syni 0.3.8

Flutter SDK for Syni — adaptive on-device LLM inference with hybrid local/cloud chat, structured persona conditioning, and a streaming chat API designed for the UI thread.

0.3.8

Changed

• iOS: the podspec links SyniRuntime.xcframework from synheart/vendor/syni-runtime/ios/, where syni-runtime 0.6.0 and later install it, beside android/jniLibs/. The old location at the package root (synheart/vendor/syni-runtime/SyniRuntime.xcframework, syni-runtime before 0.6.0) still works in this release: pod install uses it and prints a notice to re-run synheart install runtime syni. The next release reads only ios/. When both exist, ios/ is used.

Fixed

• Android: the plugin now reads libsyni_ffi.so from synheart/vendor/syni-runtime/android/jniLibs, where synheart install syni puts it and where the iOS podspec already looks. It used to read synheart/vendor/syni/android/jniLibs, so apps set up with the CLI failed to load the runtime unless they added the directory to their own sourceSets. An app module that adds this directory itself can drop that line; keeping it is harmless. An app module that adds a different directory (such as the old synheart/vendor/syni/android/jniLibs) takes precedence over the plugin, so point it at syni-runtime or remove it. The syni.runtime.android.jniLibs property still overrides the path. When the directory is missing, the build logs a warning naming it.

0.3.7

Added

• SyniAgent.turnStream(), the streaming form of turn(): emits SyniTurnDelta text as the model writes, then exactly one SyniTurnFinal with the same SyniTurnResult that turn() returns. On slow devices the first words arrive as soon as the prompt has been read.
• Cancelling the subscription of an on-device chatStream() or turnStream() stops the generation, so a turn the host gave up on no longer holds the model. The runtime reports tokens only once it is writing, so a turn cancelled while still reading its prompt stops at its first token.

Changed

• Bump crypto from 3.0.7 to 3.0.8.

0.3.6

Added

• SyniAgent.turn() for structured on-device conversations.
• Typed turn contracts: SyniTurnRequest, SyniTurnMessage, SyniTaskContract, SyniGenerationProfile, and SyniTurnResult.
• Runtime API 2.0 capability negotiation.
• Request IDs, Runtime provenance, and structured generation controls.

Changed

• Preserve user and assistant roles instead of flattening conversations into one prompt.
• Expose request identity through on-device telemetry.
• Keep existing chat APIs backward compatible.

0.3.5

Added

• Fallback metadata on SyniChatResponse / SyniRuntimeResponse: isFallback, fallbackReason, underlyingErrorCode, retryable.
• SyniRuntimeError now carries stable code + retryable (exported from package:syni/agent.dart).
• SyniAgent.telemetry() — drain on-device inference metrics / fallback diagnostics (SyniInferenceMetric, SyniFallbackDiagnostics, SyniAttemptDiagnostic).

0.3.4

Added

• SyniModelSpec for Qwen2.5-3B Instruct (Q4_K_M).

Fixed

• Close the cloud HTTP client when the agent is disposed (resource leak in the hybrid local/cloud chat path); value-equality on SyniCloudConfig.

0.3.3

Changed

• OSS launch cleanup: scrubbed internal-implementation references (e.g. "Rust core" → "native core"), aligned all install commands to synheart install runtime syni, corrected podspec author email, added LICENSE/CONTRIBUTING/SECURITY parity files, added .pubignore. No functional change to the public API.
• iOS podspec version synced to 0.3.3 (was drifted at 1.0.0).

0.3.2

Added

• SyniAgent.bindPersona(SyniPersona) — bind a persona without installing a local model. Cloud chat only needs a persona plus a configured SyniCloudConfig, so hosts that pick SyniExecutionMode.cloudOnly no longer have to download a model just to attach a persona.

Fixed

• chat / chatStream no longer throw "Syni is not installed" in cloudOnly mode. The chat gate now requires the install state only when the request would actually run on device (local / localFirst); cloudOnly proceeds as long as a persona is bound. localFirst still wants a model on disk — its fallback-to-cloud path is unchanged.

0.3.1

Added

• New bundled spec persona: life.companion.v1 — a reflective behavioral-insights companion. Non-prescriptive by default, hedged probabilistic language, no emotional or medical labels, frames live signals as the user's CURRENT state. Sister to the existing focus.coach.v1 (which prescribes); useful for conversational / daily-companion apps where reflection is the goal.

Changed

• focus.coach.v1 tightened against framing-language drift: live signal values are now explicitly framed as the user's CURRENT state. Session language (recent session, last session, session today) is only used when the provided context includes a non-empty history.recent_sessions array. If the user asks about a session but that array is missing or empty, the model says so plainly instead of dressing live state in session language.

Both changes ship via tool/sync_personas.sh from syni-spec. No Dart API changes — consumers pick the new persona up via SyniSpecPersona.load('life.companion.v1'), and Focus Coach consumers benefit from the tightening automatically.

0.3.0

Changed (breaking)

• SyniCloudConfig.authToken (a static Future<String?> Function() bearer provider) is replaced by authHeaders, a request-aware Future<Map<String, String>> Function(String method, String url). Cloud auth headers are now resolved per request — required for X-Synheart-Proof device attestation, which signs the request method and URL.
• Host SDKs that previously passed a static token provider must update to the new shape; the SDK never holds or stores credentials itself.

Fixed

• SyniChatResponse.fromCloudReply now extracts the JSON payload from replies wrapped in Markdown code fences (e.g. ```{...}```) — some cloud models emit structured-output JSON inside fences. Plain-text replies and unfenced JSON continue to parse as before.

0.2.0

Changed (breaking)

• iOS podspec switched from static .a + force_load to a vendored dynamic framework via prepare_command symlink. Consumer apps install SyniRuntime.xcframework via synheart install runtime syni; the podspec walks up from POD_DIR (or honors SYNHEART_APP_ROOT) to find it.
• library_loader.dart uses DynamicLibrary.process() on iOS so symbols resolve from the auto-loaded embedded framework (matches the pattern in core-flutter v0.5.0).

Other

• .gitignore excludes the symlink the podspec creates at pod-install time so it doesn't show up as untracked.

0.1.0

First public release.

Agent layer (package:syni/agent.dart)

• SyniAgent — install lifecycle, model catalog, persona binding, chat / chatStream orchestration, hybrid local/cloud routing via SyniExecutionMode.
• SyniInstaller — model download, tokenizer fetch, SHA-256 verification, cold-start restore from disk.
• SyniCloudClient — HTTP + SSE client for the Syni cloud chat endpoint, sticky session id, HSI forwarded as request context.
• SyniSpecPersona.load(id) — resolves persona JSON from the bundled spec assets so the same persona id consistently produces the same behavior on both client and server.

Runtime layer (package:syni/runtime.dart)

• SyniRuntime — worker-isolate wrapper over the on-device inference engine. Engine load + token generation run on a worker so the UI thread is free.
• SyniRuntimeRequest / SyniRuntimeResponse for direct inference.

Models

• SyniModels catalog with two pre-pinned entries:
  • qwen25_15bInstructQ4 (Qwen 2.5 1.5B Instruct Q4_K_M, ~1.1 GB).
  • gemma3_1bInstructQ4 (Gemma 3 1B Instruct Q4_K_M, ~770 MB, Synheart-hosted mirror).
• Both ship with pinned SHA-256 verified by SyniInstaller at install time.

Personas (bundled assets)

Five spec personas mirrored from syni-spec and bundled under assets/personas/prod/:

• focus.coach.v1
• stress.coach.v1
• cognitive.companion.v1
• performance.coach.v1
• wellness.guide.v1

Refresh via tool/sync_personas.sh.

Notes

• The legacy Syni.initialize / Syni.generate platform-channel API remains exported for backwards compatibility but new code should use the agent layer.