Skip to main content

Current status

SkellySpeak is a working proof of concept. The source of truth for application version is native/Cargo.toml; this page deliberately does not duplicate the number.

Implemented​

  • Guided, streamed conversation and the meaning-domain Skill tree.
  • English, French, Spanish, Arabic, and Mandarin as symmetric target/native languages, including RTL, segmentation, dialect, and romanization metadata from native/src/languages.rs.
  • Per-turn token glosses, translation, grammar mechanics, reply scaffolds, and private coach feedback.
  • A persistent, interactive coach thread that the conversation partner never sees.
  • Teaching-plan/profile memory updated by a non-overlapping background observer.
  • Multiple resumable conversations per language pairing and custom personas.
  • Hosted Google sign-in, bring-your-own OpenRouter/Groq keys, and custom OpenAI-compatible chat servers.
  • Native credential storage, atomic JSON persistence, strict request routing, bounded retries, and surfaced errors.
  • Microphone input, cloud or OS speech playback, playback-rate control, caching, and cancellation.
  • Bounded local traces across restarts, per-attempt requests and effective parameters, reply/suggestion prompt blocks, request comparison and scoped export.
  • A generated execution graph, reconciliation, and pause/resume/step controls.
  • Topic explanations/examples and an edit-time reference to original coach feedback.
  • Desktop update checks and CI workflows for desktop, Android, iOS, server, and documentation builds.

Known limitations​

  • Tokenization, glossing, translation, and grammar analysis still use model calls. The planned local dictionary layer is described in Future Work.
  • There is no vocabulary/SRS subsystem or dedicated first-run onboarding flow.
  • Automated frontend tests mock the Tauri IPC boundary. There is no end-to-end suite driving a packaged app and the real Rust core.
  • Native microphone, credential-vault, signing, installation, and update flows require platform/device verification. Desktop unit tests cannot establish those claims.
  • Model traces contain prompt and output text and are intended for local diagnostics; they are not anonymized telemetry.
  • Prompt editing/overrides and complete operation provenance remain proposed work; see Observability.
  • The Skill tree displays a local target-language profile with evidence-derived XP, checks/stars and saved practice focus. Previous rubric evidence remains inspectable without new-skill credit. The evaluator is not calibrated as a proficiency assessment. Browser mode uses labeled fixtures; profile switching remains future work. See the progression design.

Verification​

The maintained checks are defined in .github/workflows/ci.yml:

npm test
npm run build

cd native
cargo clippy --lib -- -D warnings
cargo test --lib

cd ../server
uv run --frozen --group dev pytest -q

cd ../skellyspeak-docs
npm run build

Firestore-emulator, native-device, signing, and live-service checks are separate because they need their corresponding environments.

Planning​

The live roadmap is generated from GitHub on the site's Roadmap page. Dated audit reports are historical evidence, not the current backlog.