@elizaos/shared 2.0.0-beta.1 → 2.0.3-beta.3

package/local-inference/providers-types.d.ts

@@ -0,0 +1,56 @@
+/**
+ * Provider-registry **data-only** contracts shared between the agent
+ * server and UI clients. Everything in this file must remain free of
+ * runtime-only capabilities (no methods, no callables) so the same types
+ * can be referenced by mobile/browser UI bundles without implying Node.js
+ * I/O.
+ *
+ * The runtime-side `ProviderDefinition` (which adds a callable
+ * `getEnableState()` reading env vars / fs / device-bridge sockets)
+ * extends `ProviderMeta` and lives in
+ * `@elizaos/app-core/src/services/local-inference/providers.ts` — it is
+ * the authoritative source for `/api/local-inference/providers`.
+ *
+ * UI consumers (`client-local-inference.ts`, `ios-local-agent-kernel.ts`)
+ * only see `ProviderStatus` (the response shape) and `ProviderMeta` /
+ * `ProviderEnableState` / `ProviderId` — never the runtime definition.
+ */
+import type { AgentModelSlot } from "./types.js";
+export type ProviderId = "eliza-local-inference" | "eliza-device-bridge" | "capacitor-llama" | "anthropic-subscription" | "openai-codex" | "gemini-cli" | "zai-coding" | "kimi-coding" | "deepseek-coding" | "anthropic" | "openai" | "deepseek" | "nearai" | "zai" | "moonshot" | "grok" | "elizacloud" | "google" | "mistral";
+export interface ProviderEnableState {
+    enabled: boolean;
+    /** Short reason, e.g. "API key set", "Device connected", "No API key". */
+    reason: string;
+}
+/**
+ * Data-only provider descriptor. UI-safe — contains no runtime methods.
+ * The server-side `ProviderDefinition` (in `@elizaos/app-core`) extends
+ * this with a `getEnableState()` callable.
+ */
+export interface ProviderMeta {
+    id: ProviderId;
+    label: string;
+    kind: "cloud-api" | "cloud-subscription" | "local" | "device-bridge";
+    /** Short blurb shown in the UI. */
+    description: string;
+    /** Agent slots this provider can plausibly serve. */
+    supportedSlots: AgentModelSlot[];
+    /**
+     * Link to the settings UI where enable/configure actually happens.
+     * UI sends the user here via anchor-scroll when they click "Configure".
+     * `null` means the provider has no separate config surface.
+     */
+    configureHref: string | null;
+}
+export interface ProviderStatus {
+    id: ProviderId;
+    label: string;
+    kind: ProviderMeta["kind"];
+    description: string;
+    supportedSlots: AgentModelSlot[];
+    configureHref: string | null;
+    enableState: ProviderEnableState;
+    /** Registered model types this provider has handlers for, right now. */
+    registeredSlots: string[];
+}
+//# sourceMappingURL=providers-types.d.ts.map

package/local-inference/providers-types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"providers-types.d.ts","sourceRoot":"","sources":["../../src/local-inference/providers-types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,MAAM,UAAU,GAClB,uBAAuB,GACvB,qBAAqB,GACrB,iBAAiB,GACjB,wBAAwB,GACxB,cAAc,GACd,YAAY,GACZ,YAAY,GACZ,aAAa,GACb,iBAAiB,GACjB,WAAW,GACX,QAAQ,GACR,UAAU,GACV,QAAQ,GACR,KAAK,GACL,UAAU,GACV,MAAM,GACN,YAAY,GACZ,QAAQ,GACR,SAAS,CAAC;AAEd,MAAM,WAAW,mBAAmB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,0EAA0E;IAC1E,MAAM,EAAE,MAAM,CAAC;CAChB;AAED;;;;GAIG;AACH,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,UAAU,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,WAAW,GAAG,oBAAoB,GAAG,OAAO,GAAG,eAAe,CAAC;IACrE,mCAAmC;IACnC,WAAW,EAAE,MAAM,CAAC;IACpB,qDAAqD;IACrD,cAAc,EAAE,cAAc,EAAE,CAAC;IACjC;;;;OAIG;IACH,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAED,MAAM,WAAW,cAAc;IAC7B,EAAE,EAAE,UAAU,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC;IAC3B,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,cAAc,EAAE,CAAC;IACjC,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,WAAW,EAAE,mBAAmB,CAAC;IACjC,wEAAwE;IACxE,eAAe,EAAE,MAAM,EAAE,CAAC;CAC3B"}

package/local-inference/providers-types.js

@@ -0,0 +1,19 @@
+/**
+ * Provider-registry **data-only** contracts shared between the agent
+ * server and UI clients. Everything in this file must remain free of
+ * runtime-only capabilities (no methods, no callables) so the same types
+ * can be referenced by mobile/browser UI bundles without implying Node.js
+ * I/O.
+ *
+ * The runtime-side `ProviderDefinition` (which adds a callable
+ * `getEnableState()` reading env vars / fs / device-bridge sockets)
+ * extends `ProviderMeta` and lives in
+ * `@elizaos/app-core/src/services/local-inference/providers.ts` — it is
+ * the authoritative source for `/api/local-inference/providers`.
+ *
+ * UI consumers (`client-local-inference.ts`, `ios-local-agent-kernel.ts`)
+ * only see `ProviderStatus` (the response shape) and `ProviderMeta` /
+ * `ProviderEnableState` / `ProviderId` — never the runtime definition.
+ */
+export {};
+//# sourceMappingURL=providers-types.js.map

package/local-inference/providers-types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"providers-types.js","sourceRoot":"","sources":["../../src/local-inference/providers-types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG"}

package/local-inference/routing-preferences.d.ts

@@ -0,0 +1,42 @@
+/**
+ * Per-model-type user override: "for TEXT_LARGE, prefer this provider".
+ *
+ * Persisted to `$STATE_DIR/local-inference/routing.json` and read by the
+ * router-handler (see `router-handler.ts`) to pick a provider at dispatch
+ * time. When a slot has no override, the runtime's native priority order
+ * wins — i.e. this is layered over the existing registration priority
+ * rather than replacing it.
+ */
+import type { AgentModelSlot } from "./types.js";
+export type RoutingPolicy = "manual" | "auto" | "local-only" | "cloud-only" | "cheapest" | "fastest" | "prefer-local" | "round-robin";
+/**
+ * The full set of selectable policies, in display order. Kept as a runtime
+ * value (not just a type) so route-layer validation and the settings UI share
+ * one source of truth for "which policies are accepted".
+ *
+ * `local-only` / `cloud-only` are the canonical per-slot replacements for the
+ * global `ELIZA_LOCAL_ONLY` env hack: `local-only` is a hard guarantee (never
+ * falls through to cloud), `cloud-only` never dispatches on-device.
+ */
+export declare const ROUTING_POLICIES: readonly RoutingPolicy[];
+export declare function isRoutingPolicy(value: unknown): value is RoutingPolicy;
+export declare const DEFAULT_ROUTING_POLICY: RoutingPolicy;
+export interface RoutingPreferences {
+    /**
+     * Explicit provider override per agent slot. Empty record = no overrides,
+     * runtime picks the highest-priority registered handler.
+     */
+    preferredProvider: Partial<Record<AgentModelSlot, string>>;
+    /**
+     * Per-slot policy. "manual" honours `preferredProvider` verbatim;
+     * everything else lets the router-handler compute a winner from the
+     * policy rule set. Absent = "prefer-local" so local models and
+     * subscriptions are tried before direct paid APIs and managed cloud.
+     */
+    policy: Partial<Record<AgentModelSlot, RoutingPolicy>>;
+}
+export declare function readRoutingPreferences(): Promise<RoutingPreferences>;
+export declare function writeRoutingPreferences(prefs: RoutingPreferences): Promise<void>;
+export declare function setPreferredProvider(slot: AgentModelSlot, provider: string | null): Promise<RoutingPreferences>;
+export declare function setPolicy(slot: AgentModelSlot, policy: RoutingPolicy | null): Promise<RoutingPreferences>;
+//# sourceMappingURL=routing-preferences.d.ts.map

package/local-inference/routing-preferences.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"routing-preferences.d.ts","sourceRoot":"","sources":["../../src/local-inference/routing-preferences.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAKH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,MAAM,aAAa,GACrB,QAAQ,GACR,MAAM,GACN,YAAY,GACZ,YAAY,GACZ,UAAU,GACV,SAAS,GACT,cAAc,GACd,aAAa,CAAC;AAElB;;;;;;;;GAQG;AACH,eAAO,MAAM,gBAAgB,EAAE,SAAS,aAAa,EAS3C,CAAC;AAEX,wBAAgB,eAAe,CAAC,KAAK,EAAE,OAAO,GAAG,KAAK,IAAI,aAAa,CAKtE;AAED,eAAO,MAAM,sBAAsB,EAAE,aAA8B,CAAC;AAEpE,MAAM,WAAW,kBAAkB;IACjC;;;OAGG;IACH,iBAAiB,EAAE,OAAO,CAAC,MAAM,CAAC,cAAc,EAAE,MAAM,CAAC,CAAC,CAAC;IAC3D;;;;;OAKG;IACH,MAAM,EAAE,OAAO,CAAC,MAAM,CAAC,cAAc,EAAE,aAAa,CAAC,CAAC,CAAC;CACxD;AAiBD,wBAAsB,sBAAsB,IAAI,OAAO,CAAC,kBAAkB,CAAC,CAY1E;AAED,wBAAsB,uBAAuB,CAC3C,KAAK,EAAE,kBAAkB,GACxB,OAAO,CAAC,IAAI,CAAC,CAMf;AAED,wBAAsB,oBAAoB,CACxC,IAAI,EAAE,cAAc,EACpB,QAAQ,EAAE,MAAM,GAAG,IAAI,GACtB,OAAO,CAAC,kBAAkB,CAAC,CAa7B;AAED,wBAAsB,SAAS,CAC7B,IAAI,EAAE,cAAc,EACpB,MAAM,EAAE,aAAa,GAAG,IAAI,GAC3B,OAAO,CAAC,kBAAkB,CAAC,CAa7B"}

package/local-inference/routing-preferences.js

@@ -0,0 +1,96 @@
+/**
+ * Per-model-type user override: "for TEXT_LARGE, prefer this provider".
+ *
+ * Persisted to `$STATE_DIR/local-inference/routing.json` and read by the
+ * router-handler (see `router-handler.ts`) to pick a provider at dispatch
+ * time. When a slot has no override, the runtime's native priority order
+ * wins — i.e. this is layered over the existing registration priority
+ * rather than replacing it.
+ */
+import fs from "node:fs/promises";
+import path from "node:path";
+import { localInferenceRoot } from "./paths.js";
+/**
+ * The full set of selectable policies, in display order. Kept as a runtime
+ * value (not just a type) so route-layer validation and the settings UI share
+ * one source of truth for "which policies are accepted".
+ *
+ * `local-only` / `cloud-only` are the canonical per-slot replacements for the
+ * global `ELIZA_LOCAL_ONLY` env hack: `local-only` is a hard guarantee (never
+ * falls through to cloud), `cloud-only` never dispatches on-device.
+ */
+export const ROUTING_POLICIES = [
+    "manual",
+    "auto",
+    "local-only",
+    "cloud-only",
+    "cheapest",
+    "fastest",
+    "prefer-local",
+    "round-robin",
+];
+export function isRoutingPolicy(value) {
+    return (typeof value === "string" &&
+        ROUTING_POLICIES.includes(value));
+}
+export const DEFAULT_ROUTING_POLICY = "prefer-local";
+const EMPTY = { preferredProvider: {}, policy: {} };
+function routingPath() {
+    return path.join(localInferenceRoot(), "routing.json");
+}
+async function ensureRoot() {
+    await fs.mkdir(localInferenceRoot(), { recursive: true });
+}
+export async function readRoutingPreferences() {
+    try {
+        const raw = await fs.readFile(routingPath(), "utf8");
+        const parsed = JSON.parse(raw);
+        if (parsed?.version !== 1 || !parsed.preferences)
+            return EMPTY;
+        return {
+            preferredProvider: parsed.preferences.preferredProvider,
+            policy: parsed.preferences.policy,
+        };
+    }
+    catch {
+        return EMPTY;
+    }
+}
+export async function writeRoutingPreferences(prefs) {
+    await ensureRoot();
+    const payload = { version: 1, preferences: prefs };
+    const tmp = `${routingPath()}.tmp`;
+    await fs.writeFile(tmp, JSON.stringify(payload, null, 2), "utf8");
+    await fs.rename(tmp, routingPath());
+}
+export async function setPreferredProvider(slot, provider) {
+    const current = await readRoutingPreferences();
+    const next = {
+        preferredProvider: { ...current.preferredProvider },
+        policy: { ...current.policy },
+    };
+    if (provider) {
+        next.preferredProvider[slot] = provider;
+    }
+    else {
+        delete next.preferredProvider[slot];
+    }
+    await writeRoutingPreferences(next);
+    return next;
+}
+export async function setPolicy(slot, policy) {
+    const current = await readRoutingPreferences();
+    const next = {
+        preferredProvider: { ...current.preferredProvider },
+        policy: { ...current.policy },
+    };
+    if (policy) {
+        next.policy[slot] = policy;
+    }
+    else {
+        delete next.policy[slot];
+    }
+    await writeRoutingPreferences(next);
+    return next;
+}
+//# sourceMappingURL=routing-preferences.js.map

package/local-inference/routing-preferences.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"routing-preferences.js","sourceRoot":"","sources":["../../src/local-inference/routing-preferences.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EAAE,MAAM,kBAAkB,CAAC;AAClC,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,kBAAkB,EAAE,MAAM,YAAY,CAAC;AAahD;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAA6B;IACxD,QAAQ;IACR,MAAM;IACN,YAAY;IACZ,YAAY;IACZ,UAAU;IACV,SAAS;IACT,cAAc;IACd,aAAa;CACL,CAAC;AAEX,MAAM,UAAU,eAAe,CAAC,KAAc;IAC5C,OAAO,CACL,OAAO,KAAK,KAAK,QAAQ;QACxB,gBAAsC,CAAC,QAAQ,CAAC,KAAK,CAAC,CACxD,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,MAAM,sBAAsB,GAAkB,cAAc,CAAC;AAsBpE,MAAM,KAAK,GAAuB,EAAE,iBAAiB,EAAE,EAAE,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC;AAExE,SAAS,WAAW;IAClB,OAAO,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,EAAE,cAAc,CAAC,CAAC;AACzD,CAAC;AAED,KAAK,UAAU,UAAU;IACvB,MAAM,EAAE,CAAC,KAAK,CAAC,kBAAkB,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;AAC5D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,sBAAsB;IAC1C,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,MAAM,EAAE,CAAC,QAAQ,CAAC,WAAW,EAAE,EAAE,MAAM,CAAC,CAAC;QACrD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAgB,CAAC;QAC9C,IAAI,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW;YAAE,OAAO,KAAK,CAAC;QAC/D,OAAO;YACL,iBAAiB,EAAE,MAAM,CAAC,WAAW,CAAC,iBAAiB;YACvD,MAAM,EAAE,MAAM,CAAC,WAAW,CAAC,MAAM;SAClC,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,KAAK,CAAC;IACf,CAAC;AACH,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAC3C,KAAyB;IAEzB,MAAM,UAAU,EAAE,CAAC;IACnB,MAAM,OAAO,GAAgB,EAAE,OAAO,EAAE,CAAC,EAAE,WAAW,EAAE,KAAK,EAAE,CAAC;IAChE,MAAM,GAAG,GAAG,GAAG,WAAW,EAAE,MAAM,CAAC;IACnC,MAAM,EAAE,CAAC,SAAS,CAAC,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;IAClE,MAAM,EAAE,CAAC,MAAM,CAAC,GAAG,EAAE,WAAW,EAAE,CAAC,CAAC;AACtC,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,IAAoB,EACpB,QAAuB;IAEvB,MAAM,OAAO,GAAG,MAAM,sBAAsB,EAAE,CAAC;IAC/C,MAAM,IAAI,GAAuB;QAC/B,iBAAiB,EAAE,EAAE,GAAG,OAAO,CAAC,iBAAiB,EAAE;QACnD,MAAM,EAAE,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE;KAC9B,CAAC;IACF,IAAI,QAAQ,EAAE,CAAC;QACb,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC;IAC1C,CAAC;SAAM,CAAC;QACN,OAAO,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACtC,CAAC;IACD,MAAM,uBAAuB,CAAC,IAAI,CAAC,CAAC;IACpC,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,IAAoB,EACpB,MAA4B;IAE5B,MAAM,OAAO,GAAG,MAAM,sBAAsB,EAAE,CAAC;IAC/C,MAAM,IAAI,GAAuB;QAC/B,iBAAiB,EAAE,EAAE,GAAG,OAAO,CAAC,iBAAiB,EAAE;QACnD,MAAM,EAAE,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE;KAC9B,CAAC;IACF,IAAI,MAAM,EAAE,CAAC;QACX,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,MAAM,CAAC;IAC7B,CAAC;SAAM,CAAC;QACN,OAAO,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IAC3B,CAAC;IACD,MAAM,uBAAuB,CAAC,IAAI,CAAC,CAAC;IACpC,OAAO,IAAI,CAAC;AACd,CAAC"}

package/local-inference/runtime-class.d.ts

@@ -0,0 +1,29 @@
+/**
+ * Local-inference runtime-class discriminator.
+ *
+ * The local stack is Eliza-1 only: every text model is an Eliza-1 bundle served
+ * by the fused `libelizainference` (`desktop-fused-ffi-backend-runtime.ts`). The
+ * fused context is anchored at a *bundle root* (`ffi.create(bundleRoot)` →
+ * `<bundleRoot>/text/*.gguf`) and the full optimization stack rides along:
+ * embedded draft head speculative decoding, stock q8_0 KV cache, native
+ * tokenization over the resident Gemma 4 (eliza-1) vocab, and fused
+ * voice/vision. It cannot load an arbitrary single GGUF — the vocab and the KV
+ * cache layout are tokenizer/format-specific to Eliza-1.
+ *
+ * The discriminator is retained as a single-valued type so external importers
+ * keep resolving; there is no longer a second runtime class to branch on.
+ */
+import type { CatalogModel, InstalledModel } from "./types.js";
+/**
+ * Which desktop text runtime a model is served by. The local stack is Eliza-1
+ * only, so this is a single constant (`"fused-eliza1"`); the generic-GGUF
+ * runtime class was removed with the multi-model machinery (#8808).
+ */
+export type RuntimeClass = "fused-eliza1";
+/** Every catalog entry in the Eliza-1-only stack is served by the fused runtime. */
+export declare function classifyCatalogModelRuntimeClass(_model: Pick<CatalogModel, "id" | "bundleManifestFile" | "runtimeRole">): RuntimeClass;
+/** Every installed model in the Eliza-1-only stack is served by the fused runtime. */
+export declare function classifyInstalledModelRuntimeClass(_model: Pick<InstalledModel, "id" | "bundleRoot" | "runtimeClass">): RuntimeClass;
+/** Backfill `runtimeClass` on an installed-model entry that predates the field. */
+export declare function withRuntimeClass(model: InstalledModel): InstalledModel;
+//# sourceMappingURL=runtime-class.d.ts.map

package/local-inference/runtime-class.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"runtime-class.d.ts","sourceRoot":"","sources":["../../src/local-inference/runtime-class.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAE/D;;;;GAIG;AACH,MAAM,MAAM,YAAY,GAAG,cAAc,CAAC;AAE1C,oFAAoF;AACpF,wBAAgB,gCAAgC,CAC9C,MAAM,EAAE,IAAI,CAAC,YAAY,EAAE,IAAI,GAAG,oBAAoB,GAAG,aAAa,CAAC,GACtE,YAAY,CAEd;AAED,sFAAsF;AACtF,wBAAgB,kCAAkC,CAChD,MAAM,EAAE,IAAI,CAAC,cAAc,EAAE,IAAI,GAAG,YAAY,GAAG,cAAc,CAAC,GACjE,YAAY,CAEd;AAED,mFAAmF;AACnF,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,cAAc,GAAG,cAAc,CAGtE"}

package/local-inference/runtime-class.js

@@ -0,0 +1,30 @@
+/**
+ * Local-inference runtime-class discriminator.
+ *
+ * The local stack is Eliza-1 only: every text model is an Eliza-1 bundle served
+ * by the fused `libelizainference` (`desktop-fused-ffi-backend-runtime.ts`). The
+ * fused context is anchored at a *bundle root* (`ffi.create(bundleRoot)` →
+ * `<bundleRoot>/text/*.gguf`) and the full optimization stack rides along:
+ * embedded draft head speculative decoding, stock q8_0 KV cache, native
+ * tokenization over the resident Gemma 4 (eliza-1) vocab, and fused
+ * voice/vision. It cannot load an arbitrary single GGUF — the vocab and the KV
+ * cache layout are tokenizer/format-specific to Eliza-1.
+ *
+ * The discriminator is retained as a single-valued type so external importers
+ * keep resolving; there is no longer a second runtime class to branch on.
+ */
+/** Every catalog entry in the Eliza-1-only stack is served by the fused runtime. */
+export function classifyCatalogModelRuntimeClass(_model) {
+    return "fused-eliza1";
+}
+/** Every installed model in the Eliza-1-only stack is served by the fused runtime. */
+export function classifyInstalledModelRuntimeClass(_model) {
+    return "fused-eliza1";
+}
+/** Backfill `runtimeClass` on an installed-model entry that predates the field. */
+export function withRuntimeClass(model) {
+    if (model.runtimeClass)
+        return model;
+    return { ...model, runtimeClass: "fused-eliza1" };
+}
+//# sourceMappingURL=runtime-class.js.map

package/local-inference/runtime-class.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"runtime-class.js","sourceRoot":"","sources":["../../src/local-inference/runtime-class.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAWH,oFAAoF;AACpF,MAAM,UAAU,gCAAgC,CAC9C,MAAuE;IAEvE,OAAO,cAAc,CAAC;AACxB,CAAC;AAED,sFAAsF;AACtF,MAAM,UAAU,kCAAkC,CAChD,MAAkE;IAElE,OAAO,cAAc,CAAC;AACxB,CAAC;AAED,mFAAmF;AACnF,MAAM,UAAU,gBAAgB,CAAC,KAAqB;IACpD,IAAI,KAAK,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IACrC,OAAO,EAAE,GAAG,KAAK,EAAE,YAAY,EAAE,cAAc,EAAE,CAAC;AACpD,CAAC"}

package/local-inference/throughput.d.ts

@@ -0,0 +1,74 @@
+/**
+ * Per-generation throughput differencing for the on-device inference path.
+ *
+ * The device bridge carries the raw counters for every on-device generation
+ * (`promptTokens`, `outputTokens`, `durationMs`, and — when the device measured
+ * it — `ttftMs`, the time to the first decoded token). This module turns those
+ * counters into the two throughput numbers a profiling harness actually wants:
+ *
+ *   - **prefill tokens/sec** — prompt-processing throughput. The model reads the
+ *     whole prompt before emitting the first token, so `ttftMs` *is* the prefill
+ *     wall-clock; prefill tok/s = `promptTokens / (ttftMs / 1000)`.
+ *   - **decode tokens/sec** — generation throughput. The decode phase runs from
+ *     the first token to completion, so decode tok/s =
+ *     `outputTokens / ((durationMs - ttftMs) / 1000)`.
+ *
+ * When the device could not measure `ttftMs` (the non-streaming Capacitor llama
+ * path before per-token timing lands), prefill and decode cannot be separated —
+ * those fields are reported as `null` and only the combined throughput
+ * (`outputTokens / durationMs`) is filled. Per the repo architecture rules
+ * (AGENTS.md §3/§7) a quantity that could not be measured is recorded as `null`,
+ * never as a fabricated `0`.
+ *
+ * Shared between the agent-side device bridge
+ * (`@elizaos/plugin-local-inference/services/device-bridge`) and the UI client
+ * (`@elizaos/ui/services/local-inference/device-bridge`) — both hand-synced
+ * copies import this single source rather than re-implementing the math.
+ */
+/** Raw per-generation counters as carried by the device-bridge wire. */
+export interface GenerationCounters {
+    /** Prompt tokens the device processed during prefill. */
+    promptTokens: number;
+    /** Tokens the device decoded (generated). */
+    outputTokens: number;
+    /** Total wall-clock for the generation, in milliseconds. */
+    durationMs: number;
+    /**
+     * Time to first decoded token, in milliseconds, when the device measured it.
+     * Equals the prefill wall-clock. Absent on the non-streaming path.
+     */
+    ttftMs?: number | null;
+}
+/** Differenced throughput for a single on-device generation. */
+export interface GenerationThroughput {
+    /** Prefill (prompt-processing) throughput, tok/s, or null when unmeasurable. */
+    prefillTokensPerSecond: number | null;
+    /** Decode (generation) throughput, tok/s, or null when unmeasurable. */
+    decodeTokensPerSecond: number | null;
+    /**
+     * Whole-generation throughput, tok/s = outputTokens / durationMs. Always
+     * filled when at least one token was decoded in a positive duration, even
+     * when `ttftMs` is missing — this is the fallback signal for the
+     * non-streaming path.
+     */
+    combinedTokensPerSecond: number | null;
+    /** Echoed TTFT, ms, or null when the device did not measure it. */
+    ttftMs: number | null;
+    /** Derived decode wall-clock (durationMs − ttftMs), ms, or null. */
+    decodeMs: number | null;
+}
+/**
+ * Difference raw generation counters into prefill / decode / combined tok/s.
+ *
+ * Pure and synchronous. Every output field is `null` unless the inputs make it
+ * genuinely computable — never a fabricated zero.
+ */
+export declare function computeGenerationThroughput(counters: GenerationCounters): GenerationThroughput;
+/**
+ * Validate the raw counter shape before differencing. The wire is hand-synced
+ * between the agent and the device client, so a malformed frame should be
+ * rejected at the boundary (and the metric dropped) rather than silently
+ * producing garbage throughput.
+ */
+export declare function isGenerationCounters(value: unknown): value is GenerationCounters;
+//# sourceMappingURL=throughput.d.ts.map

package/local-inference/throughput.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"throughput.d.ts","sourceRoot":"","sources":["../../src/local-inference/throughput.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,wEAAwE;AACxE,MAAM,WAAW,kBAAkB;IACjC,yDAAyD;IACzD,YAAY,EAAE,MAAM,CAAC;IACrB,6CAA6C;IAC7C,YAAY,EAAE,MAAM,CAAC;IACrB,4DAA4D;IAC5D,UAAU,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CACxB;AAED,gEAAgE;AAChE,MAAM,WAAW,oBAAoB;IACnC,gFAAgF;IAChF,sBAAsB,EAAE,MAAM,GAAG,IAAI,CAAC;IACtC,wEAAwE;IACxE,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC;;;;;OAKG;IACH,uBAAuB,EAAE,MAAM,GAAG,IAAI,CAAC;IACvC,mEAAmE;IACnE,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,oEAAoE;IACpE,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;CACzB;AAYD;;;;;GAKG;AACH,wBAAgB,2BAA2B,CACzC,QAAQ,EAAE,kBAAkB,GAC3B,oBAAoB,CAmCtB;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAClC,KAAK,EAAE,OAAO,GACb,KAAK,IAAI,kBAAkB,CAS7B"}

package/local-inference/throughput.js

@@ -0,0 +1,84 @@
+/**
+ * Per-generation throughput differencing for the on-device inference path.
+ *
+ * The device bridge carries the raw counters for every on-device generation
+ * (`promptTokens`, `outputTokens`, `durationMs`, and — when the device measured
+ * it — `ttftMs`, the time to the first decoded token). This module turns those
+ * counters into the two throughput numbers a profiling harness actually wants:
+ *
+ *   - **prefill tokens/sec** — prompt-processing throughput. The model reads the
+ *     whole prompt before emitting the first token, so `ttftMs` *is* the prefill
+ *     wall-clock; prefill tok/s = `promptTokens / (ttftMs / 1000)`.
+ *   - **decode tokens/sec** — generation throughput. The decode phase runs from
+ *     the first token to completion, so decode tok/s =
+ *     `outputTokens / ((durationMs - ttftMs) / 1000)`.
+ *
+ * When the device could not measure `ttftMs` (the non-streaming Capacitor llama
+ * path before per-token timing lands), prefill and decode cannot be separated —
+ * those fields are reported as `null` and only the combined throughput
+ * (`outputTokens / durationMs`) is filled. Per the repo architecture rules
+ * (AGENTS.md §3/§7) a quantity that could not be measured is recorded as `null`,
+ * never as a fabricated `0`.
+ *
+ * Shared between the agent-side device bridge
+ * (`@elizaos/plugin-local-inference/services/device-bridge`) and the UI client
+ * (`@elizaos/ui/services/local-inference/device-bridge`) — both hand-synced
+ * copies import this single source rather than re-implementing the math.
+ */
+const MS_PER_SECOND = 1000;
+function isPositive(value) {
+    return typeof value === "number" && Number.isFinite(value) && value > 0;
+}
+function isNonNegativeCount(value) {
+    return typeof value === "number" && Number.isFinite(value) && value >= 0;
+}
+/**
+ * Difference raw generation counters into prefill / decode / combined tok/s.
+ *
+ * Pure and synchronous. Every output field is `null` unless the inputs make it
+ * genuinely computable — never a fabricated zero.
+ */
+export function computeGenerationThroughput(counters) {
+    const { promptTokens, outputTokens, durationMs } = counters;
+    const ttftMs = isPositive(counters.ttftMs) && counters.ttftMs < durationMs
+        ? counters.ttftMs
+        : null;
+    // Combined throughput needs a positive duration and ≥1 decoded token. Zero
+    // decoded tokens means there was no decode to measure — report null, not 0.
+    const combinedTokensPerSecond = isPositive(durationMs) && isPositive(outputTokens)
+        ? outputTokens / (durationMs / MS_PER_SECOND)
+        : null;
+    // Prefill throughput needs a measured TTFT and ≥1 prompt token.
+    const prefillTokensPerSecond = ttftMs !== null && isPositive(promptTokens)
+        ? promptTokens / (ttftMs / MS_PER_SECOND)
+        : null;
+    // Decode throughput needs a measured TTFT (so we can subtract prefill) and a
+    // positive remaining window with ≥1 decoded token.
+    const decodeMs = ttftMs !== null ? durationMs - ttftMs : null;
+    const decodeTokensPerSecond = decodeMs !== null && isPositive(decodeMs) && isPositive(outputTokens)
+        ? outputTokens / (decodeMs / MS_PER_SECOND)
+        : null;
+    return {
+        prefillTokensPerSecond,
+        decodeTokensPerSecond,
+        combinedTokensPerSecond,
+        ttftMs,
+        decodeMs: decodeMs !== null && isPositive(decodeMs) ? decodeMs : null,
+    };
+}
+/**
+ * Validate the raw counter shape before differencing. The wire is hand-synced
+ * between the agent and the device client, so a malformed frame should be
+ * rejected at the boundary (and the metric dropped) rather than silently
+ * producing garbage throughput.
+ */
+export function isGenerationCounters(value) {
+    if (!value || typeof value !== "object")
+        return false;
+    const v = value;
+    return (isNonNegativeCount(v.promptTokens) &&
+        isNonNegativeCount(v.outputTokens) &&
+        isNonNegativeCount(v.durationMs) &&
+        (v.ttftMs == null || isNonNegativeCount(v.ttftMs)));
+}
+//# sourceMappingURL=throughput.js.map

package/local-inference/throughput.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"throughput.js","sourceRoot":"","sources":["../../src/local-inference/throughput.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAoCH,MAAM,aAAa,GAAG,IAAI,CAAC;AAE3B,SAAS,UAAU,CAAC,KAAgC;IAClD,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC;AAC1E,CAAC;AAED,SAAS,kBAAkB,CAAC,KAAgC;IAC1D,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;AAC3E,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,2BAA2B,CACzC,QAA4B;IAE5B,MAAM,EAAE,YAAY,EAAE,YAAY,EAAE,UAAU,EAAE,GAAG,QAAQ,CAAC;IAC5D,MAAM,MAAM,GACV,UAAU,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,QAAQ,CAAC,MAAM,GAAG,UAAU;QACzD,CAAC,CAAC,QAAQ,CAAC,MAAM;QACjB,CAAC,CAAC,IAAI,CAAC;IAEX,2EAA2E;IAC3E,4EAA4E;IAC5E,MAAM,uBAAuB,GAC3B,UAAU,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,YAAY,CAAC;QAChD,CAAC,CAAC,YAAY,GAAG,CAAC,UAAU,GAAG,aAAa,CAAC;QAC7C,CAAC,CAAC,IAAI,CAAC;IAEX,gEAAgE;IAChE,MAAM,sBAAsB,GAC1B,MAAM,KAAK,IAAI,IAAI,UAAU,CAAC,YAAY,CAAC;QACzC,CAAC,CAAC,YAAY,GAAG,CAAC,MAAM,GAAG,aAAa,CAAC;QACzC,CAAC,CAAC,IAAI,CAAC;IAEX,6EAA6E;IAC7E,mDAAmD;IACnD,MAAM,QAAQ,GAAG,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,UAAU,GAAG,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC;IAC9D,MAAM,qBAAqB,GACzB,QAAQ,KAAK,IAAI,IAAI,UAAU,CAAC,QAAQ,CAAC,IAAI,UAAU,CAAC,YAAY,CAAC;QACnE,CAAC,CAAC,YAAY,GAAG,CAAC,QAAQ,GAAG,aAAa,CAAC;QAC3C,CAAC,CAAC,IAAI,CAAC;IAEX,OAAO;QACL,sBAAsB;QACtB,qBAAqB;QACrB,uBAAuB;QACvB,MAAM;QACN,QAAQ,EAAE,QAAQ,KAAK,IAAI,IAAI,UAAU,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI;KACtE,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,oBAAoB,CAClC,KAAc;IAEd,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IACtD,MAAM,CAAC,GAAG,KAAgC,CAAC;IAC3C,OAAO,CACL,kBAAkB,CAAC,CAAC,CAAC,YAAsB,CAAC;QAC5C,kBAAkB,CAAC,CAAC,CAAC,YAAsB,CAAC;QAC5C,kBAAkB,CAAC,CAAC,CAAC,UAAoB,CAAC;QAC1C,CAAC,CAAC,CAAC,MAAM,IAAI,IAAI,IAAI,kBAAkB,CAAC,CAAC,CAAC,MAAgB,CAAC,CAAC,CAC7D,CAAC;AACJ,CAAC"}