Skip to content

RTerm v3.9.8

Latest
Latest

Choose a tag to compare

View all tags
@DrOlu DrOlu released this 09 Oct 12:14
v3.9.8
bd8f105

RTerm v3.9.8 — AI-Native Terminal App by Hyperspace Technologies

v3.9.8 (2026-10-01)

neuralOS plugin: top-K probe retrieval productized — neuralos_ask now
hands the 121M engine a FOCUSED menu, with a deterministic enum fast path
that can skip the engine entirely.

The problem (observed live this session)

The plugin passed each instance's FULL probe menu to the engine for
selection. That works well on small menus but degrades as menus grow —
and produced two live failures:

  • cyberbank (60 probes): "failed transactions by failure reason" — the
    engine abstained entirely (no call selected) with every candidate in
    context;
  • outlook-mail (14 probes): a "read latest mail" phrasing fired the WRONG
    probe (mail_forward) with the full 14-probe menu, while ask.py's
    8-probe focused context picked mail_read_latest correctly.

The instances' own ask.py lab tool already had the answer — lexical
retrieval with a focused context — but it was CLI-only.

The fix (retrieval.mjs, productized + unit-tested)

neuralos_ask now, for every instance:

  1. Scores the question against the menu with the ask.py weights
    (trigger overlap ×3, name ×1, description ×0.3; command-verb
    stopwords dropped — list/show/all discriminate nothing).
  2. Hands the engine only the top-K most relevant probes (default 8)
    as a real filtered menu file — the engine's --tools argument stays
    a file, written per-process and cleaned up after the ask.
  3. Deterministic enum fast path: when the rank-1 probe's REQUIRED
    arguments are all enum-caged AND every value appears verbatim in the
    question, the probe fires DIRECTLY — no engine spin, no grounding
    risk, instant. (count messages in sentitems → mail_count {folder: "sentitems"} without consulting the 121M model.)
  4. Never hides: menus below the retrieval threshold (13 probes) go
    to the engine whole, and a question that matches nothing still gets
    the full menu. Retrieval can only focus, never make a probe
    unreachable — that is the one failure mode ask.py had and this port
    fixes.

Enabled by default; neuralos.retrieval=false (or
NEURALOS_RETRIEVAL=0) restores the pre-3.9.8 full-menu behavior.
Threshold and top-K are settings (retrievalMenuThreshold,
retrievalTopK).

Structure: the contract lives in plugins/neuralos/retrieval.mjs (pure,
except the menu-file IO), re-exported by neuralosEngine.mjs so the
plugin keeps one import surface.

Spec: retrieval.extreme.spec.mjs — 31 cases (scoring weights,
threshold, never-hide, top-K capping, fast-path firing rules including
case-insensitivity and the required-non-enum rejection, menu-file round
trip, and a real-shaped chinook-style ranking sanity check). Registered
in the layout-ui chain. Full plugin spec 51/51 still green.

Verified live on the running daemon: "read the latest email" (outlook-
mail) → mail_read_latest; "count messages in sentitems" →
mail_count {folder: "sentitems"}; and on cyberbank the previously-
abstaining question now retrieves
failed_transactions_by_failure_reason into the focused menu (the
engine may still honestly abstain on genuinely ambiguous phrasings —
retrieval narrowed its world correctly).

Downloads

Platform File
macOS ARM64 RTerm-3.9.8-arm64.dmg
Windows x64 (Portable) RTerm.Portable.3.9.8-x64.exe
Windows x64 (Setup) RTerm.Setup.3.9.8-x64.exe
Linux x64 (AppImage) RTerm-3.9.8-x64.AppImage
Linux x64 (deb) RTerm-3.9.8-x64.deb
Linux x64 (pacman) RTerm-3.9.8-x64.pacman
Linux x64 (rpm) RTerm-3.9.8-x64.rpm
Linux ARM64 (AppImage) RTerm-3.9.8-arm64.AppImage
Linux ARM64 (deb) RTerm-3.9.8-arm64.deb
Linux ARM64 (rpm) RTerm-3.9.8-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