RTerm v3.2.12

RTerm v3.2.12 — AI-Native Terminal App by Hyperspace Technologies

v3.2.12 (2026-08-22)

Bug fixes — chat console: response cleared and re-answered after completion

The reported bug: after the agent delivered a full response, the chat would
clear that response and start responding again to the already-completed task
unless the user clicked Stop. Root cause found in the task-completion guard,
plus two related defects fixed alongside.

• Guard continue loop unbounded (the core bug). After the model produced
  its final answer, the task_completion_guard node asked a thinking-model
  "auditor" whether the task was truly done. The auditor prompt instructed it
  to reject stopping "if there are reasonable alternative attempts/tools
  left" — for open-ended tasks it can always invent one more verification
  step, so it kept answering false. Each false removed the assistant's
  answer from the UI and forced the model to re-answer from scratch,
  indefinitely. Live logs confirmed it ("Triggered continue. reason=…never
  verified… optionally round-trip a test request…"). The guard is now bounded
  to at most TASK_COMPLETION_GUARD_MAX_CONTINUES (1) forced continues per
  run; after that the run ends with the answer it produced. New graph state
  guardContinueCount (reducer-backed) tracks it.

• Guard deleted the completed answer. On a forced continue the guard
  called emitRemoveMessageIfPresent on the assistant's last message — the
  "clears that response" half of the bug. It no longer removes anything: the
  answer stays in the transcript and the continue instruction arrives as a
  new turn.

• Auditor prompt biased toward "not done". Rewrote
  createTaskCompletionDecisionUserPrompt: true when the user's request was
  answered/work performed; explicitly forbids inventing optional follow-up
  work (extra checks, alternative approaches, further reading) as a reason to
  continue; false only when a stated requirement is unmet or the answer is
  factually wrong.

• Duplicate DONE per run. Both the agent's final_output node and
  GatewayService.dispatchTask's finally block emitted done, so every run
  finished twice in the UI. The agent now records its runId in
  doneEmittedForRunIds; the gateway checks emittedDoneForRun(runId) and
  skips its duplicate (still emitting when the agent threw before
  final_output, so failed runs still produce exactly one DONE).

• FP: queued-insertion echo re-armed the spinner. A late
  ADD_MESSAGE(user) with a non-normal inputKind (inserted/queued echo
  arriving after DONE) set isThinking=true with no run behind it, leaving
  the chat stuck "thinking". The ChatStore handler now ignores non-normal
  input kinds; the backend's run events drive the real state.

Tests

New v3212ChatFixes.extreme.spec.ts (12/12): guard bound (first continue
allowed, second blocked, missing count = 0, over-strict auditor terminates),
answer preservation on continue, DONE dedupe (exactly one per successful AND
failed run), and the UI re-arm guard (normal arms, inserted/queued echoes
don't, busy state unchanged, no-metadata treated as normal). Registered in
test:backend-unit-extreme. Full backend-unit / layout-ui / automation
suites green.

Downloads

Platform	File
macOS ARM64	RTerm-3.2.12-arm64.dmg
Windows x64 (Portable)	RTerm.Portable.3.2.12-x64.exe
Windows x64 (Setup)	RTerm.Setup.3.2.12-x64.exe
Linux x64 (AppImage)	RTerm-3.2.12-x64.AppImage
Linux x64 (deb)	RTerm-3.2.12-x64.deb
Linux x64 (pacman)	RTerm-3.2.12-x64.pacman
Linux x64 (rpm)	RTerm-3.2.12-x64.rpm
Linux ARM64 (AppImage)	RTerm-3.2.12-arm64.AppImage
Linux ARM64 (deb)	RTerm-3.2.12-arm64.deb
Linux ARM64 (rpm)	RTerm-3.2.12-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