Prepare 0.8.20 release

CHANGELOG.md

@@ -1,47 +1,42 @@
-## Unreleased
+## 0.8.20
+
+* Added `code_assets` 2.x compatibility while retaining support for 1.x native
+  asset toolchains.
 
 * Updated the default llama.cpp native runtime pin to
   `leehack/llamadart-native@b10514`, adding BailingMoE3,
   GraniteSWA/GraniteMoeSWA, speculators-format DSpark checkpoints, and current
-  upstream multimodal/backend fixes and performance improvements. Regenerated
-  matching Dart FFI bindings, refreshed the `llamadart_llama_cpp_flutter`
-  Apple SwiftPM checksum, and aligned current native-runtime documentation.
+  upstream multimodal/backend improvements. Matching Dart FFI bindings and the
+  `llamadart_llama_cpp_flutter` Apple SwiftPM artifact were refreshed.
 
 * Updated the default WebGPU bridge assets to `v0.1.37`, embedding llama.cpp
   `b10514` to restore native/Web parity. The bridge provisions an explicit 1 MiB
-  Wasm stack for both wasm32 and memory64, preventing the `b10514` graph
-  parameter growth from aborting Qwen3-ASR memory64 context construction.
+  Wasm stack for wasm32 and memory64, preventing Qwen3-ASR memory64 context
+  construction from overflowing the default stack.
 
 * Improved Web microphone transcription by warming up browser capture before
   showing the recording-ready state, trimming the warmup silence, and
   rejecting too-short, silent, or unsupported PCM WAV captures before
   inference.
 
 * Added logical batch-size (`n_batch`) and micro-batch-size (`n_ubatch`)
-  controls for llama.cpp/WebGPU models in the Flutter chat example. Android
-  Auto now probes the packaged Vulkan device before memory planning so capable
-  models do not fall back to CPU merely because the dynamic GPU module was not
-  registered yet.
-
-* Updated the native LiteRT-LM runtime to `v0.16.0-native.2` and added an
-  experimental dedicated `SpeechToTextEngine.liteRtLm` path with runtime
-  capability discovery, worker-isolated CPU inference, bounded mono 16 kHz
-  float PCM input, partial/final transcript events, finalization, and
-  cooperative cancellation. The low-level synchronous ASR session remains
-  available for advanced integrations. The Apple companion packages the iOS
-  Gemma constraint provider and Metal accelerator/sampler plugins required by
-  the published runtime.
+  controls for llama.cpp/WebGPU models in the Flutter chat example.
+
+* Improved Android Auto backend selection by probing the packaged Vulkan device
+  before memory planning, avoiding unnecessary CPU fallback on capable models.
+
+* Updated the native LiteRT-LM runtime to `v0.16.0-native.2`. The Apple companion
+  packages the iOS Gemma constraint provider and Metal accelerator/sampler
+  plugins required by the published runtime.
+
+* Added an experimental `SpeechToTextEngine.liteRtLm` path with capability
+  discovery, worker-isolated CPU inference, bounded mono 16 kHz float PCM,
+  partial/final transcript events, finalization, and cancellation.
 
 * Added experimental live dictation to the native Flutter chat example for
-  chat models, including generic audio-chat models. The app offers the
-  recommended 54 MB Moonshine Tiny sidecar and an optional higher-capacity,
-  heavier 615 MB Parakeet TDT 0.6B sidecar with checksum-pinned assets,
-  persistent selection, explicit download progress, cancellation, and retry.
-  It captures mono 16 kHz PCM, uses the public worker-isolated streaming speech
-  API, renders confirmed and pending text, and returns the finalized transcript
-  to the editable composer. A persisted settings switch explains and enables
-  or disables the optional workflow. The first path is CPU-only,
-  English-only, capped at five minutes, and unavailable on Linux and Web.
+  chat models using selectable Moonshine Tiny and Parakeet TDT 0.6B sidecars.
+  Live dictation is CPU-only, English-only, capped at five minutes, and
+  unavailable on Linux and Web.
 
 * Improved Flutter chat example model downloads with bounded retries for
   transient network failures, safe resume after truncated responses, and a
@@ -52,44 +47,26 @@
   and stopped streaming responses from pulling users away from chat history.
 
 * Added an experimental typed `TextToSpeechEngine` for native llama.cpp and
-  WebGPU bridge assets `v0.1.33+` with Qwen3-TTS models, including capability
-  discovery, language and optional speaker
-  reference input, cancellable progress, complete 24 kHz PCM output, and WAV
-  encoding. The Flutter chat example adds a checksum-pinned Qwen3-TTS preset
-  with file or microphone speaker references, automatic playback, replay, and
-  WAV save controls. Web accepts byte-backed speaker references and complete
-  PCM/WAV output, with a bounded example-app utterance length for browser
-  stability. The example pins bridge assets `v0.1.37`, retaining the `v0.1.34`
-  recovery that retries a failed worker WebGPU synthesis once on CPU; current
-  LiteRT-LM artifacts fail explicitly as unsupported. Apple companion builds
-  now discover the native TTS wrapper ABI from the embedded llama framework.
+  WebGPU with Qwen3-TTS models, including capability discovery, speaker
+  references, cancellable progress, complete 24 kHz PCM output, and WAV
+  encoding. The Flutter chat example adds synthesis, playback, replay, and WAV
+  save controls. Apple apps discover the TTS ABI in the embedded llama
+  framework; current LiteRT-LM artifacts fail explicitly as unsupported.
 
 * Added an experimental typed `SpeechToTextEngine` with an explicit Qwen3-ASR
   adapter profile for whole-file llama.cpp transcription. The Flutter chat
-  example includes a SHA-256-verified Qwen3-ASR 0.6B preset plus separate
-  **Transcribe Audio** and foreground microphone recording flows on native and
-  WebGPU. Web requires validated bridge assets `v0.1.30+`, pins `v0.1.37` (with
-  the short-speech recovery introduced in `v0.1.32` and the `b10514` memory64
-  stack fix) to
-  recover short speech that would otherwise terminate empty, accepts WAV bytes
-  only, and remains separate from native LiteRT-LM live dictation. The native
-  flow has been exercised on a physical Pixel with CPU inference and in the
-  iOS Simulator.
+  example includes a checksum-pinned Qwen3-ASR 0.6B preset plus file and
+  microphone transcription on native and WebGPU. Web accepts WAV bytes only;
+  native LiteRT-LM live dictation remains a separate implementation.
 
 * Added **Ask with voice** to the native Flutter chat example for Gemma 4 E2B
-  through both the LiteRT-LM direct-media bundle and the audio-capable GGUF +
-  projector path. The app sends a microphone recording through normal
-  multimodal chat so the model can answer the spoken request; it remains
-  separate from typed speech-to-text and is code-supported on Android, iOS,
-  macOS, and Windows, with Linux recording and Web excluded. Current packaged
-  microphone validation covers LiteRT-LM on macOS; llama.cpp audio input is
-  experimental and has engine-level Metal evidence only. Android, iOS, and
-  Windows device validation remains outstanding.
+  through LiteRT-LM direct media and audio-capable GGUF + projector paths. It
+  sends a microphone recording through multimodal chat and remains separate
+  from typed speech-to-text.
 
 * Added experimental, opt-in native llama.cpp DSpark speculative decoding
-  through `SpeculativeDecodingConfig.draftDspark(...)`, including exact
-  `draft-dspark` mapping, external draft-model validation, typed unsupported
-  failures, and benchmark coverage.
+  through `SpeculativeDecodingConfig.draftDspark(...)` with a compatible
+  external draft model.
 
 ## 0.8.19
 

README.md

@@ -55,17 +55,17 @@ For Dart or Flutter apps:
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
+  llamadart: ^0.8.20
 ```
 
 Flutter iOS/macOS apps that should link Apple XCFrameworks through Swift
 Package Manager should also add the runtime companion packages they need:
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
-  llamadart_llama_cpp_flutter: ^0.0.13 # GGUF / llama.cpp
-  llamadart_litert_lm_flutter: ^0.0.9 # Apple .litertlm / LiteRT-LM targets
+  llamadart: ^0.8.20
+  llamadart_llama_cpp_flutter: ^0.0.14 # GGUF / llama.cpp
+  llamadart_litert_lm_flutter: ^0.0.10 # Apple .litertlm / LiteRT-LM targets
 ```
 
 The LiteRT-LM companion manifest includes the complete iOS SwiftPM runtime

example/chat_app/pubspec.lock

@@ -413,21 +413,21 @@ packages:
       path: "../.."
       relative: true
     source: path
-    version: "0.8.19"
+    version: "0.8.20"
   llamadart_litert_lm_flutter:
     dependency: "direct main"
     description:
       path: "../../packages/llamadart_litert_lm_flutter"
       relative: true
     source: path
-    version: "0.0.9"
+    version: "0.0.10"
   llamadart_llama_cpp_flutter:
     dependency: "direct main"
     description:
       path: "../../packages/llamadart_llama_cpp_flutter"
       relative: true
     source: path
-    version: "0.0.13"
+    version: "0.0.14"
   logging:
     dependency: transitive
     description:

packages/llamadart_litert_lm_flutter/CHANGELOG.md

@@ -1,10 +1,11 @@
-## Unreleased
+## 0.0.10
 
 * Updated Apple SwiftPM native pin to `leehack/litert-lm-native@v0.16.0-native.2`
   and packaged the iOS Gemma constraint provider and Metal accelerator/sampler
   plugins required by that runtime.
 * Kept macOS on the complete hook-managed runtime bundle instead of linking an
   incomplete SwiftPM subset.
+* Updated the install example for `llamadart` 0.8.20.
 
 ## 0.0.9
 

packages/llamadart_litert_lm_flutter/README.md

@@ -15,8 +15,8 @@ also request an x86_64 Simulator slice must exclude x86_64 for LiteRT-LM builds.
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
-  llamadart_litert_lm_flutter: ^0.0.9
+  llamadart: ^0.8.20
+  llamadart_litert_lm_flutter: ^0.0.10
 ```
 
 This package has no runtime Dart API of its own. Import `package:llamadart`

packages/llamadart_litert_lm_flutter/pubspec.yaml

@@ -1,6 +1,6 @@
 name: llamadart_litert_lm_flutter
 description: Flutter Apple SwiftPM runtime companion package for llamadart LiteRT-LM support.
-version: 0.0.9
+version: 0.0.10
 repository: https://github.com/leehack/llamadart/tree/main/packages/llamadart_litert_lm_flutter
 homepage: https://github.com/leehack/llamadart
 issue_tracker: https://github.com/leehack/llamadart/issues

packages/llamadart_llama_cpp_flutter/CHANGELOG.md

@@ -1,4 +1,4 @@
-## Unreleased
+## 0.0.14
 
 * Updated Apple SwiftPM native pin to `leehack/llamadart-native@b10514`.
 

packages/llamadart_llama_cpp_flutter/README.md

@@ -9,8 +9,8 @@ core package's native-assets fallback.
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
-  llamadart_llama_cpp_flutter: ^0.0.13
+  llamadart: ^0.8.20
+  llamadart_llama_cpp_flutter: ^0.0.14
 ```
 
 This package has no runtime Dart API of its own. Import `package:llamadart`

packages/llamadart_llama_cpp_flutter/pubspec.yaml

@@ -1,6 +1,6 @@
 name: llamadart_llama_cpp_flutter
 description: Flutter Apple SwiftPM runtime companion package for llamadart llama.cpp and GGUF support.
-version: 0.0.13
+version: 0.0.14
 repository: https://github.com/leehack/llamadart/tree/main/packages/llamadart_llama_cpp_flutter
 homepage: https://github.com/leehack/llamadart
 issue_tracker: https://github.com/leehack/llamadart/issues

pubspec.yaml

@@ -1,6 +1,6 @@
 name: llamadart
 description: Dart and Flutter local LLM inference with llama.cpp GGUF and LiteRT-LM across native platforms and web.
-version: 0.8.19
+version: 0.8.20
 homepage: https://github.com/leehack/llamadart
 repository: https://github.com/leehack/llamadart
 issue_tracker: https://github.com/leehack/llamadart/issues
@@ -29,7 +29,7 @@ dependencies:
   http: ^1.1.0
   crypto: ^3.0.0
   web: ^1.0.0
-  code_assets: ^1.0.0
+  code_assets: '>=1.0.0 <3.0.0'
   hooks: '>=1.0.0 <3.0.0'
   logging: ^1.3.0
   dinja: ^1.0.0

website/docs/changelog/recent-releases.md

@@ -7,65 +7,59 @@ For canonical full release notes, use:
 
 - [`CHANGELOG.md`](https://github.com/leehack/llamadart/blob/main/CHANGELOG.md)
 
-## Unreleased
+## 0.8.20
 
 - Updated WebGPU bridge assets to `v0.1.37` (llama.cpp `b10514`), restoring
   native/Web parity and provisioning an explicit 1 MiB Wasm stack for wasm32
-  and memory64 so `b10514` graph-parameter growth does not abort Qwen3-ASR
-  memory64 context construction.
+  and memory64 so Qwen3-ASR memory64 context construction does not overflow the
+  default stack.
 
 - Improved Web microphone transcription with browser-capture warmup trimming
   and early short, silent, and unsupported PCM WAV diagnostics.
 
 - Added logical and micro-batch controls for llama.cpp/WebGPU models to the
-  Flutter chat example, and made Android Auto probe the packaged Vulkan device
-  before choosing GPU offload.
+  Flutter chat example.
 
-- Updated the native LiteRT-LM runtime to `v0.16.0-native.2` and added an
-  experimental dedicated `SpeechToTextEngine.liteRtLm` path with bounded mono
-  16 kHz float PCM input, partial/final transcript events, worker-isolated CPU
-  inference, backpressure, and cancellation. The Apple companion also packages
-  the iOS Gemma constraint provider and Metal plugins required by the published
-  runtime.
+- Made Android Auto probe the packaged Vulkan device before choosing GPU
+  offload, avoiding unnecessary CPU fallback on capable models.
+
+- Updated the native LiteRT-LM runtime to `v0.16.0-native.2`; the Apple companion
+  packages the iOS Gemma constraint provider and Metal plugins required by the
+  published runtime.
+
+- Added an experimental `SpeechToTextEngine.liteRtLm` path with bounded mono
+  16 kHz float PCM, partial/final transcript events, worker-isolated CPU
+  inference, backpressure, and cancellation.
 
 - Added experimental live English dictation to native Flutter chat models,
   including generic audio-chat models, using selectable checksum-pinned
   Moonshine Tiny (recommended, 54 MB) and Parakeet TDT 0.6B (optional, 615 MB)
-  LiteRT sidecars. The UI reports determinate download progress and supports
-  cancellation/retry; the public worker-isolated streaming STT API produces
-  confirmed/pending text and editable composer finalization. A persisted
-  settings switch explains and enables or disables the optional workflow.
-  Audio-chat models retain
-  **Ask with voice** as a separate action.
+  LiteRT sidecars. Live dictation is CPU-only, English-only, capped at five
+  minutes, and unavailable on Linux and Web. Audio-chat models retain **Ask
+  with voice** as a separate action.
 
 - Improved Flutter chat example model downloads with bounded retries for
   transient network failures, safe resume after truncated responses, and a
-  distinct integrity-verification state after transfer reaches 100%.
+  distinct integrity-verification state after transfer reaches 100%. The
+  redesigned onboarding and Lab surfaces preserve model-card position while
+  downloads reorder and keep streaming responses from pulling users away from
+  chat history.
 
 - Added experimental typed Qwen3-TTS synthesis on native llama.cpp and WebGPU
-  bridge assets `v0.1.33+`, including
-  capability discovery, cancellation, complete PCM/WAV output, and a dedicated
-  synthesize workflow with file or microphone speaker references, automatic
-  playback, replay, and WAV export in the Flutter chat example. Web accepts
-  byte-backed speaker references and complete PCM/WAV output, with a bounded
-  example-app utterance length for browser stability. The example pins bridge
-  assets `v0.1.37`, retaining the `v0.1.34` recovery that retries a failed
-  worker WebGPU synthesis once on CPU; Apple companion builds discover the
-  native TTS wrapper ABI from the embedded llama framework, while current
-  LiteRT-LM artifacts remain unsupported.
+  with capability discovery, cancellation, speaker references, complete
+  PCM/WAV output, and synthesis/playback/export controls in the Flutter chat
+  example. Apple apps discover the TTS ABI in the embedded llama framework;
+  current LiteRT-LM artifacts remain unsupported.
 
 - Updated the native llama.cpp runtime to `b10514`, adding BailingMoE3,
   GraniteSWA/GraniteMoeSWA, speculators-format DSpark checkpoints, and current
   upstream multimodal/backend fixes and performance improvements. Matching Dart
   FFI bindings and the Apple SwiftPM artifact checksum were refreshed.
 
 - Added an experimental typed Qwen3-ASR whole-file transcription workflow, a
-  SHA-256-verified Qwen3-ASR 0.6B native-and-Web chat-app preset, and foreground
-  microphone recording. WebGPU requires validated bridge assets `v0.1.30+`,
-  pins `v0.1.37` with the `v0.1.32` short-speech recovery and the `b10514`
-  memory64 stack fix,
-  and accepts WAV bytes only; native LiteRT-LM live dictation remains a
-  separate implementation.
+  checksum-pinned Qwen3-ASR 0.6B native-and-Web chat-app preset, and file and
+  microphone transcription. Web accepts WAV bytes only; native LiteRT-LM live
+  dictation remains a separate implementation.
 
 - Added **Ask with voice** to the native Flutter chat example for Gemma 4 E2B
   LiteRT-LM and audio-capable GGUF models. It sends a short microphone recording

website/docs/getting-started/installation.md

@@ -27,17 +27,17 @@ In Xcode, set `IPHONEOS_DEPLOYMENT_TARGET = 16.4` or
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
+  llamadart: ^0.8.20
 ```
 
 For Flutter iOS/macOS apps that should link Apple XCFrameworks through Swift
 Package Manager, also add the runtime companion packages you need:
 
 ```yaml
 dependencies:
-  llamadart: ^0.8.19
-  llamadart_llama_cpp_flutter: ^0.0.13 # GGUF / llama.cpp
-  llamadart_litert_lm_flutter: ^0.0.9 # Apple .litertlm / LiteRT-LM targets
+  llamadart: ^0.8.20
+  llamadart_llama_cpp_flutter: ^0.0.14 # GGUF / llama.cpp
+  llamadart_litert_lm_flutter: ^0.0.10 # Apple .litertlm / LiteRT-LM targets
 ```
 
 The companion packages are published independently from the `packages/`