@elizaos/agent 2.0.0-alpha.520 → 2.0.0-alpha.521

package/packages/app-core/src/services/local-inference/handler-registry.js

@@ -1,160 +0,0 @@
-/**
- * Side-registry of model handlers registered on an AgentRuntime.
- *
- * The elizaOS core exposes `runtime.registerModel(type, handler, provider,
- * priority)` but no way to list who registered what. This module intercepts
- * `registerModel` at runtime to record every registration in a Map keyed by
- * model type, plus fires status listeners so the UI can render a live
- * [ModelType × Provider] routing table.
- *
- * Because we monkey-patch `registerModel` we also keep the original
- * handler reference — the router-handler (see `router-handler.ts`) uses
- * this to dispatch inference calls by policy without going through
- * `runtime.useModel` (which would loop back to us and recurse).
- */
-import { AgentRuntime } from "@elizaos/core";
-class HandlerRegistry {
-    registrations = new Map();
-    listeners = new Set();
-    installedOn = new WeakSet();
-    /**
-     * Snapshot of all registrations grouped by model type, sorted by
-     * priority descending inside each group (matches core's selection
-     * order). Callers must not mutate the returned array.
-     */
-    getAll() {
-        const out = [];
-        for (const list of this.registrations.values()) {
-            out.push(...list);
-        }
-        return out;
-    }
-    /** All registrations for a given model type, sorted by priority desc. */
-    getForType(modelType) {
-        const list = this.registrations.get(modelType);
-        return list ? [...list] : [];
-    }
-    /**
-     * Registrations excluding a specific provider. Used by the router-handler
-     * to find "all providers except me" when dispatching.
-     */
-    getForTypeExcluding(modelType, excludeProvider) {
-        return this.getForType(modelType).filter((r) => r.provider !== excludeProvider);
-    }
-    subscribe(listener) {
-        this.listeners.add(listener);
-        return () => {
-            this.listeners.delete(listener);
-        };
-    }
-    emit() {
-        const snapshot = this.getAll();
-        for (const listener of this.listeners) {
-            try {
-                listener(snapshot);
-            }
-            catch {
-                this.listeners.delete(listener);
-            }
-        }
-    }
-    record(reg) {
-        const existing = this.registrations.get(reg.modelType) ?? [];
-        // Replace any prior registration from the same provider for this
-        // model type. Core allows multiple providers per type but only one
-        // registration per (type, provider) pair — last write wins.
-        const filtered = existing.filter((r) => r.provider !== reg.provider);
-        filtered.push(reg);
-        filtered.sort((a, b) => b.priority - a.priority);
-        this.registrations.set(reg.modelType, filtered);
-        this.emit();
-    }
-    /**
-     * Install the interception on a runtime. Idempotent per runtime instance.
-     * For most boot paths the prototype-level patch below already covers the
-     * runtime before any plugin registers; this method is the belt-and-braces
-     * fallback for runtimes constructed before the patch ran.
-     */
-    installOn(runtime) {
-        installPrototypePatch();
-        const rt = runtime;
-        if (typeof rt.registerModel !== "function")
-            return;
-        if (this.installedOn.has(rt))
-            return;
-        this.installedOn.add(rt);
-        // If the runtime already inherited the patched prototype method the
-        // prototype handles every call. Nothing to do per-instance.
-        const protoMethod = Object.getPrototypeOf(rt)?.registerModel;
-        if (protoMethod &&
-            protoMethod[PATCH_MARK]) {
-            return;
-        }
-        // Per-instance wrap only for legacy runtimes whose prototype pre-dates
-        // our prototype patch (shouldn't happen in practice).
-        const original = rt.registerModel.bind(runtime);
-        rt.registerModel = ((modelType, handler, provider, priority) => {
-            this.record({
-                modelType: String(modelType),
-                provider: String(provider),
-                priority: typeof priority === "number" ? priority : 0,
-                registeredAt: new Date().toISOString(),
-                handler,
-            });
-            return original(modelType, handler, provider, priority);
-        });
-    }
-    /** Exposed so the prototype patch can record through the singleton. */
-    recordFromPrototype(reg) {
-        this.record(reg);
-    }
-}
-const PATCH_MARK = Symbol.for("milady.local-inference.registerModel.patched");
-let prototypePatched = false;
-/**
- * One-shot patch of `AgentRuntime.prototype.registerModel` so every runtime
- * instance — including ones constructed later in boot — records through
- * the singleton handler registry. Idempotent.
- */
-function installPrototypePatch() {
-    if (prototypePatched)
-        return;
-    const proto = AgentRuntime.prototype;
-    const original = proto.registerModel;
-    if (typeof original !== "function")
-        return;
-    if (original[PATCH_MARK]) {
-        prototypePatched = true;
-        return;
-    }
-    const patched = function patchedRegisterModel(modelType, handler, provider, priority) {
-        try {
-            handlerRegistry.recordFromPrototype({
-                modelType: String(modelType),
-                provider: String(provider),
-                priority: typeof priority === "number" ? priority : 0,
-                registeredAt: new Date().toISOString(),
-                handler,
-            });
-        }
-        catch {
-            // Never let registry bookkeeping break the registration path.
-        }
-        original.call(this, modelType, handler, provider, priority);
-    };
-    patched[PATCH_MARK] = true;
-    proto.registerModel = patched;
-    prototypePatched = true;
-}
-// Install at module-import time. This is a process-wide side effect but a
-// benign one: the patch is idempotent and forwards to the original.
-installPrototypePatch();
-export const handlerRegistry = new HandlerRegistry();
-export function toPublicRegistration(reg) {
-    return {
-        modelType: reg.modelType,
-        provider: reg.provider,
-        priority: reg.priority,
-        registeredAt: reg.registeredAt,
-    };
-}

package/packages/app-core/src/services/local-inference/hardware.d.ts

@@ -1,26 +0,0 @@
-/**
- * Hardware probe for local inference sizing.
- *
- * Uses `node-llama-cpp` when available to read GPU backend + VRAM. Falls back
- * to Node's `os` module when the binding isn't installed — we don't require
- * the plugin to be loaded for the probe endpoint to return useful data.
- *
- * Dynamic import is intentional: the binding pulls a native prebuilt that we
- * don't want eagerly required at module-load time (breaks CI environments
- * without the trusted-dependency flag).
- */
-import type { HardwareProbe } from "./types";
-/**
- * Read current system + GPU state. Cheap enough to call per-request; no
- * internal caching so the UI always reflects live VRAM usage.
- */
-export declare function probeHardware(): Promise<HardwareProbe>;
-/**
- * Compatibility assessment for a specific model given current hardware.
- *
- * Green/fits: comfortable headroom (model < 70% of effective memory).
- * Yellow/tight: will run but may swap or stutter under load.
- * Red/wontfit: exceeds available memory.
- */
-export declare function assessFit(probe: HardwareProbe, modelSizeGb: number, minRamGb: number): "fits" | "tight" | "wontfit";
-//# sourceMappingURL=hardware.d.ts.map

package/packages/app-core/src/services/local-inference/hardware.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"hardware.d.ts","sourceRoot":"","sources":["../../../../../../../app-core/src/services/local-inference/hardware.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,KAAK,EAAE,aAAa,EAAe,MAAM,SAAS,CAAC;AA+D1D;;;GAGG;AACH,wBAAsB,aAAa,IAAI,OAAO,CAAC,aAAa,CAAC,CAiE5D;AAED;;;;;;GAMG;AACH,wBAAgB,SAAS,CACvB,KAAK,EAAE,aAAa,EACpB,WAAW,EAAE,MAAM,EACnB,QAAQ,EAAE,MAAM,GACf,MAAM,GAAG,OAAO,GAAG,SAAS,CAW9B"}

package/packages/app-core/src/services/local-inference/hardware.js

@@ -1,139 +0,0 @@
-/**
- * Hardware probe for local inference sizing.
- *
- * Uses `node-llama-cpp` when available to read GPU backend + VRAM. Falls back
- * to Node's `os` module when the binding isn't installed — we don't require
- * the plugin to be loaded for the probe endpoint to return useful data.
- *
- * Dynamic import is intentional: the binding pulls a native prebuilt that we
- * don't want eagerly required at module-load time (breaks CI environments
- * without the trusted-dependency flag).
- */
-import os from "node:os";
-const BYTES_PER_GB = 1024 ** 3;
-function bytesToGb(bytes) {
-    return Math.round((bytes / BYTES_PER_GB) * 10) / 10;
-}
-/**
- * Pick a default bucket based on total available memory and architecture.
- *
- * On Apple Silicon the GPU shares system RAM, so shared memory acts as VRAM.
- * On discrete-GPU x86 boxes we weight VRAM higher than system RAM.
- */
-function recommendBucket(totalRamGb, vramGb, appleSilicon) {
-    const effective = appleSilicon
-        ? totalRamGb
-        : vramGb > 0
-            ? Math.max(vramGb * 1.25, totalRamGb * 0.5)
-            : totalRamGb * 0.5;
-    if (effective >= 36)
-        return "xl";
-    if (effective >= 18)
-        return "large";
-    if (effective >= 9)
-        return "mid";
-    return "small";
-}
-async function loadLlamaBinding() {
-    try {
-        const mod = (await import("node-llama-cpp"));
-        if (mod &&
-            typeof mod === "object" &&
-            "getLlama" in mod &&
-            typeof mod.getLlama === "function") {
-            return mod;
-        }
-        return null;
-    }
-    catch {
-        // Binding not installed or prebuilt missing. That's an expected case when
-        // the local-ai plugin is not enabled; we return null so the probe falls
-        // back to OS-level detection.
-        return null;
-    }
-}
-/**
- * Read current system + GPU state. Cheap enough to call per-request; no
- * internal caching so the UI always reflects live VRAM usage.
- */
-export async function probeHardware() {
-    const totalRamBytes = os.totalmem();
-    const freeRamBytes = os.freemem();
-    const cpuCores = os.cpus().length;
-    const platform = process.platform;
-    const arch = process.arch;
-    const appleSilicon = platform === "darwin" && arch === "arm64";
-    const binding = await loadLlamaBinding();
-    if (!binding) {
-        // OS-only fallback: we cannot detect GPU without the binding, so treat as
-        // CPU-only. On Apple Silicon shared memory still makes mid-sized models
-        // viable, which `recommendBucket` handles.
-        const totalRamGb = bytesToGb(totalRamBytes);
-        return {
-            totalRamGb,
-            freeRamGb: bytesToGb(freeRamBytes),
-            gpu: null,
-            cpuCores,
-            platform,
-            arch,
-            appleSilicon,
-            recommendedBucket: recommendBucket(totalRamGb, 0, appleSilicon),
-            source: "os-fallback",
-        };
-    }
-    const llama = await binding.getLlama({ gpu: "auto" });
-    const totalRamGb = bytesToGb(totalRamBytes);
-    const freeRamGb = bytesToGb(freeRamBytes);
-    if (llama.gpu === false) {
-        return {
-            totalRamGb,
-            freeRamGb,
-            gpu: null,
-            cpuCores,
-            platform,
-            arch,
-            appleSilicon,
-            recommendedBucket: recommendBucket(totalRamGb, 0, appleSilicon),
-            source: "node-llama-cpp",
-        };
-    }
-    const vram = await llama.getVramState();
-    const totalVramGb = bytesToGb(vram.total);
-    const freeVramGb = bytesToGb(vram.free);
-    return {
-        totalRamGb,
-        freeRamGb,
-        gpu: {
-            backend: llama.gpu,
-            totalVramGb,
-            freeVramGb,
-        },
-        cpuCores,
-        platform,
-        arch,
-        appleSilicon,
-        recommendedBucket: recommendBucket(totalRamGb, totalVramGb, appleSilicon),
-        source: "node-llama-cpp",
-    };
-}
-/**
- * Compatibility assessment for a specific model given current hardware.
- *
- * Green/fits: comfortable headroom (model < 70% of effective memory).
- * Yellow/tight: will run but may swap or stutter under load.
- * Red/wontfit: exceeds available memory.
- */
-export function assessFit(probe, modelSizeGb, minRamGb) {
-    const effectiveGb = probe.appleSilicon
-        ? probe.totalRamGb
-        : probe.gpu
-            ? Math.max(probe.gpu.totalVramGb, probe.totalRamGb * 0.5)
-            : probe.totalRamGb * 0.5;
-    if (effectiveGb < minRamGb)
-        return "wontfit";
-    if (modelSizeGb > effectiveGb * 0.9)
-        return "wontfit";
-    if (modelSizeGb > effectiveGb * 0.7)
-        return "tight";
-    return "fits";
-}

package/packages/app-core/src/services/local-inference/paths.d.ts

@@ -1,24 +0,0 @@
-/**
- * Path resolution for the local-inference service.
- *
- * All Milady-owned files live under `$STATE_DIR/local-inference/` to match
- * the convention established by `plugin-installer.ts` and the rest of
- * app-core. We never write to paths outside of this root.
- *
- * The state dir is resolved in `MILADY_STATE_DIR` → `ELIZA_STATE_DIR` →
- * `~/.eliza` order. The `.eliza` fallback is preserved for desktop
- * backward-compat with existing installs; on AOSP `MILADY_STATE_DIR` is
- * set by `MiladyAgentService.java` to `/data/data/<pkg>/files/.milady`,
- * so models land at `<that>/local-inference/models/` and not under a
- * stray homedir-derived path.
- */
-export declare function localInferenceRoot(): string;
-/** Directory for models Milady downloaded itself. Safe to delete. */
-export declare function miladyModelsDir(): string;
-/** JSON file tracking installed-model metadata (downloaded + discovered). */
-export declare function registryPath(): string;
-/** Partial-download staging directory; files here are resume candidates. */
-export declare function downloadsStagingDir(): string;
-/** True when `target` is inside Milady's local-inference root. */
-export declare function isWithinMiladyRoot(target: string): boolean;
-//# sourceMappingURL=paths.d.ts.map

package/packages/app-core/src/services/local-inference/paths.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"paths.d.ts","sourceRoot":"","sources":["../../../../../../../app-core/src/services/local-inference/paths.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAKH,wBAAgB,kBAAkB,IAAI,MAAM,CAK3C;AAED,qEAAqE;AACrE,wBAAgB,eAAe,IAAI,MAAM,CAExC;AAED,6EAA6E;AAC7E,wBAAgB,YAAY,IAAI,MAAM,CAErC;AAED,4EAA4E;AAC5E,wBAAgB,mBAAmB,IAAI,MAAM,CAE5C;AAED,kEAAkE;AAClE,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAK1D"}

package/packages/app-core/src/services/local-inference/paths.js

@@ -1,41 +0,0 @@
-/**
- * Path resolution for the local-inference service.
- *
- * All Milady-owned files live under `$STATE_DIR/local-inference/` to match
- * the convention established by `plugin-installer.ts` and the rest of
- * app-core. We never write to paths outside of this root.
- *
- * The state dir is resolved in `MILADY_STATE_DIR` → `ELIZA_STATE_DIR` →
- * `~/.eliza` order. The `.eliza` fallback is preserved for desktop
- * backward-compat with existing installs; on AOSP `MILADY_STATE_DIR` is
- * set by `MiladyAgentService.java` to `/data/data/<pkg>/files/.milady`,
- * so models land at `<that>/local-inference/models/` and not under a
- * stray homedir-derived path.
- */
-import os from "node:os";
-import path from "node:path";
-export function localInferenceRoot() {
-    const stateDir = process.env.MILADY_STATE_DIR?.trim() || process.env.ELIZA_STATE_DIR?.trim();
-    const base = stateDir || path.join(os.homedir(), ".eliza");
-    return path.join(base, "local-inference");
-}
-/** Directory for models Milady downloaded itself. Safe to delete. */
-export function miladyModelsDir() {
-    return path.join(localInferenceRoot(), "models");
-}
-/** JSON file tracking installed-model metadata (downloaded + discovered). */
-export function registryPath() {
-    return path.join(localInferenceRoot(), "registry.json");
-}
-/** Partial-download staging directory; files here are resume candidates. */
-export function downloadsStagingDir() {
-    return path.join(localInferenceRoot(), "downloads");
-}
-/** True when `target` is inside Milady's local-inference root. */
-export function isWithinMiladyRoot(target) {
-    const root = path.resolve(localInferenceRoot());
-    const resolved = path.resolve(target);
-    if (resolved === root)
-        return false;
-    return resolved.startsWith(`${root}${path.sep}`);
-}

package/packages/app-core/src/services/local-inference/providers.d.ts

@@ -1,61 +0,0 @@
-/**
- * Provider registry.
- *
- * Treats every inference source the same way — cloud subscription, cloud
- * API, local llama.cpp engine, paired-device bridge, Capacitor on-device
- * — each is a `ProviderDefinition` with an `id`, a human label, a set of
- * supported model slots, and a pluggable `getEnableState()` that inspects
- * whatever underlying gate controls it (API key presence, subscription
- * status, env flag, file on disk).
- *
- * The cloud-provider status readers are intentionally permissive: they
- * report what they can introspect without depending on the specific
- * cloud-plugin internals, and hand off to the existing ProviderSwitcher
- * UI for actual enable/disable via `configureHref`. That avoids the
- * "combined enable matrix is an architectural project" problem by making
- * configuration navigable rather than centralised.
- */
-import type { AgentModelSlot } from "./types";
-export type ProviderId = "milady-local-inference" | "milady-device-bridge" | "capacitor-llama" | "anthropic-subscription" | "openai-codex" | "anthropic" | "openai" | "grok" | "elizacloud" | "google" | "mistral";
-export interface ProviderEnableState {
-    enabled: boolean;
-    /** Short reason, e.g. "API key set", "Device connected", "No API key". */
-    reason: string;
-}
-export interface ProviderDefinition {
-    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[];
-    /**
-     * Read the current enable state. For cloud providers we inspect env
-     * vars or config fragments; for local we check file presence; for
-     * device-bridge we check connected-device count.
-     */
-    getEnableState(): Promise<ProviderEnableState>;
-    /**
-     * 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;
-}
-/** Resolve which slots have at least one registered handler from this provider. */
-export declare function getRegisteredSlotsForProvider(providerId: string): string[];
-export declare const BUILT_IN_PROVIDERS: readonly ProviderDefinition[];
-export interface ProviderStatus {
-    id: ProviderId;
-    label: string;
-    kind: ProviderDefinition["kind"];
-    description: string;
-    supportedSlots: AgentModelSlot[];
-    configureHref: string | null;
-    enableState: ProviderEnableState;
-    /** Registered model types this provider has handlers for, right now. */
-    registeredSlots: string[];
-}
-export declare function snapshotProviders(): Promise<ProviderStatus[]>;
-//# sourceMappingURL=providers.d.ts.map

package/packages/app-core/src/services/local-inference/providers.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"providers.d.ts","sourceRoot":"","sources":["../../../../../../../app-core/src/services/local-inference/providers.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAOH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,SAAS,CAAC;AAE9C,MAAM,MAAM,UAAU,GAClB,wBAAwB,GACxB,sBAAsB,GACtB,iBAAiB,GACjB,wBAAwB,GACxB,cAAc,GACd,WAAW,GACX,QAAQ,GACR,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,MAAM,WAAW,kBAAkB;IACjC,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,cAAc,IAAI,OAAO,CAAC,mBAAmB,CAAC,CAAC;IAC/C;;;;OAIG;IACH,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAED,mFAAmF;AACnF,wBAAgB,6BAA6B,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,EAAE,CAO1E;AAiOD,eAAO,MAAM,kBAAkB,EAAE,SAAS,kBAAkB,EAY3D,CAAC;AAqBF,MAAM,WAAW,cAAc;IAC7B,EAAE,EAAE,UAAU,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,kBAAkB,CAAC,MAAM,CAAC,CAAC;IACjC,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;AAED,wBAAsB,iBAAiB,IAAI,OAAO,CAAC,cAAc,EAAE,CAAC,CAiBnE"}