feat(sse): real LLMLingua-2 ONNX compression engine (stable) (#4014)

Integrated into release/v3.8.27. Adjustments before merge: - Synced with the current release tip (was 11 commits behind). - Added the 3 LLMLingua-2 ONNX optional-runtime deps to dependency-allowlist.json (@atjsh/llmlingua-2, @tensorflow/tfjs, js-tiktoken) — the only gate that was red. - socks was allowlisted directly on release (separate fix d7db5c73d; it was declared by #4004 but never allowlisted, leaving check:deps red release-wide). Verified locally: check:deps OK, file-size OK, public-creds OK, provider-consistency 161/232/0, typecheck:core clean, 24/24 LLMLingua tests pass. The only remaining Fast-QG red is the pre-existing #3972 orphan test (request-logger-autorefresh-visibility-3972.test.tsx), which is release-wide and unrelated to this PR.
2026-08-06 23:32:12 +03:00 · 2026-06-16 16:06:18 -03:00
parent d7db5c73d1
commit d691f11230
13 changed files with 1524 additions and 126 deletions
--- a/config/quality/dependency-allowlist.json
+++ b/config/quality/dependency-allowlist.json
@@ -1,6 +1,7 @@
 {
  "_comment": "Allowlist anti-slopsquatting (check-deps.mjs). Toda dep nova exige adicao EXPLICITA aqui apos verificar que e legitima.",
  "allowed": [
+    "@atjsh/llmlingua-2",
    "@aws-sdk/client-bedrock-runtime",
    "@cyclonedx/cyclonedx-npm",
    "@dnd-kit/core",
@@ -18,6 +19,7 @@
    "@stryker-mutator/tap-runner",
    "@swc/helpers",
    "@tailwindcss/postcss",
+    "@tensorflow/tfjs",
    "@testing-library/jest-dom",
    "@testing-library/react",
    "@types/bcryptjs",
@@ -64,6 +66,7 @@
    "ink-text-input",
    "ioredis",
    "jose",
+    "js-tiktoken",
    "js-yaml",
    "jscpd",
    "jsdom",
--- a/open-sse/services/compression/engines/llmlingua/constants.ts
+++ b/open-sse/services/compression/engines/llmlingua/constants.ts
@@ -0,0 +1,55 @@
+/**
+ * LLMLingua real-engine constants — pure data + types.
+ *
+ * NO imports of native deps (transformers.js, onnxruntime, etc). This module is
+ * safe to import from anywhere (main thread, worker, tests) without pulling in
+ * the heavy ONNX runtime.
+ *
+ * The real backend uses `@atjsh/llmlingua-2` (ONNX via `@huggingface/transformers`),
+ * which downloads models from the HuggingFace Hub into a cache dir. Only the two
+ * models PROVEN to work end-to-end are registered here.
+ */
+
+export type LlmlinguaFactory = "WithBERTMultilingual" | "WithXLMRoBERTa";
+
+export interface LlmlinguaModelEntry {
+  /** config value, e.g. "tinybert" */
+  id: string;
+  /** HuggingFace Hub repo id */
+  hfRepo: string;
+  factory: LlmlinguaFactory;
+  dtype: "fp32";
+  /** transformers.js subfolder option; "" for both proven models */
+  subfolder: string;
+  sizeMB: number;
+  label: string;
+}
+
+export const DEFAULT_LLMLINGUA_MODEL = "tinybert";
+
+/** Registry keyed by config `model` value. Only the two PROVEN models. */
+export const LLMLINGUA_MODELS: Record<string, LlmlinguaModelEntry> = {
+  tinybert: {
+    id: "tinybert",
+    hfRepo: "atjsh/llmlingua-2-js-tinybert-meetingbank",
+    factory: "WithBERTMultilingual",
+    dtype: "fp32",
+    subfolder: "",
+    sizeMB: 57,
+    label: "TinyBERT (57MB, fast — default)",
+  },
+  "bert-base": {
+    id: "bert-base",
+    hfRepo: "Arcoldd/llmlingua4j-bert-base-onnx",
+    factory: "WithBERTMultilingual",
+    dtype: "fp32",
+    subfolder: "",
+    sizeMB: 710,
+    label: "BERT-base (710MB, higher accuracy)",
+  },
+};
+
+/** Per-call worker reply timeout → fail-open. */
+export const LLMLINGUA_WORKER_TIMEOUT_MS = 5000;
+/** Terminate the idle worker after this long to free model RAM. */
+export const LLMLINGUA_WORKER_IDLE_MS = 300000;
--- a/open-sse/services/compression/engines/llmlingua/index.ts
+++ b/open-sse/services/compression/engines/llmlingua/index.ts
@@ -7,7 +7,9 @@
 * ## Design
 *
 * ### Backend abstraction
- * `LlmlinguaBackend` is a simple `(text: string) => Promise<string>` contract.
+ * `LlmlinguaBackend` is a `(text: string, opts?: LlmlinguaBackendOptions) =>
+ * Promise<string>` contract (the opts carry model selection / compression rate /
+ * offline model-path override; single-arg fakes remain assignable).
 * Tests inject a fake backend via `setLlmlinguaBackend()`. Production code uses
 * `workerBackend` from `./worker.ts` (a stub today — see that file for the L1
 * VPS-validation follow-up before the real ONNX model is wired).
@@ -42,7 +44,7 @@
 * for the exact spec.
 */

-import { createCompressionStats } from "../../stats.ts";
+import { createCompressionStats, estimateCompressionTokens } from "../../stats.ts";
 import { extractPreservedBlocks } from "../../preservation.ts";
 import type {
  CompressionEngine,
@@ -52,14 +54,22 @@ import type {
 } from "../types.ts";
 import type { CompressionResult } from "../../types.ts";
 import { workerBackend } from "./worker.ts";
+import { LLMLINGUA_MODELS, DEFAULT_LLMLINGUA_MODEL } from "./constants.ts";

 // ─── backend abstraction ──────────────────────────────────────────────────────

+/** Options the real backend needs (model selection + compression rate + offline override). */
+export interface LlmlinguaBackendOptions {
+  model?: string;
+  compressionRate?: number;
+  modelPath?: string;
+}
+
 /**
- * A backend takes a prose text segment and returns a compressed version.
+ * A backend takes a prose text segment (+ optional config) and returns a compressed version.
 * Any rejection or error MUST be caught by the caller; the engine fail-opens.
 */
-export type LlmlinguaBackend = (text: string) => Promise<string>;
+export type LlmlinguaBackend = (text: string, opts?: LlmlinguaBackendOptions) => Promise<string>;

 /** Module-level injectable backend (null = use default production backend). */
 let _backend: LlmlinguaBackend | null = null;
@@ -140,11 +150,12 @@ type MessageLike = {
 */
 async function compressProseText(
  text: string,
-  backend: LlmlinguaBackend
+  backend: LlmlinguaBackend,
+  opts?: LlmlinguaBackendOptions
 ): Promise<{ text: string; didCompress: boolean }> {
  if (!text.trim()) return { text, didCompress: false };
  try {
-    const compressed = await backend(text);
+    const compressed = await backend(text, opts);
    // Accept only if it actually gets shorter (reject no-ops or expansions)
    if (typeof compressed === "string" && compressed.length < text.length) {
      return { text: compressed, didCompress: true };
@@ -165,7 +176,8 @@ async function compressProseText(
 */
 async function compressMessageText(
  text: string,
-  backend: LlmlinguaBackend
+  backend: LlmlinguaBackend,
+  opts?: LlmlinguaBackendOptions
 ): Promise<{ text: string; didCompress: boolean }> {
  const segments = splitProseAndPreserved(text);
  let anyCompressed = false;
@@ -176,7 +188,7 @@ async function compressMessageText(
      // Never send preserved content (code, math, etc.) to the backend
      parts.push(seg.text);
    } else {
-      const { text: out, didCompress } = await compressProseText(seg.text, backend);
+      const { text: out, didCompress } = await compressProseText(seg.text, backend, opts);
      parts.push(out);
      if (didCompress) anyCompressed = true;
    }
@@ -191,7 +203,8 @@ async function compressMessageText(
 */
 async function processMessages(
  messages: MessageLike[],
-  backend: LlmlinguaBackend
+  backend: LlmlinguaBackend,
+  opts?: LlmlinguaBackendOptions
 ): Promise<{ messages: MessageLike[]; compressedCount: number }> {
  let compressedCount = 0;
  const result: MessageLike[] = [];
@@ -205,7 +218,7 @@ async function processMessages(

    try {
      if (typeof msg.content === "string") {
-        const { text, didCompress } = await compressMessageText(msg.content, backend);
+        const { text, didCompress } = await compressMessageText(msg.content, backend, opts);
        if (didCompress) {
          compressedCount++;
          result.push({ ...msg, content: text });
@@ -219,7 +232,8 @@ async function processMessages(
          if (part["type"] === "text" && typeof part["text"] === "string") {
            const { text, didCompress } = await compressMessageText(
              part["text"] as string,
-              backend
+              backend,
+              opts
            );
            if (didCompress) {
              changed = true;
@@ -248,19 +262,65 @@ async function processMessages(
 // ─── config schema ────────────────────────────────────────────────────────────

 const LLMLINGUA_SCHEMA: EngineConfigField[] = [
+  { key: "enabled", type: "boolean", label: "Enabled", defaultValue: true },
  {
-    key: "enabled",
-    type: "boolean",
-    label: "Enabled",
-    defaultValue: true,
+    key: "model",
+    type: "select",
+    label: "Model",
+    defaultValue: DEFAULT_LLMLINGUA_MODEL,
+    options: Object.values(LLMLINGUA_MODELS).map((m) => ({ value: m.id, label: m.label })),
  },
+  {
+    key: "minTokens",
+    type: "number",
+    label: "Min tokens (floor)",
+    defaultValue: 2000,
+    min: 0,
+    max: 100000,
+  },
+  {
+    key: "compressionRate",
+    type: "number",
+    label: "Compression rate (keep ratio)",
+    defaultValue: 0.5,
+    min: 0.1,
+    max: 0.9,
+  },
+  { key: "modelPath", type: "string", label: "Model path (offline override)", defaultValue: "" },
 ];

 function validateLlmlinguaConfig(config: Record<string, unknown>): EngineValidationResult {
  const errors: string[] = [];
+
  if (config["enabled"] !== undefined && typeof config["enabled"] !== "boolean") {
    errors.push("enabled must be a boolean");
  }
+
+  if (config["model"] !== undefined) {
+    const model = config["model"];
+    if (typeof model !== "string" || !(model in LLMLINGUA_MODELS)) {
+      errors.push("model must be one of: " + Object.keys(LLMLINGUA_MODELS).join(", "));
+    }
+  }
+
+  if (config["minTokens"] !== undefined) {
+    const minTokens = config["minTokens"];
+    if (typeof minTokens !== "number" || Number.isNaN(minTokens) || minTokens < 0) {
+      errors.push("minTokens must be a number >= 0");
+    }
+  }
+
+  if (config["compressionRate"] !== undefined) {
+    const rate = config["compressionRate"];
+    if (typeof rate !== "number" || Number.isNaN(rate) || rate < 0.1 || rate > 0.9) {
+      errors.push("compressionRate must be a number between 0.1 and 0.9");
+    }
+  }
+
+  if (config["modelPath"] !== undefined && typeof config["modelPath"] !== "string") {
+    errors.push("modelPath must be a string");
+  }
+
  return { valid: errors.length === 0, errors };
 }

@@ -275,8 +335,8 @@ export const llmlinguaEngine: CompressionEngine = {
    "Async semantic token pruning via LLMLingua-2 (ONNX/worker-thread backend). " +
    "Compresses prose in non-system messages; fenced code blocks and other preserved " +
    "constructs are never altered. Fail-opens on any backend error. Production backend: " +
-    "vendored @atjsh/llmlingua-2 (MobileBERT 99 MB) in a worker thread — see " +
-    "./worker.ts for the L1 follow-up spec (VPS validation required per Hard Rule #18).",
+    "@atjsh/llmlingua-2 (TinyBERT 57 MB default, BERT-base optional) in a worker thread; " +
+    "model lazy-downloaded to DATA_DIR. Optional deps — fail-opens if not installed.",
  icon: "brain",
  targets: ["messages"],
  stackable: true,
@@ -294,7 +354,10 @@ export const llmlinguaEngine: CompressionEngine = {
    inputScope: "messages",
    targetLatencyMs: 200,
    supportsPreview: false,
-    stable: false,
+    // Promoted to stable after VPS validation (2026-06-16): the deployed worker
+    // compressed real prose (209→107 ch, ok=true), and the bundle's walk-up
+    // resolution + optional-deps gate were confirmed against the live install.
+    stable: true,
  },

  /**
@@ -334,12 +397,41 @@ export const llmlinguaEngine: CompressionEngine = {
      return { body, compressed: false, stats: null };
    }

+    // minTokens floor: skip the model entirely on small prompts (avoid paying
+    // model latency when there is little to gain). 0 disables the floor.
+    const minTokens =
+      typeof stepConfig["minTokens"] === "number" ? (stepConfig["minTokens"] as number) : 2000;
+    if (minTokens > 0) {
+      const nonSystemText = (messages as MessageLike[])
+        .filter((m) => m.role !== "system")
+        .map((m) => (typeof m.content === "string" ? m.content : JSON.stringify(m.content ?? "")))
+        .join("\n");
+      if (estimateCompressionTokens(nonSystemText) < minTokens) {
+        // Below the floor — skip compression for this small prompt.
+        return { body, compressed: false, stats: null };
+      }
+    }
+
+    // Backend options threaded from stepConfig (model selection / rate / offline override).
+    const backendOpts: LlmlinguaBackendOptions = {
+      model: typeof stepConfig["model"] === "string" ? (stepConfig["model"] as string) : undefined,
+      compressionRate:
+        typeof stepConfig["compressionRate"] === "number"
+          ? (stepConfig["compressionRate"] as number)
+          : undefined,
+      modelPath:
+        typeof stepConfig["modelPath"] === "string" && stepConfig["modelPath"]
+          ? (stepConfig["modelPath"] as string)
+          : undefined,
+    };
+
    try {
      const backend = resolveBackend();
      const start = performance.now();
      const { messages: newMessages, compressedCount } = await processMessages(
        messages as MessageLike[],
-        backend
+        backend,
+        backendOpts
      );

      if (compressedCount === 0) {
--- a/open-sse/services/compression/engines/llmlingua/modelStore.ts
+++ b/open-sse/services/compression/engines/llmlingua/modelStore.ts
@@ -0,0 +1,67 @@
+/**
+ * LLMLingua model store — thin path/config resolver.
+ *
+ * transformers.js owns the actual model download (from the HuggingFace Hub into
+ * its `cacheDir`). This module only resolves the cache directory, maps config
+ * model ids to registry entries, and configures a transformers.js `env` object
+ * for either Hub download (default) or a local modelPath override.
+ *
+ * Deliberately does NOT import the native `@huggingface/transformers` dep — it
+ * accepts a minimal structural `env` so the heavy runtime stays out of this path.
+ */
+
+import os from "node:os";
+import path from "node:path";
+import fs from "node:fs";
+import {
+  DEFAULT_LLMLINGUA_MODEL,
+  LLMLINGUA_MODELS,
+  type LlmlinguaModelEntry,
+} from "./constants.ts";
+
+/** A minimal structural type for the transformers.js `env` object (avoids importing the native dep here). */
+export interface TransformersEnvLike {
+  cacheDir?: string;
+  localModelPath?: string;
+  allowRemoteModels?: boolean;
+  [key: string]: unknown;
+}
+
+/** Base data dir. Mirrors rtk's getDataDir() at engines/rtk/filterLoader.ts. */
+function getDataDir(): string {
+  return process.env.DATA_DIR || path.join(os.homedir(), ".omniroute");
+}
+
+/** Resolve (and ensure) the model cache dir: `${DATA_DIR}/models/llmlingua`. Mirrors rtk's getDataDir(). */
+export function getLlmlinguaModelCacheDir(): string {
+  const dir = path.join(getDataDir(), "models", "llmlingua");
+  try {
+    fs.mkdirSync(dir, { recursive: true });
+  } catch {
+    // Ignore mkdir errors — fail-open philosophy: transformers.js will surface a
+    // clearer error if the dir is genuinely unusable, and callers fail-open anyway.
+  }
+  return dir;
+}
+
+/** Resolve a config model id to its registry entry; falls back to the default for unknown/empty ids. */
+export function resolveLlmlinguaModel(modelId: string | undefined | null): LlmlinguaModelEntry {
+  if (typeof modelId === "string" && modelId.length > 0 && LLMLINGUA_MODELS[modelId]) {
+    return LLMLINGUA_MODELS[modelId];
+  }
+  return LLMLINGUA_MODELS[DEFAULT_LLMLINGUA_MODEL];
+}
+
+/** Configure a transformers.js `env` for either Hub download (default) or a local modelPath override. */
+export function configureTransformersEnv(
+  env: TransformersEnvLike,
+  opts: { modelPath?: string }
+): void {
+  env.cacheDir = getLlmlinguaModelCacheDir();
+  if (typeof opts.modelPath === "string" && opts.modelPath.length > 0) {
+    env.localModelPath = opts.modelPath;
+    env.allowRemoteModels = false;
+  } else {
+    env.allowRemoteModels = true;
+  }
+}
--- a/open-sse/services/compression/engines/llmlingua/onnxWorker.ts
+++ b/open-sse/services/compression/engines/llmlingua/onnxWorker.ts
@@ -0,0 +1,117 @@
+/**
+ * LLMLingua-2 ONNX worker-thread entry point.
+ *
+ * Runs inside a `worker_threads.Worker` (spawned by `./worker.ts`). The heavy
+ * optional deps (`@atjsh/llmlingua-2`, `@huggingface/transformers`, `js-tiktoken`)
+ * are imported LAZILY/dynamically inside the message handler so this module LOADS
+ * even when those deps are absent. The only static imports are present-by-default
+ * modules: `node:worker_threads`, `./constants.ts`, `./modelStore.ts`.
+ *
+ * Protocol (request → reply over the worker MessageChannel):
+ *   in : { id, text, model, compressionRate, modelPath }
+ *   out: { id, ok: true,  text: <compressed> }   on success
+ *        { id, ok: false, text: <original> }     on ANY failure (fail-open)
+ *
+ * Fail-open contract: missing deps, model download failure, inference error — all
+ * resolve to the ORIGINAL text with `ok:false`. The parent treats either reply as
+ * the value to return, so a failed compression is transparently the original prose.
+ *
+ * Code blocks NEVER reach this worker: the engine (index.ts) tombstones preserved
+ * constructs before calling the backend; this worker sees prose-only segments.
+ */
+
+import { parentPort } from "node:worker_threads";
+import {
+  resolveLlmlinguaModel,
+  configureTransformersEnv,
+  type TransformersEnvLike,
+} from "./modelStore.ts";
+import type { LlmlinguaModelEntry } from "./constants.ts";
+
+/**
+ * Dynamic-import indirection. These four deps are OPTIONAL and not installed by
+ * default, so a static `import(...)` of a literal specifier would make `tsc` fail
+ * with TS2307. Routing the specifier through a runtime variable keeps the module
+ * type-checkable while still loading the dep at runtime when present.
+ */
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+function dynamicImport(specifier: string): Promise<any> {
+  return import(/* @vite-ignore */ specifier);
+}
+
+/** Inbound message shape from the parent. */
+interface WorkerRequest {
+  id: number;
+  text: string;
+  model?: string;
+  compressionRate?: number;
+  modelPath?: string;
+}
+
+/**
+ * Cache of built prompt-compressors keyed by `${factory}:${hfRepo}:${modelPath||""}`.
+ * Values are Promises so concurrent first-calls share one in-flight build; a failed
+ * build deletes its key so a later call can retry.
+ * Typed `any` — the heavy lib has no static types here (no-explicit-any is warn-only).
+ */
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+const compressorCache = new Map<string, Promise<any>>();
+
+function cacheKey(entry: LlmlinguaModelEntry, modelPath?: string): string {
+  return `${entry.factory}:${entry.hfRepo}:${modelPath || ""}`;
+}
+
+/**
+ * Build (or reuse) the LLMLingua-2 prompt compressor for a model entry.
+ * All heavy imports are dynamic so this only runs the deps are actually present.
+ */
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+async function getCompressor(entry: LlmlinguaModelEntry, modelPath?: string): Promise<any> {
+  const { env } = await dynamicImport("@huggingface/transformers");
+  configureTransformersEnv(env as TransformersEnvLike, { modelPath });
+
+  const { LLMLingua2 } = await dynamicImport("@atjsh/llmlingua-2");
+  const { Tiktoken } = await dynamicImport("js-tiktoken/lite");
+  const o200k_base = (await dynamicImport("js-tiktoken/ranks/o200k_base")).default;
+  const oai = new Tiktoken(o200k_base);
+
+  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+  const { promptCompressor } = await (LLMLingua2 as any)[entry.factory](entry.hfRepo, {
+    transformerJSConfig: { device: "cpu", dtype: entry.dtype },
+    oaiTokenizer: oai,
+    modelSpecificOptions: { subfolder: entry.subfolder },
+    // MUST silence — the lib console.logs huge objects otherwise.
+    logger: () => {},
+  });
+
+  return promptCompressor;
+}
+
+if (parentPort) {
+  parentPort.on("message", async (msg: WorkerRequest) => {
+    const { id, text } = msg;
+    try {
+      const entry = resolveLlmlinguaModel(msg.model);
+      const key = cacheKey(entry, msg.modelPath);
+
+      let pending = compressorCache.get(key);
+      if (!pending) {
+        pending = getCompressor(entry, msg.modelPath);
+        compressorCache.set(key, pending);
+        // If the build rejects, evict the key so a later call can retry.
+        pending.catch(() => {
+          compressorCache.delete(key);
+        });
+      }
+
+      const compressor = await pending;
+      const rate = typeof msg.compressionRate === "number" ? msg.compressionRate : 0.5;
+      const out: string = await compressor.compress(text, { rate });
+
+      parentPort!.postMessage({ id, ok: true, text: out });
+    } catch {
+      // Fail-open: ANY error → return the ORIGINAL text with ok:false.
+      parentPort!.postMessage({ id, ok: false, text });
+    }
+  });
+}
--- a/open-sse/services/compression/engines/llmlingua/worker.ts
+++ b/open-sse/services/compression/engines/llmlingua/worker.ts
@@ -1,48 +1,342 @@
 /**
- * LLMLingua-2 worker-thread stub (production path — NOT loaded in tests).
+ * LLMLingua-2 worker-thread backend (production path).
 *
- * PRODUCTION FOLLOW-UP (L1):
- *   Replace the stub body below with the real MobileBERT ONNX inference via
- *   the vendored/pinned `@atjsh/llmlingua-2` package running in a worker thread.
- *   The package must be vendored (pinned) at a specific version and ONNX model
- *   hash verified before loading. The worker MUST remain fail-open — any failure
- *   to load the model, initialise the pipeline, or classify tokens MUST result
- *   in returning the original text unchanged (not throwing to the caller).
+ * Real MobileBERT/BERT ONNX inference via `@atjsh/llmlingua-2` running in a
+ * `worker_threads.Worker` (`./onnxWorker.ts`). The backend stays STRICTLY
+ * fail-open: any failure to resolve deps, spawn the worker, load the model, or
+ * classify tokens returns the ORIGINAL text unchanged — never throws to the caller.
 *
- *   VPS validation (Hard Rule #18): before enabling the real model, deploy to
- *   root@192.168.0.15 and run a documented live test confirming:
- *     (a) prose text is shorter after compression,
- *     (b) a message containing a fenced code block produces identical code bytes,
- *     (c) OOM / missing model → fail-open (original text returned, no crash).
+ * ## Fail-open paths
+ *  1. Optional-deps gate: if any of `@atjsh/llmlingua-2`, `@huggingface/transformers`,
+ *     `@tensorflow/tfjs`, `js-tiktoken` does not resolve, return `text` immediately —
+ *     NO worker spawn. This is the default in CI / most installs (deps are OPTIONAL).
+ *  2. Per-call timeout: first call for a model gets `FIRST_CALL_TIMEOUT_MS` (one-time
+ *     model load); warm calls get `LLMLINGUA_WORKER_TIMEOUT_MS`. On timeout → original
+ *     text (the worker keeps loading and will warm for the next call).
+ *  3. Worker error/exit → resolve all pending with their original text + respawn next.
 *
- *   NEVER apply to code blocks — the caller (index.ts) tombstones code blocks
- *   before calling this backend; this worker sees prose-only segments.
+ * ## Serialization
+ *  ONNX/tfjs are not reentrant — calls are queued FIFO and only one message is
+ *  in-flight at a time (the next is posted after the previous reply or its timeout).
 *
- * CURRENT STATE (stub):
- *   This module exports a LlmlinguaBackend function that always fail-opens
- *   (returns the original text unchanged), so the engine can be registered and
- *   used in stacked pipelines without any ONNX dependency. It will produce
- *   compressed:false for every call, which is the correct safe default until the
- *   real model is wired up.
+ * ## Idle eviction
+ *  After `LLMLINGUA_WORKER_IDLE_MS` with no calls, the worker is terminated and the
+ *  singleton reset (next call respawns). The idle timer is `unref`'d so it never keeps
+ *  the process alive.
+ *
+ * Code blocks NEVER reach this backend — the engine (index.ts) tombstones preserved
+ * constructs first; this backend sees prose-only segments.
+ *
+ * VPS validation (Hard Rule #18): the real model is exercised behind RUN_LLMLINGUA_INT.
 */

+import { Worker } from "node:worker_threads";
+import { createRequire } from "node:module";
+import { fileURLToPath } from "node:url";
+import path from "node:path";
+import fs from "node:fs";
+
+import {
+  LLMLINGUA_WORKER_TIMEOUT_MS,
+  LLMLINGUA_WORKER_IDLE_MS,
+} from "./constants.ts";
+import { resolveLlmlinguaModel } from "./modelStore.ts";
 import type { LlmlinguaBackend } from "./index.ts";

+/** One-time model-load budget on the first call for a given model (tinybert ~2s, bert-base ~27s). */
+const FIRST_CALL_TIMEOUT_MS = 60000;
+
 /**
- * Production worker backend stub.
- *
- * Currently fail-opens unconditionally.  When the real `@atjsh/llmlingua-2`
- * package is vendored and validated on the VPS, replace the body below with
- * the actual worker-thread dispatch:
- *
- * ```ts
- * // 1. Spawn / reuse a worker_threads.Worker running the ONNX pipeline.
- * // 2. Post the text to the worker via MessageChannel.
- * // 3. Await the reply with a per-call timeout (e.g. 5 000 ms).
- * // 4. On any error / timeout → return text (fail-open).
- * ```
+ * Gate probe: `@atjsh/llmlingua-2` is the entry package that declares the others
+ * (`@huggingface/transformers`, `@tensorflow/tfjs`, `js-tiktoken`) as peers. We probe
+ * ONLY it because the peers are ESM-only — e.g. `@huggingface/transformers@3.5.2`'s
+ * `exports` has no `require`/`default` condition, so `require.resolve()` throws
+ * `MODULE_NOT_FOUND` for it even when it is installed and `import()`-able (verified on
+ * the VPS). Gating on all four would therefore always fail-open. The worker still
+ * fail-opens if a peer is genuinely missing at `import()` time.
 */
-export const workerBackend: LlmlinguaBackend = async (text: string): Promise<string> => {
-  // Stub: model not yet wired — fail-open by returning the original text.
-  return text;
+const GATE_DEP = "@atjsh/llmlingua-2";
+
+// ─── optional-deps gate (memoized) ──────────────────────────────────────────────
+
+let _depsAvailable: boolean | null = null;
+
+/** Lazily (and once) check whether the optional LLMLingua dependency stack is installed. */
+function depsAvailable(): boolean {
+  if (_depsAvailable !== null) return _depsAvailable;
+  try {
+    createRequire(import.meta.url).resolve(GATE_DEP);
+    _depsAvailable = true;
+  } catch {
+    _depsAvailable = false;
+  }
+  return _depsAvailable;
+}
+
+// ─── worker reply / queue plumbing ──────────────────────────────────────────────
+
+interface WorkerReply {
+  id: number;
+  ok: boolean;
+  text: string;
+}
+
+interface PendingEntry {
+  resolve: (s: string) => void;
+  timer: NodeJS.Timeout;
+  /** Stored so error/exit/reset handlers can fail-open with the ORIGINAL text. */
+  originalText: string;
+  /** Resolved model id — used to mark the model warm ONLY on a genuine success. */
+  modelKey: string;
+}
+
+interface QueueItem {
+  text: string;
+  opts: Parameters<LlmlinguaBackend>[1];
+  resolve: (s: string) => void;
+}
+
+let worker: Worker | null = null;
+let nextId = 1;
+const pending = new Map<number, PendingEntry>();
+const queue: QueueItem[] = [];
+let busy = false;
+/** Model keys (resolved `id`) that have completed at least one successful load. */
+const warmedModels = new Set<string>();
+let idleTimer: NodeJS.Timeout | null = null;
+
+/**
+ * Resolve the worker entry file across dev and prod.
+ *
+ * Dev: `onnxWorker.ts` sits next to this file and runs via the tsx loader.
+ *
+ * Prod: this module is collapsed into a `.next` chunk and the worker is esbuild'd to
+ * `<appRoot>/open-sse/services/compression/engines/llmlingua/onnxWorker.js`
+ * (scripts/build/prepublish.ts) + kept by the pack-artifact allowlist. The process
+ * `cwd` is NOT the app root (pm2 starts it from `/root`), so cwd-relative resolution
+ * is unreliable — we instead WALK UP from this module's location (`import.meta.url`,
+ * which in the standalone bundle is a real `<appRoot>/.next/...` path) until we find
+ * an ancestor that actually contains the worker at its known relative path. cwd
+ * candidates remain as a last-resort fallback. First existing candidate wins; a `.ts`
+ * choice gets the tsx loader, a `.js` choice runs natively.
+ */
+function resolveWorkerFile(): { workerFile: string; execArgv: string[] } {
+  const moduleDir = path.dirname(fileURLToPath(import.meta.url));
+  const rel = path.join(
+    "open-sse",
+    "services",
+    "compression",
+    "engines",
+    "llmlingua",
+    "onnxWorker.js"
+  );
+
+  // 1. Dev: sibling source/compiled file next to this module.
+  const devTs = path.join(moduleDir, "onnxWorker.ts");
+  if (fs.existsSync(devTs)) return { workerFile: devTs, execArgv: ["--import", "tsx/esm"] };
+  const devJs = path.join(moduleDir, "onnxWorker.js");
+  if (fs.existsSync(devJs)) return { workerFile: devJs, execArgv: [] };
+
+  // 2. Prod: walk up from the bundled module location, then cwd, looking for the
+  //    esbuild'd worker at <root>/open-sse/.../onnxWorker.js.
+  const roots: string[] = [];
+  let dir = moduleDir;
+  for (let i = 0; i < 12; i++) {
+    roots.push(dir);
+    const parent = path.dirname(dir);
+    if (parent === dir) break;
+    dir = parent;
+  }
+  roots.push(process.cwd(), path.join(process.cwd(), "app"));
+  for (const root of roots) {
+    const candidate = path.join(root, rel);
+    if (fs.existsSync(candidate)) return { workerFile: candidate, execArgv: [] };
+  }
+
+  // 3. Nothing found — return the sibling .js path; the spawn will fail-open.
+  return { workerFile: devJs, execArgv: [] };
+}
+
+/** Reset the idle eviction timer; terminates the worker after the idle window. */
+function bumpIdleTimer(): void {
+  if (idleTimer) clearTimeout(idleTimer);
+  idleTimer = setTimeout(() => {
+    resetWorker();
+  }, LLMLINGUA_WORKER_IDLE_MS);
+  // Never keep the process alive just for idle eviction.
+  if (typeof idleTimer.unref === "function") idleTimer.unref();
+}
+
+/** Tear down the worker + all runtime state (fail-open any pending). Next call respawns. */
+function resetWorker(): void {
+  const w = worker;
+  worker = null;
+  if (idleTimer) {
+    clearTimeout(idleTimer);
+    idleTimer = null;
+  }
+  // Fail-open every in-flight call with its ORIGINAL text.
+  for (const [, entry] of pending) {
+    clearTimeout(entry.timer);
+    entry.resolve(entry.originalText);
+  }
+  pending.clear();
+  busy = false;
+  warmedModels.clear();
+  if (w) {
+    try {
+      void w.terminate();
+    } catch {
+      // ignore terminate errors
+    }
+  }
+}
+
+/** Spawn the singleton worker and wire its message/error/exit handlers. */
+function ensureWorker(): Worker {
+  if (worker) return worker;
+
+  const { workerFile, execArgv } = resolveWorkerFile();
+  const w = new Worker(workerFile, { execArgv });
+
+  w.on("message", (reply: WorkerReply) => {
+    const entry = pending.get(reply.id);
+    if (entry) {
+      clearTimeout(entry.timer);
+      pending.delete(reply.id);
+      // Only a genuine success warms the model so later calls use the short timeout.
+      // A timeout/error/fail-open MUST NOT warm it (else a still-loading model would
+      // be starved of its one-time load budget on the next call).
+      if (reply.ok) warmedModels.add(entry.modelKey);
+      // ok:false already carries the ORIGINAL text → resolving with it IS fail-open.
+      entry.resolve(reply.text);
+    }
+    busy = false;
+    pump();
+  });
+
+  const failOpenAndRespawn = () => {
+    // Resolve every pending entry fail-open, then drop the worker so the next call respawns.
+    failAllPending();
+    if (worker === w) worker = null;
+    busy = false;
+  };
+
+  w.on("error", failOpenAndRespawn);
+  w.on("exit", failOpenAndRespawn);
+
+  worker = w;
+  return w;
+}
+
+/** Resolve all pending entries with their stored fail-open value (original text). */
+function failAllPending(): void {
+  for (const [id, entry] of pending) {
+    clearTimeout(entry.timer);
+    pending.delete(id);
+    entry.resolve(entry.originalText);
+  }
+}
+
+/** Post the next queued item to the worker (one in-flight at a time). */
+function pump(): void {
+  if (busy) return;
+  const item = queue.shift();
+  if (!item) return;
+
+  busy = true;
+  bumpIdleTimer();
+
+  let w: Worker;
+  try {
+    w = ensureWorker();
+  } catch {
+    // Spawn failed → fail-open this item and continue draining the queue.
+    busy = false;
+    item.resolve(item.text);
+    pump();
+    return;
+  }
+
+  const id = nextId++;
+  const modelKey = resolveLlmlinguaModel(item.opts?.model).id;
+  const warm = warmedModels.has(modelKey);
+  const timeoutMs = warm ? LLMLINGUA_WORKER_TIMEOUT_MS : FIRST_CALL_TIMEOUT_MS;
+
+  const timer = setTimeout(() => {
+    // Timeout → fail-open with the ORIGINAL text; drop the pending entry but keep the
+    // worker (it may still be loading the model and will warm for the next call).
+    const entry = pending.get(id);
+    if (entry) {
+      pending.delete(id);
+      entry.resolve(item.text);
+    }
+    busy = false;
+    pump();
+  }, timeoutMs);
+  if (typeof timer.unref === "function") timer.unref();
+
+  pending.set(id, {
+    // Warming is decided by the reply handler (success only) — not here, so a
+    // timeout/error fail-open never marks the model warm.
+    resolve: item.resolve,
+    timer,
+    originalText: item.text,
+    modelKey,
+  });
+
+  try {
+    w.postMessage({
+      id,
+      text: item.text,
+      model: item.opts?.model,
+      compressionRate: item.opts?.compressionRate,
+      modelPath: item.opts?.modelPath,
+    });
+  } catch {
+    // postMessage failed → fail-open this item and respawn.
+    clearTimeout(timer);
+    pending.delete(id);
+    item.resolve(item.text);
+    if (worker === w) worker = null;
+    busy = false;
+    pump();
+  }
+}
+
+// ─── public backend ─────────────────────────────────────────────────────────────
+
+/**
+ * Production worker backend. Two-arg `LlmlinguaBackend` (text, opts).
+ *
+ * Returns the compressed prose on success; the ORIGINAL `text` on any failure
+ * (missing deps, spawn error, model-load/inference error, timeout). NEVER throws.
+ */
+export const workerBackend: LlmlinguaBackend = async (text, opts) => {
+  // Fail-open WITHOUT spawning when the optional deps are not installed (the common case).
+  if (!depsAvailable()) {
+    return text;
+  }
+
+  return new Promise<string>((resolve) => {
+    queue.push({ text, opts, resolve });
+    pump();
+  });
 };
+
+// ─── test-only reset ────────────────────────────────────────────────────────────
+
+/**
+ * Internal: terminate the worker (if any) and reset all module state so the
+ * process can exit cleanly after tests. Not part of the public contract.
+ */
+export function __resetLlmlinguaWorkerForTests(): void {
+  // Drain the queue fail-open so no callers hang.
+  while (queue.length) {
+    const item = queue.shift()!;
+    item.resolve(item.text);
+  }
+  resetWorker();
+  _depsAvailable = null;
+  nextId = 1;
+}
--- a/package-lock.json
+++ b/package-lock.json
@@ -17,7 +17,7 @@
        "@dnd-kit/core": "^6.3.1",
        "@dnd-kit/sortable": "^10.0.0",
        "@dnd-kit/utilities": "^3.2.2",
-        "@huggingface/transformers": "^4.2.0",
+        "@huggingface/transformers": "3.5.2",
        "@lobehub/icons": "^5.8.0",
        "@modelcontextprotocol/sdk": "^1.29.0",
        "@monaco-editor/react": "^4.7.0",
@@ -99,7 +99,7 @@
        "@testing-library/react": "^16.3.2",
        "@types/bcryptjs": "^3.0.0",
        "@types/better-sqlite3": "^7.6.13",
-        "@types/bun": "*",
+        "@types/bun": "latest",
        "@types/keytar": "^4.4.2",
        "@types/node": "^25.9.1",
        "@types/react": "^19.2.15",
@@ -137,7 +137,10 @@
        "node": ">=22.0.0 <23 || >=24.0.0 <27"
      },
      "optionalDependencies": {
+        "@atjsh/llmlingua-2": "2.0.3",
+        "@tensorflow/tfjs": "4.22.0",
        "better-sqlite3": "^12.10.0",
+        "js-tiktoken": "^1.0.20",
        "keytar": "^7.9.0",
        "tls-client-node": "^0.2.0",
        "wreq-js": "^2.3.1"
@@ -271,6 +274,21 @@
      "dev": true,
      "license": "MIT"
    },
+    "node_modules/@atjsh/llmlingua-2": {
+      "version": "2.0.3",
+      "resolved": "https://registry.npmjs.org/@atjsh/llmlingua-2/-/llmlingua-2-2.0.3.tgz",
+      "integrity": "sha512-UJJFMbzYldkZ4qX5CrSZtmytOnXf6aXhmr1sBhbpVMHdmQG+7GCnrx5rIwPSOmozXD9KiPv5nnV6pvzxdtHdYQ==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "es-toolkit": "^1.38.0"
+      },
+      "peerDependencies": {
+        "@huggingface/transformers": "*",
+        "@tensorflow/tfjs": "*",
+        "js-tiktoken": "*"
+      }
+    },
    "node_modules/@aws-crypto/crc32": {
      "version": "5.2.0",
      "resolved": "https://registry.npmjs.org/@aws-crypto/crc32/-/crc32-5.2.0.tgz",
@@ -2516,31 +2534,24 @@
      }
    },
    "node_modules/@huggingface/jinja": {
-      "version": "0.5.9",
-      "resolved": "https://registry.npmjs.org/@huggingface/jinja/-/jinja-0.5.9.tgz",
-      "integrity": "sha512-uWTG+l3VJRsl7EXxYizuL3P+cCPoc3cRqbWWRcQN0FhejRfbdq0RNhCmbY/YDtnTcz9icdLYuLDjsnz4d8JMuw==",
+      "version": "0.4.1",
+      "resolved": "https://registry.npmjs.org/@huggingface/jinja/-/jinja-0.4.1.tgz",
+      "integrity": "sha512-3WXbMFaPkk03LRCM0z0sylmn8ddDm4ubjU7X+Hg4M2GOuMklwoGAFXp9V2keq7vltoB/c7McE5aHUVVddAewsw==",
      "license": "MIT",
      "engines": {
        "node": ">=18"
      }
    },
-    "node_modules/@huggingface/tokenizers": {
-      "version": "0.1.3",
-      "resolved": "https://registry.npmjs.org/@huggingface/tokenizers/-/tokenizers-0.1.3.tgz",
-      "integrity": "sha512-8rF/RRT10u+kn7YuUbUg0OF30K8rjTc78aHpxT+qJ1uWSqxT1MHi8+9ltwYfkFYJzT/oS+qw3JVfHtNMGAdqyA==",
-      "license": "Apache-2.0"
-    },
    "node_modules/@huggingface/transformers": {
-      "version": "4.2.0",
-      "resolved": "https://registry.npmjs.org/@huggingface/transformers/-/transformers-4.2.0.tgz",
-      "integrity": "sha512-8BRCoBMH0XsWaEIamuR0LrJGAfftgHAfb2Vrffy0VKlSAE/MnUJ5/h/zTfEP3fDIft+nk7TqB8xXEyABGitBjQ==",
+      "version": "3.5.2",
+      "resolved": "https://registry.npmjs.org/@huggingface/transformers/-/transformers-3.5.2.tgz",
+      "integrity": "sha512-mfRXkmcL99+ibpjM++pvZmc2h3po8i1ZgSRI5Rtgh++P15GU0lY8UQteYt/w5V+GQw+Jpao93MoipcePzh3mKg==",
      "license": "Apache-2.0",
      "dependencies": {
-        "@huggingface/jinja": "^0.5.6",
-        "@huggingface/tokenizers": "^0.1.3",
-        "onnxruntime-node": "1.24.3",
-        "onnxruntime-web": "1.26.0-dev.20260416-b7804b056c",
-        "sharp": "^0.34.5"
+        "@huggingface/jinja": "^0.4.1",
+        "onnxruntime-node": "1.21.0",
+        "onnxruntime-web": "1.22.0-dev.20250409-89f8206ba4",
+        "sharp": "^0.34.1"
      }
    },
    "node_modules/@humanfs/core": {
@@ -3510,7 +3521,6 @@
      "version": "4.0.1",
      "resolved": "https://registry.npmjs.org/@isaacs/fs-minipass/-/fs-minipass-4.0.1.tgz",
      "integrity": "sha512-wgm9Ehl2jpeqP3zw/7mo3kRHFp5MEDhqAdwy1fTGkHAwnkGOVsgpvQhL8B5n1qlb01jV3n/bI0ZfZp5lWA1k4w==",
-      "dev": true,
      "license": "ISC",
      "dependencies": {
        "minipass": "^7.0.4"
@@ -6109,12 +6119,6 @@
      "integrity": "sha512-Ddb+kVXlXst9d+R9PfTIxh1EdNkgoRe5tOX6t01f1lYWOvJnSPDBlG241QLzcyPdoNTsblLUdujGSE4RzrTZGQ==",
      "license": "BSD-3-Clause"
    },
-    "node_modules/@protobufjs/inquire": {
-      "version": "1.1.2",
-      "resolved": "https://registry.npmjs.org/@protobufjs/inquire/-/inquire-1.1.2.tgz",
-      "integrity": "sha512-pa0vFRuws4wkvaXKK1uXZMAwAX4/t8ANaJo45iw/oQHNQ9q5xUzwgFmVJGXiga2BeN+zpX7Vf9vmsiIa2J+MUw==",
-      "license": "BSD-3-Clause"
-    },
    "node_modules/@protobufjs/path": {
      "version": "1.1.2",
      "resolved": "https://registry.npmjs.org/@protobufjs/path/-/path-1.1.2.tgz",
@@ -8402,6 +8406,241 @@
        "tailwindcss": "4.3.0"
      }
    },
+    "node_modules/@tensorflow/tfjs": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs/-/tfjs-4.22.0.tgz",
+      "integrity": "sha512-0TrIrXs6/b7FLhLVNmfh8Sah6JgjBPH4mZ8JGb7NU6WW+cx00qK5BcAZxw7NCzxj6N8MRAIfHq+oNbPUNG5VAg==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "dependencies": {
+        "@tensorflow/tfjs-backend-cpu": "4.22.0",
+        "@tensorflow/tfjs-backend-webgl": "4.22.0",
+        "@tensorflow/tfjs-converter": "4.22.0",
+        "@tensorflow/tfjs-core": "4.22.0",
+        "@tensorflow/tfjs-data": "4.22.0",
+        "@tensorflow/tfjs-layers": "4.22.0",
+        "argparse": "^1.0.10",
+        "chalk": "^4.1.0",
+        "core-js": "3.29.1",
+        "regenerator-runtime": "^0.13.5",
+        "yargs": "^16.0.3"
+      },
+      "bin": {
+        "tfjs-custom-module": "dist/tools/custom_module/cli.js"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-backend-cpu": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-backend-cpu/-/tfjs-backend-cpu-4.22.0.tgz",
+      "integrity": "sha512-1u0FmuLGuRAi8D2c3cocHTASGXOmHc/4OvoVDENJayjYkS119fcTcQf4iHrtLthWyDIPy3JiPhRrZQC9EwnhLw==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "dependencies": {
+        "@types/seedrandom": "^2.4.28",
+        "seedrandom": "^3.0.5"
+      },
+      "engines": {
+        "yarn": ">= 1.3.2"
+      },
+      "peerDependencies": {
+        "@tensorflow/tfjs-core": "4.22.0"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-backend-webgl": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-backend-webgl/-/tfjs-backend-webgl-4.22.0.tgz",
+      "integrity": "sha512-H535XtZWnWgNwSzv538czjVlbJebDl5QTMOth4RXr2p/kJ1qSIXE0vZvEtO+5EC9b00SvhplECny2yDewQb/Yg==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "dependencies": {
+        "@tensorflow/tfjs-backend-cpu": "4.22.0",
+        "@types/offscreencanvas": "~2019.3.0",
+        "@types/seedrandom": "^2.4.28",
+        "seedrandom": "^3.0.5"
+      },
+      "engines": {
+        "yarn": ">= 1.3.2"
+      },
+      "peerDependencies": {
+        "@tensorflow/tfjs-core": "4.22.0"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-converter": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-converter/-/tfjs-converter-4.22.0.tgz",
+      "integrity": "sha512-PT43MGlnzIo+YfbsjM79Lxk9lOq6uUwZuCc8rrp0hfpLjF6Jv8jS84u2jFb+WpUeuF4K33ZDNx8CjiYrGQ2trQ==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "peerDependencies": {
+        "@tensorflow/tfjs-core": "4.22.0"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-core": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-core/-/tfjs-core-4.22.0.tgz",
+      "integrity": "sha512-LEkOyzbknKFoWUwfkr59vSB68DMJ4cjwwHgicXN0DUi3a0Vh1Er3JQqCI1Hl86GGZQvY8ezVrtDIvqR1ZFW55A==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "dependencies": {
+        "@types/long": "^4.0.1",
+        "@types/offscreencanvas": "~2019.7.0",
+        "@types/seedrandom": "^2.4.28",
+        "@webgpu/types": "0.1.38",
+        "long": "4.0.0",
+        "node-fetch": "~2.6.1",
+        "seedrandom": "^3.0.5"
+      },
+      "engines": {
+        "yarn": ">= 1.3.2"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-core/node_modules/@types/offscreencanvas": {
+      "version": "2019.7.3",
+      "resolved": "https://registry.npmjs.org/@types/offscreencanvas/-/offscreencanvas-2019.7.3.tgz",
+      "integrity": "sha512-ieXiYmgSRXUDeOntE1InxjWyvEelZGP63M+cGuquuRLuIKKT1osnkXjxev9B7d1nXSug5vpunx+gNlbVxMlC9A==",
+      "license": "MIT",
+      "optional": true
+    },
+    "node_modules/@tensorflow/tfjs-data": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-data/-/tfjs-data-4.22.0.tgz",
+      "integrity": "sha512-dYmF3LihQIGvtgJrt382hSRH4S0QuAp2w1hXJI2+kOaEqo5HnUPG0k5KA6va+S1yUhx7UBToUKCBHeLHFQRV4w==",
+      "license": "Apache-2.0",
+      "optional": true,
+      "dependencies": {
+        "@types/node-fetch": "^2.1.2",
+        "node-fetch": "~2.6.1",
+        "string_decoder": "^1.3.0"
+      },
+      "peerDependencies": {
+        "@tensorflow/tfjs-core": "4.22.0",
+        "seedrandom": "^3.0.5"
+      }
+    },
+    "node_modules/@tensorflow/tfjs-layers": {
+      "version": "4.22.0",
+      "resolved": "https://registry.npmjs.org/@tensorflow/tfjs-layers/-/tfjs-layers-4.22.0.tgz",
+      "integrity": "sha512-lybPj4ZNj9iIAPUj7a8ZW1hg8KQGfqWLlCZDi9eM/oNKCCAgchiyzx8OrYoWmRrB+AM6VNEeIT+2gZKg5ReihA==",
+      "license": "Apache-2.0 AND MIT",
+      "optional": true,
+      "peerDependencies": {
+        "@tensorflow/tfjs-core": "4.22.0"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/argparse": {
+      "version": "1.0.10",
+      "resolved": "https://registry.npmjs.org/argparse/-/argparse-1.0.10.tgz",
+      "integrity": "sha512-o5Roy6tNG4SL/FOkCAN6RzjiakZS25RLYFrcMttJqbdd8BWrnA+fGz57iN5Pb06pvBGvl5gQ0B48dJlslXvoTg==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "sprintf-js": "~1.0.2"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/cliui": {
+      "version": "7.0.4",
+      "resolved": "https://registry.npmjs.org/cliui/-/cliui-7.0.4.tgz",
+      "integrity": "sha512-OcRE68cOsVMXp1Yvonl/fzkQOyjLSu/8bhPDfQt0e0/Eb283TKP20Fs2MqoPsr9SwA595rRCA+QMzYc9nBP+JQ==",
+      "license": "ISC",
+      "optional": true,
+      "dependencies": {
+        "string-width": "^4.2.0",
+        "strip-ansi": "^6.0.0",
+        "wrap-ansi": "^7.0.0"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/emoji-regex": {
+      "version": "8.0.0",
+      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
+      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
+      "license": "MIT",
+      "optional": true
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/is-fullwidth-code-point": {
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
+      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
+      "license": "MIT",
+      "optional": true,
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/string-width": {
+      "version": "4.2.3",
+      "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
+      "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "emoji-regex": "^8.0.0",
+        "is-fullwidth-code-point": "^3.0.0",
+        "strip-ansi": "^6.0.1"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/strip-ansi": {
+      "version": "6.0.1",
+      "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
+      "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "ansi-regex": "^5.0.1"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/wrap-ansi": {
+      "version": "7.0.0",
+      "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-7.0.0.tgz",
+      "integrity": "sha512-YVGIj2kamLSTxw6NsZjoBxfSwsn0ycdesmc4p+Q21c5zPuZ1pl+NfxVdxPtdHvmNVOQ6XSYG4AUtyt/Fi7D16Q==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "ansi-styles": "^4.0.0",
+        "string-width": "^4.1.0",
+        "strip-ansi": "^6.0.0"
+      },
+      "engines": {
+        "node": ">=10"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/wrap-ansi?sponsor=1"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/yargs": {
+      "version": "16.2.0",
+      "resolved": "https://registry.npmjs.org/yargs/-/yargs-16.2.0.tgz",
+      "integrity": "sha512-D1mvvtDG0L5ft/jGWkLpG1+m0eQxOfaBvTNELraWj22wSVUMWxZUvYgJYcKh6jGGIkJFhH4IZPQhR4TKpc8mBw==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "cliui": "^7.0.2",
+        "escalade": "^3.1.1",
+        "get-caller-file": "^2.0.5",
+        "require-directory": "^2.1.1",
+        "string-width": "^4.2.0",
+        "y18n": "^5.0.5",
+        "yargs-parser": "^20.2.2"
+      },
+      "engines": {
+        "node": ">=10"
+      }
+    },
+    "node_modules/@tensorflow/tfjs/node_modules/yargs-parser": {
+      "version": "20.2.9",
+      "resolved": "https://registry.npmjs.org/yargs-parser/-/yargs-parser-20.2.9.tgz",
+      "integrity": "sha512-y11nGElTIV+CT3Zv9t7VKl+Q3hTQoT9a1Qzezhhl6Rp21gJ/IVTW7Z3y9EWXhuUBC2Shnf+DX0antecpAwSP8w==",
+      "license": "ISC",
+      "optional": true,
+      "engines": {
+        "node": ">=10"
+      }
+    },
    "node_modules/@testing-library/jest-dom": {
      "version": "6.9.1",
      "resolved": "https://registry.npmjs.org/@testing-library/jest-dom/-/jest-dom-6.9.1.tgz",
@@ -8904,6 +9143,13 @@
        "keytar": "*"
      }
    },
+    "node_modules/@types/long": {
+      "version": "4.0.2",
+      "resolved": "https://registry.npmjs.org/@types/long/-/long-4.0.2.tgz",
+      "integrity": "sha512-MqTGEo5bj5t157U6fA/BiDynNkn0YknVdh48CMPkTSpFTVmvao5UQmm7uEF6xBEo7qIMAlY/JSleYaE6VOdpaA==",
+      "license": "MIT",
+      "optional": true
+    },
    "node_modules/@types/mdast": {
      "version": "4.0.4",
      "resolved": "https://registry.npmjs.org/@types/mdast/-/mdast-4.0.4.tgz",
@@ -8934,6 +9180,24 @@
        "undici-types": ">=7.24.0 <7.24.7"
      }
    },
+    "node_modules/@types/node-fetch": {
+      "version": "2.6.13",
+      "resolved": "https://registry.npmjs.org/@types/node-fetch/-/node-fetch-2.6.13.tgz",
+      "integrity": "sha512-QGpRVpzSaUs30JBSGPjOg4Uveu384erbHBoT1zeONvyCfwQxIkUshLAOqN/k9EjGviPRmWTTe6aH2qySWKTVSw==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@types/node": "*",
+        "form-data": "^4.0.4"
+      }
+    },
+    "node_modules/@types/offscreencanvas": {
+      "version": "2019.3.0",
+      "resolved": "https://registry.npmjs.org/@types/offscreencanvas/-/offscreencanvas-2019.3.0.tgz",
+      "integrity": "sha512-esIJx9bQg+QYF0ra8GnvfianIY8qWB0GBx54PK5Eps6m+xTj86KLavHv6qDhzKcu5UUOgNfJ2pWaIIV7TRUd9Q==",
+      "license": "MIT",
+      "optional": true
+    },
    "node_modules/@types/parse-json": {
      "version": "4.0.2",
      "resolved": "https://registry.npmjs.org/@types/parse-json/-/parse-json-4.0.2.tgz",
@@ -8967,6 +9231,13 @@
      "dev": true,
      "license": "MIT"
    },
+    "node_modules/@types/seedrandom": {
+      "version": "2.4.34",
+      "resolved": "https://registry.npmjs.org/@types/seedrandom/-/seedrandom-2.4.34.tgz",
+      "integrity": "sha512-ytDiArvrn/3Xk6/vtylys5tlY6eo7Ane0hvcx++TKo6RxQXuVfW0AF/oeWqAj9dN29SyhtawuXstgmPlwNcv/A==",
+      "license": "MIT",
+      "optional": true
+    },
    "node_modules/@types/trusted-types": {
      "version": "2.0.7",
      "resolved": "https://registry.npmjs.org/@types/trusted-types/-/trusted-types-2.0.7.tgz",
@@ -9722,6 +9993,13 @@
      "dev": true,
      "license": "MIT"
    },
+    "node_modules/@webgpu/types": {
+      "version": "0.1.38",
+      "resolved": "https://registry.npmjs.org/@webgpu/types/-/types-0.1.38.tgz",
+      "integrity": "sha512-7LrhVKz2PRh+DD7+S+PVaFd5HxaWQvoMqBbsV9fNJO1pjUs1P8bM2vQVNfk+3URTqbuTI7gkXi0rfsN0IadoBA==",
+      "license": "BSD-3-Clause",
+      "optional": true
+    },
    "node_modules/@xyflow/react": {
      "version": "12.11.0",
      "resolved": "https://registry.npmjs.org/@xyflow/react/-/react-12.11.0.tgz",
@@ -9874,15 +10152,6 @@
        "acorn": "^6.0.0 || ^7.0.0 || ^8.0.0"
      }
    },
-    "node_modules/adm-zip": {
-      "version": "0.5.17",
-      "resolved": "https://registry.npmjs.org/adm-zip/-/adm-zip-0.5.17.tgz",
-      "integrity": "sha512-+Ut8d9LLqwEvHHJl1+PIHqoyDxFgVN847JTVM3Izi3xHDWPE4UtzzXysMZQs64DMcrJfBeS/uoEP4AD3HQHnQQ==",
-      "license": "MIT",
-      "engines": {
-        "node": ">=12.0"
-      }
-    },
    "node_modules/agent-base": {
      "version": "9.0.0",
      "resolved": "https://registry.npmjs.org/agent-base/-/agent-base-9.0.0.tgz",
@@ -11919,6 +12188,18 @@
        "node": ">=6.6.0"
      }
    },
+    "node_modules/core-js": {
+      "version": "3.29.1",
+      "resolved": "https://registry.npmjs.org/core-js/-/core-js-3.29.1.tgz",
+      "integrity": "sha512-+jwgnhg6cQxKYIIjGtAHq2nwUOolo9eoFZ4sHfUH09BLXBgxnH4gA0zEd+t+BO2cNB8idaBtZFcFTRjQJRJmAw==",
+      "hasInstallScript": true,
+      "license": "MIT",
+      "optional": true,
+      "funding": {
+        "type": "opencollective",
+        "url": "https://opencollective.com/core-js"
+      }
+    },
    "node_modules/cors": {
      "version": "2.8.6",
      "resolved": "https://registry.npmjs.org/cors/-/cors-2.8.6.tgz",
@@ -15518,9 +15799,9 @@
      }
    },
    "node_modules/global-agent/node_modules/semver": {
-      "version": "7.8.1",
-      "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.1.tgz",
-      "integrity": "sha512-rkVq3IXh+4FDGch+KwzX3aV9W3kO54GyEgpvBzSyctDA6Xtd7RJQV1xmXbeQp5v7+VzLOfVqiutSE6GICgPFvg==",
+      "version": "7.8.4",
+      "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.4.tgz",
+      "integrity": "sha512-rUCObTnP32Q08R2uuIrt7r9PlEonuTmtuXYcW6s5kjdlj3xbnwe+21yXptAUYcMAABLkYYTtnmzb3w3EDZfueA==",
      "license": "ISC",
      "bin": {
        "semver": "bin/semver.js"
@@ -17573,6 +17854,16 @@
      "dev": true,
      "license": "MIT"
    },
+    "node_modules/js-tiktoken": {
+      "version": "1.0.21",
+      "resolved": "https://registry.npmjs.org/js-tiktoken/-/js-tiktoken-1.0.21.tgz",
+      "integrity": "sha512-biOj/6M5qdgx5TKjDnFT1ymSpM5tbd3ylwDtrQvFQSu0Z7bBYko2dF+W/aUkXUPuk6IVpRxk/3Q2sHOzGlS36g==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "base64-js": "^1.5.1"
+      }
+    },
    "node_modules/js-tokens": {
      "version": "4.0.0",
      "resolved": "https://registry.npmjs.org/js-tokens/-/js-tokens-4.0.0.tgz",
@@ -19054,10 +19345,11 @@
      }
    },
    "node_modules/long": {
-      "version": "5.3.2",
-      "resolved": "https://registry.npmjs.org/long/-/long-5.3.2.tgz",
-      "integrity": "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA==",
-      "license": "Apache-2.0"
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/long/-/long-4.0.0.tgz",
+      "integrity": "sha512-XsP+KhQif4bjX1kbuSiySJFNAehNxgLb6hPRGJ9QsUr8ajHkuXGdrHmFUTUUXhDwVX2R5bY4JNZEwbUiMhV+MA==",
+      "license": "Apache-2.0",
+      "optional": true
    },
    "node_modules/longest-streak": {
      "version": "3.1.0",
@@ -20514,7 +20806,6 @@
      "version": "7.1.3",
      "resolved": "https://registry.npmjs.org/minipass/-/minipass-7.1.3.tgz",
      "integrity": "sha512-tEBHqDnIoM/1rXME1zgka9g6Q2lcoCkxHLuc7ODJ5BxbP5d4c2Z5cGgtXAku59200Cx7diuHTOYfSBD8n6mm8A==",
-      "dev": true,
      "license": "BlueOak-1.0.0",
      "engines": {
        "node": ">=16 || 14 >=14.17"
@@ -20634,7 +20925,6 @@
      "version": "3.1.0",
      "resolved": "https://registry.npmjs.org/minizlib/-/minizlib-3.1.0.tgz",
      "integrity": "sha512-KZxYo1BUkWD2TVFLr0MQoM8vUUigWD3LlD83a/75BqC+4qE0Hb1Vo5v1FgcfaNXvfXzr+5EhQ6ing/CaBijTlw==",
-      "dev": true,
      "license": "MIT",
      "dependencies": {
        "minipass": "^7.1.2"
@@ -21093,6 +21383,52 @@
        "url": "https://github.com/sponsors/ljharb"
      }
    },
+    "node_modules/node-fetch": {
+      "version": "2.6.13",
+      "resolved": "https://registry.npmjs.org/node-fetch/-/node-fetch-2.6.13.tgz",
+      "integrity": "sha512-StxNAxh15zr77QvvkmveSQ8uCQ4+v5FkvNTj0OESmiHu+VRi/gXArXtkWMElOsOUNLtUEvI4yS+rdtOHZTwlQA==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "whatwg-url": "^5.0.0"
+      },
+      "engines": {
+        "node": "4.x || >=6.0.0"
+      },
+      "peerDependencies": {
+        "encoding": "^0.1.0"
+      },
+      "peerDependenciesMeta": {
+        "encoding": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/node-fetch/node_modules/tr46": {
+      "version": "0.0.3",
+      "resolved": "https://registry.npmjs.org/tr46/-/tr46-0.0.3.tgz",
+      "integrity": "sha512-N3WMsuqV66lT30CrXNbEjx4GEwlow3v6rr4mCcv6prnfwhS01rkgyFdjPNBYd9br7LpXV1+Emh01fHnq2Gdgrw==",
+      "license": "MIT",
+      "optional": true
+    },
+    "node_modules/node-fetch/node_modules/webidl-conversions": {
+      "version": "3.0.1",
+      "resolved": "https://registry.npmjs.org/webidl-conversions/-/webidl-conversions-3.0.1.tgz",
+      "integrity": "sha512-2JAn3z8AR6rjK8Sm8orRC0h/bcl/DqL7tRPdGZ4I1CjdF+EaMLmYxBHyXuKL849eucPFhvBoxMsflfOb8kxaeQ==",
+      "license": "BSD-2-Clause",
+      "optional": true
+    },
+    "node_modules/node-fetch/node_modules/whatwg-url": {
+      "version": "5.0.0",
+      "resolved": "https://registry.npmjs.org/whatwg-url/-/whatwg-url-5.0.0.tgz",
+      "integrity": "sha512-saE57nupxk6v3HY35+jzBwYa0rKSy0XR8JSxZPwgLr7ys0IBzhGviA1/TUGJLmSVqs8pb9AnvICXEuOHLprYTw==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tr46": "~0.0.3",
+        "webidl-conversions": "^3.0.0"
+      }
+    },
    "node_modules/node-gyp": {
      "version": "12.4.0",
      "resolved": "https://registry.npmjs.org/node-gyp/-/node-gyp-12.4.0.tgz",
@@ -21652,15 +21988,15 @@
      }
    },
    "node_modules/onnxruntime-common": {
-      "version": "1.24.3",
-      "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.24.3.tgz",
-      "integrity": "sha512-GeuPZO6U/LBJXvwdaqHbuUmoXiEdeCjWi/EG7Y1HNnDwJYuk6WUbNXpF6luSUY8yASul3cmUlLGrCCL1ZgVXqA==",
+      "version": "1.21.0",
+      "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.21.0.tgz",
+      "integrity": "sha512-Q632iLLrtCAVOTO65dh2+mNbQir/QNTVBG3h/QdZBpns7mZ0RYbLRBgGABPbpU9351AgYy7SJf1WaeVwMrBFPQ==",
      "license": "MIT"
    },
    "node_modules/onnxruntime-node": {
-      "version": "1.24.3",
-      "resolved": "https://registry.npmjs.org/onnxruntime-node/-/onnxruntime-node-1.24.3.tgz",
-      "integrity": "sha512-JH7+czbc8ALA819vlTgcV+Q214/+VjGeBHDjX81+ZCD0PCVCIFGFNtT0V4sXG/1JXypKPgScQcB3ij/hk3YnTg==",
+      "version": "1.21.0",
+      "resolved": "https://registry.npmjs.org/onnxruntime-node/-/onnxruntime-node-1.21.0.tgz",
+      "integrity": "sha512-NeaCX6WW2L8cRCSqy3bInlo5ojjQqu2fD3D+9W5qb5irwxhEyWKXeH2vZ8W9r6VxaMPUan+4/7NDwZMtouZxEw==",
      "hasInstallScript": true,
      "license": "MIT",
      "os": [
@@ -21669,29 +22005,35 @@
        "linux"
      ],
      "dependencies": {
-        "adm-zip": "^0.5.16",
        "global-agent": "^3.0.0",
-        "onnxruntime-common": "1.24.3"
+        "onnxruntime-common": "1.21.0",
+        "tar": "^7.0.1"
      }
    },
    "node_modules/onnxruntime-web": {
-      "version": "1.26.0-dev.20260416-b7804b056c",
-      "resolved": "https://registry.npmjs.org/onnxruntime-web/-/onnxruntime-web-1.26.0-dev.20260416-b7804b056c.tgz",
-      "integrity": "sha512-MD6Ss4GSpQBo6zqoJzyT9LRbKYs7x/JVN23FT24EcEvlqF4VuzPOeH6X38orZPKHQDbprn7K+SBpu0/mj2CQiw==",
+      "version": "1.22.0-dev.20250409-89f8206ba4",
+      "resolved": "https://registry.npmjs.org/onnxruntime-web/-/onnxruntime-web-1.22.0-dev.20250409-89f8206ba4.tgz",
+      "integrity": "sha512-0uS76OPgH0hWCPrFKlL8kYVV7ckM7t/36HfbgoFw6Nd0CZVVbQC4PkrR8mBX8LtNUFZO25IQBqV2Hx2ho3FlbQ==",
      "license": "MIT",
      "dependencies": {
        "flatbuffers": "^25.1.24",
        "guid-typescript": "^1.0.9",
        "long": "^5.2.3",
-        "onnxruntime-common": "1.24.0-dev.20251116-b39e144322",
+        "onnxruntime-common": "1.22.0-dev.20250409-89f8206ba4",
        "platform": "^1.3.6",
        "protobufjs": "^7.2.4"
      }
    },
+    "node_modules/onnxruntime-web/node_modules/long": {
+      "version": "5.3.2",
+      "resolved": "https://registry.npmjs.org/long/-/long-5.3.2.tgz",
+      "integrity": "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA==",
+      "license": "Apache-2.0"
+    },
    "node_modules/onnxruntime-web/node_modules/onnxruntime-common": {
-      "version": "1.24.0-dev.20251116-b39e144322",
-      "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.24.0-dev.20251116-b39e144322.tgz",
-      "integrity": "sha512-BOoomdHYmNRL5r4iQ4bMvsl2t0/hzVQ3OM3PHD0gxeXu1PmggqBv3puZicEUVOA3AtHHYmqZtjMj9FOfGrATTw==",
+      "version": "1.22.0-dev.20250409-89f8206ba4",
+      "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.22.0-dev.20250409-89f8206ba4.tgz",
+      "integrity": "sha512-vDJMkfCfb0b1A836rgHj+ORuZf4B4+cc2bASQtpeoJLueuFc5DuYwjIZUBrSvx/fO5IrLjLz+oTrB3pcGlhovQ==",
      "license": "MIT"
    },
    "node_modules/open": {
@@ -22688,9 +23030,9 @@
      "license": "ISC"
    },
    "node_modules/protobufjs": {
-      "version": "7.6.1",
-      "resolved": "https://registry.npmjs.org/protobufjs/-/protobufjs-7.6.1.tgz",
-      "integrity": "sha512-4K0myLaWL5EteuSAro91EGFgcfVgxb64Jx+7oDAY6GOkXD4M69yuSEljNcInGVCA5sOPxmZ/EqDLj2x0Q0+Ygg==",
+      "version": "7.6.4",
+      "resolved": "https://registry.npmjs.org/protobufjs/-/protobufjs-7.6.4.tgz",
+      "integrity": "sha512-RJJPTTpvFfHcWLkIa2JFWK4XvtSzS0yEWDmunqHXli1h3JlkbcQZXDZdcWxv+JK3Xsl5/UFDPZ0iGm7DAengYw==",
      "hasInstallScript": true,
      "license": "BSD-3-Clause",
      "dependencies": {
@@ -22700,7 +23042,6 @@
        "@protobufjs/eventemitter": "^1.1.1",
        "@protobufjs/fetch": "^1.1.1",
        "@protobufjs/float": "^1.0.2",
-        "@protobufjs/inquire": "^1.1.2",
        "@protobufjs/path": "^1.1.2",
        "@protobufjs/pool": "^1.1.0",
        "@protobufjs/utf8": "^1.1.1",
@@ -22711,6 +23052,12 @@
        "node": ">=12.0.0"
      }
    },
+    "node_modules/protobufjs/node_modules/long": {
+      "version": "5.3.2",
+      "resolved": "https://registry.npmjs.org/long/-/long-5.3.2.tgz",
+      "integrity": "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA==",
+      "license": "Apache-2.0"
+    },
    "node_modules/proxy-addr": {
      "version": "2.0.7",
      "resolved": "https://registry.npmjs.org/proxy-addr/-/proxy-addr-2.0.7.tgz",
@@ -23476,6 +23823,13 @@
        "url": "https://github.com/sponsors/ljharb"
      }
    },
+    "node_modules/regenerator-runtime": {
+      "version": "0.13.11",
+      "resolved": "https://registry.npmjs.org/regenerator-runtime/-/regenerator-runtime-0.13.11.tgz",
+      "integrity": "sha512-kY1AZVr2Ra+t+piVaJ4gxaFaReZVH40AKNo7UCX6W+dEwBo/2oZJzqfuN1qLq1oL45o56cPaTXELwrTh8Fpggg==",
+      "license": "MIT",
+      "optional": true
+    },
    "node_modules/regex": {
      "version": "6.1.0",
      "resolved": "https://registry.npmjs.org/regex/-/regex-6.1.0.tgz",
@@ -24134,6 +24488,13 @@
      ],
      "license": "BSD-3-Clause"
    },
+    "node_modules/seedrandom": {
+      "version": "3.0.5",
+      "resolved": "https://registry.npmjs.org/seedrandom/-/seedrandom-3.0.5.tgz",
+      "integrity": "sha512-8OwmbklUNzwezjGInmZ+2clQmExQPvomqjL7LFqOYqtmuxRgQYqOD3mHaU+MvZn5FLUeVxVfQjwLZW/n/JFuqg==",
+      "license": "MIT",
+      "optional": true
+    },
    "node_modules/selfsigned": {
      "version": "5.5.0",
      "resolved": "https://registry.npmjs.org/selfsigned/-/selfsigned-5.5.0.tgz",
@@ -25583,7 +25944,6 @@
      "version": "7.5.16",
      "resolved": "https://registry.npmjs.org/tar/-/tar-7.5.16.tgz",
      "integrity": "sha512-56adEpPMouktRlBLXiaYFFzZ/3+JXa8P9n7WbR+ibIjtviN55mEaOkiysCnPnWm+7kkui1Dn8J9l+g6zV8731w==",
-      "dev": true,
      "license": "BlueOak-1.0.0",
      "dependencies": {
        "@isaacs/fs-minipass": "^4.0.0",
@@ -25630,7 +25990,6 @@
      "version": "3.0.0",
      "resolved": "https://registry.npmjs.org/chownr/-/chownr-3.0.0.tgz",
      "integrity": "sha512-+IxzY9BZOQd/XuYPRmrvEVjF/nqj5kgT4kEq7VofrDoM1MxoRjEWkrCC3EtLi59TVawxTAn+orJwFQcrqEN1+g==",
-      "dev": true,
      "license": "BlueOak-1.0.0",
      "engines": {
        "node": ">=18"
@@ -25640,7 +25999,6 @@
      "version": "5.0.0",
      "resolved": "https://registry.npmjs.org/yallist/-/yallist-5.0.0.tgz",
      "integrity": "sha512-YgvUTfwqyc7UXVMrB+SImsVYSmTS8X/tSrtdNZMImM+n7+QTriRXyXim0mBrTXNeqzVF0KWGgHPeiyViFFrNDw==",
-      "dev": true,
      "license": "BlueOak-1.0.0",
      "engines": {
        "node": ">=18"
--- a/package.json
+++ b/package.json
@@ -190,7 +190,7 @@
    "@dnd-kit/core": "^6.3.1",
    "@dnd-kit/sortable": "^10.0.0",
    "@dnd-kit/utilities": "^3.2.2",
-    "@huggingface/transformers": "^4.2.0",
+    "@huggingface/transformers": "3.5.2",
    "@lobehub/icons": "^5.8.0",
    "@modelcontextprotocol/sdk": "^1.29.0",
    "@monaco-editor/react": "^4.7.0",
@@ -258,7 +258,10 @@
    "zustand": "^5.0.13"
  },
  "optionalDependencies": {
+    "@atjsh/llmlingua-2": "2.0.3",
+    "@tensorflow/tfjs": "4.22.0",
    "better-sqlite3": "^12.10.0",
+    "js-tiktoken": "^1.0.20",
    "keytar": "^7.9.0",
    "tls-client-node": "^0.2.0",
    "wreq-js": "^2.3.1"
--- a/scripts/build/pack-artifact-policy.ts
+++ b/scripts/build/pack-artifact-policy.ts
@@ -36,6 +36,9 @@ export const APP_STAGING_ALLOWED_EXACT_PATHS: string[] = [
  "BUILD_SHA",
  "docs/reference/openapi.yaml",
  "open-sse/mcp-server/server.js",
+  // LLMLingua ONNX worker — esbuild'd standalone .js spawned via worker_threads
+  // (the Next.js bundler can't trace the computed Worker path). Kept like the MCP server.
+  "open-sse/services/compression/engines/llmlingua/onnxWorker.js",
  "package.json",
  "peer-stamp.mjs",
  "responses-ws-proxy.mjs",
--- a/scripts/build/prepublish.ts
+++ b/scripts/build/prepublish.ts
@@ -239,6 +239,54 @@ if (existsSync(mcpSrcFile)) {
  }
 }

+// ── Step 8.6: Bundle LLMLingua ONNX worker ────────────────────────────
+// The worker is spawned via worker_threads at a path the Next.js bundler cannot
+// statically trace, so it must ship as a standalone .js (mirrors the MCP-server
+// bundling above). Heavy deps (@atjsh/llmlingua-2 / @huggingface/transformers /
+// @tensorflow/tfjs / js-tiktoken) stay EXTERNAL — they are optionalDependencies,
+// dynamically imported at runtime, and the worker fail-opens if any is absent.
+const llmWorkerSrc = join(
+  ROOT,
+  "open-sse",
+  "services",
+  "compression",
+  "engines",
+  "llmlingua",
+  "onnxWorker.ts"
+);
+const llmWorkerDestDir = join(
+  DIST_DIR,
+  "open-sse",
+  "services",
+  "compression",
+  "engines",
+  "llmlingua"
+);
+if (existsSync(llmWorkerSrc)) {
+  console.log("  🔨 Bundling LLMLingua ONNX worker (TypeScript → JavaScript)...");
+  mkdirSync(llmWorkerDestDir, { recursive: true });
+  try {
+    execFileSync(
+      NPX_BIN,
+      [
+        "esbuild",
+        "open-sse/services/compression/engines/llmlingua/onnxWorker.ts",
+        "--bundle",
+        "--platform=node",
+        "--packages=external",
+        "--format=esm",
+        "--outfile=dist/open-sse/services/compression/engines/llmlingua/onnxWorker.js",
+      ],
+      { cwd: ROOT, stdio: "inherit" }
+    );
+    console.log(
+      "  ✅ LLMLingua worker bundled to dist/open-sse/services/compression/engines/llmlingua/onnxWorker.js"
+    );
+  } catch (err: any) {
+    console.warn("  ⚠️  LLMLingua worker bundle error:", err.message);
+  }
+}
+
 // ── Step 8.7: Bundle CLI Entrypoint ──────────────────────────
 const cliSrcFile = join(ROOT, "bin", "omniroute.ts");
 const cliDestFile = join(ROOT, "bin", "omniroute.mjs");
--- a/tests/unit/compression/llmlingua-failopen.test.ts
+++ b/tests/unit/compression/llmlingua-failopen.test.ts
@@ -17,8 +17,13 @@ import assert from "node:assert/strict";
 import {
  llmlinguaEngine,
  setLlmlinguaBackend,
+  type LlmlinguaBackendOptions,
 } from "../../../open-sse/services/compression/engines/llmlingua/index.ts";

+// A large prose blob comfortably above the default 2000-token floor
+// (estimateCompressionTokens ≈ length/4). ~12 k chars ⇒ ~3 k tokens.
+const LARGE_PROSE = "The quick brown fox jumps over the lazy dog. ".repeat(280);
+
 // ─── helpers ─────────────────────────────────────────────────────────────────

 function makeBody(messages: Array<{ role: string; content: string }>): Record<string, unknown> {
@@ -74,7 +79,9 @@ describe("llmlingua engine", () => {
      "This is a sufficiently long prose paragraph to ensure compression is triggered.";
    const body = makeBody([{ role: "user", content: originalContent }]);

-    const result = await llmlinguaEngine.applyAsync!(body);
+    // minTokens:0 disables the small-prompt floor so the short fixture still
+    // exercises the backend (real behavior added by the minTokens floor — Task 4c).
+    const result = await llmlinguaEngine.applyAsync!(body, { stepConfig: { minTokens: 0 } });

    assert.equal(result.compressed, true, "should be marked compressed");
    assert.notEqual(result.stats, null, "stats should be present");
@@ -99,7 +106,7 @@ describe("llmlingua engine", () => {

    let result: Awaited<ReturnType<typeof llmlinguaEngine.applyAsync>>;
    try {
-      result = await llmlinguaEngine.applyAsync!(body);
+      result = await llmlinguaEngine.applyAsync!(body, { stepConfig: { minTokens: 0 } });
    } catch (err) {
      assert.fail(`applyAsync must not throw on backend error, but threw: ${err}`);
    }
@@ -123,7 +130,7 @@ describe("llmlingua engine", () => {
    const content = `${prose}\n\n${codeBlock}\n\nMore prose follows the code block here.`;

    const body = makeBody([{ role: "user", content }]);
-    const result = await llmlinguaEngine.applyAsync!(body);
+    const result = await llmlinguaEngine.applyAsync!(body, { stepConfig: { minTokens: 0 } });

    // Code block text must be byte-identical in the output
    const outContent = (result.body.messages as Array<{ role: string; content: string }>)[0]!
@@ -159,7 +166,7 @@ describe("llmlingua engine", () => {
      { role: "user", content: userContent },
    ]);

-    const result = await llmlinguaEngine.applyAsync!(body);
+    const result = await llmlinguaEngine.applyAsync!(body, { stepConfig: { minTokens: 0 } });

    const outMessages = result.body.messages as Array<{ role: string; content: string }>;
    const outSystem = outMessages.find((m) => m.role === "system")!;
@@ -175,3 +182,131 @@ describe("llmlingua engine", () => {
    }
  });
 });
+
+// ─── Task 3/4: minTokens floor, opts threading, config schema/validation ───────
+
+describe("llmlingua engine — minTokens floor + config schema (Task 3/4)", () => {
+  // ── 1. floor skips small input (no stepConfig → default floor 2000) ──────────
+  it("minTokens floor skips small input — backend never called, compressed:false", async () => {
+    const calls: string[] = [];
+    setLlmlinguaBackend((text) => {
+      calls.push(text);
+      return Promise.resolve("X");
+    });
+
+    const body = makeBody([{ role: "user", content: "Short prose, well below the floor." }]);
+
+    const result = await llmlinguaEngine.applyAsync!(body);
+
+    assert.equal(result.compressed, false, "small input must skip compression");
+    assert.equal(result.stats, null, "stats must be null when skipped by the floor");
+    assert.equal(calls.length, 0, "backend must NOT be called below the floor");
+  });
+
+  // ── 2. floor disabled (minTokens:0) → passes through to backend ──────────────
+  it("minTokens:0 disables the floor — backend IS called, compressed:true", async () => {
+    const calls: string[] = [];
+    setLlmlinguaBackend((text) => {
+      calls.push(text);
+      return Promise.resolve("X");
+    });
+
+    const body = makeBody([{ role: "user", content: "Short prose, well below the floor." }]);
+
+    const result = await llmlinguaEngine.applyAsync!(body, { stepConfig: { minTokens: 0 } });
+
+    assert.equal(result.compressed, true, "floor disabled → backend compresses");
+    assert.ok(calls.length > 0, "backend must be called when the floor is disabled");
+  });
+
+  // ── 3. opts threading: model + compressionRate reach the backend ─────────────
+  it("threads model + compressionRate from stepConfig down to the backend opts", async () => {
+    let capturedOpts: LlmlinguaBackendOptions | undefined;
+    setLlmlinguaBackend((_text, opts) => {
+      capturedOpts = opts;
+      return Promise.resolve("X");
+    });
+
+    const body = makeBody([{ role: "user", content: "Some prose to send to the backend." }]);
+
+    await llmlinguaEngine.applyAsync!(body, {
+      stepConfig: { minTokens: 0, model: "bert-base", compressionRate: 0.3 },
+    });
+
+    assert.ok(capturedOpts, "backend must receive an opts object");
+    assert.equal(capturedOpts!.model, "bert-base", "model must be threaded into opts");
+    assert.equal(capturedOpts!.compressionRate, 0.3, "compressionRate must be threaded into opts");
+  });
+
+  // ── 3b. floor uses estimated tokens of non-system content (large → compress) ─
+  it("large input above the default floor IS compressed (no stepConfig)", async () => {
+    const calls: string[] = [];
+    setLlmlinguaBackend((text) => {
+      calls.push(text);
+      return Promise.resolve("X");
+    });
+
+    const body = makeBody([{ role: "user", content: LARGE_PROSE }]);
+
+    const result = await llmlinguaEngine.applyAsync!(body);
+
+    assert.equal(result.compressed, true, "large input must clear the floor and compress");
+    assert.ok(calls.length > 0, "backend must be called for input above the floor");
+  });
+
+  // ── 4. config validation ─────────────────────────────────────────────────────
+  it("validateConfig accepts a fully valid config", () => {
+    const res = llmlinguaEngine.validateConfig({
+      model: "tinybert",
+      compressionRate: 0.5,
+      minTokens: 2000,
+      modelPath: "",
+    });
+    assert.equal(res.valid, true, `expected valid, got errors: ${res.errors.join(", ")}`);
+  });
+
+  it("validateConfig rejects unknown model / out-of-range / wrong-type fields", () => {
+    assert.equal(
+      llmlinguaEngine.validateConfig({ model: "mobilebert" }).valid,
+      false,
+      "unknown model must be invalid"
+    );
+    assert.equal(
+      llmlinguaEngine.validateConfig({ compressionRate: 1.5 }).valid,
+      false,
+      "compressionRate > 0.9 must be invalid"
+    );
+    assert.equal(
+      llmlinguaEngine.validateConfig({ compressionRate: 0.05 }).valid,
+      false,
+      "compressionRate < 0.1 must be invalid"
+    );
+    assert.equal(
+      llmlinguaEngine.validateConfig({ minTokens: -1 }).valid,
+      false,
+      "negative minTokens must be invalid"
+    );
+    assert.equal(
+      llmlinguaEngine.validateConfig({ modelPath: 123 }).valid,
+      false,
+      "non-string modelPath must be invalid"
+    );
+  });
+
+  // ── 5. schema shape ──────────────────────────────────────────────────────────
+  it("getConfigSchema exposes model select + minTokens/compressionRate/modelPath fields", () => {
+    const schema = llmlinguaEngine.getConfigSchema();
+    const byKey = new Map(schema.map((f) => [f.key, f]));
+
+    const modelField = byKey.get("model");
+    assert.ok(modelField, "schema must include a 'model' field");
+    assert.equal(modelField!.type, "select", "model field must be a select");
+    const optionValues = (modelField!.options ?? []).map((o) => o.value);
+    assert.ok(optionValues.includes("tinybert"), "model options must include tinybert");
+    assert.ok(optionValues.includes("bert-base"), "model options must include bert-base");
+
+    assert.ok(byKey.has("minTokens"), "schema must include minTokens");
+    assert.ok(byKey.has("compressionRate"), "schema must include compressionRate");
+    assert.ok(byKey.has("modelPath"), "schema must include modelPath");
+  });
+});
--- a/tests/unit/compression/llmlingua-model-store.test.ts
+++ b/tests/unit/compression/llmlingua-model-store.test.ts
@@ -0,0 +1,131 @@
+/**
+ * TDD tests for the llmlingua model registry + model store (Tasks 1-2).
+ *
+ * Tests are written RED-first (before the implementation exists).
+ *
+ * Coverage:
+ *  1. Registry shape — tinybert + bert-base present; default points at a valid key.
+ *  2. Each entry's factory/dtype/subfolder/hfRepo invariants hold.
+ *  3. resolveLlmlinguaModel() — known id, and the unknown/undefined/empty fallbacks.
+ *  4. getLlmlinguaModelCacheDir() — path suffix + DATA_DIR honoring.
+ *  5. configureTransformersEnv() — Hub download default vs local modelPath override.
+ */
+
+import { describe, it, after } from "node:test";
+import assert from "node:assert/strict";
+import os from "node:os";
+import path from "node:path";
+import fs from "node:fs";
+
+import {
+  DEFAULT_LLMLINGUA_MODEL,
+  LLMLINGUA_MODELS,
+} from "../../../open-sse/services/compression/engines/llmlingua/constants.ts";
+import {
+  getLlmlinguaModelCacheDir,
+  resolveLlmlinguaModel,
+  configureTransformersEnv,
+  type TransformersEnvLike,
+} from "../../../open-sse/services/compression/engines/llmlingua/modelStore.ts";
+
+// ─── tests ────────────────────────────────────────────────────────────────────
+
+describe("llmlingua model registry (constants)", () => {
+  // ── 1. registry shape ──────────────────────────────────────────────────────
+  it("LLMLINGUA_MODELS has both proven models and the default is a valid key", () => {
+    assert.ok(LLMLINGUA_MODELS.tinybert, "tinybert entry must exist");
+    assert.ok(LLMLINGUA_MODELS["bert-base"], "bert-base entry must exist");
+    assert.equal(DEFAULT_LLMLINGUA_MODEL, "tinybert");
+    assert.ok(
+      Object.prototype.hasOwnProperty.call(LLMLINGUA_MODELS, DEFAULT_LLMLINGUA_MODEL),
+      "default must be a key of the registry"
+    );
+  });
+
+  // ── 2. per-entry invariants ─────────────────────────────────────────────────
+  it("every entry uses WithBERTMultilingual / fp32 / '' subfolder and a sane hfRepo", () => {
+    for (const [key, entry] of Object.entries(LLMLINGUA_MODELS)) {
+      assert.equal(entry.factory, "WithBERTMultilingual", `${key}.factory`);
+      assert.equal(entry.dtype, "fp32", `${key}.dtype`);
+      assert.equal(entry.subfolder, "", `${key}.subfolder`);
+      assert.equal(typeof entry.hfRepo, "string", `${key}.hfRepo type`);
+      assert.ok(entry.hfRepo.length > 0, `${key}.hfRepo non-empty`);
+      assert.ok(entry.hfRepo.includes("/"), `${key}.hfRepo contains "/"`);
+      assert.equal(entry.id, key, `${key}.id matches its registry key`);
+      assert.equal(typeof entry.sizeMB, "number", `${key}.sizeMB type`);
+      assert.equal(typeof entry.label, "string", `${key}.label type`);
+    }
+  });
+});
+
+describe("resolveLlmlinguaModel", () => {
+  // ── 3. resolution + fallback ────────────────────────────────────────────────
+  it("returns the requested entry for a known id", () => {
+    const resolved = resolveLlmlinguaModel("bert-base");
+    assert.equal(resolved.id, "bert-base");
+    assert.equal(resolved, LLMLINGUA_MODELS["bert-base"]);
+  });
+
+  it("falls back to the default (tinybert) for unknown / undefined / empty ids", () => {
+    const def = LLMLINGUA_MODELS[DEFAULT_LLMLINGUA_MODEL];
+    assert.equal(resolveLlmlinguaModel("nonexistent"), def);
+    assert.equal(resolveLlmlinguaModel(undefined), def);
+    assert.equal(resolveLlmlinguaModel(null), def);
+    assert.equal(resolveLlmlinguaModel(""), def);
+  });
+});
+
+describe("getLlmlinguaModelCacheDir", () => {
+  const originalDataDir = process.env.DATA_DIR;
+  let tmpDir: string | undefined;
+
+  after(() => {
+    if (originalDataDir === undefined) {
+      delete process.env.DATA_DIR;
+    } else {
+      process.env.DATA_DIR = originalDataDir;
+    }
+    if (tmpDir) {
+      try {
+        fs.rmSync(tmpDir, { recursive: true, force: true });
+      } catch {
+        /* ignore cleanup errors */
+      }
+    }
+  });
+
+  // ── 4. path suffix + DATA_DIR honoring ──────────────────────────────────────
+  it("ends with models/llmlingua and lives under DATA_DIR when set", () => {
+    tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "llm-"));
+    process.env.DATA_DIR = tmpDir;
+
+    const dir = getLlmlinguaModelCacheDir();
+    assert.ok(
+      dir.endsWith(path.join("models", "llmlingua")),
+      `cache dir should end with models/llmlingua, got: ${dir}`
+    );
+    assert.ok(dir.startsWith(tmpDir), `cache dir should be under DATA_DIR, got: ${dir}`);
+    assert.equal(dir, path.join(tmpDir, "models", "llmlingua"));
+  });
+});
+
+describe("configureTransformersEnv", () => {
+  // ── 5. Hub default vs local override ────────────────────────────────────────
+  it("Hub download default: cacheDir set, allowRemoteModels true, no localModelPath", () => {
+    const env: TransformersEnvLike = {};
+    configureTransformersEnv(env, {});
+    assert.equal(typeof env.cacheDir, "string");
+    assert.ok((env.cacheDir as string).length > 0, "cacheDir must be set");
+    assert.equal(env.allowRemoteModels, true);
+    assert.equal(env.localModelPath, undefined);
+  });
+
+  it("local modelPath override: localModelPath set, allowRemoteModels false, cacheDir still set", () => {
+    const env: TransformersEnvLike = {};
+    configureTransformersEnv(env, { modelPath: "/some/local/dir" });
+    assert.equal(env.localModelPath, "/some/local/dir");
+    assert.equal(env.allowRemoteModels, false);
+    assert.equal(typeof env.cacheDir, "string");
+    assert.ok((env.cacheDir as string).length > 0, "cacheDir must still be set");
+  });
+});
--- a/tests/unit/compression/llmlingua-worker.test.ts
+++ b/tests/unit/compression/llmlingua-worker.test.ts
@@ -0,0 +1,92 @@
+/**
+ * Tests for the real LLMLingua worker-thread backend (`worker.ts` + `onnxWorker.ts`).
+ *
+ * The four optional deps (`@atjsh/llmlingua-2`, `@huggingface/transformers`,
+ * `@tensorflow/tfjs`, `js-tiktoken`) are NOT installed in this worktree, so the
+ * default path MUST fail-open WITHOUT spawning a worker:
+ *
+ *  1. Deps absent → fail-open, no spawn (ALWAYS runs here): the backend returns the
+ *     ORIGINAL text unchanged, fast (no model load / worker spawn).
+ *  2. Type smoke: `workerBackend` is a function.
+ *  3. GATED real compression (RUN_LLMLINGUA_INT=1): real shrink with deps present;
+ *     no-op skip here (deps absent).
+ */
+
+import { test, after } from "node:test";
+import assert from "node:assert/strict";
+import { createRequire } from "node:module";
+
+import {
+  workerBackend,
+  __resetLlmlinguaWorkerForTests,
+} from "../../../open-sse/services/compression/engines/llmlingua/worker.ts";
+
+const require = createRequire(import.meta.url);
+
+/** Whether all four optional deps resolve in this environment. */
+function depsResolve(): boolean {
+  try {
+    require.resolve("@atjsh/llmlingua-2");
+    require.resolve("@huggingface/transformers");
+    require.resolve("@tensorflow/tfjs");
+    require.resolve("js-tiktoken");
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+// Let the process exit cleanly: terminate any spawned worker + reset singletons.
+after(() => {
+  __resetLlmlinguaWorkerForTests();
+});
+
+test("deps absent → fail-open, no spawn, returns original text fast", async () => {
+  if (depsResolve()) {
+    // Premise of this test is that the optional deps are NOT installed (the CI
+    // default). When they ARE present (e.g. a local integration setup), the backend
+    // legitimately spawns the worker and compresses, so this assertion no longer
+    // applies — the real path is covered by the gated test below.
+    console.log("skip: optional deps present — fail-open-when-absent test N/A");
+    return;
+  }
+  const input = "hello world this is some prose";
+
+  const start = Date.now();
+  const out1 = await workerBackend(input, {});
+  const elapsed = Date.now() - start;
+
+  // EXACT original text (fail-open), unchanged.
+  assert.equal(out1, input);
+  // Fast: no model load / worker spawn (proves the optional-deps gate short-circuits).
+  assert.ok(elapsed < 1000, `expected <1000ms, got ${elapsed}ms`);
+
+  // Second call exercises the memoized gate — still fail-open, still fast.
+  const out2 = await workerBackend(input, {});
+  assert.equal(out2, input);
+});
+
+test("type smoke: workerBackend is a function", () => {
+  assert.equal(typeof workerBackend, "function");
+});
+
+test("GATED real compression (RUN_LLMLINGUA_INT=1)", async () => {
+  if (process.env.RUN_LLMLINGUA_INT !== "1") {
+    console.log("skip: RUN_LLMLINGUA_INT!=1");
+    return;
+  }
+  if (!depsResolve()) {
+    console.log("skip: deps absent");
+    return;
+  }
+
+  // Long prose well above any practical floor so compression has room to shrink.
+  const LONG_PROSE =
+    "The quick brown fox jumps over the lazy dog while the sun sets slowly behind the distant hills. ".repeat(
+      120
+    );
+
+  const out = await workerBackend(LONG_PROSE, { model: "tinybert", compressionRate: 0.5 });
+  assert.equal(typeof out, "string");
+  assert.ok(out.length < LONG_PROSE.length, "expected a real shrink");
+});