Skip to content

RTerm v3.4.2

Choose a tag to compare

View all tags
@DrOlu DrOlu released this 30 Aug 14:55
· 54 commits to main since this release
v3.4.2
d09abd5

RTerm v3.4.2 — AI-Native Terminal App by Hyperspace Technologies

v3.4.2 (2026-08-30)

Optimization — the distributed, parallel, persistent execution loop

RTerm drives many devices, servers and OSes in parallel. This release makes
that loop safe under load and fixes the persistence bugs the new concurrency
exposed.

1. Per-host concurrency caps + a fair work queue (hostConcurrency.ts).
run_fleet_command() fanned out with bare Promise.allSettled over up to 25
targets — no per-host cap, no global cap. A fleet command over N tabs on ONE
box opened N simultaneous SSH/WinRM channels; sshd's MaxStartups (10:30:100
default) refuses around the 10th, and even when accepted the box starves.

  • Keyed semaphore: at most 3 concurrent operations per host (default),
    regardless of how many tabs or agent runs target it — two tabs on one box
    share one bucket via the new getTerminalConfig()
  • Global cap (12) so a 25-target fleet cannot open 25 channels at once
    even across distinct hosts
  • FIFO fairness per host — no starvation and no queue-jumping (the fast
    path is disabled while anyone is queued)
  • Queue depth + wait time are observable (stats(), onWaitedMs), and the
    fleet summary now reports how many ops were still queued, so a slow fleet
    is diagnosable instead of mysterious
  • Queued ops honour an AbortSignal — a cancelled run does not keep holding
    queue slots
  • collect_facts runs under the same cap: a fleet inventory over many tabs
    on one box is the same channel burst

2. Fleet tools never batch together in the parallel dispatcher. Two
run_fleet_command / collect_facts calls batched in one turn could both
write the same PTY — the duplicate-terminalId check only sees explicit
per-call targets. A batch with more than one fleet tool is now demoted to
sequential.

3. Session persistence — the streaming rewrite. The v3.4.1 append path
can only ADD rows. A message persisted while STREAMING has truncated content
on disk and must be rewritten later — but the scan keyed on the streaming
flag goes blind the moment the message FINISHES (done /
command_finished flips it to false). Its final content never landed: on
reload the session showed the truncated mid-stream text.

  • persistedStreamingMessageIds records which rows were written mid-stream
    and keeps the rewrite obligation alive across the streaming→finished
    transition, until a flush actually rewrites them
  • appendUiSessionMessages() is now a TRUNCATE-AND-APPEND: rows at
    position >= fromPosition are deleted inside the same transaction as the
    inserts, which is also what makes a rollback expressible on the
    incremental path
  • remove_message resets the append cursor (only rollback did before) —
    positions shift after a removal and the cursor was stale
  • messagesCount in the no-summary branch is the real message-list length
    (the old fromPosition + slice.length double-counted after a truncation)

4. getLastVisiblePreview — blank previews and the flush tax. The
return "" sat INSIDE the loop, so the first empty message (a just-opened
sub_tool) produced a blank session preview instead of falling back to the
last non-empty one. It also returned the FULL message body: this column is
rewritten on EVERY flush, and an uncapped preview wrote kilobytes into
ui_sessions each time — a measurable part of the v3.4.1 flush cost.
Previews are now capped at 200 chars.

5. Abort semantics in the limiter. A queued-op abort threw a plain
Error('Aborted while waiting for a host slot'), which isAbortError() did
not recognise — a user-cancelled fleet run reported N bogus per-target FAIL
rows instead of propagating one clean abort. Queue-abort and pre-abort now
carry name: 'AbortError', and the fast path checks the signal BEFORE
consuming a slot (an already-cancelled run no longer fires commands).

6. UIHistoryService.dispose(). Every instance ever constructed stayed
in the static instances set forever; the shared signal handler then flushed
(and held) dead services on every SIGINT/SIGTERM. dispose() flushes once
and deregisters. beforeExit is likewise installed once for the class, not
per instance.

Tests

  • hostConcurrency.extreme.spec.ts — 30 assertions: per-host cap, global
    cap, the double-acquire deadlock regression, FIFO fairness, abort-while-
    queued, queue timeout, host keying (same-host tabs share a bucket),
    stats observability, pre-abort fast path, AbortError naming
  • uiHistoryStreamingRewrite.extreme.spec.ts — 13 assertions: the
    finished-streaming rewrite, growth-while-streaming, preview fallback +
    cap, remove_message cursor reset, dispose() flush + deregister, no
    listener leak across 12 instances
  • Both wired into test:backend-unit-extreme (41 specs)

Downloads

Platform File
macOS ARM64 RTerm-3.4.2-arm64.dmg
Windows x64 (Portable) RTerm.Portable.3.4.2-x64.exe
Windows x64 (Setup) RTerm.Setup.3.4.2-x64.exe
Linux x64 (AppImage) RTerm-3.4.2-x64.AppImage
Linux x64 (deb) RTerm-3.4.2-x64.deb
Linux x64 (pacman) RTerm-3.4.2-x64.pacman
Linux x64 (rpm) RTerm-3.4.2-x64.rpm
Linux ARM64 (AppImage) RTerm-3.4.2-arm64.AppImage
Linux ARM64 (deb) RTerm-3.4.2-arm64.deb
Linux ARM64 (rpm) RTerm-3.4.2-arm64.rpm

Install

macOS: After installing, run xattr -cr /Applications/RTerm.app to remove quarantine.

Linux AppImage: chmod +x RTerm-*.AppImage && ./RTerm-*.AppImage

Linux deb: sudo dpkg -i RTerm-*.deb

Linux rpm: sudo rpm -i RTerm-*.rpm

Contact: rterm@hyperspace.ng

Assets 12
Loading