xybrid-ai/xybrid

Use Case The whisper.cpp fix enables suppress_nst so bracketed annotations such as [BLANK_AUDIO] do not appear in ordinary transcripts. whisper.cpp defaults this flag to false, and some callers may want its non-speech tokens instead of cleaned text. The setting is fixed inside the backend, so callers cannot choose. Proposed Solution Add an explicit ASR option, tentatively suppress_non_speech_tokens, to RunOptions or its ASR-specific configuration. Pass it through the SDK, core runtime, and foreign bindings to TranscribeParams and whisper.cpp. This is a decoding capability, so it should not add another run* method. Document one default across backends. We need to choose between: suppression on, preserving clean transcripts and the behavior introduced by the current fix; suppression off, matching whisper.cpp's upstream default. Add tests for both values. The suppressed case should keep bracketed annotations out of the transcript. The raw case should use a fixture that emits an annotati

Current behavior xybrid-ffi-facade and xybrid-bolt expose an opaque conversation context plus context-aware run and streaming calls. The generated Apple and Kotlin bindings can create a context, add turns, set a system prompt, prune history, and pass that context into inference. The React Native spec has no conversation handle or context-aware inference path. A JS caller cannot use the SDK's multi-turn conversation state. Scope Add an opaque native context handle on iOS and Android, with explicit release and cleanup during module invalidation. Add a TypeScript ConversationContext wrapper for the bolt surface: create, create with an ID, push a role-aware envelope, set the system envelope, clear history, read the ID and history length, inspect whether a system prompt is set, and set the maximum history length. Let both Model.run() and Model.runStreaming() receive a context without adding capability suffix chains to the public JS API. Map the JS context handle to XybridConversationContex

Context The Unity starter only ships XybridDemoController.cs (TTS). Community requests for ASR examples have come in. Adding a second controller gives Unity developers a clear template that mirrors what we already provide on iOS. Scope Add examples/unity/starter/Assets/Scripts/XybridAsrController.cs. Use Microphone.Start to record from the default input device. Load whisper-tiny (already in the registry). Display the transcript in an on-scene Text element. Out of scope Changes to the Unity binding itself (bindings/unity/). New scene assets beyond what's needed to wire the script. Acceptance criteria Recording produces a transcript in the Unity scene. Starter README mentions both demos. No changes outside examples/unity/starter/. References examples/unity/starter/Assets/Scripts/XybridDemoController.cs (TTS reference) examples/ios/XybridExample/ContentView.swift (iOS ASR pattern)

Context crates/xybrid-core/examples/resnet50_imagenet.rs:88 prints "Resize to 256x256 (TODO - currently passthrough)". The preprocessing pipeline doesn't actually resize, so the example produces wrong predictions unless inputs are already the right size. Visible UX bug that misleads anyone evaluating the SDK with the example. Scope Pick one of two paths: Add a Resize preprocessing step to the execution template system (preferred — benefits all vision models). Look at crates/xybrid-core/src/execution/preprocessing/image.rs for the pattern. Inline image::imageops::resize in the example before envelope creation. Either way, remove the "TODO" line in the print statement. Out of scope Adding general image preprocessing beyond what the example needs (other models can add their own steps). Replacing the image crate dependency. Acceptance criteria Example correctly classifies a sample image (top-1 in the expected ImageNet bucket). If Resize step added: documented in the metadata schem

Context PostprocessingStep::TemperatureSample is stubbed at crates/xybrid-core/src/execution/postprocessing/mod.rs:60. This is needed for non-llama.cpp logit-output models (small classification heads, custom LMs run via ONNX) where we want to sample rather than take argmax. Medium difficulty — wants someone comfortable with sampling primitives but doesn't need Xybrid internals. Scope Implement temperature scaling (logits / temperature). Implement top-k filtering. Implement top-p (nucleus) filtering. Final multinomial sample via a seedable RNG so tests are deterministic. Accept all three parameters via the step's config. Out of scope Beam search, repetition penalties, frequency penalties, or other advanced sampling tricks. Modifying llama.cpp's existing sampler path. Acceptance criteria Unit tests verify: temperature=0 ≈ argmax, top-k=1 ≈ argmax, top-p=1.0 + high temp ≈ uniform over allowed tokens. Determinism with a fixed seed. Clear error on invalid params (negative te