@elizaos/ui 2.0.3-beta.2 → 2.0.3-beta.4

package/services/local-inference/hf-search.js

@@ -1,280 +1,22 @@
+export const CUSTOM_MODEL_SEARCH_DISABLED_MESSAGE = "Custom model search is disabled; local inference uses curated Eliza-1 bundles only.";
 /**
- * HuggingFace Hub search for GGUF models.
- *
- * Calls `https://huggingface.co/api/models` with `filter=gguf` to narrow
- * results to repos that actually ship GGUF quantisations. Each matching
- * repo is expanded with `/api/models/<repo>` to pick a representative
- * quant file (preferring Q4_K_M when present). Results are shaped like
- * `CatalogModel` so the existing ModelCard renders them directly.
- *
- * We deliberately do not persist these — they're dynamic, and a curated
- * entry with the same hfRepo always takes precedence in the UI.
- */
-const HF_API = "https://huggingface.co/api";
-const SEARCH_TIMEOUT_MS = 10_000;
-const QUANT_PREFERENCE = [
-    "Q4_K_M",
-    "Q5_K_M",
-    "Q4_0",
-    "Q5_0",
-    "Q3_K_M",
-    "Q8_0",
-    "Q2_K",
-];
-function pickQuantFile(siblings) {
-    const ggufs = siblings.filter((s) => s.rfilename?.toLowerCase().endsWith(".gguf"));
-    if (ggufs.length === 0)
-        return null;
-    for (const quant of QUANT_PREFERENCE) {
-        const match = ggufs.find((s) => s.rfilename?.toUpperCase().includes(quant));
-        if (match)
-            return match;
-    }
-    // No preferred quant — pick the smallest GGUF so we don't surprise the
-    // user with a 40 GB file.
-    return [...ggufs].sort((a, b) => (a.size ?? 0) - (b.size ?? 0))[0] ?? null;
-}
-function extractQuantLabel(filename) {
-    for (const quant of QUANT_PREFERENCE) {
-        if (filename.toUpperCase().includes(quant))
-            return quant;
-    }
-    return "GGUF";
-}
-/**
- * Very rough parameter-count inference from model name / tags. We use this
- * only to pick a bucket label — not for any hard memory check.
- */
-function inferParams(name, tags) {
-    const lower = `${name} ${tags.join(" ")}`.toLowerCase();
-    const sizes = [
-        [/\b32b\b/, "32B", "xl"],
-        [/\b27b\b/, "27B", "large"],
-        [/\b24b\b/, "24B", "large"],
-        [/\b22b\b/, "22B", "large"],
-        [/\b16b\b/, "16B", "large"],
-        [/\b14b\b/, "14B", "large"],
-        [/\b13b\b/, "14B", "large"],
-        [/\b9b\b/, "9B", "mid"],
-        [/(?:^|[^0-9])0\.8b\b/, "0.8B", "small"],
-        [/\b8b\b/, "8B", "mid"],
-        [/\b7b\b/, "7B", "mid"],
-        [/\b3b\b/, "3B", "small"],
-        [/\b2b\b/, "2B", "small"],
-        [/\b1b\b/, "1B", "small"],
-    ];
-    for (const [re, params, bucket, parameterLabel] of sizes) {
-        if (re.test(lower)) {
-            return {
-                params,
-                bucket,
-                ...(parameterLabel ? { parameterLabel } : {}),
-            };
-        }
-    }
-    return { params: "7B", bucket: "mid" };
-}
-function inferCategory(tags, pipelineTag) {
-    const lowerTags = tags.map((t) => t.toLowerCase());
-    if (lowerTags.some((t) => t.includes("code") || t.includes("coder"))) {
-        return "code";
-    }
-    if (lowerTags.some((t) => t.includes("reasoning") || t === "math" || t.includes("r1"))) {
-        return "reasoning";
-    }
-    if (lowerTags.some((t) => t.includes("function") || t.includes("tool") || t.includes("hermes"))) {
-        return "tools";
-    }
-    if (pipelineTag === "text-generation") {
-        return "chat";
-    }
-    return "chat";
-}
-async function fetchWithTimeout(url, init) {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), SEARCH_TIMEOUT_MS);
-    try {
-        return await fetch(url, { ...init, signal: controller.signal });
-    }
-    finally {
-        clearTimeout(timer);
-    }
-}
-/**
- * Search HuggingFace for GGUF repos matching `query`, returning
- * catalog-shaped entries ready for the Model Hub UI.
+ * Compatibility shim for older callers. Consumer local inference no longer
+ * searches third-party hubs; setup is managed through the curated Eliza-1
+ * bundles registered in the catalog.
  */
 export async function searchHuggingFaceGguf(query, limit = 12) {
-    const trimmed = query.trim();
-    if (trimmed.length === 0)
-        return [];
-    const searchUrl = new URL(`${HF_API}/models`);
-    searchUrl.searchParams.set("search", trimmed);
-    searchUrl.searchParams.set("filter", "gguf");
-    searchUrl.searchParams.set("limit", String(Math.min(50, Math.max(1, limit * 2))));
-    searchUrl.searchParams.set("sort", "downloads");
-    searchUrl.searchParams.set("direction", "-1");
-    const searchRes = await fetchWithTimeout(searchUrl.toString(), {
-        headers: { accept: "application/json" },
-    });
-    if (!searchRes.ok) {
-        throw new Error(`Hugging Face search failed: HTTP ${searchRes.status}`);
-    }
-    const rawList = (await searchRes.json());
-    const candidates = rawList
-        .map((r) => r.id ?? r.modelId)
-        .filter((id) => typeof id === "string" && id.length > 0)
-        .slice(0, limit);
-    // Parallel detail fetches; failures drop quietly so one bad repo doesn't
-    // take down the whole search.
-    const details = await Promise.all(candidates.map(async (id) => {
-        try {
-            const res = await fetchWithTimeout(`${HF_API}/models/${encodeURIComponent(id)}`, { headers: { accept: "application/json" } });
-            if (!res.ok)
-                return null;
-            return (await res.json());
-        }
-        catch {
-            return null;
-        }
-    }));
-    const results = [];
-    for (let i = 0; i < candidates.length; i++) {
-        const id = candidates[i];
-        const detail = details[i];
-        if (!id || !detail?.siblings)
-            continue;
-        const sibling = pickQuantFile(detail.siblings);
-        if (!sibling?.rfilename)
-            continue;
-        const sizeBytes = sibling.size ?? 0;
-        const sizeGb = sizeBytes > 0 ? sizeBytes / 1024 ** 3 : 4;
-        const { params, bucket, parameterLabel } = inferParams(id, detail.tags ?? []);
-        const quant = extractQuantLabel(sibling.rfilename);
-        const category = inferCategory(detail.tags ?? [], detail.pipeline_tag);
-        const displayName = id.split("/").pop() ?? id;
-        // minRam heuristic: 2x the file size, minimum 4 GB. Close enough for
-        // the fit badge; the ModelCard's fit computation still does the work.
-        const minRamGb = Math.max(4, Math.round(sizeGb * 2));
-        results.push({
-            id: `hf:${id}::${sibling.rfilename}`,
-            displayName,
-            hfRepo: id,
-            ggufFile: sibling.rfilename,
-            params,
-            ...(parameterLabel ? { parameterLabel } : {}),
-            quant,
-            sizeGb: Math.round(sizeGb * 10) / 10,
-            minRamGb,
-            category,
-            bucket,
-            blurb: (detail.tags ?? []).slice(0, 4).join(" · ") ||
-                `${detail.downloads ?? 0} downloads · ${detail.likes ?? 0} likes`,
-        });
-    }
-    return results;
+    void query;
+    void limit;
+    return [];
 }
-function normalizeModelScopeRepoId(query) {
-    const trimmed = query.trim().replace(/^modelscope:\/\//i, "");
-    const match = /^[A-Za-z0-9_.-]+\/[A-Za-z0-9_.-]+$/.exec(trimmed);
-    return match ? trimmed : null;
-}
-function modelScopeOwnerQuery(query) {
-    const trimmed = query.trim().replace(/^modelscope:\/\//i, "");
-    const first = trimmed.split(/\s+/)[0] ?? "";
-    if (/^[A-Za-z0-9_.-]+$/.test(first))
-        return first;
-    return null;
-}
-function pickModelScopeGguf(files) {
-    return pickQuantFile(files.map((file) => ({
-        rfilename: file.Path ?? file.Name,
-        size: file.Size,
-    })));
-}
-async function fetchModelScopeJson(url, init) {
-    const headers = new Headers(init?.headers);
-    headers.set("accept", "application/json");
-    const res = await fetchWithTimeout(url, {
-        ...init,
-        headers,
-        redirect: "follow",
-    });
-    if (!res.ok) {
-        throw new Error(`ModelScope request failed: HTTP ${res.status}`);
-    }
-    const envelope = (await res.json());
-    if (envelope.Code !== 200 || envelope.Data === undefined) {
-        throw new Error(envelope.Message ?? "ModelScope request failed");
-    }
-    return envelope.Data;
-}
-async function listModelScopeOwnerModels(owner, limit) {
-    const data = await fetchModelScopeJson("https://www.modelscope.cn/api/v1/models", {
-        method: "PUT",
-        headers: { "content-type": "application/json" },
-        body: JSON.stringify({
-            Path: owner,
-            PageNumber: 1,
-            PageSize: Math.min(50, Math.max(1, limit * 3)),
-        }),
-    });
-    return (data.Models ?? [])
-        .map((m) => m.ModelId ?? m.model_id)
-        .filter((id) => typeof id === "string" && id.length > 0);
-}
-async function modelScopeResultForRepo(repoId) {
-    const filesData = await fetchModelScopeJson(`https://www.modelscope.cn/api/v1/models/${repoId}/repo/files?Revision=master&Recursive=true`);
-    const sibling = pickModelScopeGguf(filesData.Files ?? []);
-    if (!sibling?.rfilename)
-        return null;
-    const sizeBytes = sibling.size ?? 0;
-    const sizeGb = sizeBytes > 0 ? sizeBytes / 1024 ** 3 : 4;
-    const { params, bucket } = inferParams(repoId, []);
-    const quant = extractQuantLabel(sibling.rfilename);
-    const displayName = repoId.split("/").pop() ?? repoId;
-    return {
-        id: `modelscope:${repoId}::${sibling.rfilename}`,
-        displayName,
-        hub: "modelscope",
-        hfRepo: repoId,
-        ggufFile: sibling.rfilename,
-        params,
-        quant,
-        sizeGb: Math.round(sizeGb * 10) / 10,
-        minRamGb: Math.max(4, Math.round(sizeGb * 2)),
-        category: "chat",
-        bucket,
-        blurb: "ModelScope GGUF result · explicit opt-in only",
-    };
-}
-/**
- * Search ModelScope for GGUF repos. ModelScope exposes reliable owner-list
- * and repo-file APIs, so this accepts either `owner/model` or an owner name
- * and filters the returned repos for GGUF files.
- */
 export async function searchModelScopeGguf(query, limit = 12) {
-    const trimmed = query.trim();
-    if (trimmed.length === 0)
-        return [];
-    const direct = normalizeModelScopeRepoId(trimmed);
-    const repoIds = direct
-        ? [direct]
-        : await listModelScopeOwnerModels(modelScopeOwnerQuery(trimmed) ?? trimmed, limit);
-    const results = await Promise.all(repoIds.slice(0, Math.min(50, limit * 3)).map(async (id) => {
-        try {
-            return await modelScopeResultForRepo(id);
-        }
-        catch {
-            return null;
-        }
-    }));
-    return results
-        .filter((model) => Boolean(model))
-        .slice(0, limit);
+    void query;
+    void limit;
+    return [];
 }
 export async function searchModelHubGguf(query, hub = "huggingface", limit = 12) {
-    return hub === "modelscope"
-        ? searchModelScopeGguf(query, limit)
-        : searchHuggingFaceGguf(query, limit);
+    void query;
+    void hub;
+    void limit;
+    return [];
 }

package/services/local-inference/home-model-status.d.ts

@@ -1,4 +1,4 @@
-import type { LocalInferenceReadiness } from "./types";
+import type { LocalInferenceReadiness } from "./types.js";
 export type HomeModelStatusKind = "not-required" | "ready" | "downloading" | "loading" | "missing" | "error";
 export interface HomeModelStatus {
     kind: HomeModelStatusKind;

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

@@ -1,12 +1,12 @@
 export { computeGenerationThroughput, type GenerationCounters, type GenerationThroughput, isGenerationCounters, } from "@elizaos/shared/local-inference";
-export type { LocalInferenceLoader } from "./active-model";
-export { ELIZA_1_PLACEHOLDER_IDS, FIRST_RUN_DEFAULT_MODEL_ID, findCatalogModel, MODEL_CATALOG, } from "./catalog";
-export { filterSettingsDefaultLocalModels, isDefaultLocalModelFamily, isEliza1ModelFamilyId, isSettingsDefaultLocalModel, } from "./catalog-policy";
-export { DEFAULT_LOCAL_MODEL_SEARCH_PROVIDER_ID, getLocalModelSearchProvider, isLocalModelSearchProviderId, type LocalModelSearchProviderDescriptor, type LocalModelSearchProviderId, type LocalModelSearchResponse, type LocalModelSearchResult, listLocalModelSearchProviders, searchLocalModelProvider, wrapLocalModelSearchResults, } from "./custom-search";
-export { assessFit, probeHardware } from "./hardware";
-export { buildTextGenerationReadiness } from "./readiness";
-export { assessCatalogModelFit, catalogDownloadSizeBytes, catalogDownloadSizeGb, chooseSmallerFallbackModel, classifyRecommendationPlatform, type RecommendationPlatformClass, type RecommendedModelSelection, recommendForFirstRun, selectRecommendedModelForSlot, selectRecommendedModels, } from "./recommendation";
-export { type DeviceResourceSnapshot, getDeviceResourceSnapshot, normalizeResourceSnapshot, type SnapshotThermalState, } from "./resource-snapshot-bridge";
-export { LocalInferenceService, localInferenceService } from "./service";
-export type { ActiveModelState, CatalogModel, DownloadEvent, DownloadJob, DownloadState, HardwareFitLevel, HardwareProbe, InstalledModel, LocalInferenceDownloadStatus, LocalInferenceReadiness, LocalInferenceSlotReadiness, ModelBucket, ModelCategory, ModelHubSnapshot, TextGenerationSlot, } from "./types";
+export type { LocalInferenceLoader } from "./active-model.js";
+export { ELIZA_1_PLACEHOLDER_IDS, FIRST_RUN_DEFAULT_MODEL_ID, findCatalogModel, MODEL_CATALOG, } from "./catalog.js";
+export { filterSettingsDefaultLocalModels, isDefaultLocalModelFamily, isEliza1ModelFamilyId, isSettingsDefaultLocalModel, } from "./catalog-policy.js";
+export { DEFAULT_LOCAL_MODEL_SEARCH_PROVIDER_ID, getLocalModelSearchProvider, isLocalModelSearchProviderId, type LocalModelSearchProviderDescriptor, type LocalModelSearchProviderId, type LocalModelSearchResponse, type LocalModelSearchResult, listLocalModelSearchProviders, searchLocalModelProvider, wrapLocalModelSearchResults, } from "./custom-search.js";
+export { assessFit, probeHardware } from "./hardware.js";
+export { buildTextGenerationReadiness } from "./readiness.js";
+export { assessCatalogModelFit, catalogDownloadSizeBytes, catalogDownloadSizeGb, chooseSmallerFallbackModel, classifyRecommendationPlatform, type RecommendationPlatformClass, type RecommendedModelSelection, recommendForFirstRun, selectRecommendedModelForSlot, selectRecommendedModels, } from "./recommendation.js";
+export { type DeviceResourceSnapshot, getDeviceResourceSnapshot, normalizeResourceSnapshot, type SnapshotThermalState, } from "./resource-snapshot-bridge.js";
+export { LocalInferenceService, localInferenceService } from "./service.js";
+export type { ActiveModelState, CatalogModel, DownloadEvent, DownloadJob, DownloadState, HardwareFitLevel, HardwareProbe, InstalledModel, LocalInferenceDownloadStatus, LocalInferenceReadiness, LocalInferenceSlotReadiness, ModelBucket, ModelCategory, ModelHubSnapshot, TextGenerationSlot, } from "./types.js";
 //# sourceMappingURL=index.d.ts.map

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

@@ -1,4 +1,4 @@
-import type { ActiveModelState, CatalogModel, DownloadJob, InstalledModel, LocalInferenceReadiness, ModelAssignments } from "./types";
+import type { ActiveModelState, CatalogModel, DownloadJob, InstalledModel, LocalInferenceReadiness, ModelAssignments } from "./types.js";
 export declare function buildTextGenerationReadiness(input: {
     assignments: ModelAssignments;
     installed: InstalledModel[];

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

@@ -1,4 +1,4 @@
-import type { CatalogModel, HardwareFitLevel, HardwareProbe, TextGenerationSlot } from "./types";
+import type { CatalogModel, HardwareFitLevel, HardwareProbe, TextGenerationSlot } from "./types.js";
 export type RecommendationPlatformClass = "mobile" | "apple-silicon" | "linux-gpu" | "linux-cpu" | "desktop-gpu" | "desktop-cpu";
 export interface RecommendedModelSelection {
     slot: TextGenerationSlot;

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

@@ -1,21 +1,19 @@
 /**
  * On-disk registry of installed models.
  *
- * Two sources feed the registry:
- *   1. Eliza-owned downloads (source: "eliza-download") — written on
- *      successful completion by the downloader.
- *   2. External scans (source: "external-scan") — merged in at read time
- *      from `scanExternalModels()`. These are never persisted to the
- *      registry file; a rescan runs whenever we read.
- *
- * The JSON file only holds Eliza-owned entries. That way, if a user
- * cleans up LM Studio models we don't show stale ghosts.
+ * The default registry contains only Eliza-owned downloads
+ * (source: "eliza-download") written on successful completion by the
+ * curated bundle downloader. External scans are developer-only diagnostics
+ * behind `ELIZA_LOCAL_INFERENCE_ENABLE_EXTERNAL_SCAN=1`; they never enter
+ * first-run, setup, or normal Settings surfaces.
  */
-import type { InstalledModel } from "./types";
+import type { InstalledModel } from "./types.js";
 /**
- * Return all models currently usable: persisted Eliza downloads plus a
- * fresh external-tool scan. External duplicates of Eliza-owned files are
- * filtered out by path.
+ * Return models currently usable by the curated local-inference path.
+ *
+ * Normal product behavior is Eliza-1 only. The external scan remains available
+ * only to developers who explicitly opt into the old arbitrary-GGUF diagnostic
+ * path with `ELIZA_LOCAL_INFERENCE_ENABLE_EXTERNAL_SCAN=1`.
  */
 export declare function listInstalledModels(): Promise<InstalledModel[]>;
 /** Add or update a Eliza-owned entry. External entries are rejected. */

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

@@ -1 +1 @@
-{"version":3,"file":"registry.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAMH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,SAAS,CAAC;AAmC9C;;;;GAIG;AACH,wBAAsB,mBAAmB,IAAI,OAAO,CAAC,cAAc,EAAE,CAAC,CAerE;AAED,wEAAwE;AACxE,wBAAsB,gBAAgB,CAAC,KAAK,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAyB3E;AAED,gEAAgE;AAChE,wBAAsB,eAAe,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAM/D;AAED;;;;;GAKG;AACH,wBAAsB,gBAAgB,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC;IAC1D,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,CAAC,EAAE,UAAU,GAAG,WAAW,CAAC;CACnC,CAAC,CA6BD"}
+{"version":3,"file":"registry.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAMH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,SAAS,CAAC;AAyC9C;;;;;;GAMG;AACH,wBAAsB,mBAAmB,IAAI,OAAO,CAAC,cAAc,EAAE,CAAC,CAarE;AAED,wEAAwE;AACxE,wBAAsB,gBAAgB,CAAC,KAAK,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAyB3E;AAED,gEAAgE;AAChE,wBAAsB,eAAe,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAM/D;AAED;;;;;GAKG;AACH,wBAAsB,gBAAgB,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC;IAC1D,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,CAAC,EAAE,UAAU,GAAG,WAAW,CAAC;CACnC,CAAC,CA6BD"}

package/services/local-inference/registry.js

@@ -1,15 +1,11 @@
 /**
  * On-disk registry of installed models.
  *
- * Two sources feed the registry:
- *   1. Eliza-owned downloads (source: "eliza-download") — written on
- *      successful completion by the downloader.
- *   2. External scans (source: "external-scan") — merged in at read time
- *      from `scanExternalModels()`. These are never persisted to the
- *      registry file; a rescan runs whenever we read.
- *
- * The JSON file only holds Eliza-owned entries. That way, if a user
- * cleans up LM Studio models we don't show stale ghosts.
+ * The default registry contains only Eliza-owned downloads
+ * (source: "eliza-download") written on successful completion by the
+ * curated bundle downloader. External scans are developer-only diagnostics
+ * behind `ELIZA_LOCAL_INFERENCE_ENABLE_EXTERNAL_SCAN=1`; they never enter
+ * first-run, setup, or normal Settings surfaces.
  */
 import fs from "node:fs/promises";
 import path from "node:path";
@@ -38,19 +34,24 @@ async function writeElizaOwned(models) {
     await fs.writeFile(tmp, JSON.stringify(payload, null, 2), "utf8");
     await fs.rename(tmp, registryPath());
 }
+function externalScanEnabled() {
+    const value = process.env.ELIZA_LOCAL_INFERENCE_ENABLE_EXTERNAL_SCAN?.trim().toLowerCase();
+    return value === "1" || value === "true" || value === "yes";
+}
 /**
- * Return all models currently usable: persisted Eliza downloads plus a
- * fresh external-tool scan. External duplicates of Eliza-owned files are
- * filtered out by path.
+ * Return models currently usable by the curated local-inference path.
+ *
+ * Normal product behavior is Eliza-1 only. The external scan remains available
+ * only to developers who explicitly opt into the old arbitrary-GGUF diagnostic
+ * path with `ELIZA_LOCAL_INFERENCE_ENABLE_EXTERNAL_SCAN=1`.
  */
 export async function listInstalledModels() {
-    const [ownedRaw, external] = await Promise.all([
-        readElizaOwned(),
-        scanExternalModels(),
-    ]);
-    const owned = ownedRaw;
+    const owned = await readElizaOwned();
+    if (!externalScanEnabled())
+        return owned;
     // Filter out Eliza-owned files that also survived a reboot of the local
     // file and got re-detected by the scanner.
+    const external = await scanExternalModels();
     const ownedPaths = new Set(owned.map((m) => path.resolve(m.path)));
     const dedupedExternal = external.filter((m) => !ownedPaths.has(path.resolve(m.path)));
     return [...owned, ...dedupedExternal];

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

@@ -10,11 +10,15 @@
  *   - manual       — honour `preferredProvider`; when no pref set, fall
  *                    through to the runtime's native priority order
  *                    (highest registered priority wins).
+ *   - auto         — capability-driven server-side; this client-side mirror has
+ *                    no device classifier, so it previews as local-first.
  *   - 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.
+ *   - local-only   — always on-device; returns null when no local handler.
+ *   - cloud-only   — always off-device; returns null when only local handlers.
  *   - round-robin  — distribute load evenly across eligible providers.
  *
  * Latency is tracked in a ring buffer per provider per model type. Cost
@@ -22,8 +26,8 @@
  * 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";
+import type { HandlerRegistration } from "./handler-registry.js";
+import type { RoutingPolicy } from "./routing-preferences.js";
 declare class PolicyEngine {
     private stats;
     private statsFor;

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

@@ -1 +1 @@
-{"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;AA6D3D,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;IA2E9B,iCAAiC;IACjC,QAAQ,IAAI,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC,CAAC;CAW1D;AAED,eAAO,MAAM,YAAY,cAAqB,CAAC"}
+{"version":3,"file":"routing-policy.d.ts","sourceRoot":"","sources":["../../../../../../src/services/local-inference/routing-policy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAEH,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAC9D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AAwF3D,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;IA2E9B,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

@@ -10,11 +10,15 @@
  *   - manual       — honour `preferredProvider`; when no pref set, fall
  *                    through to the runtime's native priority order
  *                    (highest registered priority wins).
+ *   - auto         — capability-driven server-side; this client-side mirror has
+ *                    no device classifier, so it previews as local-first.
  *   - 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.
+ *   - local-only   — always on-device; returns null when no local handler.
+ *   - cloud-only   — always off-device; returns null when only local handlers.
  *   - round-robin  — distribute load evenly across eligible providers.
  *
  * Latency is tracked in a ring buffer per provider per model type. Cost
@@ -23,6 +27,26 @@
  * rather than dollar-accurate billing.
  */
 const RING_SIZE = 32;
+/** Provider IDs that serve inference on-device (no network round-trip). */
+const LOCAL_PROVIDERS = new Set([
+    "eliza-local-inference",
+    "capacitor-llama",
+    "eliza-device-bridge",
+]);
+function isLocalProvider(provider) {
+    return LOCAL_PROVIDERS.has(provider);
+}
+/**
+ * The first registered local handler in priority order, or null. Prefers the
+ * in-process / Capacitor backends over the device bridge.
+ */
+function findLocalCandidate(candidates) {
+    const inProcess = candidates.find((c) => c.provider === "eliza-local-inference" ||
+        c.provider === "capacitor-llama");
+    if (inProcess)
+        return inProcess;
+    return candidates.find((c) => c.provider === "eliza-device-bridge") ?? null;
+}
 class RingBuffer {
     buf = [];
     push(sample) {
@@ -159,18 +183,22 @@ class PolicyEngine {
             case "auto":
             case "prefer-local": {
                 // "auto" is capability-driven server-side (the local-inference plugin's
-                // PolicyEngine consults device tier). This client-side engine has no
-                // device classifier, so it mirrors prefer-local (local-first) — a safe,
-                // consistent fallback for any client-side routing preview.
-                const local = eligible.find((c) => c.provider === "eliza-local-inference" ||
-                    c.provider === "capacitor-llama");
+                // PolicyEngine consults device tier + live signals). This client-side
+                // engine has no device classifier, so it mirrors prefer-local
+                // (local-first) — a safe, consistent fallback for any routing preview.
+                const local = findLocalCandidate(eligible);
                 if (local)
                     return local;
-                const bridge = eligible.find((c) => c.provider === "eliza-device-bridge");
-                if (bridge)
-                    return bridge;
                 return eligible[0] ?? null;
             }
+            case "local-only": {
+                // Never leaves the device; null when no local handler is registered.
+                return findLocalCandidate(eligible);
+            }
+            case "cloud-only": {
+                // Never runs on-device; null when only local handlers exist.
+                return eligible.find((c) => !isLocalProvider(c.provider)) ?? null;
+            }
             case "round-robin": {
                 // Pick the one least-recently-picked. Ties broken by priority.
                 const ranked = [...eligible].sort((a, b) => {

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

@@ -7,10 +7,10 @@
  * stream across the process.
  */
 import type { AgentRuntime } from "@elizaos/core";
-import { type LocalModelSearchProviderId, type LocalModelSearchResponse } from "./custom-search";
-import { type RecommendedModelSelection } from "./recommendation";
-import type { ActiveModelState, AgentModelSlot, CatalogModel, DownloadEvent, DownloadJob, HardwareProbe, LocalInferenceReadiness, ModelAssignments, ModelHubSnapshot, TextGenerationSlot } from "./types";
-import { type VerifyResult } from "./verify";
+import { type LocalModelSearchProviderId, type LocalModelSearchResponse } from "./custom-search.js";
+import { type RecommendedModelSelection } from "./recommendation.js";
+import type { ActiveModelState, AgentModelSlot, CatalogModel, DownloadEvent, DownloadJob, HardwareProbe, LocalInferenceReadiness, ModelAssignments, ModelHubSnapshot, TextGenerationSlot } from "./types.js";
+import { type VerifyResult } from "./verify.js";
 export declare class LocalInferenceService {
     private readonly downloader;
     private readonly activeModel;
@@ -40,14 +40,14 @@ export declare class LocalInferenceService {
      * actually execute.
      */
     private installedBinaryKernels;
-    startDownload(modelIdOrSpec: string | CatalogModel): Promise<DownloadJob>;
+    startDownload(modelId: string): Promise<DownloadJob>;
     startSmallerFallbackDownload(currentModelId: string, slot?: TextGenerationSlot, hardware?: HardwareProbe): Promise<{
         model: CatalogModel;
         job: DownloadJob;
     } | null>;
     searchHuggingFace(query: string, limit?: number): Promise<CatalogModel[]>;
     searchModelHub(query: string, hub: "huggingface" | "modelscope", limit?: number): Promise<CatalogModel[]>;
-    getSearchProviders(): import("./custom-search").LocalModelSearchProviderDescriptor[];
+    getSearchProviders(): never[];
     searchCustomModels(providerId: LocalModelSearchProviderId, query: string, limit?: number): Promise<LocalModelSearchResponse>;
     /**
      * Verify an installed model's file integrity. When the model was a

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

@@ -1 +1 @@
-{"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;AAMlD,OAAO,EACL,KAAK,0BAA0B,EAC/B,KAAK,wBAAwB,EAG9B,MAAM,iBAAiB,CAAC;AAKzB,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;;;;OAIG;IACH,OAAO,CAAC,sBAAsB;IAIxB,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;IAIpB,cAAc,CAClB,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,aAAa,GAAG,YAAY,EACjC,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,YAAY,EAAE,CAAC;IAI1B,kBAAkB;IAIZ,kBAAkB,CACtB,UAAU,EAAE,0BAA0B,EACtC,KAAK,EAAE,MAAM,EACb,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,wBAAwB,CAAC;IAIpC;;;;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,GACd,OAAO,CAAC,gBAAgB,CAAC;IAQtB,WAAW,CAAC,OAAO,EAAE,YAAY,GAAG,IAAI,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAIpE,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"}
+{"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;AAMlD,OAAO,EAGL,KAAK,0BAA0B,EAC/B,KAAK,wBAAwB,EAC9B,MAAM,iBAAiB,CAAC;AAIzB,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;;;;OAIG;IACH,OAAO,CAAC,sBAAsB;IAIxB,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,WAAW,CAAC;IAIpD,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;IAMpB,cAAc,CAClB,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,aAAa,GAAG,YAAY,EACjC,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,YAAY,EAAE,CAAC;IAO1B,kBAAkB;IAIZ,kBAAkB,CACtB,UAAU,EAAE,0BAA0B,EACtC,KAAK,EAAE,MAAM,EACb,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,wBAAwB,CAAC;IAYpC;;;;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,GACd,OAAO,CAAC,gBAAgB,CAAC;IAQtB,WAAW,CAAC,OAAO,EAAE,YAAY,GAAG,IAAI,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAIpE,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"}

package/services/local-inference/service.js

@@ -11,10 +11,9 @@ import { readEffectiveAssignments, setAssignment } from "./assignments.js";
 import { registerBundledModels } from "./bundled-models.js";
 import { MODEL_CATALOG } from "./catalog.js";
 import { filterSettingsDefaultLocalModels } from "./catalog-policy.js";
-import { listLocalModelSearchProviders, searchLocalModelProvider, } from "./custom-search.js";
+import { CUSTOM_MODEL_SEARCH_DISABLED_MESSAGE, getLocalModelSearchProvider, } from "./custom-search.js";
 import { Downloader } from "./downloader.js";
 import { probeHardware } from "./hardware.js";
-import { searchHuggingFaceGguf, searchModelHubGguf } from "./hf-search.js";
 import { buildTextGenerationReadiness } from "./readiness.js";
 import { chooseSmallerFallbackModel, selectRecommendedModelForSlot, selectRecommendedModels, } from "./recommendation.js";
 import { listInstalledModels, removeElizaModel, upsertElizaModel, } from "./registry.js";
@@ -112,8 +111,8 @@ export class LocalInferenceService {
     installedBinaryKernels() {
         return null;
     }
-    async startDownload(modelIdOrSpec) {
-        return this.downloader.start(modelIdOrSpec);
+    async startDownload(modelId) {
+        return this.downloader.start(modelId);
     }
     async startSmallerFallbackDownload(currentModelId, slot = "TEXT_LARGE", hardware) {
         const model = chooseSmallerFallbackModel(currentModelId, hardware ?? (await this.getHardware()), slot, MODEL_CATALOG);
@@ -125,16 +124,28 @@ export class LocalInferenceService {
         };
     }
     async searchHuggingFace(query, limit) {
-        return searchHuggingFaceGguf(query, limit);
+        void query;
+        void limit;
+        return [];
     }
     async searchModelHub(query, hub, limit) {
-        return searchModelHubGguf(query, hub, limit);
+        void query;
+        void hub;
+        void limit;
+        return [];
     }
     getSearchProviders() {
-        return listLocalModelSearchProviders();
+        return [];
     }
     async searchCustomModels(providerId, query, limit) {
-        return searchLocalModelProvider(providerId, query, limit);
+        void query;
+        void limit;
+        const provider = getLocalModelSearchProvider(providerId);
+        return {
+            provider,
+            results: [],
+            unavailableMessage: provider.unavailableMessage ?? CUSTOM_MODEL_SEARCH_DISABLED_MESSAGE,
+        };
     }
     /**
      * Verify an installed model's file integrity. When the model was a