v0.2.0

Added

• GET /api/loaded gives each model a status, loaded or sleeping, and taraengine ps shows it in a STATUS
  column; taraengine load and taraengine doctor say when a model sleeps.
• An import from hf:// or a local path takes its native context window from the file's header when it is pulled,
  and its default window is no larger; imports pulled by 0.1.0 get theirs when the daemon starts.
• An import keeps the capabilities it showed while it ran, across unloads and restarts.
• pkg/client: Error.NotJSON marks a refusal whose body is not the engine's error JSON; such a body is shown as
  one line of at most 300 bytes.

Changed

• The memory figure and the capabilities of a running model are recorded whoever started it: a model that a /v1
  request brings up gets them from a later read of the model list. Reading a model's properties never makes
  llama-server load it.
• POST /api/load answers at once for a model that sleeps at the requested window, and waits for a model that is
  still loading instead of refusing the load.
• The end of llama-server's log that GET /api/models/{name} carries keeps line breaks and tabs and drops every other
  control character.
• A pre-release tag such as v0.2.0-rc.1 is released with the [Unreleased] section of this file as its notes,
  unless it has a section of its own.
• The catalog runs a q8_0 KV cache (cache-type-k and cache-type-v) for gemma4:e4b, glm-4.7-flash, qwen3.6:27b,
  qwen3.8:27b and qwen3.6:35b, which holds less memory at the same window; the README's catalog table names each
  model's KV cache. A key set in preset_global wins over a catalog model's own preset for that key, and with
  flash attention turned off there a catalog model keeps an f16 V cache, which llama.cpp quantizes only with it.

v0.2.0-rc.2

Added

• GET /api/loaded gives each model a status, loaded or sleeping, and taraengine ps shows it in a STATUS
  column; taraengine load and taraengine doctor say when a model sleeps.
• An import from hf:// or a local path takes its native context window from the file's header when it is pulled,
  and its default window is no larger; imports pulled by 0.1.0 get theirs when the daemon starts.
• An import keeps the capabilities it showed while it ran, across unloads and restarts.
• pkg/client: Error.NotJSON marks a refusal whose body is not the engine's error JSON; such a body is shown as
  one line of at most 300 bytes.

Changed

• The memory figure and the capabilities of a running model are recorded whoever started it: a model that a /v1
  request brings up gets them from a later read of the model list. Reading a model's properties never makes
  llama-server load it.
• POST /api/load answers at once for a model that sleeps at the requested window, and waits for a model that is
  still loading instead of refusing the load.
• The end of llama-server's log that GET /api/models/{name} carries keeps line breaks and tabs and drops every other
  control character.
• A pre-release tag such as v0.2.0-rc.1 is released with the [Unreleased] section of this file as its notes,
  unless it has a section of its own.
• The catalog runs a q8_0 KV cache (cache-type-k and cache-type-v) for gemma4:e4b, glm-4.7-flash, qwen3.6:27b,
  qwen3.8:27b and qwen3.6:35b, which holds less memory at the same window; the README's catalog table names each
  model's KV cache. A key set in preset_global wins over a catalog model's own preset for that key, and with
  flash attention turned off there a catalog model keeps an f16 V cache, which llama.cpp quantizes only with it.

v0.2.0-rc.1

Added

• GET /api/loaded gives each model a status, loaded or sleeping, and taraengine ps shows it in a STATUS
  column; taraengine load and taraengine doctor say when a model sleeps.
• An import from hf:// or a local path takes its native context window from the file's header when it is pulled,
  and its default window is no larger; imports pulled by 0.1.0 get theirs when the daemon starts.
• An import keeps the capabilities it showed while it ran, across unloads and restarts.
• pkg/client: Error.NotJSON marks a refusal whose body is not the engine's error JSON; such a body is shown as
  one line of at most 300 bytes.

Changed

• The memory figure and the capabilities of a running model are recorded whoever started it: a model that a /v1
  request brings up gets them from a later read of the model list. Reading a model's properties never makes
  llama-server load it.
• POST /api/load answers at once for a model that sleeps at the requested window, and waits for a model that is
  still loading instead of refusing the load.
• The end of llama-server's log that GET /api/models/{name} carries keeps line breaks and tabs and drops every other
  control character.
• A pre-release tag such as v0.2.0-rc.1 is released with the [Unreleased] section of this file as its notes,
  unless it has a section of its own.

v0.1.0

Added

• taraengine serve: supervises one llama-server (upstream build b11514, pinned and checksum-verified) in
  router mode and serves the control API on 127.0.0.1:8272 with an OpenAI-compatible /v1.
• taraengine runtime install | update | status: the pinned build per platform, CUDA 12.8 or 13.4 by driver,
  the CPU build as the fallback, Metal on Apple Silicon.
• taraengine pull | list | ps | load | unload | rm | doctor: a curated registry of eight models pinned to
  Hugging Face artifacts with their projector and draft siblings, plus hf:// and local escape hatches.
• Measured GPU memory per loaded model on NVIDIA, estimated resident memory on macOS.
• .deb and .rpm packages that install the binary under /usr/bin and carry LICENSE and NOTICE.

Security

• The API refuses web pages of other origins: only pages served from localhost, 127.0.0.1 or [::1] and the
  allowed_origins of the configuration may use it. The daemon answers preflights itself, removes the CORS headers of
  llama-server, and requires application/json on every POST under /api and /v1.
• The supervised llama-server runs with an allowlisted environment, without the API key and the Hugging Face token,
  and with an empty download cache, so it serves only the models of the catalog.
• A key in the configuration file that this version does not know is an error that names the key.
• Release workflows pin every action to a commit; CodeQL, Dependabot and goreleaser check run in CI.