@elizaos/plugin-local-inference 2.0.3-beta.2 → 2.0.3-beta.4

package/dist/services/inference-capabilities.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"inference-capabilities.d.ts","sourceRoot":"","sources":["../../src/services/inference-capabilities.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,MAAM,MAAM,YAAY,GAAG,SAAS,GAAG,MAAM,GAAG,SAAS,GAAG,UAAU,CAAC;AAEvE,MAAM,WAAW,qBAAqB;IACrC,YAAY,EAAE,OAAO,CAAC;IACtB,YAAY,EAAE,OAAO,CAAC;IACtB,kBAAkB,EAAE,OAAO,CAAC;IAC5B,eAAe,EAAE,OAAO,CAAC;IACzB,YAAY,EAAE,YAAY,CAAC;IAC3B,8CAA8C;IAC9C,QAAQ,EAAE,SAAS,GAAG,KAAK,GAAG,SAAS,GAAG,SAAS,CAAC;CACpD;AAED,wDAAwD;AACxD,MAAM,WAAW,gBAAgB;IAChC,yEAAyE;IACzE,kBAAkB,IAAI,OAAO,CAAC;IAC9B,yEAAyE;IACzE,kBAAkB,IAAI,OAAO,CAAC;IAC9B,4EAA4E;IAC5E,WAAW,IAAI,OAAO,CAAC;IACvB,mEAAmE;IACnE,cAAc,IAAI,OAAO,CAAC;IAC1B,0FAA0F;IAC1F,YAAY,IAAI,YAAY,CAAC;IAC7B,oBAAoB;IACpB,QAAQ,IAAI,SAAS,GAAG,KAAK,GAAG,SAAS,GAAG,SAAS,CAAC;CACtD;AAED;;;;;;;;;;;;GAYG;AACH,wBAAgB,iBAAiB,CAChC,MAAM,EAAE,gBAAgB,GACtB,qBAAqB,CAqBvB;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,IAAI,qBAAqB,CAS5D;AAMD;;;;;;;GAOG;AACH,MAAM,WAAW,gBAAgB;IAChC,6EAA6E;IAC7E,YAAY,EAAE,YAAY,GAAG,SAAS,CAAC;IACvC,yEAAyE;IACzE,YAAY,EAAE,OAAO,GAAG,IAAI,CAAC;IAC7B,gDAAgD;IAChD,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,gDAAgD;IAChD,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,kDAAkD;IAClD,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,qCAAqC;IACrC,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,sCAAsC;IACtC,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,uBAAuB;IACvC;;;OAGG;IACH,yBAAyB,EAAE,OAAO,CAAC;IACnC;;;OAGG;IACH,UAAU,EAAE,OAAO,CAAC;IACpB,MAAM,EAAE,MAAM,CAAC;CACf;AASD;;;;;;;;;;;;GAYG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE;IAC9C,YAAY,EAAE,YAAY,GAAG,SAAS,CAAC;IACvC,YAAY,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;CAC9B,GAAG,uBAAuB,CAsB1B"}

package/dist/services/inference-telemetry.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"inference-telemetry.d.ts","sourceRoot":"","sources":["../../src/services/inference-telemetry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAQH,UAAU,gBAAgB;IACzB,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;CACpB;AAwCD,MAAM,MAAM,aAAa,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,OAAO,CAAC,CAAC;AAEtE,qBAAa,kBAAkB;IAC9B,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAkC;IACxD,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;gBAEtB,QAAQ,SAAM;IAI1B;;;;;;;OAOG;IACH,MAAM,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,aAAa,GAAG,IAAI;IAoB/D;;;OAGG;IACH,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG,gBAAgB;IAWvC,mEAAmE;IACnE,WAAW,IAAI,MAAM,EAAE;IAIvB,mDAAmD;IACnD,KAAK,IAAI,IAAI;CAGb;AAED,sEAAsE;AACtE,eAAO,MAAM,kBAAkB,oBAA2B,CAAC"}

package/dist/services/ios-llama-streaming.d.ts

@@ -0,0 +1,119 @@
+/**
+ * iOS streaming-LLM adapter.
+ *
+ * Mirror of the AOSP plugin's `aosp-llama-streaming.ts` for the iOS
+ * platform.  Bridges the JS-side streaming surface (the same shape the
+ * `FfiStreamingRunner` expects) over the `LlamaCpp.xcframework` Swift
+ * bridge that ships with the iOS shell.
+ *
+ * Status (2026-05-12): the XCFramework currently ships without streaming-LLM
+ * support. The Swift implementation against
+ * `libelizainference.dylib` (built by `build-llama-cpp-mtp.mjs` with
+ * the `darwin-arm64-metal-fused` target) is the gating item; until then
+ * `loadIosStreamingLlmBinding` returns `null` and the runtime falls
+ * back to the cloud route.
+ *
+ * Why XCFramework + Swift bridge rather than `bun:ffi` directly:
+ *   - bun:ffi runs only inside the Bun runtime.  On iOS, Eliza ships the
+ *     ElizaBunEngine.xcframework which embeds a Bun runtime — but the
+ *     llama symbols cannot be resolved at `bun:ffi.dlopen` time on a
+ *     codesigned device build (the dynamic linker forbids loading
+ *     arbitrary `.dylib` from the app bundle on a real device).
+ *   - The Swift bridge wraps the symbols at compile time inside the
+ *     XCFramework so the codesignature covers them.  Capacitor's
+ *     `llama-cpp-capacitor` plugin already uses this pattern for the
+ *     non-streaming surface.
+ *
+ * The Swift glue gating this binding needs to:
+ *   1. Re-export the streaming-LLM symbols from `ffi-streaming-llm.h`
+ *      under the Swift module name `LlamaCpp.Streaming.*`.
+ *   2. Expose an Objective-C-bridgeable wrapper class
+ *      (`LlamaStreamingSession`) so Capacitor's plugin can register
+ *      methods on it.
+ *   3. Wire `ProcessInfo.thermalState` into a thermal-throttle hook so
+ *      iOS can bail out of speculative decoding under sustained heat.
+ *
+ * Until that lands, this file:
+ *   - declares the JS contract,
+ *   - probes for `(window as any).LlamaStreaming` (the Capacitor plugin
+ *     entry point Swift will expose),
+ *   - logs + returns null when the bridge isn't there.
+ */
+export type IosLlmStreamHandle = bigint;
+export type IosInferenceContextHandle = bigint;
+export interface IosLlmStreamConfig {
+    maxTokens: number;
+    temperature: number;
+    topP: number;
+    topK: number;
+    repeatPenalty: number;
+    slotId: number;
+    promptCacheKey: string | null;
+    draftMin: number;
+    draftMax: number;
+    mtpDrafterPath: string | null;
+    disableThinking: boolean;
+}
+export interface IosLlmStreamStep {
+    tokens: number[];
+    text: string;
+    done: boolean;
+    drafterDrafted: number;
+    drafterAccepted: number;
+}
+export interface IosStreamingLlmBinding {
+    llmStreamSupported(): boolean;
+    llmStreamOpen(args: {
+        ctx: IosInferenceContextHandle;
+        config: IosLlmStreamConfig;
+    }): IosLlmStreamHandle;
+    llmStreamPrefill(args: {
+        stream: IosLlmStreamHandle;
+        tokens: Int32Array;
+    }): void;
+    llmStreamNext(args: {
+        stream: IosLlmStreamHandle;
+        maxTokensPerStep?: number;
+        maxTextBytes?: number;
+    }): IosLlmStreamStep;
+    llmStreamCancel(stream: IosLlmStreamHandle): void;
+    llmStreamSaveSlot(args: {
+        stream: IosLlmStreamHandle;
+        filename: string;
+    }): void;
+    llmStreamRestoreSlot(args: {
+        stream: IosLlmStreamHandle;
+        filename: string;
+    }): void;
+    llmStreamClose(stream: IosLlmStreamHandle): void;
+}
+/**
+ * Try to load the iOS streaming-LLM binding.  Returns null when:
+ *   - we are not on iOS,
+ *   - the Capacitor `LlamaStreaming` plugin isn't registered,
+ *   - the bridge reports `isAvailable() === false` (e.g. the XCFramework
+ *     was built without streaming-LLM support).
+ *
+ * Returning null is NOT a failure — the runtime then falls back to the
+ * cloud route.  Throws only when the bridge is present but
+ * mis-configured (probe-and-catch is reserved for the boot path).
+ */
+export declare function loadIosStreamingLlmBinding(): Promise<IosStreamingLlmBinding | null>;
+export interface IosInferenceCapabilities {
+    /** True only when the iOS Swift bridge is present AND reports streaming-LLM. */
+    streamingLlm: boolean;
+    /** Always false on iOS until the drafter weights ship in the bundle. */
+    mtpSupported: boolean;
+    /** Whether the XCFramework reports omnivoice streaming. */
+    omnivoiceStreaming: boolean;
+    /** Phone-tier iOS devices rarely have headroom for mmproj. */
+    mmprojSupported: boolean;
+    /**
+     * `ProcessInfo.thermalState` snapshot at probe time.  Surfaced so the
+     * runtime can refuse to start speculative decoding when the device is
+     * already in `serious` / `critical`.  Always `nominal` on the
+     * web/sim fallback path.
+     */
+    thermalState: "nominal" | "fair" | "serious" | "critical";
+}
+//# sourceMappingURL=ios-llama-streaming.d.ts.map

package/dist/services/ios-llama-streaming.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ios-llama-streaming.d.ts","sourceRoot":"","sources":["../../src/services/ios-llama-streaming.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwCG;AASH,MAAM,MAAM,kBAAkB,GAAG,MAAM,CAAC;AACxC,MAAM,MAAM,yBAAyB,GAAG,MAAM,CAAC;AAE/C,MAAM,WAAW,kBAAkB;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,aAAa,EAAE,MAAM,CAAC;IACtB,MAAM,EAAE,MAAM,CAAC;IACf,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,MAAM,CAAC;IACjB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,eAAe,EAAE,OAAO,CAAC;CACzB;AAED,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,OAAO,CAAC;IACd,cAAc,EAAE,MAAM,CAAC;IACvB,eAAe,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,sBAAsB;IACtC,kBAAkB,IAAI,OAAO,CAAC;IAC9B,aAAa,CAAC,IAAI,EAAE;QACnB,GAAG,EAAE,yBAAyB,CAAC;QAC/B,MAAM,EAAE,kBAAkB,CAAC;KAC3B,GAAG,kBAAkB,CAAC;IACvB,gBAAgB,CAAC,IAAI,EAAE;QACtB,MAAM,EAAE,kBAAkB,CAAC;QAC3B,MAAM,EAAE,UAAU,CAAC;KACnB,GAAG,IAAI,CAAC;IACT,aAAa,CAAC,IAAI,EAAE;QACnB,MAAM,EAAE,kBAAkB,CAAC;QAC3B,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,YAAY,CAAC,EAAE,MAAM,CAAC;KACtB,GAAG,gBAAgB,CAAC;IACrB,eAAe,CAAC,MAAM,EAAE,kBAAkB,GAAG,IAAI,CAAC;IAClD,iBAAiB,CAAC,IAAI,EAAE;QACvB,MAAM,EAAE,kBAAkB,CAAC;QAC3B,QAAQ,EAAE,MAAM,CAAC;KACjB,GAAG,IAAI,CAAC;IACT,oBAAoB,CAAC,IAAI,EAAE;QAC1B,MAAM,EAAE,kBAAkB,CAAC;QAC3B,QAAQ,EAAE,MAAM,CAAC;KACjB,GAAG,IAAI,CAAC;IACT,cAAc,CAAC,MAAM,EAAE,kBAAkB,GAAG,IAAI,CAAC;CACjD;AA8DD;;;;;;;;;;GAUG;AACH,wBAAsB,0BAA0B,IAAI,OAAO,CAAC,sBAAsB,GAAG,IAAI,CAAC,CAuBzF;AAoCD,MAAM,WAAW,wBAAwB;IACxC,gFAAgF;IAChF,YAAY,EAAE,OAAO,CAAC;IACtB,wEAAwE;IACxE,YAAY,EAAE,OAAO,CAAC;IACtB,2DAA2D;IAC3D,kBAAkB,EAAE,OAAO,CAAC;IAC5B,8DAA8D;IAC9D,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;OAKG;IACH,YAAY,EAAE,SAAS,GAAG,MAAM,GAAG,SAAS,GAAG,UAAU,CAAC;CAC1D"}

package/dist/services/kv-spill.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"kv-spill.d.ts","sourceRoot":"","sources":["../../src/services/kv-spill.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAEzC,4EAA4E;AAC5E,eAAO,MAAM,oBAAoB,QAAQ,CAAC;AAE1C;;;;;GAKG;AACH,eAAO,MAAM,cAAc,MAAM,CAAC;AAElC;;;;;;;;GAQG;AACH,eAAO,MAAM,gCAAgC,MAAM,CAAC;AACpD,eAAO,MAAM,+BAA+B,OAAO,CAAC;AAEpD;;;;;;;;;;;;;;;;;GAiBG;AACH,QAAA,MAAM,iCAAiC;;;;;CAK7B,CAAC;AAEX,MAAM,MAAM,cAAc,GAAG,MAAM,OAAO,iCAAiC,CAAC;AAE5E;;;;;;;GAOG;AACH,MAAM,WAAW,UAAU;IAC1B,sEAAsE;IACtE,aAAa,EAAE,MAAM,CAAC;IACtB,4EAA4E;IAC5E,YAAY,EAAE,OAAO,CAAC;CACtB;AAoBD,wBAAgB,gCAAgC,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAOvE;AAED;;;;GAIG;AACH,MAAM,MAAM,WAAW,GAAG,KAAK,GAAG,MAAM,CAAC;AAEzC,MAAM,WAAW,mBAAmB;IACnC,IAAI,EAAE,UAAU,CAAC;IACjB,sEAAsE;IACtE,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,EAAE,MAAM,CAAC;CACtB;AAED,MAAM,WAAW,gBAAgB;IAChC,IAAI,EAAE,OAAO,CAAC;IACd,IAAI,EAAE,WAAW,CAAC;IAClB,yDAAyD;IACzD,aAAa,EAAE,MAAM,CAAC;IACtB,iCAAiC;IACjC,UAAU,EAAE,MAAM,CAAC;IACnB,iCAAiC;IACjC,aAAa,EAAE,MAAM,CAAC;IACtB,qCAAqC;IACrC,UAAU,EAAE,MAAM,CAAC;IACnB,qDAAqD;IACrD,YAAY,EAAE,MAAM,CAAC;IACrB,0DAA0D;IAC1D,kBAAkB,EAAE,MAAM,CAAC;IAC3B,+DAA+D;IAC/D,eAAe,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,MAAM,WAAW,GAAG,mBAAmB,GAAG,gBAAgB,CAAC;AAEjE;;;;GAIG;AACH,qBAAa,uBAAwB,SAAQ,KAAK;IACjD,QAAQ,CAAC,IAAI,0BAA0B;IACvC,QAAQ,CAAC,OAAO,EAAE;QACjB,gBAAgB,EAAE,MAAM,CAAC;QACzB,YAAY,EAAE,MAAM,CAAC;QACrB,aAAa,EAAE,MAAM,CAAC;QACtB,UAAU,EAAE,MAAM,CAAC;QACnB,kBAAkB,EAAE,MAAM,CAAC;QAC3B,eAAe,EAAE,MAAM,CAAC;QACxB,YAAY,EAAE,cAAc,CAAC;QAC7B,YAAY,EAAE,OAAO,CAAC;KACtB,CAAC;gBAEU,OAAO,EAAE,uBAAuB,CAAC,SAAS,CAAC;CAgBvD;AAED;;;;;GAKG;AACH,MAAM,WAAW,YAAY;IAC5B,gBAAgB,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,UAAU,CAAC;IACrB,qBAAqB,EAAE,MAAM,CAAC;IAC9B,YAAY,EAAE,cAAc,CAAC;IAC7B;;;OAGG;IACH,iBAAiB,EAAE,OAAO,CAAC;CAC3B;AAED;;;;;;GAMG;AACH,eAAO,MAAM,2BAA2B,OAAO,CAAC;AAEhD,wBAAgB,6BAA6B,CAAC,MAAM,EAAE,SAAS,GAAG,MAAM,CAIvE;AAMD;;;;;;;;;;GAUG;AACH,wBAAgB,WAAW,CAAC,KAAK,EAAE,YAAY,GAAG,WAAW,CAmG5D;AAED;;;;;GAKG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE;IAC9C,YAAY,EAAE,OAAO,CAAC;IACtB,cAAc,EAAE,OAAO,CAAC;CACxB,GAAG,cAAc,CAIjB"}

package/dist/services/latency-trace.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"latency-trace.d.ts","sourceRoot":"","sources":["../../src/services/latency-trace.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwDG;AAGH,OAAO,KAAK,EAAY,cAAc,EAAE,MAAM,eAAe,CAAC;AAM9D;;;;;GAKG;AACH,eAAO,MAAM,iBAAiB,qSAcpB,CAAC;AAEX,MAAM,MAAM,eAAe,GAAG,CAAC,OAAO,iBAAiB,CAAC,CAAC,MAAM,CAAC,CAAC;AAsBjE;8EAC8E;AAC9E,eAAO,MAAM,sBAAsB,0SAElC,CAAC;AAMF;;;;GAIG;AACH,MAAM,WAAW,cAAc;IAC9B,2DAA2D;IAC3D,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,iEAAiE;IACjE,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,6EAA6E;IAC7E,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,+DAA+D;IAC/D,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,yEAAyE;IACzE,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,oEAAoE;IACpE,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,6EAA6E;IAC7E,uBAAuB,EAAE,MAAM,GAAG,IAAI,CAAC;IACvC,2EAA2E;IAC3E,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,yEAAyE;IACzE,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,yEAAyE;IACzE,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,sEAAsE;IACtE,kBAAkB,EAAE,MAAM,GAAG,IAAI,CAAC;IAElC;;;;OAIG;IACH,sBAAsB,EAAE,MAAM,GAAG,IAAI,CAAC;IACtC,qDAAqD;IACrD,oCAAoC,EAAE,MAAM,GAAG,IAAI,CAAC;IACpD,kDAAkD;IAClD,6BAA6B,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;;;;OAIG;IACH,wCAAwC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxD;kEAC8D;IAC9D,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;CACpC;AAED,iDAAiD;AACjD,eAAO,MAAM,oBAAoB,ieAoBuB,CAAC;AAEzD,MAAM,MAAM,iBAAiB,GAAG,CAAC,OAAO,oBAAoB,CAAC,CAAC,MAAM,CAAC,CAAC;AAoCtE,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,eAAe,CAAC;IACtB,2EAA2E;IAC3E,GAAG,EAAE,MAAM,CAAC;IACZ,0DAA0D;IAC1D,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,YAAY;IAC5B,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,6EAA6E;IAC7E,SAAS,EAAE,MAAM,CAAC;IAClB,oEAAoE;IACpE,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,WAAW,EAAE,iBAAiB,EAAE,CAAC;IACjC,OAAO,EAAE,cAAc,CAAC;IACxB,mEAAmE;IACnE,OAAO,EAAE,eAAe,EAAE,CAAC;IAC3B,sEAAsE;IACtE,QAAQ,EAAE,OAAO,CAAC;IAClB;;;;OAIG;IACH,SAAS,EAAE,MAAM,EAAE,CAAC;CACpB;AAMD,MAAM,WAAW,gBAAgB;IAChC,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACnB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;CACpB;AAED;;;;;;;;GAQG;AACH,qBAAa,gBAAgB;IAEhB,OAAO,CAAC,QAAQ,CAAC,QAAQ;IADrC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAgB;gBACX,QAAQ,EAAE,MAAM;IAE7C,GAAG,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI;IAMxB,OAAO,IAAI,gBAAgB;CA+B3B;AAMD,MAAM,WAAW,aAAa;IAC7B,wEAAwE;IACxE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,6DAA6D;IAC7D,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B;;;;;OAKG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;CACtB;AAiBD,qBAAa,qBAAqB;IACjC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAsB;IAC3C,OAAO,CAAC,QAAQ,CAAC,IAAI,CAA+B;IACpD,OAAO,CAAC,QAAQ,CAAC,MAAM,CAA6B;IACpD,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAkD;IAC7E,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAS;IACtC,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAS;IAC3C,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAS;gBAE1B,IAAI,GAAE,aAAkB;IASpC;;;OAGG;IACH,SAAS,CAAC,IAAI,GAAE;QAAE,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;KAAO,GAAG,MAAM;IA0BzE,wEAAwE;IACxE,WAAW,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM;IAMnC;;;;OAIG;IACH,IAAI,CAAC,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI;IAmC3E,0EAA0E;IAC1E,UAAU,CACT,MAAM,EAAE,MAAM,EACd,UAAU,EAAE,eAAe,EAC3B,SAAS,CAAC,EAAE,MAAM,GAChB,IAAI;IAIP;;;;;OAKG;IACH,OAAO,CAAC,MAAM,EAAE,MAAM,GAAG,YAAY,GAAG,IAAI;IAiB5C,gEAAgE;IAChE,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,YAAY,GAAG,IAAI;IAM7C,yDAAyD;IACzD,YAAY,CAAC,CAAC,SAAoB,GAAG,YAAY,EAAE;IAKnD,uEAAuE;IACvE,kBAAkB,IAAI,MAAM,CAAC,iBAAiB,EAAE,gBAAgB,CAAC;IAgBjE,4DAA4D;IAC5D,KAAK,IAAI,IAAI;IASb,kDAAkD;IAClD,IAAI,aAAa,IAAI,MAAM,CAE1B;IAED;;;;;;;;;OASG;IACH,eAAe,CACd,MAAM,EAAE,cAAc,EACtB,IAAI,GAAE;QACL,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACvB,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,CAAC;KACjC,GACJ,MAAM,IAAI;IAgBb,OAAO,CAAC,YAAY;IA8BpB,OAAO,CAAC,cAAc;CAiBtB;AAMD;;;;;GAKG;AACH,eAAO,MAAM,kBAAkB,uBAA8B,CAAC;AAE9D;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAC/B,MAAM,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,EACjC,UAAU,EAAE,eAAe,EAC3B,SAAS,CAAC,EAAE,MAAM,GAChB,IAAI,CAgBN;AAED,8EAA8E;AAC9E,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,MAAM,GAAG,YAAY,GAAG,IAAI,CAYvE;AAMD,MAAM,WAAW,sBAAsB;IACtC,kBAAkB,EAAE,MAAM,CAAC;IAC3B,8EAA8E;IAC9E,WAAW,EAAE,aAAa,CAAC,eAAe,CAAC,CAAC;IAC5C,WAAW,EAAE,aAAa,CAAC,iBAAiB,CAAC,CAAC;IAC9C,aAAa,EAAE,MAAM,CAAC;IACtB,MAAM,EAAE,YAAY,EAAE,CAAC;IACvB,UAAU,EAAE,MAAM,CAAC,iBAAiB,EAAE,gBAAgB,CAAC,CAAC;CACxD;AAED,4DAA4D;AAC5D,wBAAgB,2BAA2B,CAC1C,MAAM,GAAE,qBAA0C,EAClD,KAAK,SAAK,GACR,sBAAsB,CASxB;AAMD;;sDAEsD;AACtD,MAAM,WAAW,gBAAgB;IAChC;+EAC2E;IAC3E,aAAa,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B;wDACoD;IACpD,WAAW,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,gCAAgC;IAChC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B;;;wCAGoC;IACpC,+BAA+B,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAChD,sEAAsE;IACtE,eAAe,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC;+CAC2C;IAC3C,WAAW,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CAC5B;AAED,MAAM,WAAW,sBAAsB;IACtC,KAAK,EAAE,MAAM,CAAC;IACd;gEAC4D;IAC5D,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,0EAA0E;IAC1E,sBAAsB,EAAE,gBAAgB,CAAC;IACzC,iEAAiE;IACjE,+BAA+B,EAAE,gBAAgB,CAAC;IAClD,wCAAwC;IACxC,eAAe,EAAE,gBAAgB,CAAC;IAClC;;iDAE6C;IAC7C,GAAG,EAAE;QACJ,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;QACvB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;QACtB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;QACrB,OAAO,EAAE,MAAM,CAAC;QAChB,aAAa,EAAE,OAAO,CAAC;QACvB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;KACxB,CAAC;CACF;AAID;;;;;;GAMG;AACH,qBAAa,eAAe;IAef,OAAO,CAAC,QAAQ,CAAC,IAAI;IAdjC,OAAO,CAAC,KAAK,CAAK;IAClB,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,UAAU,CAAK;IACvB,OAAO,CAAC,QAAQ,CAAC,cAAc,CAE7B;IACF,OAAO,CAAC,QAAQ,CAAC,WAAW,CAE1B;IACF,OAAO,CAAC,QAAQ,CAAC,UAAU,CAEzB;IACF,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAgB;gBAEd,IAAI,GAAE;QAAE,qBAAqB,CAAC,EAAE,MAAM,CAAA;KAAO;IAE1E,UAAU,CAAC,CAAC,EAAE,gBAAgB,GAAG,IAAI;IAsBrC,OAAO,IAAI,sBAAsB;IAsCjC,KAAK,IAAI,IAAI;CASb"}

package/dist/services/lib-target.d.ts

@@ -0,0 +1,55 @@
+/**
+ * Host → fused-lib target resolution for bundle delivery (#9105 / local-
+ * inference).
+ *
+ * An Eliza-1 bundle's `files.lib[]` carries the fused `libelizainference` SET
+ * (the fused lib + its ggml/llama/mtmd sibling backends) per platform `target`.
+ * The downloader fetches ONLY the set matching the host into `<bundleRoot>/lib/`,
+ * which the desktop FFI runtime resolves with no env wiring
+ * (`desktop-fused-ffi-backend-runtime.ts` `resolveFusedLibraryPath`, path #2).
+ *
+ * Pure + injectable (platform / arch / env are all overridable) so the mapping
+ * is unit-testable on any host.
+ */
+import type { Eliza1Files, Eliza1LibFileEntry } from "./manifest";
+export interface HostLibTargetOptions {
+    platform?: NodeJS.Platform;
+    arch?: string;
+    env?: NodeJS.ProcessEnv;
+    /** Prefer a GPU-accelerated target when the bundle hosts one. */
+    preferGpu?: boolean;
+}
+/**
+ * Ordered list of acceptable `target` keys for the host, best-first. The
+ * downloader picks the first target the bundle actually hosts.
+ *
+ *   - `ELIZA_INFERENCE_LIB_TARGET` pins a single target (power users / CI).
+ *   - Mobile (`ELIZA_PLATFORM=android|ios`) returns `[]` — phones ship the lib
+ *     natively (jniLibs / xcframework), never via bundle delivery.
+ *   - CPU is preferred by default: the fused lib always has GGML_CPU built in,
+ *     so the CPU set works on every host and is the smallest download. GPU sets
+ *     (`…-cuda`) are opt-in via `preferGpu` or the env pin.
+ *   - macOS arm64 maps to the `metal` set, which itself carries the CPU
+ *     fallback, so it is the canonical mac target.
+ */
+export declare function resolveHostLibTargets(opts?: HostLibTargetOptions): string[];
+export interface SelectedLibTarget {
+    readonly target: string;
+    readonly files: ReadonlyArray<Eliza1LibFileEntry>;
+}
+/**
+ * Pick the bundle's lib file set for the first host target the manifest
+ * actually carries. Returns null when the bundle has no `lib[]` or hosts no
+ * target the host accepts (⇒ the runtime falls back to other resolution paths,
+ * ultimately cloud).
+ */
+export declare function selectBundleLibFiles(manifest: {
+    files: Pick<Eliza1Files, "lib">;
+}, targets: ReadonlyArray<string>): SelectedLibTarget | null;
+/**
+ * Flat staged filename under `<bundleRoot>/lib/` for a lib entry. Always a
+ * basename (any directory component in `name`/`path` is stripped) so a manifest
+ * can never write outside the lib dir.
+ */
+export declare function libStagedName(entry: Eliza1LibFileEntry): string;
+//# sourceMappingURL=lib-target.d.ts.map

package/dist/services/lib-target.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"lib-target.d.ts","sourceRoot":"","sources":["../../src/services/lib-target.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,kBAAkB,EAAE,MAAM,YAAY,CAAC;AAElE,MAAM,WAAW,oBAAoB;IACpC,QAAQ,CAAC,EAAE,MAAM,CAAC,QAAQ,CAAC;IAC3B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,CAAC,UAAU,CAAC;IACxB,iEAAiE;IACjE,SAAS,CAAC,EAAE,OAAO,CAAC;CACpB;AAOD;;;;;;;;;;;;GAYG;AACH,wBAAgB,qBAAqB,CACpC,IAAI,GAAE,oBAAyB,GAC7B,MAAM,EAAE,CAqBV;AAED,MAAM,WAAW,iBAAiB;IACjC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,aAAa,CAAC,kBAAkB,CAAC,CAAC;CAClD;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,CACnC,QAAQ,EAAE;IAAE,KAAK,EAAE,IAAI,CAAC,WAAW,EAAE,KAAK,CAAC,CAAA;CAAE,EAC7C,OAAO,EAAE,aAAa,CAAC,MAAM,CAAC,GAC5B,iBAAiB,GAAG,IAAI,CAQ1B;AAED;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,kBAAkB,GAAG,MAAM,CAK/D"}

package/dist/services/live-signals.d.ts

@@ -0,0 +1,86 @@
+/**
+ * Live device signals for the capability-driven `auto` routing policy.
+ *
+ * `classifyDeviceTier()` is a *static* assessment — it scores the hardware once
+ * and caches the result. But a device that classifies MAX/GOOD at session start
+ * can still become a poor place to run local inference *right now*: a sustained
+ * workload pushes the SoC into thermal throttling, or decode throughput collapses
+ * under memory pressure. Those are live, time-varying signals the static tier
+ * cannot see.
+ *
+ * This module exposes the two live signals the `auto` branch demotes on:
+ *
+ *   - **thermal state** — read from the device bridge's most recent device
+ *     capabilities (`nominal | fair | serious | critical | unknown`). `serious`
+ *     or `critical` means the OS is actively throttling clocks; forcing local
+ *     inference there makes turns slower than cloud and worsens the throttle.
+ *   - **decode throughput** — the recent p50 decode tokens/sec from
+ *     `inferenceTelemetry`. When it falls below a usable budget the on-device
+ *     path is no longer competitive and `auto` routes to cloud.
+ *
+ * Both signals are *optional*: a quantity the host could not measure is reported
+ * as `null` (never a fabricated `0`), and a `null` signal never demotes — the
+ * static tier decision stands. The default source reads the real subsystems; a
+ * test or alternate host can inject its own via `setLiveDeviceSignalsSource`.
+ */
+/** Thermal pressure reported by the OS. Mirrors the device-bridge wire enum. */
+export type ThermalState = "nominal" | "fair" | "serious" | "critical" | "unknown";
+/**
+ * The telemetry metric name carrying recent on-device decode throughput
+ * (tokens/sec). The backend records this after a generation completes; until
+ * then `summary()` returns a `count: 0` row and the signal is `null`.
+ */
+export declare const DECODE_TPS_METRIC = "inference.decode_tps";
+/**
+ * Minimum decode throughput (tokens/sec) below which the local path is no longer
+ * worth keeping under `auto`. Roughly the floor where on-device generation feels
+ * slower than a cloud round-trip on a constrained mobile SoC.
+ */
+export declare const MIN_DECODE_TPS_BUDGET = 6;
+/**
+ * Approximate decoded-token count for a generated string when an exact tokenizer
+ * count is not reachable from the call site. English averages ~4 characters per
+ * token; this is deliberately coarse — it only feeds a p50 the router compares
+ * against {@link MIN_DECODE_TPS_BUDGET}, never a billing or context-window count.
+ */
+export declare function estimateDecodeTokens(text: string): number;
+/**
+ * Record one on-device decode-throughput observation into the telemetry ring the
+ * `auto` router reads via {@link DECODE_TPS_METRIC}. This is the producer that
+ * activates the dormant TPS demotion signal: backends call it after a real local
+ * generation completes with a known decoded-token count and wall-clock decode
+ * span.
+ *
+ * Drops the sample (records nothing) when it would be noise rather than signal:
+ * an empty/aborted generation, a sub-{@link MIN_TOKENS_FOR_TPS_SAMPLE} reply, or
+ * a non-positive elapsed time (a clock that did not advance). Never throws —
+ * `inferenceTelemetry.record` already guards the call site.
+ */
+export declare function recordDecodeThroughput(sample: {
+    tokens: number;
+    elapsedMs: number;
+}): void;
+/** A point-in-time snapshot of the live device signals. */
+export interface LiveDeviceSignals {
+    /** Current OS thermal pressure, or `null` when no device reported it. */
+    thermalState: ThermalState | null;
+    /** Recent p50 decode throughput in tokens/sec, or `null` when unmeasured. */
+    decodeTokensPerSecond: number | null;
+}
+/** A source of live device signals — swappable for tests and alternate hosts. */
+export type LiveDeviceSignalsSource = () => LiveDeviceSignals;
+/** Read the current live device signals. */
+export declare function readLiveDeviceSignals(): LiveDeviceSignals;
+/**
+ * Override the live-signals source. Pass `null` to restore the default
+ * (device-bridge + telemetry) source. Intended for tests and alternate hosts.
+ */
+export declare function setLiveDeviceSignalsSource(source: LiveDeviceSignalsSource | null): void;
+/**
+ * Whether the live signals say the device should be demoted off local for now,
+ * even though the static tier favours local. True when the OS is throttling
+ * (`serious`/`critical`) or measured decode throughput sits below the budget.
+ * A `null` signal never demotes.
+ */
+export declare function liveSignalsDemoteLocal(signals: LiveDeviceSignals): boolean;
+//# sourceMappingURL=live-signals.d.ts.map

package/dist/services/live-signals.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"live-signals.d.ts","sourceRoot":"","sources":["../../src/services/live-signals.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAKH,gFAAgF;AAChF,MAAM,MAAM,YAAY,GACrB,SAAS,GACT,MAAM,GACN,SAAS,GACT,UAAU,GACV,SAAS,CAAC;AAEb;;;;GAIG;AACH,eAAO,MAAM,iBAAiB,yBAAyB,CAAC;AAExD;;;;GAIG;AACH,eAAO,MAAM,qBAAqB,IAAI,CAAC;AAUvC;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEzD;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,sBAAsB,CAAC,MAAM,EAAE;IAC9C,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;CAClB,GAAG,IAAI,CAeP;AAED,2DAA2D;AAC3D,MAAM,WAAW,iBAAiB;IACjC,yEAAyE;IACzE,YAAY,EAAE,YAAY,GAAG,IAAI,CAAC;IAClC,6EAA6E;IAC7E,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;CACrC;AAED,iFAAiF;AACjF,MAAM,MAAM,uBAAuB,GAAG,MAAM,iBAAiB,CAAC;AA+B9D,4CAA4C;AAC5C,wBAAgB,qBAAqB,IAAI,iBAAiB,CAEzD;AAED;;;GAGG;AACH,wBAAgB,0BAA0B,CACzC,MAAM,EAAE,uBAAuB,GAAG,IAAI,GACpC,IAAI,CAEN;AAED;;;;;GAKG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,iBAAiB,GAAG,OAAO,CAc1E"}

package/dist/services/llama-server-metrics.d.ts

@@ -0,0 +1,114 @@
+/**
+ * Scraper for llama-server's `/metrics` (Prometheus exposition format)
+ * endpoint. Translates the running counters into the
+ * Anthropic-SDK-shaped `usage` block that callers already know how to
+ * consume from the cloud Anthropic plugin.
+ *
+ * llama-server publishes the following counters (per-process, monotonic):
+ *
+ *   llamacpp:n_decode_total           — context tokens decoded (prefill + gen)
+ *   llamacpp:n_tokens_predicted_total — output tokens
+ *   llamacpp:prompt_tokens_total      — total input tokens accepted
+ *   llamacpp:n_past_max               — high-water mark of cached past-tokens
+ *   llamacpp:n_prompt_tokens_processed_total — fresh tokens prefilled
+ *                                       (i.e. cache MISS), excludes cache hits
+ *   llamacpp:kv_cache_tokens          — current size of KV cache (gauge)
+ *   llamacpp:kv_cache_used_cells      — slots with active KV (gauge)
+ *
+ * For MTP speculative decoding, the fork additionally publishes:
+ *
+ *   llamacpp:n_drafted_total          — drafter-emitted tokens
+ *   llamacpp:n_drafted_accepted_total — accepted speculative tokens
+ *
+ * The mapping into Anthropic shape:
+ *
+ *   prompt_tokens_total                              → input_tokens
+ *   n_tokens_predicted_total                         → output_tokens
+ *   n_prompt_tokens_processed_total                  → cache_creation_input_tokens
+ *   prompt_tokens_total - n_prompt_tokens_processed_total → cache_read_input_tokens
+ *   n_drafted_total / n_drafted_accepted_total       → MTP extension fields
+ *
+ * Counters are taken as deltas across two snapshots: take one before
+ * `generate`, one after, and subtract. Losing a few samples to process
+ * restart is acceptable — the deltas are useful for the call's own
+ * usage accounting, not for global monitoring.
+ */
+export interface LlamaServerMetricSnapshot {
+    /** Wall-clock ms when the snapshot was taken; useful for diagnostics. */
+    takenAtMs: number;
+    /** True when `/metrics` was fetched and parsed. False means scrape failure. */
+    scrapeOk?: boolean;
+    /** True when the scrape included at least one generation/speculation counter. */
+    hasGenerationCounters?: boolean;
+    promptTokensTotal: number;
+    predictedTokensTotal: number;
+    /** Tokens that had to be freshly prefilled — i.e. cache MISS this turn. */
+    promptTokensProcessedTotal: number;
+    draftedTotal: number;
+    acceptedTotal: number;
+    /** Current size of the KV cache (gauge). */
+    kvCacheTokens: number;
+    /** Number of slots currently holding active KV (gauge). */
+    kvCacheUsedCells: number;
+}
+/**
+ * Parse a Prometheus exposition-format payload into a metric snapshot.
+ * Unknown or malformed lines are silently skipped — counters we don't
+ * recognise are not interesting and metric exporters add new ones over
+ * time.
+ *
+ * llama-server usually exposes one sample per metric (no labels), e.g.
+ *   `llamacpp:prompt_tokens_total 1234`
+ * Some MTP forks expose per-slot labelled samples, e.g.
+ *   `llamacpp:n_drafted_accepted_total{slot_id="0"} 12`
+ * Labelled samples are summed unless an unlabelled total exists for the same
+ * canonical field, in which case the unlabelled total wins.
+ */
+export declare function parsePrometheusMetrics(body: string, takenAtMs?: number): LlamaServerMetricSnapshot;
+/**
+ * Anthropic-SDK-shaped usage block, optionally extended with MTP
+ * speculative-decoding metrics. The cloud plugin (plugin-anthropic)
+ * emits the first three fields verbatim; local inference adds the
+ * `mtp_*` fields when speculative decoding is active. Callers that
+ * already handle the cloud `usage` shape need no change.
+ */
+export interface LocalUsageBlock {
+    [key: string]: unknown;
+    input_tokens: number;
+    output_tokens: number;
+    cache_creation_input_tokens: number;
+    cache_read_input_tokens: number;
+    mtp_drafted_tokens?: number;
+    mtp_accepted_tokens?: number;
+    /** 0..1 — proportion of drafted tokens that were accepted. */
+    mtp_acceptance_rate?: number;
+    /** 0..1 — proportion of input tokens that hit a warm slot (cache reuse). */
+    cache_hit_rate?: number;
+}
+/**
+ * Compute the Anthropic-shape usage block for a single generation by
+ * differencing two snapshots. `before` is taken just before the request,
+ * `after` just after the response was received. Negative deltas (caused
+ * by a metric reset between snapshots, e.g. server restart) are clamped
+ * to 0 — losing the sample is preferable to surfacing nonsense to the
+ * caller.
+ *
+ * Pass `responseUsage` to override input/output counts when the response
+ * payload itself reports per-call counters that are more accurate than
+ * the metric delta — llama-server's chat completion response includes
+ * `usage.{prompt,completion}_tokens` per request, which is exact while
+ * the metric delta is "everything that happened during the wall-clock
+ * window of the request."
+ */
+export declare function diffSnapshots(before: LlamaServerMetricSnapshot, after: LlamaServerMetricSnapshot, responseUsage?: {
+    prompt_tokens?: number;
+    completion_tokens?: number;
+}): LocalUsageBlock;
+/**
+ * GET `/metrics` from a running llama-server and parse it. Errors fall
+ * back to a zero-valued snapshot rather than throwing — observability
+ * MUST NOT break generation. `scrapeOk=false` tells callers that the
+ * zeros are not evidence of absent MTP/KV activity.
+ */
+export declare function fetchMetricsSnapshot(baseUrl: string, signal?: AbortSignal, timeoutMs?: number): Promise<LlamaServerMetricSnapshot>;
+//# sourceMappingURL=llama-server-metrics.d.ts.map

package/dist/services/llama-server-metrics.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"llama-server-metrics.d.ts","sourceRoot":"","sources":["../../src/services/llama-server-metrics.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AAEH,MAAM,WAAW,yBAAyB;IACzC,yEAAyE;IACzE,SAAS,EAAE,MAAM,CAAC;IAClB,+EAA+E;IAC/E,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,iFAAiF;IACjF,qBAAqB,CAAC,EAAE,OAAO,CAAC;IAChC,iBAAiB,EAAE,MAAM,CAAC;IAC1B,oBAAoB,EAAE,MAAM,CAAC;IAC7B,2EAA2E;IAC3E,0BAA0B,EAAE,MAAM,CAAC;IACnC,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,EAAE,MAAM,CAAC;IACtB,4CAA4C;IAC5C,aAAa,EAAE,MAAM,CAAC;IACtB,2DAA2D;IAC3D,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAuBD;;;;;;;;;;;;GAYG;AACH,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,MAAM,EACZ,SAAS,GAAE,MAAmB,GAC5B,yBAAyB,CAuD3B;AAED;;;;;;GAMG;AACH,MAAM,WAAW,eAAe;IAC/B,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;IACvB,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,EAAE,MAAM,CAAC;IACtB,2BAA2B,EAAE,MAAM,CAAC;IACpC,uBAAuB,EAAE,MAAM,CAAC;IAChC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,8DAA8D;IAC9D,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,4EAA4E;IAC5E,cAAc,CAAC,EAAE,MAAM,CAAC;CACxB;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,aAAa,CAC5B,MAAM,EAAE,yBAAyB,EACjC,KAAK,EAAE,yBAAyB,EAChC,aAAa,CAAC,EAAE;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;CAAE,GACpE,eAAe,CA2CjB;AAOD;;;;;GAKG;AACH,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,MAAM,EACf,MAAM,CAAC,EAAE,WAAW,EACpB,SAAS,SAAoC,GAC3C,OAAO,CAAC,yBAAyB,CAAC,CAsDpC"}

package/dist/services/llm-streaming-binding.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"llm-streaming-binding.d.ts","sourceRoot":"","sources":["../../src/services/llm-streaming-binding.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,KAAK,EACX,2BAA2B,EAC3B,iBAAiB,EACjB,eAAe,EACf,eAAe,EACf,aAAa,EACb,MAAM,sBAAsB,CAAC;AAE9B;;;;;;GAMG;AACH,MAAM,MAAM,YAAY,GAAG,2BAA2B,CAAC;AAEvD;;;;;;;;;;;GAWG;AACH,MAAM,WAAW,mBAAmB;IACnC,6EAA6E;IAC7E,kBAAkB,IAAI,OAAO,CAAC;IAC9B;;;;OAIG;IACH,aAAa,CAAC,IAAI,EAAE;QACnB,GAAG,EAAE,YAAY,CAAC;QAClB,MAAM,EAAE,eAAe,CAAC;KACxB,GAAG,eAAe,CAAC;IACpB,2EAA2E;IAC3E,gBAAgB,CAAC,IAAI,EAAE;QAAE,MAAM,EAAE,eAAe,CAAC;QAAC,MAAM,EAAE,UAAU,CAAA;KAAE,GAAG,IAAI,CAAC;IAC9E;;;;OAIG;IACH,aAAa,CAAC,IAAI,EAAE;QACnB,MAAM,EAAE,eAAe,CAAC;QACxB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,YAAY,CAAC,EAAE,MAAM,CAAC;KACtB,GAAG,aAAa,CAAC;IAClB,uEAAuE;IACvE,eAAe,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;IAC/C,kFAAkF;IAClF,cAAc,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;IAC9C,8DAA8D;IAC9D,iBAAiB,CAAC,CAAC,IAAI,EAAE;QAAE,MAAM,EAAE,eAAe,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,CAAC;IAC9E,0DAA0D;IAC1D,oBAAoB,CAAC,CAAC,IAAI,EAAE;QAC3B,MAAM,EAAE,eAAe,CAAC;QACxB,QAAQ,EAAE,MAAM,CAAC;KACjB,GAAG,IAAI,CAAC;CACT;AAED;;;;;;;;;GASG;AACH,wBAAgB,qBAAqB,CACpC,GAAG,EAAE,iBAAiB,GACpB,mBAAmB,CAmCrB"}

package/dist/services/load-args.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"load-args.d.ts","sourceRoot":"","sources":["../../src/services/load-args.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH;;;;;;;;GAQG;AACH,MAAM,MAAM,aAAa,GAAG,KAAK,GAAG,KAAK,GAAG,OAAO,GAAG;IAAE,SAAS,EAAE,MAAM,CAAA;CAAE,CAAC;AAE5E;;;;;GAKG;AACH,MAAM,WAAW,sBAAsB;IACtC,SAAS,EAAE,MAAM,CAAC;IAClB;;;;OAIG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;;;;OAKG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;;;OAKG;IACH,SAAS,CAAC,EAAE,aAAa,CAAC;IAC1B,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,IAAI,CAAC,EAAE,OAAO,CAAC;IACf,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB;;;;;;;;;;;;;;OAcG;IACH,UAAU,CAAC,EAAE,MAAM,CAAC;CACpB"}

package/dist/services/manifest/index.d.ts

@@ -0,0 +1,4 @@
+export { ELIZA_1_BACKENDS, ELIZA_1_KERNELS, ELIZA_1_MANIFEST_SCHEMA_URL, ELIZA_1_MANIFEST_SCHEMA_VERSION, ELIZA_1_PROVENANCE_SLOTS, ELIZA_1_RELEASE_CHANNELS, ELIZA_1_RELEASE_STATES, ELIZA_1_TEXT_RUNTIME_CLASSES, ELIZA_1_TIERS, ELIZA_1_TOKENIZER_FAMILY, ELIZA_1_TOKENIZER_VOCAB_SIZE, ELIZA_1_VOICE_CAPABILITIES, ELIZA_1_VOICE_MANIFEST_VERSION, ELIZA1_TO_RUNTIME_KERNEL, Eliza1BackendEnumSchema, Eliza1EvalsSchema, Eliza1FileEntrySchema, Eliza1FilesSchema, Eliza1KernelEnumSchema, Eliza1KernelsSchema, Eliza1LibFileEntrySchema, Eliza1LineageSchema, Eliza1ManifestSchema, Eliza1ProvenanceSchema, Eliza1RamBudgetSchema, Eliza1TextRuntimeClassEnumSchema, Eliza1TierEnumSchema, Eliza1VerifiedBackendStatusSchema, Eliza1VoiceSchema, OPTIONAL_KERNELS_BY_TIER, REQUIRED_KERNELS_BY_TIER, RUNTIME_TO_ELIZA1_KERNEL, SUPPORTED_BACKENDS_BY_TIER, VOICE_PRESET_CACHE_PATH, } from "./schema";
+export type { Eliza1Backend, Eliza1DeviceCaps, Eliza1Evals, Eliza1FileEntry, Eliza1Files, Eliza1Kernel, Eliza1Kernels, Eliza1LibFileEntry, Eliza1Lineage, Eliza1Manifest, Eliza1Provenance, Eliza1ProvenanceSlot, Eliza1RamBudget, Eliza1ReleaseChannel, Eliza1ReleaseState, Eliza1TextRuntimeClass, Eliza1Tier, Eliza1VerifiedBackendStatus, Eliza1Voice, } from "./types";
+export { canSetAsDefault, missingRequiredKernels, parseManifestOrThrow, type ValidationErr, type ValidationOk, type ValidationResult, validateManifest, } from "./validator";
+//# sourceMappingURL=index.d.ts.map

package/dist/services/manifest/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/services/manifest/index.ts"],"names":[],"mappings":"AAIA,OAAO,EACN,gBAAgB,EAChB,eAAe,EACf,2BAA2B,EAC3B,+BAA+B,EAC/B,wBAAwB,EACxB,wBAAwB,EACxB,sBAAsB,EACtB,4BAA4B,EAC5B,aAAa,EACb,wBAAwB,EACxB,4BAA4B,EAC5B,0BAA0B,EAC1B,8BAA8B,EAC9B,wBAAwB,EACxB,uBAAuB,EACvB,iBAAiB,EACjB,qBAAqB,EACrB,iBAAiB,EACjB,sBAAsB,EACtB,mBAAmB,EACnB,wBAAwB,EACxB,mBAAmB,EACnB,oBAAoB,EACpB,sBAAsB,EACtB,qBAAqB,EACrB,gCAAgC,EAChC,oBAAoB,EACpB,iCAAiC,EACjC,iBAAiB,EACjB,wBAAwB,EACxB,wBAAwB,EACxB,wBAAwB,EACxB,0BAA0B,EAC1B,uBAAuB,GACvB,MAAM,UAAU,CAAC;AAElB,YAAY,EACX,aAAa,EACb,gBAAgB,EAChB,WAAW,EACX,eAAe,EACf,WAAW,EACX,YAAY,EACZ,aAAa,EACb,kBAAkB,EAClB,aAAa,EACb,cAAc,EACd,gBAAgB,EAChB,oBAAoB,EACpB,eAAe,EACf,oBAAoB,EACpB,kBAAkB,EAClB,sBAAsB,EACtB,UAAU,EACV,2BAA2B,EAC3B,WAAW,GACX,MAAM,SAAS,CAAC;AAEjB,OAAO,EACN,eAAe,EACf,sBAAsB,EACtB,oBAAoB,EACpB,KAAK,aAAa,EAClB,KAAK,YAAY,EACjB,KAAK,gBAAgB,EACrB,gBAAgB,GAChB,MAAM,aAAa,CAAC"}