abstractframework 0.6.1__tar.gz → 0.6.3__tar.gz

{abstractframework-0.6.1 → abstractframework-0.6.3}/CHANGELOG.md

@@ -4,6 +4,254 @@ All notable changes to AbstractFramework will be documented in this file.
 
 ## [Unreleased]
 
+## [0.6.3] - 2026-09-29
+
+The `gpu` setting on NVIDIA machines, speech-to-text in the fresh setup, and loaded image models
+that are reused. Pins: AbstractGateway 0.7.4, AbstractCore 2.19.2, AbstractRuntime 0.7.3,
+AbstractVoice 0.13.2, AbstractVision 0.3.33 and Abstract3D 0.3.2; the gateway's terminal console is
+`abstractgateway-console` 0.11.2, and the browser client AbstractCode web is 0.6.2. Every other
+version, and the other terminal clients, are unchanged from 0.6.2.
+
+### Added
+
+- **The `gpu` setting on Linux + NVIDIA, rehearsed on real hardware with this release.** On
+  Ubuntu 26.04 with a Quadro RTX 5000 (16 GB, driver 595.91.07, CUDA 13.2; 4 vCPU, 26 GB RAM),
+  `install.sh` installed this release's candidate versions from PyPI in 3 min 1 s (AbstractGateway
+  0.7.3, AbstractCore 2.19.1, AbstractRuntime 0.7.2, AbstractVoice 0.13.1 and AbstractVision 0.3.32;
+  the patch releases listed under Fixed came after this run). The previous
+  install had been removed with `uninstall.sh --yes --purge` and uv's download cache cleared; model
+  weights, LM Studio and a Rust toolchain were already on the machine. The Python packages took 39 s
+  (about 5.8 GB of wheels), llama.cpp's `cu130` build 18 s (726 MB), compiling the two terminal
+  clients with cargo 110 s, and the start and health check 5 s. The install added 14.2 GB on disk.
+  Its checks reported PyTorch 2.11.0 on CUDA 13.0, llama.cpp 0.3.35 (`cu130`) with GPU offload and
+  Whisper on the GPU. Run again, it reported "already up to date" in 10 s. With **Use recommended
+  defaults** and the model weights already downloaded:
+  - text with LM Studio (`qwen/qwen3.5-9b`, Q4_K_M, 6.5 GB on the GPU): the first gateway request
+    took 23 s including loading the model, the next 12 s for a 318-token answer (308 of them
+    reasoning); from AbstractCore in Python about 41 tokens/s;
+  - text-to-speech with Supertonic, on the processor: 1.1 to 1.4 s per sentence (3.2 s for the
+    first gateway request);
+  - image generation with FLUX.2 [klein] 4B through Diffusers, with model CPU offload (7.8 GB
+    peak): 41 s to load the model, then 15 to 16 s per 1024x1024 image. Each gateway image request
+    loads the model again, so one request takes 54 to 59 s;
+  - speech-to-text did not work as installed; see Known issues below.
+
+  vLLM 0.22.1 installs with the gpu setting but was not validated: it needs a C compiler
+  (`build-essential`) the first time it starts a model, and on GPUs older than compute capability
+  8.0 the `--attention-backend TRITON_ATTN` option. See
+  [GPU on Linux (NVIDIA)](docs/install.md#gpu-on-linux-nvidia).
+- **`install.sh` on Linux + NVIDIA** installs llama.cpp's CUDA build that matches PyTorch's CUDA
+  (PyPI's PyTorch for Linux is a CUDA 13 build, which needs NVIDIA driver 580 or newer; with an
+  older driver PyTorch and llama.cpp run on the processor) and keeps it only when it
+  loads and offloads to the GPU; otherwise the CPU build goes back and the summary says why. At the
+  end it checks that PyTorch sees the GPU and which device Whisper uses (`PyTorch:`, `GGUF:` and
+  `Voice:` summary lines), and warns when no C compiler is found for vLLM.
+- **`install.ps1` on Windows + NVIDIA** picks the GPU stack from `nvidia-smi` (CUDA 13 with driver
+  580 or newer and compute capability 7.5 or newer, CUDA 12 with driver 525 or newer, otherwise the
+  CPU), installs PyTorch's CUDA build and llama.cpp's matching prebuilt build (CUDA, then Vulkan,
+  then CPU), and ends with GPU checks that replace any part that does not work with one that does.
+  The Windows gpu setting is implemented and tested with simulated drivers and wheel-only
+  resolves, but it has **not yet been validated on real NVIDIA hardware**. See
+  [GPU on Windows (NVIDIA)](docs/install.md#gpu-on-windows-nvidia).
+- **Live installer progress.** Both installers show uv's output as it comes, with a
+  `still working (… elapsed; last: …)` line after 15 seconds of silence, so a multi-GB download no
+  longer looks frozen. `install.ps1` announces the big downloads before they start.
+
+### Changed
+
+- **Only the three install settings in hints.** Every install hint across the framework names
+  light (`pip install -U abstractcore`), `abstractcore[apple]` or `abstractcore[gpu]`, or says that
+  a capability is not available on this machine, never a bare package or a plugin's own extra
+  (AbstractCore 2.19.1, AbstractGateway 0.7.3, AbstractVoice 0.13.1, AbstractRuntime 0.7.2,
+  Abstract3D 0.3.2).
+- **Smaller gpu setting.** The gpu setting no longer installs MLX-Gen (AbstractVision 0.3.32):
+  about 2.1 GB less on Linux, and it installs on Linux distributions with an older glibc.
+- **`abstractcore[gpu]` installs on Windows with wheels only** (vLLM is Linux only;
+  stable-diffusion.cpp and llama.cpp are left to the installer's prebuilt builds on Windows).
+- **Recommendations on NVIDIA machines.** **Use recommended defaults** picks FLUX.2 [klein] 4B
+  through Diffusers for image generation on an NVIDIA GPU (it fits a 16 GB card), and the LM Studio
+  text download is `qwen/qwen3.5-9b@q4_k_m`, which LM Studio offers on Linux and Windows.
+- **AbstractRuntime depends on AbstractCore's three settings** (`abstractcore`, `[apple]`, `[gpu]`),
+  and `openai` 2.x is allowed across the stack, so the gpu setting resolves a current vLLM.
+
+### Fixed
+
+- **The approval gate shows in every client.** When the agent's tool call waits for approval, a
+  browser or phone that opens the same conversation later now shows the same Allow / Deny card and
+  "Approval needed" as the client that started the turn, instead of "Running a tool" and a Steer box
+  (AbstractCode web 0.6.2 with `@abstractframework/panel-chat` 0.1.20; two-browser end-to-end test).
+- **Speech-to-text without an OpenAI key.** Transcription runs on the configured `input.voice`
+  route (for example local Whisper) and no longer needs TTS credentials (AbstractCore 2.19.1,
+  AbstractVoice 0.13.1).
+- **Speech-to-text in the fresh setup.** The fresh-install defaults and **Use recommended defaults**
+  set speech input (`input.voice`: faster-whisper, model `base`) with text, voice and images, so a
+  new install transcribes locally without an OpenAI key (AbstractCore 2.19.2, AbstractGateway
+  0.7.4). An install that earlier release defaults seeded gains the route once, when it is empty; a
+  route you set yourself is kept. PyAV is held below 19, which faster-whisper 1.2.1 needs to read
+  audio files (AbstractVoice 0.13.2).
+- **Use recommended defaults saves every route it reports**, including on a fresh install right
+  after the console read the routes (AbstractCore 2.19.2).
+- **Whisper on NVIDIA GPUs.** faster-whisper uses CUDA only when CUDA 12 cuBLAS loads (installed
+  with the gpu setting on Linux and Windows), and otherwise runs on the CPU with a warning.
+- **llama.cpp on CUDA** loads without importing PyTorch first on Linux, and a GGUF model too large
+  for the GPU at its full context tries smaller contexts on the GPU before falling back to the CPU.
+- **Image generation on smaller GPUs.** A Diffusers pipeline that does not fit the GPU's free
+  memory loads with model CPU offload instead of failing with CUDA out of memory. When even the
+  largest component does not fit (for example next to a loaded text model), it falls back to
+  sequential CPU offload, which uses much less GPU memory and is slower per step (AbstractVision
+  0.3.33). On the Quadro RTX 5000, FLUX.2 [klein] 4B at 768x768 and 4 steps peaked near 1.4 GB and
+  took 10.9 to 15.6 s, against 13.4 to 18.5 s with model CPU offload; at about 2.3 s per step
+  against 1.5 s, models that run many steps are slower with it.
+- **A loaded image model is reused.** After **Load** in the console (or `POST /models/load`), image
+  and video requests for that model run on the loaded pipeline instead of loading the model again
+  for each request (AbstractRuntime 0.7.3, AbstractGateway 0.7.4). On the rehearsal machine,
+  FLUX.2 [klein] 4B images through the gateway took 17 to 19 s each instead of 54 to 59 s. A model
+  that was not loaded still runs each request in its own process.
+- **Unloading an image model frees its memory.** On CUDA, FLUX.2 [klein] 4B with model CPU offload
+  left 16.3 GB held after an unload; it now leaves 0.86 GB (AbstractVision 0.3.33).
+
+### Security
+
+- **Email connections verify TLS.** AbstractCore's mail tools (`list_emails`, `read_email`,
+  `send_email`) and the gateway's email bridge check the server certificate and host name, and refuse
+  before login when the check fails; a server signed by a private CA is trusted with the new
+  `ca_file` field of an account (AbstractCore 2.19.2, AbstractGateway 0.7.4).
+- **Automations no longer email anyone the model chooses.** With the default automatic tool
+  approval, message-sending tools are no longer pre-approved: a `send_email` to your registered
+  email runs unattended, any other recipient waits for your approval (AbstractRuntime 0.7.3,
+  AbstractGateway 0.7.4).
+- A `password_env_var` / `imap_password_env_var` that is not an environment variable name is refused
+  instead of being read as the password (AbstractCore 2.19.2, AbstractGateway 0.7.4).
+
+### Found in the rehearsal
+
+Found in the rehearsal on the Quadro RTX 5000 described above, which ran the candidate versions. The
+Fixed entries above address each one.
+
+- **Speech-to-text failed on a fresh gpu install**, for two reasons:
+  - **Use recommended defaults** did not set the speech input route (`input.voice`). With no route,
+    transcription used OpenAI and failed with "OpenAI audio requires OPENAI_API_KEY".
+  - PyAV 19.0.0, released on 2026-09-29, removed an option that faster-whisper 1.2.1 uses to read
+    audio files. Every transcription through the gateway then failed with `open() got an unexpected
+    keyword argument 'metadata_errors'`.
+
+  With the route set to faster-whisper `base` by hand and PyAV 18, faster-whisper transcribed on the
+  GPU with no OpenAI key: 1.0 to 2.3 s per gateway request, and 0.3 GB of GPU memory.
+- **An image next to a loaded text model ran out of GPU memory on a 16 GB card.** LM Studio's
+  `qwen/qwen3.5-9b` keeps 6.5 GB on the GPU while it is loaded. FLUX.2 [klein] needed 7.8 GB even
+  with model CPU offload, so the image request failed with CUDA out of memory after three attempts
+  (about 2.5 minutes). With the text model unloaded (`lms unload --all`, or the console's model
+  list), the same request worked. In this case AbstractVision 0.3.33 falls back to sequential CPU
+  offload (see Fixed above).
+- **Use recommended defaults could skip the image route on a fresh install.** On an NVIDIA machine,
+  if the button was used within a few seconds of the console reading the routes, it could report the
+  image route as already set without saving it. Using it again saved it.
+
+## [0.6.2] - 2026-09-29
+
+One line installs, upgrades and repairs AbstractFramework, and **Update** in the consoles and the
+menu-bar icon runs that same line. Pins: AbstractGateway 0.7.2, AbstractCore 2.19.0,
+AbstractVision 0.3.31, Abstract3D 0.3.1 (new in the matrix: AbstractCore now installs it) and the
+terminal consoles `abstractgateway-console` 0.11.1 and
+`abstractcore-console` 0.4.1; every other version is unchanged from 0.6.1.
+
+### Changed
+
+- **One line installs, upgrades and repairs.** Re-running the install line (or double-clicking
+  **Install AbstractFramework.command** again) finds the existing install and says what it does:
+  `AbstractFramework <yours> found: upgrading to AbstractFramework <latest>` or `already up to
+  date` for installs from 0.6.2 on, which record their release. An install made by 0.6.1 or
+  earlier recorded none, so its first upgrade says `AbstractFramework found (abstractgateway 0.7.1;
+  its release was not recorded): upgrading to AbstractFramework 0.6.2`.
+  The summary lists what changed, old -> new, under **Changes**, and ends with `Upgraded: …` or
+  `Already up to date: …; nothing changed.` `install.ps1` does the same on Windows. See
+  [Upgrade](docs/install.md#upgrade).
+- **Update in the web console, the terminal console and the menu-bar icon runs the installer**
+  (AbstractGateway 0.7.2). For an install made by the AbstractFramework installer, the check
+  compares your AbstractFramework release with the latest one, and **Update** runs the latest
+  `install.sh` with `--no-start`. Before you confirm, it shows the installer's address, commit,
+  sha256 and the exact command. The result says what moved and offers the restart, says
+  **already up to date**, or shows why it didn't finish. On Windows, the check shows the PowerShell
+  line to paste. See [From the console or the menu-bar icon](docs/install.md#from-the-console-or-the-menu-bar-icon).
+- **Your install options are remembered.** A re-run keeps the profile, port, start at login, a
+  custom `--data-dir`, and `--no-console`, `--no-code-cli`, `--no-core-cli`, `--no-tray` and
+  `--full`, and says which ones it kept. `--with-console`, `--with-code-cli`, `--with-core-cli`,
+  `--with-tray` and `--no-full` (Windows: `-WithConsole`, `-WithCodeCli`, `-WithCoreCli`,
+  `-WithTray`, `-NoFull`) turn one back.
+- **The libraries land on the release's tested versions.** The installer installs AbstractCore,
+  AbstractRuntime, AbstractAgent, AbstractSkill, AbstractMemory, AbstractSemantics and the voice,
+  vision, music and 3D packages at exactly the versions of the AbstractFramework release, not only at
+  the gateway's minimums. `--print-versions` (`-PrintVersions`) lists the release and that matrix.
+  `--pin latest` still installs the newest gateway and libraries, and records no release.
+- **The gateway restarts whenever anything changed**, a library alone included: the macOS login
+  item through launchd, a running Linux `systemd --user` service with `systemctl --user restart`,
+  and a background gateway by starting it again. On Windows the installer stops the gateway before
+  it changes any file and starts it again afterwards. AbstractGateway 0.7.2 fixes the macOS
+  `5: Input/output error` when the login item is replaced: it waits until launchd has removed the
+  previous job and retries. When the login item cannot be registered, the gateway starts in the
+  background instead, so an upgrade never leaves it stopped.
+- **Upgrading from 0.6.1 or earlier.** Those installs did not record their options, so the first
+  re-run reads them from what is installed (the terminal console, `abstractcode`, the tray extra,
+  the AbstractCore commands, a `--full` build, a custom data directory through the gateway
+  pointer), says what it found and records it; you can also repeat your options once. The Update
+  button of AbstractGateway 0.7.1 and earlier cannot upgrade an installer install: re-run the line
+  once. See [Upgrading from 0.6.1 or earlier](docs/install.md#upgrading-from-061-or-earlier).
+- **`--no-start` (`-NoStart`)** upgrades without starting or restarting the gateway and says when a
+  restart is due. It keeps this install's port even when a program the installer did not start
+  holds it.
+- **One installer at a time.** A second installer for the same data directory (the line re-run
+  while a console's Update runs, or the other way round) stops with `another AbstractFramework
+  installer is already running` and changes nothing. The lock of an interrupted run is taken over
+  by the next one.
+- **Install AbstractFramework.command runs the latest installer** from GitHub, so double-clicking
+  it again upgrades. It uses the copy it came with only when GitHub cannot be reached, and says so.
+- **`--pin latest` (`-Pin latest`) upgrades an existing install** with the same profile, voice and
+  llama.cpp setup (`uv tool install --upgrade`), and the summary's upgrade lines give the commands
+  that work.
+- **AbstractCore has three install settings** (AbstractCore 2.19.0): `pip install abstractcore`
+  (light: every remote provider, the built-in tools, media inputs, the server and console, and the
+  voice, vision, music and 3D plugins with their remote backends; about 600 MB for AbstractCore
+  alone in a fresh Python 3.12 environment), `pip install
+  "abstractcore[apple]"` (adds every local engine an Apple silicon Mac runs) and `pip install
+  "abstractcore[gpu]"` (adds every local engine an NVIDIA or AMD machine runs). The earlier extra
+  names still install as deprecated aliases; see AbstractCore's
+  [Installation](https://github.com/lpalbou/abstractcore/blob/main/docs/installation.md#deprecated-aliases).
+  `abstract3d` now comes with AbstractCore in every profile, pinned at 0.3.1.
+- **Recommendations for every capability** (AbstractCore 2.19.0).
+  `abstractcore models recommendations` shows the recommended model for text, image input, speech
+  output, speech input, image, video and music on every kind of machine (`--host`: this one), with
+  the engine, the download size, the memory need and whether it fits. On Apple silicon the fit
+  estimate follows the measured GPU memory limit (about 75% of unified memory by default). The text
+  tiers are unchanged: Qwen3.5 9B below 24 GB (on an 8 GB Mac it reads as tight, with a small context), Qwen3.8 27B
+  from 24 GB, Qwen3.8 Flash-Next from 128 GB. A model that runs only with a small context, such as
+  Qwen3.8 27B on a 24 GB Mac, says so and prints the `sudo sysctl iogpu.wired_limit_mb=…` command
+  that gives it more (20480 on 24 GB, 114688 for Flash-Next on 128 GB); nothing runs it for you.
+  The recommended image model is memory-gated: on an 8 GB Mac the gateway's first-run guide shows
+  it as not available here, with the reason, and a saved image or video route that does not fit is
+  flagged. See AbstractCore's
+  [Recommended models](https://github.com/lpalbou/abstractcore/blob/main/docs/recommended-models.md).
+- **Faster video by default** (AbstractVision 0.3.31). Each Wan2.2 model has its own
+  speed-oriented default size: 832x480 (TI2V-5B: 121 frames at 24 fps; T2V-A14B and I2V-A14B: 81
+  frames at 16 fps), and the video is decoded in overlapping tiles, so the decode no longer sets the
+  memory peak. Measured on Apple silicon (MLX peaks at 832x480 with tiled decode), TI2V-5B needs
+  about 16.3 GiB for text-to-video and 16.6 GiB for image-to-video (about 60 GiB at the previous
+  1280x704 default), so the recommended video route is written on Macs with 32 GB of unified memory
+  or more, and a 24 GB Mac is told the GPU memory limit that makes it fit. T2V-A14B and I2V-A14B
+  (about 38.3 and 38.4 GiB, measured) fit a 48 GB Mac once its GPU memory limit is raised. Pass `width`, `height`, `num_frames`, `fps` and `steps` for
+  another size. The memory figures are AbstractVision/mlx-gen's measurements (the engine keeps the
+  text encoder and VAE in memory), not the model's own requirement.
+- **Left and Right switch screens in both terminal consoles** (`abstractgateway-console` 0.11.1,
+  `abstractcore-console` 0.4.1): the previous or next screen in browse mode, wrapping at both ends,
+  like `Ctrl+P` / `Ctrl+N`. A focused element that uses the arrows keeps them: a text field moves
+  its caret, a list or tabs bar changes its selection, a dialog keeps every key, and a focused
+  scrolling pane scrolls.
+
+### Documentation
+
+- **[Upgrade](docs/install.md#upgrade)**: upgrading everything, the gateway only, from the consoles
+  or the menu-bar icon, restarting the gateway, upgrading the apps and checking versions. README and
+  Getting started lead with an **Install** section: the Mac package and the one-line commands.
+
 ## [0.6.1] - 2026-09-28
 
 Every terminal command lands on your PATH, AbstractCode's terminal client says how to sign in, and

{abstractframework-0.6.1/abstractframework.egg-info → abstractframework-0.6.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: abstractframework
-Version: 0.6.1
+Version: 0.6.3
 Summary: Unified installer and documentation hub for the AbstractFramework ecosystem
 Author-email: Laurent-Philippe Albou <contact@abstractframework.ai>
 License: MIT
@@ -23,22 +23,23 @@ Classifier: Typing :: Typed
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: abstractgateway==0.7.1
+Requires-Dist: abstractgateway==0.7.4
 Requires-Dist: abstractassistant==0.9.1
-Requires-Dist: abstractcore==2.18.0
-Requires-Dist: AbstractRuntime==0.7.1
+Requires-Dist: abstractcore==2.19.2
+Requires-Dist: AbstractRuntime==0.7.3
 Requires-Dist: abstractagent==0.3.17
 Requires-Dist: abstractskill==0.3.0
 Requires-Dist: AbstractMemory==0.3.0
 Requires-Dist: abstractsemantics==0.0.5
-Requires-Dist: abstractvoice==0.13.0
-Requires-Dist: abstractvision==0.3.30
+Requires-Dist: abstractvoice==0.13.2
+Requires-Dist: abstractvision==0.3.33
 Requires-Dist: abstractmusic==0.1.15
+Requires-Dist: abstract3d==0.3.2
 Provides-Extra: apple
-Requires-Dist: abstractgateway[apple]==0.7.1; extra == "apple"
+Requires-Dist: abstractgateway[apple]==0.7.4; extra == "apple"
 Requires-Dist: abstractassistant[apple]==0.9.1; platform_system == "Darwin" and extra == "apple"
 Provides-Extra: gpu
-Requires-Dist: abstractgateway[gpu]==0.7.1; extra == "gpu"
+Requires-Dist: abstractgateway[gpu]==0.7.4; extra == "gpu"
 Requires-Dist: abstractassistant[gpu]==0.9.1; extra == "gpu"
 Dynamic: license-file
 
@@ -61,17 +62,7 @@ Think of it as an **agentic OS**: durable runs + replay-first observability + mu
 
 ---
 
-## Quick start
-
-The installer sets up the gateway in your user account (no admin password, no system Python),
-with local voice (Supertonic text-to-speech and Whisper speech-to-text), asks whether to start it
-at login (Enter = yes; an unattended install leaves it off and says how to turn it on), starts it
-on `127.0.0.1:8080`, and opens its web console in your browser already signed in. A first-run guide
-then sets up a local engine (Ollama, LM Studio, MLX, llama.cpp), downloads a model that fits your
-machine, and lists the apps; the gateway serves them on its own address at `/apps/<app>/`. The
-installer also builds the terminal console, `abstractgateway-console`, which runs the same guide on
-a server without a browser and offers itself at the end of an install over SSH
-([Headless or remote machine](docs/install.md#headless-or-remote-machine)).
+## Install
 
 - **Mac, no Terminal:** download and double-click
   [AbstractFramework-Installer.pkg](https://github.com/lpalbou/AbstractFramework/releases/latest/download/AbstractFramework-Installer.pkg).
@@ -90,7 +81,32 @@ a server without a browser and offers itself at the end of an install over SSH
   powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/lpalbou/AbstractFramework/main/scripts/install.ps1 | iex"
   ```
 
-Every failure says what to do next; running the installer again repairs or upgrades in place. To
+To upgrade, run the same line again (on a Mac, double-clicking **Install AbstractFramework.command**
+again does the same), or, on macOS and Linux, press **Update** in the web console, the terminal console or the menu-bar
+icon: each runs the same installer, which installs the latest AbstractFramework release, keeps your
+settings, options and data, and lists what changed. Coming from 0.6.1 or earlier, re-run the line
+once: it reads your earlier options from what is installed (or repeat them once), and the Update
+button of gateway 0.7.1 cannot upgrade. Gateway only, apps, and version checks:
+[Upgrade](docs/install.md#upgrade).
+
+On Linux or Windows with an NVIDIA GPU (`nvidia-smi` works), the installer picks the `gpu` profile:
+PyTorch and llama.cpp on CUDA, Whisper on the GPU, Diffusers image generation. On Linux it needs
+NVIDIA driver 580 or newer. Its release candidate was rehearsed on a Quadro RTX 5000 on Ubuntu
+26.04; the measured results, and what the final versions changed, are in GPU on Linux. The Windows gpu install is
+implemented but not yet validated on real NVIDIA hardware. See [GPU on Linux](docs/install.md#gpu-on-linux-nvidia) and
+[GPU on Windows](docs/install.md#gpu-on-windows-nvidia).
+
+The installer sets up the gateway in your user account (no admin password, no system Python),
+with local voice (Supertonic text-to-speech and Whisper speech-to-text), asks whether to start it
+at login (Enter = yes; an unattended install leaves it off and says how to turn it on), starts it
+on `127.0.0.1:8080`, and opens its web console in your browser already signed in. A first-run guide
+then sets up a local engine (Ollama, LM Studio, MLX, llama.cpp), downloads a model that fits your
+machine, and lists the apps; the gateway serves them on its own address at `/apps/<app>/`. The
+installer also builds the terminal console, `abstractgateway-console`, which runs the same guide on
+a server without a browser and offers itself at the end of an install over SSH
+([Headless or remote machine](docs/install.md#headless-or-remote-machine)).
+
+Every failure says what to do next; running the installer again repairs the install in place. To
 remove it, double-click `Uninstall AbstractFramework.command` (in
 `~/Library/Application Support/AbstractFramework/Installer` after a package install), or run
 `curl -LsSf https://raw.githubusercontent.com/lpalbou/AbstractFramework/main/scripts/uninstall.sh | sh`.
@@ -125,7 +141,9 @@ Start here if you need a lightweight LLM library for scripts, notebooks, or exis
 - Multimodal via capability plugins (Voice, Vision, Music)
 
 ```bash
-pip install abstractcore
+pip install abstractcore              # light: every remote provider (OpenAI, Anthropic, OpenRouter, Ollama, LM Studio, ...)
+pip install "abstractcore[apple]"     # + the local engines of an Apple silicon Mac (MLX, voice, image, video, music)
+pip install "abstractcore[gpu]"       # + the local engines of an NVIDIA or AMD machine (vLLM, voice, image, video, music)
 ```
 
 ```python
@@ -198,7 +216,7 @@ abstractgateway-console --gateway-url http://127.0.0.1:8080 --token <admin token
 ```
 
 Container images are published for the gateway and the AbstractCore server:
-`ghcr.io/lpalbou/abstractgateway:0.7.1` and `ghcr.io/lpalbou/abstractcore-server:2.18.0`.
+`ghcr.io/lpalbou/abstractgateway:0.7.4` and `ghcr.io/lpalbou/abstractcore-server:2.19.2`.
 
 For artifact and runtime-resource investigation, see
 [Runtime artifacts and retrieval](docs/guide/runtime-artifacts.md).
@@ -342,9 +360,9 @@ pip install "abstractframework[gpu]"
 |---|---|---|---|
 | Light | `pip install abstractframework` | macOS, Linux, Windows | 3.10–3.13 |
 | Apple | `pip install "abstractframework[apple]"` | macOS 14+ on Apple Silicon | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
-| GPU | `pip install "abstractframework[gpu]"` | Linux / Windows with a CUDA or ROCm GPU | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
+| GPU | `pip install "abstractframework[gpu]"` | Linux with a CUDA or ROCm GPU; on Windows, use the one-line installer (a plain pip install gets PyPI's CPU-only PyTorch and no llama.cpp) | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
 
-### Release matrix (abstractframework 0.6.1)
+### Release matrix (abstractframework 0.6.3)
 
 `abstractframework` pins every Python package with `==`, so one version of the
 meta-package always installs the same stack. The browser apps and Rust tools are
@@ -353,31 +371,32 @@ and tested together.
 
 | Registry | Package | Version |
 |---|---|---|
-| PyPI | `abstractgateway` | 0.7.1 |
+| PyPI | `abstractgateway` | 0.7.4 |
 | PyPI | `abstractassistant` | 0.9.1 |
-| PyPI | `abstractcore` | 2.18.0 |
-| PyPI | `AbstractRuntime` | 0.7.1 |
+| PyPI | `abstractcore` | 2.19.2 |
+| PyPI | `AbstractRuntime` | 0.7.3 |
 | PyPI | `abstractagent` | 0.3.17 |
 | PyPI | `abstractskill` | 0.3.0 |
 | PyPI | `AbstractMemory` | 0.3.0 |
 | PyPI | `abstractsemantics` | 0.0.5 |
-| PyPI | `abstractvoice` | 0.13.0 |
-| PyPI | `abstractvision` | 0.3.30 |
+| PyPI | `abstractvoice` | 0.13.2 |
+| PyPI | `abstractvision` | 0.3.33 |
 | PyPI | `abstractmusic` | 0.1.15 |
+| PyPI | `abstract3d` | 0.3.2 |
 | npm | `@abstractframework/flow` | 0.4.0 |
-| npm | `@abstractframework/code` | 0.6.1 |
+| npm | `@abstractframework/code` | 0.6.2 |
 | npm | `@abstractframework/observer` | 0.2.1 |
 | npm | `@abstractframework/continuum` | 0.4.0 |
 | npm | `@abstractframework/entity` | 0.3.0 |
 | crates.io | `abstractcode` | 0.7.1 |
-| crates.io | `abstractgateway-console` | 0.11.0 |
-| crates.io | `abstractcore-console` | 0.4.0 |
+| crates.io | `abstractgateway-console` | 0.11.2 |
+| crates.io | `abstractcore-console` | 0.4.1 |
 | crates.io | `abstracttui` | 0.6.0 |
-| GHCR | `ghcr.io/lpalbou/abstractgateway` | 0.7.1 (`gpu-latest` / `<version>-gpu` experimental) |
-| GHCR | `ghcr.io/lpalbou/abstractcore-server` | 2.18.0 |
+| GHCR | `ghcr.io/lpalbou/abstractgateway` | 0.7.4 (`gpu-latest` / `<version>-gpu` experimental) |
+| GHCR | `ghcr.io/lpalbou/abstractcore-server` | 2.19.2 |
 
-Optional add-ons that are not part of any profile install separately:
-`pip install abstract3d` (0.3.1) and `pip install abstractcamera` (0.2.0). `abstractskill` (pinned above)
+`abstract3d` (pinned above) comes with AbstractCore in every profile. The optional add-on that is
+not part of any profile installs separately: `pip install abstractcamera` (0.2.1). `abstractskill` (pinned above)
 comes with the gateway, which carries its curated skill shelf; install it on its own with
 `pip install abstractskill`.
 

{abstractframework-0.6.1 → abstractframework-0.6.3}/README.md

@@ -17,17 +17,7 @@ Think of it as an **agentic OS**: durable runs + replay-first observability + mu
 
 ---
 
-## Quick start
-
-The installer sets up the gateway in your user account (no admin password, no system Python),
-with local voice (Supertonic text-to-speech and Whisper speech-to-text), asks whether to start it
-at login (Enter = yes; an unattended install leaves it off and says how to turn it on), starts it
-on `127.0.0.1:8080`, and opens its web console in your browser already signed in. A first-run guide
-then sets up a local engine (Ollama, LM Studio, MLX, llama.cpp), downloads a model that fits your
-machine, and lists the apps; the gateway serves them on its own address at `/apps/<app>/`. The
-installer also builds the terminal console, `abstractgateway-console`, which runs the same guide on
-a server without a browser and offers itself at the end of an install over SSH
-([Headless or remote machine](docs/install.md#headless-or-remote-machine)).
+## Install
 
 - **Mac, no Terminal:** download and double-click
   [AbstractFramework-Installer.pkg](https://github.com/lpalbou/AbstractFramework/releases/latest/download/AbstractFramework-Installer.pkg).
@@ -46,7 +36,32 @@ a server without a browser and offers itself at the end of an install over SSH
   powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/lpalbou/AbstractFramework/main/scripts/install.ps1 | iex"
   ```
 
-Every failure says what to do next; running the installer again repairs or upgrades in place. To
+To upgrade, run the same line again (on a Mac, double-clicking **Install AbstractFramework.command**
+again does the same), or, on macOS and Linux, press **Update** in the web console, the terminal console or the menu-bar
+icon: each runs the same installer, which installs the latest AbstractFramework release, keeps your
+settings, options and data, and lists what changed. Coming from 0.6.1 or earlier, re-run the line
+once: it reads your earlier options from what is installed (or repeat them once), and the Update
+button of gateway 0.7.1 cannot upgrade. Gateway only, apps, and version checks:
+[Upgrade](docs/install.md#upgrade).
+
+On Linux or Windows with an NVIDIA GPU (`nvidia-smi` works), the installer picks the `gpu` profile:
+PyTorch and llama.cpp on CUDA, Whisper on the GPU, Diffusers image generation. On Linux it needs
+NVIDIA driver 580 or newer. Its release candidate was rehearsed on a Quadro RTX 5000 on Ubuntu
+26.04; the measured results, and what the final versions changed, are in GPU on Linux. The Windows gpu install is
+implemented but not yet validated on real NVIDIA hardware. See [GPU on Linux](docs/install.md#gpu-on-linux-nvidia) and
+[GPU on Windows](docs/install.md#gpu-on-windows-nvidia).
+
+The installer sets up the gateway in your user account (no admin password, no system Python),
+with local voice (Supertonic text-to-speech and Whisper speech-to-text), asks whether to start it
+at login (Enter = yes; an unattended install leaves it off and says how to turn it on), starts it
+on `127.0.0.1:8080`, and opens its web console in your browser already signed in. A first-run guide
+then sets up a local engine (Ollama, LM Studio, MLX, llama.cpp), downloads a model that fits your
+machine, and lists the apps; the gateway serves them on its own address at `/apps/<app>/`. The
+installer also builds the terminal console, `abstractgateway-console`, which runs the same guide on
+a server without a browser and offers itself at the end of an install over SSH
+([Headless or remote machine](docs/install.md#headless-or-remote-machine)).
+
+Every failure says what to do next; running the installer again repairs the install in place. To
 remove it, double-click `Uninstall AbstractFramework.command` (in
 `~/Library/Application Support/AbstractFramework/Installer` after a package install), or run
 `curl -LsSf https://raw.githubusercontent.com/lpalbou/AbstractFramework/main/scripts/uninstall.sh | sh`.
@@ -81,7 +96,9 @@ Start here if you need a lightweight LLM library for scripts, notebooks, or exis
 - Multimodal via capability plugins (Voice, Vision, Music)
 
 ```bash
-pip install abstractcore
+pip install abstractcore              # light: every remote provider (OpenAI, Anthropic, OpenRouter, Ollama, LM Studio, ...)
+pip install "abstractcore[apple]"     # + the local engines of an Apple silicon Mac (MLX, voice, image, video, music)
+pip install "abstractcore[gpu]"       # + the local engines of an NVIDIA or AMD machine (vLLM, voice, image, video, music)
 ```
 
 ```python
@@ -154,7 +171,7 @@ abstractgateway-console --gateway-url http://127.0.0.1:8080 --token <admin token
 ```
 
 Container images are published for the gateway and the AbstractCore server:
-`ghcr.io/lpalbou/abstractgateway:0.7.1` and `ghcr.io/lpalbou/abstractcore-server:2.18.0`.
+`ghcr.io/lpalbou/abstractgateway:0.7.4` and `ghcr.io/lpalbou/abstractcore-server:2.19.2`.
 
 For artifact and runtime-resource investigation, see
 [Runtime artifacts and retrieval](docs/guide/runtime-artifacts.md).
@@ -298,9 +315,9 @@ pip install "abstractframework[gpu]"
 |---|---|---|---|
 | Light | `pip install abstractframework` | macOS, Linux, Windows | 3.10–3.13 |
 | Apple | `pip install "abstractframework[apple]"` | macOS 14+ on Apple Silicon | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
-| GPU | `pip install "abstractframework[gpu]"` | Linux / Windows with a CUDA or ROCm GPU | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
+| GPU | `pip install "abstractframework[gpu]"` | Linux with a CUDA or ROCm GPU; on Windows, use the one-line installer (a plain pip install gets PyPI's CPU-only PyTorch and no llama.cpp) | 3.10–3.13 (F5-TTS voice cloning needs 3.11+) |
 
-### Release matrix (abstractframework 0.6.1)
+### Release matrix (abstractframework 0.6.3)
 
 `abstractframework` pins every Python package with `==`, so one version of the
 meta-package always installs the same stack. The browser apps and Rust tools are
@@ -309,31 +326,32 @@ and tested together.
 
 | Registry | Package | Version |
 |---|---|---|
-| PyPI | `abstractgateway` | 0.7.1 |
+| PyPI | `abstractgateway` | 0.7.4 |
 | PyPI | `abstractassistant` | 0.9.1 |
-| PyPI | `abstractcore` | 2.18.0 |
-| PyPI | `AbstractRuntime` | 0.7.1 |
+| PyPI | `abstractcore` | 2.19.2 |
+| PyPI | `AbstractRuntime` | 0.7.3 |
 | PyPI | `abstractagent` | 0.3.17 |
 | PyPI | `abstractskill` | 0.3.0 |
 | PyPI | `AbstractMemory` | 0.3.0 |
 | PyPI | `abstractsemantics` | 0.0.5 |
-| PyPI | `abstractvoice` | 0.13.0 |
-| PyPI | `abstractvision` | 0.3.30 |
+| PyPI | `abstractvoice` | 0.13.2 |
+| PyPI | `abstractvision` | 0.3.33 |
 | PyPI | `abstractmusic` | 0.1.15 |
+| PyPI | `abstract3d` | 0.3.2 |
 | npm | `@abstractframework/flow` | 0.4.0 |
-| npm | `@abstractframework/code` | 0.6.1 |
+| npm | `@abstractframework/code` | 0.6.2 |
 | npm | `@abstractframework/observer` | 0.2.1 |
 | npm | `@abstractframework/continuum` | 0.4.0 |
 | npm | `@abstractframework/entity` | 0.3.0 |
 | crates.io | `abstractcode` | 0.7.1 |
-| crates.io | `abstractgateway-console` | 0.11.0 |
-| crates.io | `abstractcore-console` | 0.4.0 |
+| crates.io | `abstractgateway-console` | 0.11.2 |
+| crates.io | `abstractcore-console` | 0.4.1 |
 | crates.io | `abstracttui` | 0.6.0 |
-| GHCR | `ghcr.io/lpalbou/abstractgateway` | 0.7.1 (`gpu-latest` / `<version>-gpu` experimental) |
-| GHCR | `ghcr.io/lpalbou/abstractcore-server` | 2.18.0 |
+| GHCR | `ghcr.io/lpalbou/abstractgateway` | 0.7.4 (`gpu-latest` / `<version>-gpu` experimental) |
+| GHCR | `ghcr.io/lpalbou/abstractcore-server` | 2.19.2 |
 
-Optional add-ons that are not part of any profile install separately:
-`pip install abstract3d` (0.3.1) and `pip install abstractcamera` (0.2.0). `abstractskill` (pinned above)
+`abstract3d` (pinned above) comes with AbstractCore in every profile. The optional add-on that is
+not part of any profile installs separately: `pip install abstractcamera` (0.2.1). `abstractskill` (pinned above)
 comes with the gateway, which carries its curated skill shelf; install it on its own with
 `pip install abstractskill`.
 

{abstractframework-0.6.1 → abstractframework-0.6.3}/abstractframework/__init__.py

@@ -10,21 +10,22 @@ Most implementation functionality still lives in component projects.
 
 from __future__ import annotations
 
-__version__ = "0.6.1"
+__version__ = "0.6.3"
 __author__ = "Laurent-Philippe Albou"
 __license__ = "MIT"
 
 RELEASE_VERSIONS: dict[str, str] = {
-    "abstractcore": "2.18.0",
-    "abstractruntime": "0.7.1",
+    "abstractcore": "2.19.2",
+    "abstractruntime": "0.7.3",
     "abstractagent": "0.3.17",
-    "abstractgateway": "0.7.1",
+    "abstractgateway": "0.7.4",
     "abstractskill": "0.3.0",
     "abstractmemory": "0.3.0",
     "abstractsemantics": "0.0.5",
-    "abstractvoice": "0.13.0",
-    "abstractvision": "0.3.30",
+    "abstractvoice": "0.13.2",
+    "abstractvision": "0.3.33",
     "abstractmusic": "0.1.15",
+    "abstract3d": "0.3.2",
     "abstractassistant": "0.9.1",
 }
 
@@ -39,12 +40,13 @@ PACKAGE_DISTRIBUTIONS: dict[str, str] = {
     "abstractvoice": "abstractvoice",
     "abstractvision": "abstractvision",
     "abstractmusic": "abstractmusic",
+    "abstract3d": "abstract3d",
     "abstractassistant": "abstractassistant",
 }
 
 NPM_RELEASE_VERSIONS: dict[str, str] = {
     "@abstractframework/flow": "0.4.0",
-    "@abstractframework/code": "0.6.1",
+    "@abstractframework/code": "0.6.2",
     "@abstractframework/observer": "0.2.1",
     "@abstractframework/continuum": "0.4.0",
     "@abstractframework/entity": "0.3.0",
@@ -53,21 +55,16 @@ NPM_RELEASE_VERSIONS: dict[str, str] = {
 # Terminal tools published on crates.io alongside this release (installed with `cargo install`,
 # or by the bootstrap scripts: the gateway console and abstractcode by default).
 CRATE_RELEASE_VERSIONS: dict[str, str] = {
-    "abstractgateway-console": "0.11.0",
-    "abstractcore-console": "0.4.0",
+    "abstractgateway-console": "0.11.2",
+    "abstractcore-console": "0.4.1",
     "abstractcode": "0.7.1",
     "abstracttui": "0.6.0",
 }
 
-CORE_DEFAULT_EXTRAS = [
-    "remote",
-    "tools",
-    "media",
-    "vision",
-    "voice",
-    "audio",
-    "music",
-]
+# The light profile installs plain `abstractcore` (no extra): AbstractCore's light setting already
+# carries every remote provider, the tools, media and the capability plugins. The apple and gpu
+# profiles get AbstractCore's local engines through the gateway's profile.
+CORE_DEFAULT_EXTRAS: list[str] = []
 
 # Convenience re-exports (AbstractCore is a base dependency of this meta-package).
 # Keep this import lightweight: do not import optional tool/media deps here.