Repository navigation
Releases: facebey/Glyvex-Group-AI-Suite-Open-Source
Release list
GlyvexAI Suite v0.7.7-beta
GlyvexAI Suite v0.7.7-beta
Language: Español
Re-release (2026-10-06, same version): the initial 0.7.7-beta build renamed the NSIS installer identity, which broke in-place upgrades from 0.7.6. This release replaces the assets with the corrected installer (same version, tag moved) and adds automatic recovery from stale-chunk errors after an update.
The quick-release bundle: cumulative cache-hit and MTP acceptance charts in
the Monitor, --n-cpu-moe in the Launcher, native Browse buttons for every
path, and the display name unified to GlyvexAI Suite across the app UI.
Added
- Monitor: cumulative cache-hit and MTP accept rate charts: two new
persisted metrics (llm.cache_hit_pct_total,llm.spec_accept_pct_total)
computed from the llama-server/metricscounters, shown as a fourth
chart. - Launcher:
--n-cpu-moeparameter: offloads MoE experts to CPU through
llama-server's--n-cpu-moe; the binary probe filters it out on older
builds that do not know the flag. - Native Browse buttons: a Tauri dialog file/folder picker for the
backend binary paths, model directories and GGUF/draft/mmproj/LORA files;
in browser mode it falls back to manual input.
Changed
- Display name unified to "GlyvexAI Suite" (no space) in the window
title, onboarding and About dialog; the executable is nowglyvexai.exe.
The NSIS installer identity stays"Glyvex AI Suite"(with space, Start
Menu folder "Glyvex-AI-Suite") so in-place upgrades from 0.7.6 keep
working — see Fixed. - Ollama no longer pre-fills
/usr/bin/ollama: the default is now empty
on Windows and the launcher resolvesollamafrom PATH.
Fixed
- NSIS in-place upgrade was broken by the name change: the initial
0.7.7-beta build renamed the installer identity ("Glyvex AI Suite" →
"GlyvexAI Suite"), which moves the NSIS default install dir and the
uninstall registry key, so the passive updater could not find the 0.7.6
install and placed 0.7.7 side-by-side.productNamereverted to
"Glyvex AI Suite" andstartMenuFolderto "Glyvex-AI-Suite"; the
"GlyvexAI Suite" branding remains in the window title and in-app UI. - Stale-chunk crash after an update: a router
errorElement
(ChunkReloadError) now detects "Failed to fetch dynamically imported
module" (hashed chunks 404 when a stale instance serves the previous
build's assets) and does a one-time guarded hard reload instead of showing
a broken page.
Validation
- Full pytest suite (606) green.
- Frontend build + signed Tauri rebuild: NSIS installer 96.51 MiB with
minisign.sig(re-release build). latest.jsonverified: version, minisign signature, plugin versions and
the download URL pointing to this public repo.
Installation
Windows x64, NSIS installer (re-release assets — replace the initial
0.7.7-beta files):
| Asset | Size | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.7-beta_x64-setup.exe |
96.51 MiB | DCEF0613F071CFEF0A3AD9199CA4F402979F738F6BC2ADC0ADC9F707ED8381D9 |
Glyvex.AI.Suite_0.7.7-beta_x64-setup.exe.sig |
460 B | 95E02796FC91031B4AB129E4BB8823C883BAE4D90DD0451E11C1794C9788A9AA |
latest.json |
991 B | F59721AFCCD8C1CF63DDFA7CEFFD47063C60D913103A3D411203033B1D166609 |
Español
Re-release (2026-10-06, misma versión): el build inicial de 0.7.7-beta
renombró la identidad del instalador NSIS, lo que rompía los upgrades
in-place desde 0.7.6. Esta release reemplaza los assets con el instalador
corregido (misma versión, tag movido) y añade recuperación automática ante
errores de chunks stale tras un update.
El quick-release bundle: gráficas acumuladas de cache hit rate y MTP
acceptance en el Monitor, --n-cpu-moe en el Launcher, botones Browse
nativos para todas las rutas y el nombre de display unificado a
GlyvexAI Suite en la UI.
Añadido
- Monitor: gráficas acumuladas de cache hit rate y MTP acceptance: dos
métricas persistidas nuevas (llm.cache_hit_pct_total,
llm.spec_accept_pct_total) calculadas desde los contadores/metricsde
llama-server, mostradas como cuarta gráfica. - Launcher: parámetro
--n-cpu-moe: offload de expertos MoE a CPU vía
--n-cpu-moede llama-server; el probe del binario lo descarta en builds
viejas que no conocen el flag. - Botones Browse nativos: selector de archivos/carpeta (diálogo Tauri)
para las rutas de los binarios de backend, los directorios de modelos y
los archivos GGUF/draft/mmproj/LORA; en modo navegador cae a entrada
manual.
Cambiado
- Nombre de display unificado a "GlyvexAI Suite" (sin espacio) en el
título de ventana, onboarding y diálogo Acerca de; el ejecutable ahora es
glyvexai.exe. La identidad del instalador NSIS se mantiene como
"Glyvex AI Suite"(con espacio, carpeta "Glyvex-AI-Suite" en el menú
Inicio) para que los upgrades in-place desde 0.7.6 sigan funcionando —
ver Corregido. - Ollama ya no preselecciona
/usr/bin/ollama: el default ahora queda
vacío en Windows y el launcher resuelveollamadesde PATH.
Corregido
- El cambio de nombre rompía el upgrade in-place de NSIS: el build
inicial de 0.7.7-beta renombró la identidad del instalador ("Glyvex AI
Suite" → "GlyvexAI Suite"), lo que mueve el directorio de instalación por
defecto de NSIS y la key de desinstalación del registro, de modo que el
updater pasivo no encontraba la instalación 0.7.6 y colocaba 0.7.7
lado a lado.productNamerevertido a "Glyvex AI Suite" y
startMenuFoldera "Glyvex-AI-Suite"; el branding "GlyvexAI Suite" se
mantiene en el título de ventana y en la UI. - Crash por chunks stale tras un update: un
errorElementdel router
(ChunkReloadError) detecta "Failed to fetch dynamically imported module"
(chunks hasheados en 404 cuando una instancia vieja sirve los assets de la
build anterior) y hace un hard reload único con guarda en vez de mostrar
una página rota.
Validación
- Suite completa pytest (606) en verde.
- Build del frontend + rebuild firmada de Tauri: instalador NSIS de 96.51
MiB con.sigde minisign (build de la re-release). latest.jsonverificado: versión, firma minisign, versiones de plugins y
la URL de descarga apuntando a este repositorio público.
Instalación
Windows x64, instalador NSIS (assets de la re-release — reemplazan los
archivos iniciales de 0.7.7-beta):
| Asset | Tamaño | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.7-beta_x64-setup.exe |
96.51 MiB | DCEF0613F071CFEF0A3AD9199CA4F402979F738F6BC2ADC0ADC9F707ED8381D9 |
Glyvex.AI.Suite_0.7.7-beta_x64-setup.exe.sig |
460 B | 95E02796FC91031B4AB129E4BB8823C883BAE4D90DD0451E11C1794C9788A9AA |
latest.json |
991 B | F59721AFCCD8C1CF63DDFA7CEFFD47063C60D913103A3D411203033B1D166609 |
Glyvex AI Suite v0.7.6-beta
Glyvex AI Suite v0.7.6-beta
Language: Español
The app is now fully branded: Glyvex Sans (variable, wght 100-900), a
typeface based on Exo 2 with the brand diamond glyphs, replaces Exo 2 across
the UI. Every font now ships inside the installer — zero requests to Google
Fonts.
Changed
- Brand font — Glyvex Sans (variable, wght 100-900): replaces Exo 2 in
the whole UI. Same Exo 2 base plus the brand diamond glyphs oni,j,
0,oandX, so the look stays consistent while every character
carries the brand. - 100% local typography: Exo 2, JetBrains Mono and Glyvex Sans ship as
self-hosted WOFF2 files inside the installer — no font requests over the
network, works fully offline.
Fixed
- Dead Exo 2 assets removed from
frontend/public/fonts/(superseded by the
brand font; a reference copy stays inassets/brand/fonts/).
Validation
- Full pytest suite (604) green.
- Frontend build + signed Tauri build: NSIS installer 96.44 MiB with
minisign.sig. - Sidecar smoke test on the packaged bundle:
/api/healthOK, SPA served. latest.jsonverified: version, minisign signature, plugin versions and
the download URL pointing to this public repo.- Browser check: the
Glyvex Sansface loads (100-900), the UI applies it,
0 remote font requests.
Installation
Windows x64, NSIS installer:
| Asset | Size | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.6-beta_x64-setup.exe |
96.44 MiB | A27D16679E8315E1EB2D89CBA24BF5B06ED68807241BF3028404EF770FF69A59 |
Glyvex.AI.Suite_0.7.6-beta_x64-setup.exe.sig |
460 B | C173EF9830837CDAF67AA55A4773F07197DF61206080B51AC6F8425509015E2B |
latest.json |
892 B | 17634B5A99103C996E1102474ADA9C24E91ADF6E9004B4B926819AB752A6B6B3 |
Español
La app queda totalmente con marca: Glyvex Sans (variable, wght 100-900),
una tipografía basada en Exo 2 con los glifos de rombos de la marca,
reemplaza a Exo 2 en toda la UI. Todas las fuentes ahora viajan dentro del
instalador — cero peticiones a Google Fonts.
Cambiado
- Fuente de marca — Glyvex Sans (variable, wght 100-900): reemplaza a
Exo 2 en toda la UI. Misma base de Exo 2 más los glifos de rombos de la
marca eni,j,0,oyX, manteniendo el look mientras cada
carácter lleva la marca. - Tipografía 100% local: Exo 2, JetBrains Mono y Glyvex Sans viajan como
WOFF2 self-hosted dentro del instalador — sin peticiones de fuentes por
red, funciona sin conexión.
Corregido
- Activos muertos de Exo 2 fuera de
frontend/public/fonts/(los sustituye
la fuente de marca; una copia de referencia sigue enassets/brand/fonts/).
Validación
- Suite completa pytest (604) en verde.
- Build del frontend + build firmada de Tauri: instalador NSIS de 96.44 MiB
con.sigde minisign. - Smoke test del sidecar sobre el bundle empaquetado:
/api/healthOK, SPA
servida. latest.jsonverificado: versión, firma minisign, versiones de plugins y
la URL de descarga apuntando a este repositorio público.- Verificación en navegador: la fuente
Glyvex Sanscarga (100-900), la UI
la aplica y hay 0 peticiones remotas de fuentes.
Instalación
Windows x64, instalador NSIS:
| Asset | Tamaño | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.6-beta_x64-setup.exe |
96.44 MiB | A27D16679E8315E1EB2D89CBA24BF5B06ED68807241BF3028404EF770FF69A59 |
Glyvex.AI.Suite_0.7.6-beta_x64-setup.exe.sig |
460 B | C173EF9830837CDAF67AA55A4773F07197DF61206080B51AC6F8425509015E2B |
latest.json |
892 B | 17634B5A99103C996E1102474ADA9C24E91ADF6E9004B4B926819AB752A6B6B3 |
Glyvex AI Suite v0.7.5-beta
Glyvex AI Suite v0.7.5-beta
Language: Español
The auto-update channel is now reachable from public installs: the updater
endpoint and the latest.json download URL point to this public repository
(previously they referenced the private source repo, so installs from the
public release could never find or download an update).
Fixed
- Auto-update — public endpoint: the Tauri updater endpoint
(tauri.conf.json) and the download URL generated inlatest.json
(build-latest-json.ps1) now point to
facebey/Glyvex-Group-AI-Suite-Open-Source. Installs from the public
release can now detect, download and install updates via
releases/latest. - Version files no longer drift (PUB-7):
APP_VERSION,
frontend/package.json,src-tauri/Cargo.tomlandtauri.conf.json
(plus both lockfiles) all carry 0.7.5-beta together; the npm and Cargo
files had been stuck at 0.7.1-beta since the 0.7.1 release.
Validation
- Full pytest suite (604) green.
- Sidecar smoke test on the packaged bundle:
/api/healthOK, SPA served. latest.jsonverified: version, minisign signature, plugin versions and
the download URL pointing to the public repo.- End-to-end update from a public install: pending first validation — this
is the first release reachable through the public endpoint.
Installation
Windows x64, NSIS installer:
| Asset | Size | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.5-beta_x64-setup.exe |
96.30 MiB | 0E0FC9D198E72001CF98D0EA4D45E2BB761D41C6775111C7C4E9350974B14A2B |
Glyvex.AI.Suite_0.7.5-beta_x64-setup.exe.sig |
460 B | 118608BCB4FD42ED421C1557DEBB2A8391C90BE6F49B1B6215D49B9551ABA8A9 |
latest.json |
839 B | F004B5E257FC86EF8AFAB9DA00B1A49A51AF075C73E4358847F15B8E32BC70F0 |
Español
El canal de auto-update ahora es alcanzable desde las instalaciones
públicas: el endpoint del updater y la URL de descarga de latest.json
apuntan a este repositorio público (antes referenciaban el repo privado de
origen, así que las instalaciones de la release pública nunca podían
encontrar ni descargar una actualización).
Corregido
- Auto-update — endpoint público: el endpoint del updater de Tauri
(tauri.conf.json) y la URL de descarga generada enlatest.json
(build-latest-json.ps1) ahora apuntan a
facebey/Glyvex-Group-AI-Suite-Open-Source. Las instalaciones de la
release pública ya pueden detectar, descargar e instalar actualizaciones
víareleases/latest. - Los archivos de versión ya no se desfasan (PUB-7):
APP_VERSION,
frontend/package.json,src-tauri/Cargo.tomlytauri.conf.json
(más ambos lockfiles) llevan 0.7.5-beta juntos; los archivos de npm y
Cargo se habían quedado en 0.7.1-beta desde la release 0.7.1.
Validación
- Suite completa pytest (604) en verde.
- Smoke test del sidecar sobre el bundle empaquetado:
/api/healthOK,
SPA servida. latest.jsonverificado: versión, firma minisign, versiones de plugins y
la URL de descarga apuntando al repo público.- Update de punta a punta desde una instalación pública: pendiente la
primera validación — esta es la primera release alcanzable por el
endpoint público.
Instalación
Windows x64, instalador NSIS:
| Asset | Tamaño | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.5-beta_x64-setup.exe |
96.30 MiB | 0E0FC9D198E72001CF98D0EA4D45E2BB761D41C6775111C7C4E9350974B14A2B |
Glyvex.AI.Suite_0.7.5-beta_x64-setup.exe.sig |
460 B | 118608BCB4FD42ED421C1557DEBB2A8391C90BE6F49B1B6215D49B9551ABA8A9 |
latest.json |
839 B | F004B5E257FC86EF8AFAB9DA00B1A49A51AF075C73E4358847F15B8E32BC70F0 |
Glyvex AI Suite v0.7.4-beta
Glyvex AI Suite v0.7.4-beta
Language: Español
Third round of the validation cycle over v0.7.1-beta: the installer details
reported during the local validation (T6.2) are resolved, along with the
auto-update fixes from the intermediate rounds (0.7.2/0.7.3, validated
locally and not published on GitHub).
Fixed
- Installer — language selector: the NSIS installer was not showing the
Spanish/English language option (Tauri'sdisplayLanguageSelectordefaults
tofalse). It now appears; the default language remains English. - Start Menu — folder: shortcuts are now installed in the
Glyvex-AI-Suite folder (previously "Glyvex AI Suite"), consistent with
the rest of the brand. - Auto-update — sidecar: the Tauri updater no longer keeps the old
backend sidecar running during update installation (the process blocked the
binary replacement). - Auto-update — permissions: the ACL for the custom commands
(sidecar.toml) needed to restart with the new version was added. - Local config: the local-only full config variant
(tauri.update-local*.conf.json, channelhttp://127.0.0.1:9735) was kept
out of the repo so it is not synced to the public mirror.
Validation
- Full pytest suite (604) green.
- Validated in the packaged exe by the author: the NSIS installer shows the
language selector and creates the "Glyvex-AI-Suite" folder in the Start
Menu. - Auto-update validated locally (T6.2 channel on
127.0.0.1:9735):
detection, download, installation and restart all correct.
Installation
Windows x64, NSIS installer:
| Asset | Size | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.4-beta_x64-setup.exe |
96.34 MiB | 5824AB34A58E1459ED5F06A0C9E5E472B710AD1E6C79F0231D933E3A7BC26810 |
Español
Tercera ronda del ciclo de validación sobre v0.7.1-beta: quedan resueltos
los detalles del instalador reportados en la validación local (T6.2) y los
fixes de auto-update de las rondas intermedias (0.7.2/0.7.3, validados en
local y no publicados en GitHub).
Corregido
- Instalador — selector de idioma: el instalador NSIS no mostraba la
opción de idioma español/inglés (eldisplayLanguageSelectorde Tauri
quedafalsepor defecto). Ahora aparece; el idioma por defecto sigue
siendo inglés. - Menú Inicio — carpeta: los accesos directos se instalan ahora en la
carpeta Glyvex-AI-Suite (antes "Glyvex AI Suite"), consistente con el
resto de la marca. - Auto-update — sidecar: el updater de Tauri ya no levanta el sidecar
del backend viejo durante la instalación de la actualización (el proceso
bloqueaba el reemplazo del binario). - Auto-update — permisos: se añadió el ACL para los custom commands
(sidecar.toml) necesarios al reiniciar con la nueva versión. - Config local: la variante completa de config solo-local
(tauri.update-local*.conf.json, canalhttp://127.0.0.1:9735) quedó
fuera del repo para que no se sync al mirror público.
Validación
- Suite completa pytest (604) en verde.
- Validado en el exe empaquetado por el autor: el instalador NSIS muestra
el selector de idioma y crea la carpeta "Glyvex-AI-Suite" en el menú
Inicio. - Auto-update validado en local (canal T6.2 en
127.0.0.1:9735):
detección, descarga, instalación y reinicio correctos.
Instalación
Windows x64, instalador NSIS:
| Asset | Tamaño | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.4-beta_x64-setup.exe |
96.34 MiB | 5824AB34A58E1459ED5F06A0C9E5E472B710AD1E6C79F0231D933E3A7BC26810 |
Glyvex AI Suite v0.7.1-beta
Glyvex AI Suite v0.7.1-beta
Language: Español
Second round of the T7.1 validation cycle over v0.7.0-beta: the three
reported details (slow Config detection, manual endpoint not persisting and
TTS audio) are resolved.
Fixed
- TTS — playback: back to the
<audio>element (the WebAudio rewrite
left audio muted in some Chrome/Edge/WebView2 environments). A playback or
autoplay failure is now shown in a notice, instead of silence or a stuck
button. - TTS — WAV cache in the backend (LRU 32): re-listening to the same
response is instant and does not re-synthesize. - Config: Web Search / STT / TTS detection in Settings no longer takes
30–45 s (60 s TTL cache per config signature; the "Check" button forces a
re-probe). - Chat: the last 5 custom endpoints are persisted (localStorage) and
appear in the selector, no need to retype the URL. - TTS — parallel downloads (TTS-3): independent progress bar per
download (Piper voice + Kokoro model at the same time). - Packaging — refresh: refreshing (F5 or "Reload" from the right-click
menu) in a module other than Chat showed{"detail":"Not Found"}. The
backend now servesindex.htmlfor any app route; the API still answers
404 JSON on non-existent routes.
Validation
- Full pytest suite (594) + vitest (14) + frontend build green.
- Validated in the packaged exe by the author: Bluetooth speaker and two
desktop speakers, correct TTS playback on both.
Installation
Windows x64, NSIS installer:
| Asset | Size | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.1-beta_x64-setup.exe |
95.36 MiB | 516EAF6BDE20FB480C04877AE0B3F510571F6611CBDBB02C15A39AE3C5B15E87 |
Español
Segunda ronda del ciclo de validación T7.1 sobre v0.7.0-beta: los tres
detalles reportados (detección de Config lenta, endpoint manual que no
persistía y audio TTS) quedaron resueltos.
Corregido
- TTS — reproducción: vuelve al elemento
<audio>(el rewrite a WebAudio
dejaba el audio mudo en algunos entornos Chrome/Edge/WebView2). Un fallo de
reproducción o de autoplay se muestra ahora en un aviso, en vez de silencio
o botón atascado. - TTS — caché de WAV en el backend (LRU 32): re-escuchar la misma respuesta
es instantáneo y no resintetiza. - Config: la detección de Web Search / STT / TTS en Opciones ya no tarda
30-45 s (caché TTL 60 s por firma de config; el botón "Comprobar" fuerza el
re-probe). - Chat: los últimos 5 endpoints personalizados se persisten (localStorage)
y aparecen en el selector, sin volver a tipear la URL. - TTS — descargas en paralelo (TTS-3): barra de progreso independiente por
descarga (voz Piper + modelo Kokoro a la vez). - Empaquetado — refresco: al refrescar (F5 o "Actualizar" del menú de
clic derecho) en un módulo distinto del Chat salía{"detail":"Not Found"}.
El backend ahora sirveindex.htmlpara cualquier ruta de la app; la API
sigue respondiendo 404 JSON en rutas inexistentes.
Validación
- Suite completa pytest (594) + vitest (14) + build de frontend en verde.
- Validado en el exe empaquetado por el autor: parlante Bluetooth y dos
parlantes de mesa, reproducción de TTS correcta en ambos.
Instalación
Windows x64, instalador NSIS:
| Asset | Tamaño | SHA-256 |
|---|---|---|
Glyvex.AI.Suite_0.7.1-beta_x64-setup.exe |
95.36 MiB | 516EAF6BDE20FB480C04877AE0B3F510571F6611CBDBB02C15A39AE3C5B15E87 |
Glyvex AI Suite 0.7.0-beta
Language: Español
0.7.0-beta (prerelease)
Open source distribution synchronized with the main repository.
Added
- Neural TTS (Piper + Kokoro) packaged in the sidecar: the neural chain
(piper, kokoro-onnx, onnxruntime, espeak-ng) ships inside
glyvex-backend. Voices and model are downloaded in Settings → TTS; the
autoengine picks Kokoro → Piper → SAPI depending on what is ready. - Provisioning screen — TTS step: direct pointer to configure the neural
TTS in Settings.
Changed
- Dependency bumps: uvicorn 0.54, sqlalchemy 2.1.1, nvidia-ml-py
13.615.71 (backend) and vite 8.3.1 (frontend). - numpy floor 2.2 in
requirements.txt: 2.5 required Python >=3.12 and
broke the install in the 3.11 CI job.
Fixed
- CI on Linux: 3 tests assumed Windows (STT/TTS status) and now force the
platform via monkeypatch; the suite passes on 3.11 and 3.12.
The binary installer (NSIS x64) is published on the main repository
facebey/glyvex-ai-suite (release v0.7.0-beta); this repo is code only.
Español
0.7.0-beta (prerelease)
Distribución open source sincronizada con el repo principal.
Nuevo
- TTS neuronal (Piper + Kokoro) empaquetado en el sidecar: la cadena neural
(piper, kokoro-onnx, onnxruntime, espeak-ng) viaja dentro deglyvex-backend.
Voces y modelo se descargan en Ajustes → TTS; el motorautoelige Kokoro →
Piper → SAPI según qué esté listo. - Pantalla de provisión — paso TTS: puntero directo para configurar el TTS
neuronal en Ajustes.
Cambiado
- Bumps de dependencias: uvicorn 0.54, sqlalchemy 2.1.1, nvidia-ml-py
13.615.71 (backend) y vite 8.3.1 (frontend). - Piso de numpy 2.2 en
requirements.txt: el 2.5 exigía Python >=3.12 y
rompía la install del job de CI 3.11.
Corregido
- CI en Linux: 3 tests asumían Windows (status STT/TTS) y ahora fuerzan la
plataforma vía monkeypatch; la suite pasa en 3.11 y 3.12.
El instalador binario (NSIS x64) se publica en el repo principal
facebey/glyvex-ai-suite (release v0.7.0-beta); este repo es solo el código.
Glyvex AI Suite 0.6.2-beta
Language: Español
0.6.2-beta (prerelease)
Major frontend dependency migration (one branch per library, build and tests
green at every step):
| Dependency | Before | After |
|---|---|---|
react / react-dom |
^18.3.0 | ^19.3.0 |
lucide-react |
^0.400.0 | ^1.47.0 |
react-markdown |
^9.0.1 | ^10.1.0 |
recharts |
^2.12.0 | ^3.10.0 |
Code changes (minimal, only what each major required):
- lucide-react 1.x: dropped the brand icons. The footer
Githubicon
becameExternalLink(frontend/src/App.jsx). - react-markdown 10: no code changes — the custom
rehypeFilenameplugin
already used the unified 11plugin() → transformer()signature. - recharts 3: no code changes — the Monitor (line) and Benchmark
(bar + scatter) charts use a compatible API.
Verification: production build OK, full pytest suite green, manual smoke of
the 6 pages (Chat, Launcher, Benchmark, Monitor, Reports, Config) confirmed.
Note: no MSI on this tag — the installer was not changed. When the next
bundle is built, the MSI will be 0.6.2.
Español
0.6.2-beta (prerelease)
Migración mayor de dependencias del frontend (una rama por librería, build y
tests en verde en cada paso):
| Dependencia | Antes | Después |
|---|---|---|
react / react-dom |
^18.3.0 | ^19.3.0 |
lucide-react |
^0.400.0 | ^1.47.0 |
react-markdown |
^9.0.1 | ^10.1.0 |
recharts |
^2.12.0 | ^3.10.0 |
Cambios de código (mínimos, solo lo que exigió cada mayor):
- lucide-react 1.x: eliminó los iconos de marca. El ícono
Githubdel
footer pasó aExternalLink(frontend/src/App.jsx). - react-markdown 10: sin cambios de código — el plugin custom
rehypeFilenameya usaba la firmaplugin() → transformer()de unified 11. - recharts 3: sin cambios de código — los charts de Monitor (line) y
Benchmark (bar + scatter) usan API compatible.
Verificación: build de producción OK, suite pytest completa en verde, smoke
manual de las 6 páginas (Chat, Launcher, Benchmark, Monitor, Reports, Config)
confirmado.
Nota: sin MSI en este tag — el instalador no cambió. Cuando se arme el
siguiente bundle, el MSI será 0.6.2.
v0.6.1-beta1
Language: Español
Glyvex AI Suite — v0.6.1-beta1
Fixed MSI installer over v0.6.0-beta1 (previous rotated beta).
Fixed (packaging)
- Benchmark — empty step 2: the 6 prompt sets (
coder,infra_network,
logic,math,science,spanish) were not bundled; they are now
(_internal/backend/prompts)./api/benchmark/sets→ 6. - Uninstall — error 2911: the backend kept running and blocked the
uninstall; aKillBackendcustom action now terminates it first. - Install layout: picking a drive root created
glyvex-backend\; now a
flat layout in%LOCALAPPDATA%\Glyvex-AI-Suite(picking a drive root no
longer createsD:\glyvex-backend). - Slow load: server-side GZip + frontend code splitting (React.lazy +
per-library chunks). - Benchmark: visible error state + retry if the sets load fails.
- Installer: v3 wizard branding (banner, ExitDialog image and a new icon
in wizard / Control Panel /ARPPRODUCTICON).
Infra (OSS repo)
- CodeQL / GitHub code scanning: security analysis workflow (python +
javascript) enabled on the public repository.
Assets
| File | Size | sha256 |
|---|---|---|
Glyvex-AI-Suite-Setup-0.6.1-beta1.msi |
51,081,179 B (48.7 MB) | ba6dd3a2af4ecf0f4d65edb37dd94128a8ce2378238a4402dc57270fc89e9bf5 |
Español
Glyvex AI Suite — v0.6.1-beta1
Instalador MSI corregido sobre v0.6.0-beta1 (beta anterior rotada).
Corregido (empaquetado)
- Benchmark — step 2 vacío: los 6 sets de prompts (
coder,infra_network,
logic,math,science,spanish) no entraban al bundle; ahora sí
(_internal/backend/prompts)./api/benchmark/sets→ 6. - Uninstall — error 2911: el backend seguía corriendo y bloqueaba el
desinstalado; una custom actionKillBackendlo termina antes. - Layout de instalación: al elegir una raíz de disco se creaba
glyvex-backend\; ahora layout plano en%LOCALAPPDATA%\Glyvex-AI-Suite
(elegir una raíz de disco ya no creaD:\glyvex-backend). - Carga lenta: GZip en el servidor + code splitting del frontend
(React.lazy + chunks por lib). - Benchmark: estado de error visible + retry si la carga de sets falla.
- Instalador: branding v3 del wizard (banner, imagen del ExitDialog y
nuevo icono en wizard / Panel de Control /ARPPRODUCTICON).
Infra (repo OSS)
- CodeQL / GitHub code scanning: workflow de análisis de seguridad
(python + javascript) activo en el repo público.
Assets
| Archivo | Tamaño | sha256 |
|---|---|---|
Glyvex-AI-Suite-Setup-0.6.1-beta1.msi |
51,081,179 B (48.7 MB) | ba6dd3a2af4ecf0f4d65edb37dd94128a8ce2378238a4402dc57270fc89e9bf5 |
v0.5.0
Language: Español
Added
- Embedded inference runtime: on Windows the suite downloads its own
tested build of llama.cpp (pin b11009), verified by sha256, with
nothing else to install. 2-level split: base engine (~19 MB, runs on
any hardware) + NVIDIA acceleration (~531 MB, CUDA 13.4), chosen by the
detected GPU family (nvidia|amd|intel|cpu). 2-stage download with SSE
progress; if the acceleration fails it degrades to CPU ("degraded" state),
never an error. Sources by preference: the suite's own asset (this release)
→ official ggml-org/llama.cpp build (fallback). - Binary cascade at launch:
binary_path(expert mode) always wins →
managed runtime ifready→ otherwise the launch is rejected with
runtime_missingand the UI offers to download it. - Runtime API:
GET /api/runtime/status(GPU + family, state, version),
POST /api/runtime/download(SSE with progress),POST /api/runtime/reset. - Runtime UI: "Download runtime" step in the onboarding (detected GPU,
estimated size per family, progress, states), "Runtime" card in Config
(state, reinstall) and runtime chip in the Launcher. On Windows the base
engine is offered for every GPU; AMD/Intel run on CPU with a "acceleration
coming soon" note. - M5 — Monitor: TDP control (power limit) for the GPU on the GPU card
(nvidia-ml-py). - Footer with a link to the public GitHub repository (OSS Apache 2.0).
- Full ES/EN i18n for the new runtime views.
Changed
- Dependencies:
pynvml(deprecated) replaced bynvidia-ml-py
(NVIDIA's official library). - The runtime archives are resolved from the open source public repository
release (this repo).
Tests
- New runtime test suite:
runtime.pycore (download verified against a fake
server, extract, states, source cascade), API (status/download SSE/reset)
and launcher cascade. Full suite: 356 tests green.
Assets
| File | Size | sha256 |
|---|---|---|
glyvex-runtime-base-win-x64-b11009.zip |
~18.6 MB | 76d7c36366cee940662a2bcaf294517255517117b322a64123f64a5d01e3da79 |
glyvex-runtime-accel-nvidia-win-x64-b11009.zip |
~530.6 MB | d61c6371fca42b018b7f2a40d35c95f78304c1fdc6015433c33e37b242ee30f4 |
Español
Agregado
- Runtime de inferencia embebido: en Windows la suite descarga su propia
build probada de llama.cpp (pin b11009), verificada por sha256, sin
instalar nada aparte. Split en 2 niveles: motor base (~19 MB, corre en
cualquier hardware) + aceleración NVIDIA (~531 MB, CUDA 13.4), elegida
por familia de GPU detectada (nvidia|amd|intel|cpu). Download en 2 etapas
con progreso por SSE; si la aceleración falla, degrada a CPU (estado
"degradado"), nunca es error. Fuentes con preferencia: asset propio de la
suite (esta release) → build oficial de ggml-org/llama.cpp (fallback). - Cascada de binario al lanzar:
binary_path(modo experto) gana
siempre → runtime gestionado si estáready→ si no, el launch se rechaza
conruntime_missingy la UI ofrece descargarlo. - API runtime:
GET /api/runtime/status(GPU + familia, estado,
versión),POST /api/runtime/download(SSE con progreso),
POST /api/runtime/reset. - UI runtime: paso "Descargar runtime" en el onboarding (GPU detectada,
tamaño estimado por familia, progreso, estados), card "Runtime" en Config
(estado, reinstalar) y chip de runtime en el Launcher. En Windows el motor
base se ofrece a toda la GPU; AMD/Intel corren en CPU con nota "la
aceleración llega próximamente". - M5 — Monitor: control TDP (power limit) de la GPU en la tarjeta GPU
(nvidia-ml-py). - Footer con link al repo público de GitHub (OSS Apache 2.0).
- i18n ES/EN completo para las vistas nuevas de runtime.
Cambiado
- Dependencias:
pynvml(deprecated) reemplazado pornvidia-ml-py
(lib oficial de NVIDIA). - Los archives del runtime se resuelven desde la release del repo open source
público (este repo).
Tests
- Nuevo test suite del runtime: núcleo
runtime.py(download verificado
contra server fake, extract, estados, cascada de fuentes), API
(status/download SSE/reset) y cascada del launcher. Suite completa:
356 tests en verde.
Assets
| Archivo | Tamaño | sha256 |
|---|---|---|
glyvex-runtime-base-win-x64-b11009.zip |
~18.6 MB | 76d7c36366cee940662a2bcaf294517255517117b322a64123f64a5d01e3da79 |
glyvex-runtime-accel-nvidia-win-x64-b11009.zip |
~530.6 MB | d61c6371fca42b018b7f2a40d35c95f78304c1fdc6015433c33e37b242ee30f4 |
v0.4.3
Language: Español
Added
- A3 — ES/EN i18n: react-i18next with a language selector in the UI
(Spanish default). Covers Launcher, Chat, Monitor, Benchmark, Reports and
Config (706 keys per language); benchmark set names and descriptions are
bilingual (JSON fallback);frontend/package-lock.jsonversioned for
reproducible installs. - P8 — Flags reference:
docs/LAUNCH-FLAGS.md(EN) +
docs/LAUNCH-FLAGS.es.md(ES) — ~45 flags in 14 groups (what it does,
impact, app default, when to change it); official per-flag labels in the
Launcher UI via the probe (flag_help). - Docs (P7): central "Documentation" index in the README (EN/ES),
docs/QUICKSTART.md,docs/ARQUITECTURA.md,docs/PRIVACIDAD.md,
.githubtemplates (bug/feature issue, config, PR template) and
CONTRIBUTORS.md+ up-for-grabs inCONTRIBUTING.md. - Dependabot:
.github/dependabot.yml(pip + npm, weekly, cap of 10 open
PRs per ecosystem).
Fixed
- K3: benchmark persistence race — the final status (completed/error) is
saved to the DB before being exposed in memory. --fittoggle: OFF now emits--fit offinstead of omitting the flag
(which left the build'sondefault).
Baseline
- 297 tests green.
Español
Agregado
- A3 — i18n ES/EN: react-i18next con selector de idioma en la UI (default
español). Cubre Launcher, Chat, Monitor, Benchmark, Reports y Config (706
keys por idioma); nombres y descripciones de sets de benchmark bilingües
(fallback al JSON);frontend/package-lock.jsonversionado para installs
reproducibles. - P8 — Referencia de flags:
docs/LAUNCH-FLAGS.md(EN) +
docs/LAUNCH-FLAGS.es.md(ES) — ~45 flags en 14 grupos (qué hace, impacto,
default de la app, cuándo cambiarlo); etiquetas oficiales por flag en la UI
del Launcher vía el probe (flag_help). - Docs (P7): índice central "Documentación" en README (EN/ES),
docs/QUICKSTART.md,docs/ARQUITECTURA.md,docs/PRIVACIDAD.md,
plantillas.github(issue bug/feature, config, PR template) y
CONTRIBUTORS.md+ up-for-grabs enCONTRIBUTING.md. - Dependabot:
.github/dependabot.yml(pip + npm, semanal, tope de 10 PRs
abiertas por ecosistema).
Corregido
- K3: race de persistencia del benchmark — el status final (completed/
error) se guarda en la DB antes de exponerlo en memoria. --fittoggle: OFF ahora emite--fit offen vez de omitir el flag
(que dejaba el defaultonde la build).
Baseline
- 297 tests en verde.