et@0.0.10

Add debug diagnostics for dropped server sessions (3110e1f)

ET_DEBUG=1 now gives etserver a durable machine-local log destination,
non-silent logging, and default verbosity 2. Operators can also use
ET_LOGDIR and ET_VERBOSE when CLI or INI settings do not override them.

Server diagnostics now record client accept/reject reasons, reconnect and
recovery outcomes, terminal disconnects, and session removal, making long-lived
session drops diagnosable from the server host. Client debug logs omit session
credentials, malformed handshakes are bounded by a five-second deadline and
128-connection pre-auth capacity, and completed connection handlers no longer
accumulate after rapid failed handshakes.

Pull request & contributors

• #23 feat(server): ET_DEBUG mode and session diagnostic logs by @minpeter

et@0.0.9

Recover from macOS post-sleep socket timeouts (8dd9dc9)

Changing networks or waking a MacBook can leave the existing TCP connection in
an unusable state. macOS reports some of those stale connections as
ETIMEDOUT (Operation timed out, os error 60) rather than EOF or a reset.
The client previously treated that live-transport error as fatal and exited,
which also exposed any terminal mode left by the remote program.

Every socket I/O failure from the live ET transport now enters the normal
reconnect path. Recovery retains the same remote session and keeps retrying
while Wi-Fi or routing returns; protocol, crypto, framing, backpressure, and
local-terminal failures remain fatal. This matches upstream Eternal Terminal's
behavior of reconnecting after any socket read/write error.

Pull request & contributors

• #21 fix: reconnect after macOS live socket timeouts by @minpeter

et@0.0.8

Trim replay backups with delivery acknowledgements on keep-alives (4b97fe8)

Replay backups previously only shrank at a fixed cap, so a single burst of
output (cat a large file) pinned the cap's worth of memory per session for
the rest of the daemon's uptime, and idle sessions slowly accumulated
keep-alive echoes.

Keep-alives now carry the sender's reader sequence as an 8-byte payload.
Every implementation ignores keep-alive payloads on receipt (verified
against upstream C++ TerminalServer/TerminalClient and released et.rs),
so the extension is invisible to legacy peers. An et.rs receiver uses it to
drop backup packets the peer has already consumed, keeping only the
unacknowledged tail plus a fixed slack — a session's replay backup now
returns to near-empty within one keep-alive interval instead of pinning the
connected cap forever.

Acknowledgements are strictly per-hop. Jumphosts relay packets verbatim
(upstream and et.rs alike), so the jumphost etserver consumes the client's
acknowledgement for its own connection and forwards a payload-less
keep-alive, and etterminal --jump attaches its own sequence towards the
destination. A foreign sequence that still arrives through a legacy C++
jumphost is absorbed by the retention slack.

Pull request & contributors

• #19 perf: trim replay backups with keep-alive delivery acknowledgements by @minpeter

Stop the daemon from retaining 64 MiB of replay history per live session (d70352b)

Every packet written to a client is kept in a per-session replay backup so a
reconnecting peer can catch up. That backup was only trimmed at upstream's
64 MiB / 262,144-packet cap and never expired otherwise, so a long-lived
etserver slowly pinned up to 64 MiB per session — with a couple dozen
sessions the daemon grew towards multiple gigabytes, exactly like the C++
server it replaces.

While the transport is connected, a reconnecting peer can only be missing
data that was in flight, which is bounded by kernel socket buffering
(≤ 4 MiB on default Linux autotuning). The connected backup is now trimmed
to 8 MiB / 32,768 packets and the deque returns its slack capacity after a
reconnect. The disconnected catch-up buffer keeps upstream's full 64 MiB,
and nothing on the wire changes: recovery, the handshake, and the
receive-side catch-up validation limits all stay byte-compatible with
upstream C++ peers.

Pull request & contributors

• #16 perf: trim the connected replay backup to 8 MiB per session by @minpeter

et@0.0.7

Survive network outages during reconnect and restore leaked terminal modes (048244d)

A laptop waking from sleep has no route to the server for several seconds.
The client attempted exactly one reconnect and treated its failure as fatal,
so a live session died with "could not reach the ET server: connection timed
out" the moment the first attempt raced the returning Wi-Fi. Transient
network failures (unreachable endpoint, DNS outages, connect/transport I/O
errors) are now retried every second until the link returns — matching
upstream ET, which keeps a session alive until the server ends it. The
client announces the retry loop once, and Ctrl-C gives up (raw mode turns
ISIG off, so the byte is read from stdin); protocol mismatches and server
rejections still fail immediately.

Separately, exiting only restored the local termios. Terminal modes a remote
application had enabled in the local emulator — the kitty keyboard protocol,
bracketed paste, mouse and focus reporting, the alternate screen — survived
the client, leaving the shell prompt printing key reports like 2618;9u as
garbage text. The client now emits the matching reset sequences when it
leaves raw mode; every reset is a no-op when the mode is off or unsupported.

Pull request & contributors

• #14 fix: retry reconnects through network outages and reset leaked terminal modes by @minpeter

et@0.0.6

Fix a port-forwarding deadlock under bidirectional load (66f38b0)

The session loops handed inbound tunnel packets to the forwarding worker with
a blocking send, while the worker hands outbound tunnel packets back through a
bounded queue that only those same session loops drain. Under sustained
bulk transfers in both directions the two bounded queues could fill at the
same time, leaving the session loop and the worker each waiting on the other —
a permanent wedge that also stopped keepalives and made reconnects impossible.

Session loops (client, server bridge, on both Unix and Windows) now hand
packets to the worker without blocking: when the worker is at capacity the
packet is held, no further session packets are read (preserving order), and
the loop keeps draining the worker's outbound queue — which is exactly what
frees the worker — retrying on a 10ms cadence.

Pull request & contributors

• #13 fix: break the forwarder channel deadlock cycle in session loops by @minpeter

Fix session teardown under local IPC backpressure (9c327b7)

set_nonblocking(true) applies to the socket, not the handle: every
try_clone() of a local stream shares the flag. The readiness-driven session
loops flip their local streams to non-blocking for reads, which silently made
the packet writers on cloned handles non-blocking too. When the kernel socket
buffer filled (a shell that stops reading input during a large paste, or a
bridge pausing terminal output while a client reconnects), write_all failed
with WouldBlock and tore the whole session down — killing the shell, failing
the server bridge, or silently ending a jump-host relay — and could leave a
truncated frame that corrupted the local packet stream.

write_local_packet now treats WouldBlock as backpressure: it resumes from
the partial write and retries until the peer drains, matching upstream's
blocking-write semantics on both Unix and Windows.

Pull request & contributors

• #11 fix: treat WouldBlock as backpressure in local packet writes by @minpeter

et@0.0.5

Fix macOS reconnect and PTY end-to-end failures (87486da)

Fix the remaining macOS portability failures in reconnect, terminal shutdown,
and background process-group cleanup.

The terminal launcher now resolves busybox-style symlinks before re-exec,
client and router loops detect Darwin socket closure reliably, and the PTY
fixtures no longer depend on interactive-shell timing. macOS CI now runs the
affected reconnect and PTY/process integration tests explicitly.

Pull request & contributors

• #9 Add macOS PTY fix release note by @minpeter

et@0.0.4

Fix client dying with "polling terminal streams: Interrupted system call (os error 4)" (768fd8f)

The client installs a process-wide SIGWINCH handler for terminal resizes, and
poll() is never auto-restarted by SA_RESTART: any signal delivered to the
thread blocked in poll() (a window resize, SIGCONT after job control,
SIGINFO from Ctrl-T on macOS) made it fail with EINTR, which the client
treated as fatal and tore the session down mid-use.

The terminal loop now retries poll() on EINTR, recomputing the keepalive
timeout from the absolute deadline so timing stays correct. The same retry was
applied to every other interruptible poll() in the workspace: the server
session bridge, the PTY worker, the jump-host bridge, and the port-forward
acceptor (which previously exited silently on EINTR).

Pull request & contributors

• #7 fix: retry poll() on EINTR so signals do not kill live sessions by @minpeter

et@0.0.3

Identify the et.rs port in --version output (ca7c581)

--version on every role (et, etserver, etterminal, htm, htmd) now
prints the et.rs identity and project URL:

et version 0.0.3 (et.rs)
A Rust port of Eternal Terminal
https://github.com/minpeter/et.rs

-V keeps the short upstream-compatible et version X.Y.Z line for scripts.
etterminal gains --version/-V support.

et@0.0.2

Fix reconnect against upstream C++ peers rejecting the session with "invalid recovery proof" (ad30f8b)

Reconnecting to an upstream C++ etserver (or accepting an upstream C++ et
client) failed with server sent an invalid recovery proof whenever the peer's
first packet after recovery was regular session traffic, such as terminal
output. et.rs required the first post-recovery packet to be an empty
keep-alive, a convention upstream does not follow.

Recovery authentication now accepts any packet that decrypts with the session
key (which is the actual proof) and requeues it for the session loop instead
of discarding it, so no traffic is lost.

Also fixes a port-forwarding race surfaced by this change: closing a forwarded
socket could drop data still queued for its writer thread. Teardown now drains
queued writes before shutting the socket down.

et@0.0.1

Initial release (e351bb4)

First automated release of et.rs: a single telemetry-free EternalTerminal binary (Rust port).

One binary serves every role: et (client), etserver, etterminal, htm, htmd.