@tduncan / Cortana
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
md 3.34 KB
View source Raw

Issue #78 and imported-model/TTS review fixes

Validated on 2026-09-08 against LocalMind 1.7.9 (84).

Changes

  • Preserve the existing Ollama fix that removes trailing empty assistant placeholders and sends native think controls.
  • Keep native message.thinking and message.content authoritative and separate, including when the thinking field is empty.
  • Handle legacy leading <think>...</think> content across stream boundaries. Tags later in an ordinary answer are preserved. Unclosed reasoning remains reasoning rather than being guessed into a final answer.
  • Preserve Ollama's final frame without a trailing newline, including UTF-8 text split across network chunks.
  • Add saved Model default / On / Off thinking controls to imported GGUF model cards in Local Models. Apply changes at the next generation, including when the model is already loaded. Explicit Off reaches llama.cpp through llamadart's enableThinking option. The model's template must support the option.
  • Preserve the system TTS engine's voice/language instead of forcing en-US. Follow changes to Android's default TTS engine, await engine initialization, surface failed playback, and avoid starting speech after cancellation during setup.
  • Declare Android TTS service discovery. Translate the new model controls into every supported locale.

Reasoning auto-collapse continues to respect the user's existing setting and manual expand/collapse choices.

Automated verification

  • fvm flutter test: 382 passed, one environment-dependent test skipped.
  • fvm flutter --suppress-analytics analyze: no issues.
  • fvm flutter --suppress-analytics build apk --debug --target-platform android-arm64: succeeded.
  • git diff --check: clean.

Regression coverage includes native/legacy Ollama channel separation, split tags, UTF-8 boundaries, final frames without newlines, preference persistence and UI editing, actual forwarding to the inference API, default-engine changes, and TTS initialization failures.

Android runtime verification

Installed the final debug APK on the Pixel 9 Pro Android 16 emulator and verified successful startup. The merged manifest contains android.intent.action.TTS_SERVICE.

Installed the Sherpa 1.13.7 arm64 English Piper lessac-low engine linked from the official Sherpa APK catalog. Selected it as Android's system TTS engine, then used LocalMind's existing Text To Speech → Speak sample.

Native logs confirmed:

  • Binding to com.k2fsa.sherpa.onnx.tts.engine.
  • Sherpa received and synthesized the sample text.
  • LocalMind's utterance started at 21:49:23 and completed at 21:49:31.
  • Restoring the original system engine without restarting LocalMind bound to com.google.android.tts; the next sample started at 21:50:33 and completed at 21:50:41.

The emulator was launched without host audio; this verifies synthesis and playback completion callbacks, not subjective audio quality. The temporary system-engine override was removed after testing.

Remaining environment limits

The reporter's remote Ollama server and qwen3.5:9b model were not available for a live reproduction. Ollama behavior is verified with captured-request and streaming regression tests. Android 12 hardware was not available; the device smoke test used Android 16. This work does not publish a store release or modify the server's model template.