 'open' event had been dispatched before the handler was assigned and never fired, leaving that side with no heartbeat, no watchdog and no file-transfer init. The peer whose network was fine kept showing "connected" indefinitely because nothing was running to notice. - Answering a heartbeat required the peer to have finished verifying, but the two sides confirm a SAS code at different moments; for that whole window one of them could not reply and was declared dead on a healthy connection. - Sending on a channel that was not ready returned in silence: the text stayed in the box, nothing was transmitted, and nothing said why. - The send path gated on navigator.onLine and the offline/online events, which report whether an interface exists rather than whether anything is reachable. A tab the OS froze misses the 'online' edge, and this side then queued every message forever: one tick on everything it sent, while incoming messages kept arriving. Sending is now decided by the data channel, and queues drain by polling rather than on an edge, so a missed event cannot strand them. - A false offline modal appeared on a working session, because the offline event was taken at face value. tests/session-recovery.test.mjs covers the state machine, the backoff and its serialisation, the offline hold, the sleeping-peer discriminator and the identity check. |