@elizaos/app-core 2.0.0-beta.3 → 2.0.3-beta.3

package/services/local-inference/router-handler.js

@@ -1,165 +0,0 @@
-/**
- * Top-priority router handler.
- *
- * Registers a model handler for every `AgentModelSlot` at priority
- * `Number.MAX_SAFE_INTEGER`, which guarantees the runtime dispatches to
- * us first. At dispatch time we:
- *
- *   1. Read the user's per-slot policy + preferred-provider choice from
- *      `routing-preferences.ts`.
- *   2. Ask the `policyEngine` to pick a provider from the handler
- *      registry's current set (excluding ourselves).
- *   3. Invoke that provider's original handler directly — bypassing
- *      `runtime.useModel` which would recurse into us.
- *   4. Record the observed latency so future "fastest" picks have data.
- *
- * If no other handler exists we throw a clear error rather than return
- * garbage — the caller is meant to see "no provider configured" so they
- * know to set one up.
- *
- * Because the router sits at the top of the priority stack, the user's
- * preference is always authoritative regardless of what plugins register
- * at lower priorities. This is the mechanism that unifies cloud + local
- * + device-bridge routing from one settings panel.
- */
-import { logger, ModelType } from "@elizaos/core";
-import { readEffectiveAssignments } from "./assignments";
-import { localInferenceEngine } from "./engine";
-import { handlerRegistry } from "./handler-registry";
-import { policyEngine } from "./routing-policy";
-import { DEFAULT_ROUTING_POLICY, readRoutingPreferences, } from "./routing-preferences";
-import { AGENT_MODEL_SLOTS } from "./types";
-export const ROUTER_PROVIDER = "eliza-router";
-/**
- * Max safe integer keeps us at the top even if a plugin registers with
- * a very high priority. If someone deliberately wants to outrank us,
- * they can register with Infinity — unlikely in practice.
- */
-const ROUTER_PRIORITY = Number.MAX_SAFE_INTEGER;
-function slotToModelType(slot) {
-    switch (slot) {
-        case "TEXT_SMALL":
-            return ModelType.TEXT_SMALL;
-        case "TEXT_LARGE":
-            return ModelType.TEXT_LARGE;
-        case "TEXT_EMBEDDING":
-            return ModelType.TEXT_EMBEDDING;
-    }
-}
-function modelTypeToSlot(modelType) {
-    for (const slot of AGENT_MODEL_SLOTS) {
-        if (slotToModelType(slot) === modelType)
-            return slot;
-    }
-    return null;
-}
-function shouldForceLocalInference(policy, preferredProvider) {
-    return policy === "manual" && preferredProvider === "eliza-local-inference";
-}
-export function filterUnavailableLocalInferenceCandidates(candidates, localInferenceAvailable, forceLocalInference) {
-    if (forceLocalInference || localInferenceAvailable) {
-        return candidates;
-    }
-    return candidates.filter((candidate) => candidate.provider !== "eliza-local-inference");
-}
-export async function filterUnavailableLocalInference(slot, policy, preferredProvider, candidates) {
-    const hasLocalInference = candidates.some((candidate) => candidate.provider === "eliza-local-inference");
-    if (!hasLocalInference) {
-        return candidates;
-    }
-    const assignments = await readEffectiveAssignments();
-    return filterUnavailableLocalInferenceCandidates(candidates, Boolean(assignments[slot]) || localInferenceEngine.hasLoadedModel(), shouldForceLocalInference(policy, preferredProvider));
-}
-function makeRouterHandler(slot) {
-    return async (runtime, params) => {
-        const modelType = slotToModelType(slot);
-        if (!modelType) {
-            throw new Error(`[router] Unknown agent slot: ${slot}`);
-        }
-        // Read the user's policy for this slot. Absent = local-first fallback.
-        const prefs = await readRoutingPreferences();
-        const policy = prefs.policy[slot] ?? DEFAULT_ROUTING_POLICY;
-        const preferred = prefs.preferredProvider[slot] ?? null;
-        // Ask the policy engine which handler to dispatch to. For automatic
-        // policies, honor the documented fallback behaviour: if the selected
-        // provider throws, try the next eligible provider instead of surfacing a
-        // local/model-specific failure while cloud providers are available.
-        const candidates = await filterUnavailableLocalInference(slot, policy, preferred, handlerRegistry.getForTypeExcluding(modelType, ROUTER_PROVIDER));
-        const failedProviders = new Set();
-        let lastError = null;
-        while (true) {
-            const remaining = candidates.filter((candidate) => !failedProviders.has(candidate.provider));
-            const pick = policyEngine.pickProvider({
-                modelType,
-                policy,
-                preferredProvider: preferred,
-                candidates: remaining,
-                selfProvider: ROUTER_PROVIDER,
-            });
-            if (!pick) {
-                if (lastError) {
-                    throw lastError;
-                }
-                throw new Error(`[router] No provider registered for ${slot}. Configure a cloud provider, enable local inference, or pair a device.`);
-            }
-            policyEngine.recordPick(pick.provider, modelType);
-            const start = Date.now();
-            try {
-                const result = await pick.handler(runtime, params);
-                policyEngine.recordLatency(pick.provider, modelType, Date.now() - start);
-                return result;
-            }
-            catch (err) {
-                // Record the timing even on failure so "fastest" doesn't silently
-                // prefer providers that error out fast.
-                policyEngine.recordLatency(pick.provider, modelType, Date.now() - start);
-                const manualPreferred = policy === "manual" &&
-                    preferred !== null &&
-                    pick.provider === preferred;
-                const hasAlternative = remaining.some((candidate) => candidate.provider !== pick.provider);
-                if (manualPreferred || !hasAlternative) {
-                    throw err;
-                }
-                failedProviders.add(pick.provider);
-                lastError = err;
-                logger.warn(`[router] Provider ${pick.provider} failed for ${slot}; trying fallback provider (${err instanceof Error ? err.message : String(err)})`);
-            }
-        }
-    };
-}
-/**
- * Install the router as the top-priority handler for every slot.
- *
- * Idempotent per-runtime via the handler-registry's "last write wins"
- * behaviour — re-registering our handlers just refreshes them in place.
- * Called from `ensure-local-inference-handler.ts` after `handlerRegistry`
- * has been installed on the runtime.
- */
-export function installRouterHandler(runtime) {
-    const rt = runtime;
-    if (typeof rt.registerModel !== "function")
-        return;
-    for (const slot of AGENT_MODEL_SLOTS) {
-        const modelType = slotToModelType(slot);
-        if (!modelType)
-            continue;
-        rt.registerModel(modelType, makeRouterHandler(slot), ROUTER_PROVIDER, ROUTER_PRIORITY);
-    }
-}
-/** Public helper — useful for diagnostics endpoints. */
-export function describeCurrentRouting() {
-    const out = [];
-    for (const slot of AGENT_MODEL_SLOTS) {
-        const modelType = slotToModelType(slot);
-        if (!modelType)
-            continue;
-        const candidates = handlerRegistry
-            .getForTypeExcluding(modelType, ROUTER_PROVIDER)
-            .map((c) => ({ provider: c.provider, priority: c.priority }));
-        out.push({ slot, modelType, candidates });
-    }
-    return out;
-}
-// Re-export so the handler registry can tell whether it's looking at a
-// recursive router registration when filtering.
-export { modelTypeToSlot };

package/services/local-inference/routing-policy.d.ts

@@ -1,55 +0,0 @@
-/**
- * Policy engine + latency + cost tracking for cross-provider routing.
- *
- * The policy engine sits on top of the `handlerRegistry` and, given a
- * model type and a user-selected policy, decides which provider's handler
- * should serve the next request. The router-handler (registered at top
- * priority) calls `pickProvider` to make that decision.
- *
- * Policies:
- *   - manual       — honour `preferredProvider`; when no pref set, fall
- *                    through to the runtime's native priority order
- *                    (highest registered priority wins).
- *   - cheapest     — pick the provider with the lowest per-token cost.
- *   - fastest      — pick the provider with the lowest tracked p50 latency
- *                    (needs at least a few samples; falls back to native).
- *   - prefer-local — try local first; if it fails or has no handler,
- *                    fall through to the next-best non-local.
- *   - round-robin  — distribute load evenly across eligible providers.
- *
- * Latency is tracked in a ring buffer per provider per model type. Cost
- * is a static table of published per-million-token rates; local providers
- * are $0. Neither is exact — the goal is "good enough to discriminate"
- * rather than dollar-accurate billing.
- */
-import type { HandlerRegistration } from "./handler-registry";
-import type { RoutingPolicy } from "./routing-preferences";
-declare class PolicyEngine {
-    private stats;
-    private statsFor;
-    recordLatency(provider: string, modelType: string, durationMs: number): void;
-    recordPick(provider: string, modelType: string): void;
-    p50(provider: string, modelType: string): number | null;
-    lastPicked(provider: string, modelType: string): number | null;
-    costOf(provider: string): number | null;
-    /**
-     * Pick a provider for this (modelType, policy) given the registry.
-     * Returns the HandlerRegistration whose handler the router-handler
-     * should dispatch to, or null if no eligible handler exists.
-     *
-     * `preferredProvider` is only honoured for policy === "manual".
-     */
-    pickProvider(args: {
-        modelType: string;
-        policy: RoutingPolicy;
-        preferredProvider: string | null;
-        candidates: HandlerRegistration[];
-        /** Provider ID of the router itself — always excluded from candidates. */
-        selfProvider: string;
-    }): HandlerRegistration | null;
-    /** For tests and diagnostics. */
-    snapshot(): Record<string, Record<string, number | null>>;
-}
-export declare const policyEngine: PolicyEngine;
-export {};
-//# sourceMappingURL=routing-policy.d.ts.map

package/services/local-inference/routing-policy.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"routing-policy.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/routing-policy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAEH,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAC9D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AA4D3D,cAAM,YAAY;IAChB,OAAO,CAAC,KAAK,CAAmD;IAEhE,OAAO,CAAC,QAAQ;IAShB,aAAa,CAAC,QAAQ,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,IAAI;IAU5E,UAAU,CAAC,QAAQ,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,GAAG,IAAI;IAIrD,GAAG,CAAC,QAAQ,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI;IAIvD,UAAU,CAAC,QAAQ,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI;IAI9D,MAAM,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI;IAQvC;;;;;;OAMG;IACH,YAAY,CAAC,IAAI,EAAE;QACjB,SAAS,EAAE,MAAM,CAAC;QAClB,MAAM,EAAE,aAAa,CAAC;QACtB,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;QACjC,UAAU,EAAE,mBAAmB,EAAE,CAAC;QAClC,0EAA0E;QAC1E,YAAY,EAAE,MAAM,CAAC;KACtB,GAAG,mBAAmB,GAAG,IAAI;IAsE9B,iCAAiC;IACjC,QAAQ,IAAI,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC,CAAC;CAW1D;AAED,eAAO,MAAM,YAAY,cAAqB,CAAC"}

package/services/local-inference/routing-policy.js

@@ -1,195 +0,0 @@
-/**
- * Policy engine + latency + cost tracking for cross-provider routing.
- *
- * The policy engine sits on top of the `handlerRegistry` and, given a
- * model type and a user-selected policy, decides which provider's handler
- * should serve the next request. The router-handler (registered at top
- * priority) calls `pickProvider` to make that decision.
- *
- * Policies:
- *   - manual       — honour `preferredProvider`; when no pref set, fall
- *                    through to the runtime's native priority order
- *                    (highest registered priority wins).
- *   - cheapest     — pick the provider with the lowest per-token cost.
- *   - fastest      — pick the provider with the lowest tracked p50 latency
- *                    (needs at least a few samples; falls back to native).
- *   - prefer-local — try local first; if it fails or has no handler,
- *                    fall through to the next-best non-local.
- *   - round-robin  — distribute load evenly across eligible providers.
- *
- * Latency is tracked in a ring buffer per provider per model type. Cost
- * is a static table of published per-million-token rates; local providers
- * are $0. Neither is exact — the goal is "good enough to discriminate"
- * rather than dollar-accurate billing.
- */
-const RING_SIZE = 32;
-class RingBuffer {
-    buf = [];
-    push(sample) {
-        this.buf.push(sample);
-        if (this.buf.length > RING_SIZE)
-            this.buf.shift();
-    }
-    p50() {
-        if (this.buf.length === 0)
-            return null;
-        const sorted = [...this.buf].map((s) => s.durationMs).sort((a, b) => a - b);
-        return sorted[Math.floor(sorted.length / 2)] ?? null;
-    }
-    size() {
-        return this.buf.length;
-    }
-}
-/**
- * Relative per-million-token costs. Keep conservative: the policy only
- * needs the order to be right for product defaults.
- * Local / device-bridge = 0 because the user already paid for the hardware.
- * Subscriptions get a small marginal cost, direct APIs sit above that,
- * and Eliza Cloud is last because managed fallback is the most expensive
- * path for the user.
- */
-const COST_PER_MILLION_TOKENS = {
-    "eliza-local-inference": { input: 0, output: 0 },
-    "eliza-device-bridge": { input: 0, output: 0 },
-    "capacitor-llama": { input: 0, output: 0 },
-    "anthropic-subscription": { input: 0.1, output: 0.1 },
-    "openai-codex": { input: 0.1, output: 0.1 },
-    "openai-subscription": { input: 0.1, output: 0.1 },
-    anthropic: { input: 3, output: 15 },
-    openai: { input: 2.5, output: 10 },
-    grok: { input: 5, output: 15 },
-    google: { input: 1.25, output: 5 },
-    "google-genai": { input: 1.25, output: 5 },
-    moonshot: { input: 1.25, output: 5 },
-    kimi: { input: 1.25, output: 5 },
-    zai: { input: 1.25, output: 5 },
-    glm: { input: 1.25, output: 5 },
-    mistral: { input: 2, output: 6 },
-    elizacloud: { input: 30, output: 60 },
-};
-class PolicyEngine {
-    stats = new Map();
-    statsFor(provider) {
-        let s = this.stats.get(provider);
-        if (!s) {
-            s = { latency: new Map(), lastPicked: new Map() };
-            this.stats.set(provider, s);
-        }
-        return s;
-    }
-    recordLatency(provider, modelType, durationMs) {
-        const s = this.statsFor(provider);
-        let buf = s.latency.get(modelType);
-        if (!buf) {
-            buf = new RingBuffer();
-            s.latency.set(modelType, buf);
-        }
-        buf.push({ durationMs, at: Date.now() });
-    }
-    recordPick(provider, modelType) {
-        this.statsFor(provider).lastPicked.set(modelType, Date.now());
-    }
-    p50(provider, modelType) {
-        return this.statsFor(provider).latency.get(modelType)?.p50() ?? null;
-    }
-    lastPicked(provider, modelType) {
-        return this.statsFor(provider).lastPicked.get(modelType) ?? null;
-    }
-    costOf(provider) {
-        const c = COST_PER_MILLION_TOKENS[provider];
-        if (!c)
-            return null;
-        // Weighted sum (3:1 output:input is a typical chat ratio). Treat missing
-        // output pricing as same as input.
-        return c.input * 0.25 + c.output * 0.75;
-    }
-    /**
-     * Pick a provider for this (modelType, policy) given the registry.
-     * Returns the HandlerRegistration whose handler the router-handler
-     * should dispatch to, or null if no eligible handler exists.
-     *
-     * `preferredProvider` is only honoured for policy === "manual".
-     */
-    pickProvider(args) {
-        const eligible = args.candidates
-            .filter((c) => c.provider !== args.selfProvider)
-            .slice()
-            // Defensive sort — real callers already sort, but test fixtures and
-            // non-registry callers might not, and a silent "pick-wrong" would be
-            // worse than the extra O(n log n).
-            .sort((a, b) => b.priority - a.priority);
-        if (eligible.length === 0)
-            return null;
-        switch (args.policy) {
-            case "manual": {
-                if (args.preferredProvider) {
-                    const match = eligible.find((c) => c.provider === args.preferredProvider);
-                    if (match)
-                        return match;
-                }
-                // Fallback: highest native priority.
-                return eligible[0] ?? null;
-            }
-            case "cheapest": {
-                const ranked = [...eligible].sort((a, b) => {
-                    const ca = this.costOf(a.provider) ?? Number.POSITIVE_INFINITY;
-                    const cb = this.costOf(b.provider) ?? Number.POSITIVE_INFINITY;
-                    if (ca !== cb)
-                        return ca - cb;
-                    return b.priority - a.priority;
-                });
-                return ranked[0] ?? null;
-            }
-            case "fastest": {
-                const ranked = [...eligible].sort((a, b) => {
-                    const la = this.p50(a.provider, args.modelType);
-                    const lb = this.p50(b.provider, args.modelType);
-                    // Untracked providers get Infinity → deprioritised until we
-                    // have samples. First call always falls through to native
-                    // priority via the tie-break.
-                    const va = la ?? Number.POSITIVE_INFINITY;
-                    const vb = lb ?? Number.POSITIVE_INFINITY;
-                    if (va !== vb)
-                        return va - vb;
-                    return b.priority - a.priority;
-                });
-                return ranked[0] ?? null;
-            }
-            case "prefer-local": {
-                const local = eligible.find((c) => c.provider === "eliza-local-inference" ||
-                    c.provider === "capacitor-llama");
-                if (local)
-                    return local;
-                const bridge = eligible.find((c) => c.provider === "eliza-device-bridge");
-                if (bridge)
-                    return bridge;
-                return eligible[0] ?? null;
-            }
-            case "round-robin": {
-                // Pick the one least-recently-picked. Ties broken by priority.
-                const ranked = [...eligible].sort((a, b) => {
-                    const la = this.lastPicked(a.provider, args.modelType) ?? 0;
-                    const lb = this.lastPicked(b.provider, args.modelType) ?? 0;
-                    if (la !== lb)
-                        return la - lb;
-                    return b.priority - a.priority;
-                });
-                return ranked[0] ?? null;
-            }
-        }
-    }
-    /** For tests and diagnostics. */
-    snapshot() {
-        const out = {};
-        for (const [provider, stats] of this.stats) {
-            out[provider] = {};
-            for (const [modelType, buf] of stats.latency) {
-                const row = out[provider];
-                if (row)
-                    row[modelType] = buf.p50();
-            }
-        }
-        return out;
-    }
-}
-export const policyEngine = new PolicyEngine();

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

@@ -1,8 +0,0 @@
-/**
- * Re-export of the shared routing-preferences module. The canonical
- * implementation lives in `@elizaos/shared/local-inference` because both
- * the server (`@elizaos/app-core`) and the UI client (`@elizaos/ui`)
- * read/write the same routing.json with identical semantics.
- */
-export { DEFAULT_ROUTING_POLICY, type RoutingPolicy, type RoutingPreferences, readRoutingPreferences, setPolicy, setPreferredProvider, writeRoutingPreferences, } from "@elizaos/shared";
-//# sourceMappingURL=routing-preferences.d.ts.map

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

@@ -1 +0,0 @@
-{"version":3,"file":"routing-preferences.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/routing-preferences.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,OAAO,EACL,sBAAsB,EACtB,KAAK,aAAa,EAClB,KAAK,kBAAkB,EACvB,sBAAsB,EACtB,SAAS,EACT,oBAAoB,EACpB,uBAAuB,GACxB,MAAM,iBAAiB,CAAC"}

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

@@ -1,7 +0,0 @@
-/**
- * Re-export of the shared routing-preferences module. The canonical
- * implementation lives in `@elizaos/shared/local-inference` because both
- * the server (`@elizaos/app-core`) and the UI client (`@elizaos/ui`)
- * read/write the same routing.json with identical semantics.
- */
-export { DEFAULT_ROUTING_POLICY, readRoutingPreferences, setPolicy, setPreferredProvider, writeRoutingPreferences, } from "@elizaos/shared";

package/services/local-inference/service.d.ts

@@ -1,88 +0,0 @@
-/**
- * Public facade for the local-inference service.
- *
- * Single entry point used by the API routes, the settings UI, and any
- * future orchestration code. Holds singleton instances of the downloader
- * and active-model coordinator so subscribers receive the same event
- * stream across the process.
- */
-import type { AgentRuntime } from "@elizaos/core";
-import { type LocalInferenceLoadOverrides } from "./active-model";
-import type { CacheStatsEntry } from "./cache-bridge";
-import { type RecommendedModelSelection } from "./recommendation";
-import type { ActiveModelState, AgentModelSlot, CatalogModel, DownloadEvent, DownloadJob, HardwareProbe, LocalInferenceReadiness, ModelAssignments, ModelHubSnapshot, TextGenerationSlot } from "./types";
-import { type VerifyResult } from "./verify";
-export declare class LocalInferenceService {
-    private readonly downloader;
-    private readonly activeModel;
-    private bundledBootstrap;
-    getCatalog(): CatalogModel[];
-    /**
-     * Register any bundled GGUF files staged by the AOSP build (or any
-     * other install path that drops a `manifest.json` next to the model
-     * files) into the registry. Runs at most once per process; the
-     * promise is cached so concurrent first callers wait on the same
-     * work.
-     */
-    private bootstrapBundled;
-    getInstalled(): Promise<import("@elizaos/shared").InstalledModel[]>;
-    getHardware(): Promise<HardwareProbe>;
-    getDownloads(): DownloadJob[];
-    getActive(): ActiveModelState;
-    getAssignments(): Promise<ModelAssignments>;
-    setSlotAssignment(slot: AgentModelSlot, modelId: string | null): Promise<ModelAssignments>;
-    snapshot(): Promise<ModelHubSnapshot>;
-    getTextReadiness(): Promise<LocalInferenceReadiness>;
-    getRecommendedModel(slot: TextGenerationSlot, hardware?: HardwareProbe): Promise<RecommendedModelSelection>;
-    getRecommendedModels(hardware?: HardwareProbe): Promise<Record<TextGenerationSlot, RecommendedModelSelection>>;
-    /**
-     * Pull the kernels map from CAPABILITIES.json next to the installed
-     * llama-server binary. Null when the file is absent or when DFlash isn't
-     * enabled. Surfaces to the recommender so we don't recommend a model the
-     * installed binary can't actually run.
-     */
-    private installedBinaryKernels;
-    startDownload(modelIdOrSpec: string | CatalogModel): Promise<DownloadJob>;
-    startSmallerFallbackDownload(currentModelId: string, slot?: TextGenerationSlot, hardware?: HardwareProbe): Promise<{
-        model: CatalogModel;
-        job: DownloadJob;
-    } | null>;
-    searchHuggingFace(query: string, limit?: number): Promise<CatalogModel[]>;
-    /**
-     * Verify an installed model's file integrity. When the model was a
-     * Eliza-download and there was no stored sha256 yet (legacy entry), the
-     * computed hash is persisted so subsequent verifies have a baseline.
-     */
-    verifyModel(id: string): Promise<VerifyResult>;
-    cancelDownload(modelId: string): boolean;
-    subscribeDownloads(listener: (event: DownloadEvent) => void): () => void;
-    subscribeActive(listener: (state: ActiveModelState) => void): () => void;
-    setActive(runtime: AgentRuntime | null, modelId: string, overrides?: LocalInferenceLoadOverrides): Promise<ActiveModelState>;
-    clearActive(runtime: AgentRuntime | null): Promise<ActiveModelState>;
-    /**
-     * Diagnostic snapshot of the local prefix-cache state. Returns:
-     *   - `dflash`: per-slot files persisted by the out-of-process
-     *     llama-server (size + mtime + age in ms).
-     *   - `engine`: in-process session-pool size and live cache keys.
-     * Used by the API layer to render a "local cache" debug panel.
-     */
-    getLocalCacheStats(): Promise<{
-        dflash: {
-            modelHash: string | null;
-            slotDir: string | null;
-            parallel: number;
-            files: CacheStatsEntry[];
-        };
-        engine: {
-            size: number;
-            maxSize: number;
-            keys: string[];
-        } | null;
-    }>;
-    uninstall(modelId: string): Promise<{
-        removed: boolean;
-        reason?: "external" | "not-found";
-    }>;
-}
-export declare const localInferenceService: LocalInferenceService;
-//# sourceMappingURL=service.d.ts.map

package/services/local-inference/service.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"service.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/service.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAClD,OAAO,EAEL,KAAK,2BAA2B,EACjC,MAAM,gBAAgB,CAAC;AAGxB,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AAQtD,OAAO,EAEL,KAAK,yBAAyB,EAG/B,MAAM,kBAAkB,CAAC;AAM1B,OAAO,KAAK,EACV,gBAAgB,EAChB,cAAc,EACd,YAAY,EACZ,aAAa,EACb,WAAW,EACX,aAAa,EACb,uBAAuB,EACvB,gBAAgB,EAChB,gBAAgB,EAChB,kBAAkB,EACnB,MAAM,SAAS,CAAC;AACjB,OAAO,EAAE,KAAK,YAAY,EAAwB,MAAM,UAAU,CAAC;AAEnE,qBAAa,qBAAqB;IAChC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAoB;IAC/C,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAgC;IAC5D,OAAO,CAAC,gBAAgB,CAA8B;IAEtD,UAAU;IAIV;;;;;;OAMG;IACH,OAAO,CAAC,gBAAgB;IASlB,YAAY;IAKZ,WAAW,IAAI,OAAO,CAAC,aAAa,CAAC;IAI3C,YAAY,IAAI,WAAW,EAAE;IAI7B,SAAS,IAAI,gBAAgB;IAIvB,cAAc,IAAI,OAAO,CAAC,gBAAgB,CAAC;IAI3C,iBAAiB,CACrB,IAAI,EAAE,cAAc,EACpB,OAAO,EAAE,MAAM,GAAG,IAAI,GACrB,OAAO,CAAC,gBAAgB,CAAC;IAKtB,QAAQ,IAAI,OAAO,CAAC,gBAAgB,CAAC;IAyBrC,gBAAgB,IAAI,OAAO,CAAC,uBAAuB,CAAC;IAcpD,mBAAmB,CACvB,IAAI,EAAE,kBAAkB,EACxB,QAAQ,CAAC,EAAE,aAAa,GACvB,OAAO,CAAC,yBAAyB,CAAC;IAS/B,oBAAoB,CACxB,QAAQ,CAAC,EAAE,aAAa,GACvB,OAAO,CAAC,MAAM,CAAC,kBAAkB,EAAE,yBAAyB,CAAC,CAAC;IAQjE;;;;;OAKG;IACH,OAAO,CAAC,sBAAsB;IAKxB,aAAa,CACjB,aAAa,EAAE,MAAM,GAAG,YAAY,GACnC,OAAO,CAAC,WAAW,CAAC;IAIjB,4BAA4B,CAChC,cAAc,EAAE,MAAM,EACtB,IAAI,GAAE,kBAAiC,EACvC,QAAQ,CAAC,EAAE,aAAa,GACvB,OAAO,CAAC;QAAE,KAAK,EAAE,YAAY,CAAC;QAAC,GAAG,EAAE,WAAW,CAAA;KAAE,GAAG,IAAI,CAAC;IActD,iBAAiB,CACrB,KAAK,EAAE,MAAM,EACb,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,YAAY,EAAE,CAAC;IAI1B;;;;OAIG;IACG,WAAW,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAoCpD,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO;IAIxC,kBAAkB,CAAC,QAAQ,EAAE,CAAC,KAAK,EAAE,aAAa,KAAK,IAAI,GAAG,MAAM,IAAI;IAIxE,eAAe,CAAC,QAAQ,EAAE,CAAC,KAAK,EAAE,gBAAgB,KAAK,IAAI,GAAG,MAAM,IAAI;IAIlE,SAAS,CACb,OAAO,EAAE,YAAY,GAAG,IAAI,EAC5B,OAAO,EAAE,MAAM,EACf,SAAS,CAAC,EAAE,2BAA2B,GACtC,OAAO,CAAC,gBAAgB,CAAC;IAQtB,WAAW,CAAC,OAAO,EAAE,YAAY,GAAG,IAAI,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAI1E;;;;;;OAMG;IACG,kBAAkB,IAAI,OAAO,CAAC;QAClC,MAAM,EAAE;YACN,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;YACzB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;YACvB,QAAQ,EAAE,MAAM,CAAC;YACjB,KAAK,EAAE,eAAe,EAAE,CAAC;SAC1B,CAAC;QACF,MAAM,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,EAAE,CAAA;SAAE,GAAG,IAAI,CAAC;KAClE,CAAC;IAOI,SAAS,CACb,OAAO,EAAE,MAAM,GACd,OAAO,CAAC;QAAE,OAAO,EAAE,OAAO,CAAC;QAAC,MAAM,CAAC,EAAE,UAAU,GAAG,WAAW,CAAA;KAAE,CAAC;CAQpE;AAED,eAAO,MAAM,qBAAqB,uBAA8B,CAAC"}