AbstractCore v2.18.0

Capability routes say when their engine is not installed, Apple silicon fits account for the GPU
memory limit macOS can raise, the saved OpenAI key reaches AbstractVoice (never for a request the
server has not authenticated), and neither abstractcore engines open nor the terminal console
(abstractcore-console 0.4.0) tries to open a browser on a machine without a display.

Added

• engine_missing: {engine, name, reason, install[, engine_row]} on capability routes whose
  in-process engine is not installed in this Python environment: MLX (mlx-lm), MLX-Gen
  (abstractvision + mlx-gen), llama.cpp (GGUF models on the huggingface provider), Transformers
  (other huggingface models) and every local voice engine (Supertonic, faster-whisper, ...), whose
  answer is AbstractVoice's own abstractvoice.engine_runtime probe (an AbstractVoice older than
  0.13.0 is itself the missing engine of those routes, with its upgrade command, and the rest of
  the grid is unaffected). install targets this Python interpreter, like the Engines screen's
  rows (uv pip install --python <interpreter> ... in a pip-less uv environment, else
  <interpreter> -m pip install ...). It appears on
  config defaults rows, apply-recommended entries and the recommended download plan, and the
  CLI prints it. It is distinct from route_unavailable (this host cannot run the engine at all)
  and from weights that are not downloaded. See Centralized config.
• Fit verdict needs_gpu_limit (Apple silicon): the model does not fit the current GPU memory
  limit but fits under one macOS can grant; fit.gpu_limit carries the exact
  sudo sysctl iogpu.wired_limit_mb=<MB> command, the value and the verdict after it. A raised
  iogpu.wired_limit_mb is the ceiling in the light host reading too. The 128 GiB tier's
  recommended text model reports this verdict with its command instead of too_large, and the
  --fits filter keeps such models. See Local models.
• The OpenAI API key saved in AbstractCore's config reaches the voice plugin as AbstractVoice's host
  setting voice_openai_api_key: from the config file an instance was created for (the gateway's
  core config), else the global config. An explicit create_llm(..., voice_openai_api_key=...)
  wins; with no key saved nothing is set. In the server's audio routes the key is read on every
  request and handed over only to server-authenticated requests whose voice engine is openai or
  openai-compatible, on a capability host built for that request (never cached): a rotated key
  is used on the next call, a removed one is no longer sent. See Server.
• Wan2.2 T2V-A14B and I2V-A14B 8-bit carry their measured run-time memory (resident): about
  72 GiB each at AbstractVision's default canvas (1280x720, 81 frames), so their fit reads
  too_large on 64 GiB Macs, needs_gpu_limit on 96 GiB Macs and fits from 128 GiB. The
  recommended video route is unchanged (Wan2.2 TI2V-5B from about 96 GiB).
• The Supertonic voice artifact (supertonic-3) carries its download size (401,276,744 bytes).
• CI job local-models: the whole suite with CPU torch, Transformers and llama.cpp installed; the
  release workflow requires it.

Changed

• Server audio routes without a provider/model (the AbstractVoice path: speech, speech stream,
  transcriptions, voice clone, and their /{provider}/v1/... forms) apply the same rule as
  model=openai/...: when the engine that would run is remote in AbstractVoice's own words
  (abstractvoice.engine_runtime: OpenAI or an OpenAI-compatible server, under every spelling it
  accepts, such as remote, compatible or proxy), or is an engine AbstractVoice does not know,
  and the server holds an OpenAI key (saved, or OPENAI_API_KEY, which AbstractVoice would fall
  back to), a request that is not server-authenticated gets 401 (also with
  ABSTRACTCORE_SERVER_ALLOW_UNAUTHENTICATED=1); an explicit X-AbstractCore-Provider-API-Key
  is then the only key used. Local engines are unaffected. The audio discovery routes
  (/v1/audio/voices, .../providers, .../models) follow the same rule: the saved key counts as
  server-held, ABSTRACTCORE_SERVER_ALLOW_UNAUTHENTICATED=1 no longer exempts them, and they pass
  the saved key to AbstractVoice only for authenticated requests.
• The vision routes guarded by server-held OpenAI credentials (/v1/vision/providers/,
  /v1/vision/models and the other vision discovery routes, and image/video generation, edits and
  upscale through OpenAI or an OpenAI-compatible upstream) apply the same rule, through the one
  guard the audio discovery routes use (abstractcore.server.credentials): an unauthenticated
  request gets 401 when the server holds an OpenAI key (saved or OPENAI_API_KEY), also with
  ABSTRACTCORE_SERVER_ALLOW_UNAUTHENTICATED=1; an explicit X-AbstractCore-Provider-API-Key is
  spent instead.
• /v1/audio/music and /{provider}/v1/audio/music: while the server holds a key AbstractMusic
  reads (ACEMUSIC_API_KEY, ELEVENLABS_API_KEY, in the environment or as the matching
  music_*_api_key route option), a request that is not server-authenticated gets 401 whatever
  backend it names, also with ABSTRACTCORE_SERVER_ALLOW_UNAUTHENTICATED=1; an explicit
  X-AbstractCore-Provider-API-Key is handed to the remote music backends in place of the
  server's keys. Local music on a server holding none of these keys works unauthenticated. This is
  an interim rule until AbstractMusic can say which of its backends are remote.
• The capability discovery routes (/v1/capabilities, /v1/capabilities/{capability}/providers,
  /v1/capabilities/{capability}/models, /v1/audio/music/providers, /v1/audio/music/models,
  /v1/audio/music/provider-details) follow the same rules: they call into the plugins, which
  probe their remote providers with the keys they read, so a request that is not
  server-authenticated gets 401 while the server holds a key the capability could spend (OpenAI
  for voice, audio and vision; the music keys for music; both for /v1/capabilities and for any
  other capability). A caller's own key is handed to the plugins in place of the server's.
• Music requests: an unknown backend selector (body provider or route prefix, e.g.
  /elevenlabs/v1/audio/music; AbstractMusic's ElevenLabs selector is elevenlabs-music) is
  refused with 400 instead of running another backend; a selected backend that is not installed
  is refused (501) instead of falling back to another one; a caller's own key reaches only the
  backend the request selected.
• abstractcore engines open prints the link and launches nothing without a display (SSH session;
  Linux/BSD without DISPLAY or WAYLAND_DISPLAY), with the reason (not_opened in --json):
  the console crate's rule, now also abstractcore.utils.display in Python.
• The voice extras (voice, audio, all, all-apple, all-gpu, all-non-mlx, full-dev)
  require abstractvoice>=0.13.0, the first release with the public engine runtime probe.
• vLLM's engine row on a host without an NVIDIA GPU reads "vLLM needs Linux with an NVIDIA GPU
  (CUDA); use a remote vLLM server instead".
• abstractcore-console 0.4.0: o never launches a browser without a display (SSH
  sessions; Linux/BSD without DISPLAY or WAYLAND_DISPLAY) and prints the link instead; routes
  whose engine is not installed read engine not installed with the install command; models that
  need a raised GPU limit show the command; remote-only engines read remote only.

Fixed

• Five local-model tests (Transformers cached lane, GGUF control plane) match the provider's
  current signatures and cache contracts.
• The Transformers snapshot-lane tests load their tokenizer at test time from the cached snapshot
  directory, and skip without network when it is not cached: loading it by repo id at collection
  made huggingface_hub 1.33 contact huggingface.co, and the test network guard failed the
  local-models job.
• Clearing a configured API key (set_api_key(provider, "")) restores the environment variable
  value it had shadowed (or unsets it), so the removed key stops being used in this process.