@atlaskit/editor-plugin-autocomplete 3.4.1 → 3.6.0

package/CHANGELOG.md

@@ -1,5 +1,31 @@
 # @atlaskit/editor-plugin-autocomplete
 
+## 3.6.0
+
+### Minor Changes
+
+- [`9f6b6c9fffc50`](https://bitbucket.org/atlassian/atlassian-frontend-monorepo/commits/9f6b6c9fffc50) -
+  Add analytics for the on-device (local LLM) autocomplete slow lane: fire a `localModelLoaded`
+  track event when the engine initialises successfully (with load duration and GPU info) and a
+  `localModelLoadFailed` track event when it fails, categorising the reason and capturing WebGPU
+  capability diagnostics to surface user-machine limitations.
+
+### Patch Changes
+
+- Updated dependencies
+
+## 3.5.0
+
+### Minor Changes
+
+- [`52c7f5f973025`](https://bitbucket.org/atlassian/atlassian-frontend-monorepo/commits/52c7f5f973025) -
+  Switch the CTC autocomplete debug toggle off `localStorage` and onto a storage-free mechanism.
+  Debug logging is now enabled via the `__atlCtcDebug__.enable()` / `.disable()` console API for the
+  current session, or by appending `?atlCtcDebug=1` to the URL to have it active from initial load
+  (and survive reloads). This avoids browser-storage consent controls (BSC) that can block
+  uncategorized `localStorage` writes in some products. The `isAutocompleteDebugEnabled()` API is
+  unchanged for callers.
+
 ## 3.4.1
 
 ### Patch Changes

package/dist/cjs/pm-plugins/autocomplete-plugin.js

@@ -247,12 +247,49 @@ var createAutocompletePlugin = exports.createAutocompletePlugin = function creat
       }
     });
   };
-  var fireSuggestionInsertedAnalytics = function fireSuggestionInsertedAnalytics(ghostText) {
+  var fireLocalModelLoadedAnalytics = function fireLocalModelLoadedAnalytics(info) {
     var _api$analytics2;
+    api === null || api === void 0 || (_api$analytics2 = api.analytics) === null || _api$analytics2 === void 0 || _api$analytics2.actions.fireAnalyticsEvent({
+      action: _analytics.ACTION.LOCAL_MODEL_LOADED,
+      actionSubject: _analytics.ACTION_SUBJECT.CONTEXTUAL_TYPEAHEAD,
+      eventType: _analytics.EVENT_TYPE.TRACK,
+      attributes: {
+        modelId: info.modelId,
+        embeddingModelId: info.embeddingModelId,
+        loadDurationMs: info.loadDurationMs,
+        gpuVendor: info.capabilities.vendor,
+        gpuArchitecture: info.capabilities.architecture
+      }
+    });
+  };
+  var fireLocalModelLoadFailedAnalytics = function fireLocalModelLoadFailedAnalytics(error) {
+    var _api$analytics3;
+    var capabilities = error.capabilities;
+    api === null || api === void 0 || (_api$analytics3 = api.analytics) === null || _api$analytics3 === void 0 || _api$analytics3.actions.fireAnalyticsEvent({
+      action: _analytics.ACTION.LOCAL_MODEL_LOAD_FAILED,
+      actionSubject: _analytics.ACTION_SUBJECT.CONTEXTUAL_TYPEAHEAD,
+      eventType: _analytics.EVENT_TYPE.TRACK,
+      attributes: {
+        reason: error.reason,
+        message: error.message,
+        modelId: error.modelId,
+        embeddingModelId: error.embeddingModelId,
+        webgpuAvailable: capabilities.available,
+        adapterAvailable: capabilities.adapterAvailable,
+        shaderF16Supported: capabilities.shaderF16Supported,
+        maxBufferSizeMB: capabilities.maxBufferSizeMB,
+        maxStorageBufferBindingSizeMB: capabilities.maxStorageBufferBindingSizeMB,
+        gpuVendor: capabilities.vendor,
+        gpuArchitecture: capabilities.architecture
+      }
+    });
+  };
+  var fireSuggestionInsertedAnalytics = function fireSuggestionInsertedAnalytics(ghostText) {
+    var _api$analytics4;
     var typedLength = lastSuggestionTypedLength;
     var suggestionLength = lastSuggestionLength || typedLength + ghostText.length;
     var kssDelta = suggestionLength - typedLength;
-    api === null || api === void 0 || (_api$analytics2 = api.analytics) === null || _api$analytics2 === void 0 || _api$analytics2.actions.fireAnalyticsEvent({
+    api === null || api === void 0 || (_api$analytics4 = api.analytics) === null || _api$analytics4 === void 0 || _api$analytics4.actions.fireAnalyticsEvent({
       action: _analytics.ACTION.SUGGESTION_INSERTED,
       actionSubject: _analytics.ACTION_SUBJECT.CONTEXTUAL_TYPEAHEAD,
       eventType: _analytics.EVENT_TYPE.TRACK,
@@ -265,7 +302,9 @@ var createAutocompletePlugin = exports.createAutocompletePlugin = function creat
     });
   };
   var slowLaneClient = options !== null && options !== void 0 && options.useLocalModel ? (0, _localSlowLaneClient.createLocalSlowLaneClient)({
-    debounceMs: LOCAL_SLOW_LANE_DEBOUNCE_MS
+    debounceMs: LOCAL_SLOW_LANE_DEBOUNCE_MS,
+    onLoadSuccess: fireLocalModelLoadedAnalytics,
+    onLoadError: fireLocalModelLoadFailedAnalytics
   }) : (0, _slowLaneClient.createSlowLaneClient)({
     baseUrl: '',
     debounceMs: NETWORK_SLOW_LANE_DEBOUNCE_MS
@@ -419,9 +458,9 @@ var createAutocompletePlugin = exports.createAutocompletePlugin = function creat
           lastSuggestionLength = typedLength + prediction.length;
           showGhostText(view, prediction, selection.from);
           if (prediction !== lastShownGhostText) {
-            var _api$analytics3;
+            var _api$analytics5;
             lastShownGhostText = prediction;
-            api === null || api === void 0 || (_api$analytics3 = api.analytics) === null || _api$analytics3 === void 0 || _api$analytics3.actions.fireAnalyticsEvent({
+            api === null || api === void 0 || (_api$analytics5 = api.analytics) === null || _api$analytics5 === void 0 || _api$analytics5.actions.fireAnalyticsEvent({
               action: _analytics.ACTION.SUGGESTION_VIEWED,
               actionSubject: _analytics.ACTION_SUBJECT.CONTEXTUAL_TYPEAHEAD,
               eventType: _analytics.EVENT_TYPE.TRACK,

package/dist/cjs/pm-plugins/debug-mode.js

@@ -7,18 +7,59 @@ exports.isAutocompleteDebugEnabled = void 0;
 /**
  * Contextual Typeahead Completions (CTC) debug logging utility.
  *
- * Logs are silent by default. To enable in any environment (dev, staging, prod):
+ * Logs are silent by default. Enable in any environment (dev, staging, prod):
  *
- *   localStorage.setItem('atl-ctc-dbg', '1')
+ *   // Live, for the current session (no reload required):
+ *   __atlCtcDebug__.enable()
  *
- * Then reload the page. To disable:
+ *   // To disable:
+ *   __atlCtcDebug__.disable()
  *
- *   localStorage.removeItem('atl-ctc-dbg')
+ *   // From initial page load (survives reload) — append to the URL:
+ *   ?atlCtcDebug=1
+ *
+ * Storage-free by design: avoids browser-storage consent controls (BSC), which can
+ * block uncategorized localStorage/sessionStorage/cookie writes in some products.
  */
-var isAutocompleteDebugEnabled = exports.isAutocompleteDebugEnabled = function isAutocompleteDebugEnabled() {
+
+var hasUrlDebugFlag = function hasUrlDebugFlag() {
+  if (typeof window === 'undefined') {
+    return false;
+  }
   try {
-    return typeof localStorage !== 'undefined' && localStorage.getItem('atl-ctc-dbg') === '1';
+    return new URLSearchParams(window.location.search).get('atlCtcDebug') === '1';
   } catch (_unused) {
     return false;
   }
-};
+};
+
+// State lives on the window object (not a module closure) so duplicate copies of this
+// module across separate bundles/realms share one source of truth and the console API
+// controls them all. In-memory only; persisted across reload via the URL flag.
+var getDebugApi = function getDebugApi() {
+  if (typeof window === 'undefined') {
+    return undefined;
+  }
+  if (!window.__atlCtcDebug__) {
+    var debugEnabled = hasUrlDebugFlag();
+    window.__atlCtcDebug__ = {
+      enable: function enable() {
+        debugEnabled = true;
+      },
+      disable: function disable() {
+        debugEnabled = false;
+      },
+      isEnabled: function isEnabled() {
+        return debugEnabled;
+      }
+    };
+  }
+  return window.__atlCtcDebug__;
+};
+var isAutocompleteDebugEnabled = exports.isAutocompleteDebugEnabled = function isAutocompleteDebugEnabled() {
+  var _getDebugApi$isEnable, _getDebugApi;
+  return (_getDebugApi$isEnable = (_getDebugApi = getDebugApi()) === null || _getDebugApi === void 0 ? void 0 : _getDebugApi.isEnabled()) !== null && _getDebugApi$isEnable !== void 0 ? _getDebugApi$isEnable : false;
+};
+
+// Eagerly install so the console API is available on load, regardless of call order.
+getDebugApi();

package/dist/cjs/pm-plugins/local-slow-lane-client.js

@@ -18,7 +18,7 @@ var _debugMode = require("./debug-mode");
 var _slowLaneClient = require("./slow-lane-client");
 function ownKeys(e, r) { var t = Object.keys(e); if (Object.getOwnPropertySymbols) { var o = Object.getOwnPropertySymbols(e); r && (o = o.filter(function (r) { return Object.getOwnPropertyDescriptor(e, r).enumerable; })), t.push.apply(t, o); } return t; }
 function _objectSpread(e) { for (var r = 1; r < arguments.length; r++) { var t = null != arguments[r] ? arguments[r] : {}; r % 2 ? ownKeys(Object(t), !0).forEach(function (r) { (0, _defineProperty2.default)(e, r, t[r]); }) : Object.getOwnPropertyDescriptors ? Object.defineProperties(e, Object.getOwnPropertyDescriptors(t)) : ownKeys(Object(t)).forEach(function (r) { Object.defineProperty(e, r, Object.getOwnPropertyDescriptor(t, r)); }); } return e; }
-function _interopRequireWildcard(e, t) { if ("function" == typeof WeakMap) var r = new WeakMap(), n = new WeakMap(); return (_interopRequireWildcard = function _interopRequireWildcard(e, t) { if (!t && e && e.__esModule) return e; var o, i, f = { __proto__: null, default: e }; if (null === e || "object" != _typeof3(e) && "function" != typeof e) return f; if (o = t ? n : r) { if (o.has(e)) return o.get(e); o.set(e, f); } for (var _t3 in e) "default" !== _t3 && {}.hasOwnProperty.call(e, _t3) && ((i = (o = Object.defineProperty) && Object.getOwnPropertyDescriptor(e, _t3)) && (i.get || i.set) ? o(f, _t3, i) : f[_t3] = e[_t3]); return f; })(e, t); }
+function _interopRequireWildcard(e, t) { if ("function" == typeof WeakMap) var r = new WeakMap(), n = new WeakMap(); return (_interopRequireWildcard = function _interopRequireWildcard(e, t) { if (!t && e && e.__esModule) return e; var o, i, f = { __proto__: null, default: e }; if (null === e || "object" != _typeof3(e) && "function" != typeof e) return f; if (o = t ? n : r) { if (o.has(e)) return o.get(e); o.set(e, f); } for (var _t6 in e) "default" !== _t6 && {}.hasOwnProperty.call(e, _t6) && ((i = (o = Object.defineProperty) && Object.getOwnPropertyDescriptor(e, _t6)) && (i.get || i.set) ? o(f, _t6, i) : f[_t6] = e[_t6]); return f; })(e, t); }
 function _createForOfIteratorHelper(r, e) { var t = "undefined" != typeof Symbol && r[Symbol.iterator] || r["@@iterator"]; if (!t) { if (Array.isArray(r) || (t = _unsupportedIterableToArray(r)) || e && r && "number" == typeof r.length) { t && (r = t); var _n = 0, F = function F() {}; return { s: F, n: function n() { return _n >= r.length ? { done: !0 } : { done: !1, value: r[_n++] }; }, e: function e(r) { throw r; }, f: F }; } throw new TypeError("Invalid attempt to iterate non-iterable instance.\nIn order to be iterable, non-array objects must have a [Symbol.iterator]() method."); } var o, a = !0, u = !1; return { s: function s() { t = t.call(r); }, n: function n() { var r = t.next(); return a = r.done, r; }, e: function e(r) { u = !0, o = r; }, f: function f() { try { a || null == t.return || t.return(); } finally { if (u) throw o; } } }; }
 function _unsupportedIterableToArray(r, a) { if (r) { if ("string" == typeof r) return _arrayLikeToArray(r, a); var t = {}.toString.call(r).slice(8, -1); return "Object" === t && r.constructor && (t = r.constructor.name), "Map" === t || "Set" === t ? Array.from(r) : "Arguments" === t || /^(?:Ui|I)nt(?:8|16|32)(?:Clamped)?Array$/.test(t) ? _arrayLikeToArray(r, a) : void 0; } }
 function _arrayLikeToArray(r, a) { (null == a || a > r.length) && (a = r.length); for (var e = 0, n = Array(a); e < a; e++) n[e] = r[e]; return n; } /**
@@ -61,6 +61,19 @@ function _arrayLikeToArray(r, a) { (null == a || a > r.length) && (a = r.length)
 
 // Same return type as createSlowLaneClient for drop-in compatibility
 
+/**
+ * Why the local engine failed to load/start.
+ *
+ * The first three are user-machine limitations (WebGPU missing, no compatible
+ * GPU adapter, GPU lacks the `shader-f16` feature the model needs);
+ * `insufficient_memory` is hit when weights don't fit in VRAM. The rest cover
+ * delivery/runtime failures unrelated to hardware.
+ */
+
+/** Snapshot of the machine's WebGPU support, used to explain hardware limits. */
+
+// Minimal WebGPU shape: lib.dom types aren't guaranteed in this build target.
+
 // ─── Constants ───────────────────────────────────────────────────────────────
 
 var DEFAULT_DEBOUNCE_MS = 300;
@@ -611,6 +624,8 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
     debounceMs = _config$debounceMs === void 0 ? DEFAULT_DEBOUNCE_MS : _config$debounceMs,
     onUpdate = config.onUpdate,
     onStatus = config.onStatus,
+    onLoadError = config.onLoadError,
+    onLoadSuccess = config.onLoadSuccess,
     _config$modelId = config.modelId,
     modelId = _config$modelId === void 0 ? LOCAL_MLC_CAUSAL_MODEL_ID : _config$modelId,
     customModelConfig = config.customModelConfig;
@@ -657,30 +672,191 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
     }
     onStatus === null || onStatus === void 0 || onStatus(message);
   };
-  var initEngine = /*#__PURE__*/function () {
+  var bytesToMB = function bytesToMB(bytes) {
+    return typeof bytes === 'number' ? Math.round(bytes / (1024 * 1024)) : undefined;
+  };
+
+  /**
+   * Inspect the machine's WebGPU support so a load failure can be attributed
+   * to a concrete hardware/browser limitation rather than a generic error.
+   */
+  var probeWebGpuCapabilities = /*#__PURE__*/function () {
     var _ref8 = (0, _asyncToGenerator2.default)( /*#__PURE__*/_regenerator.default.mark(function _callee2() {
-      var _yield$Promise$all3, _yield$Promise$all4, _yield$Promise$all4$, MLCEngineCtor, prebuiltAppConfig, customModelRecord, appConfig, newEngine, errorMsg, _t;
+      var gpu, _adapter$limits, _adapter$limits2, adapter, vendor, architecture, _adapter$info, _adapter$requestAdapt, info, _t, _t2, _t3;
       return _regenerator.default.wrap(function (_context2) {
         while (1) switch (_context2.prev = _context2.next) {
           case 0:
-            _context2.prev = 0;
-            if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
-              // eslint-disable-next-line no-console
-              console.log("%c[LocalSlowLane] %c\uD83D\uDE80 Initialising MLC engine with models: ".concat(modelId, " (LM) + ").concat(LOCAL_MLC_EMBEDDING_MODEL_ID, " (embedder)"), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
-            }
-            onStatus === null || onStatus === void 0 || onStatus("Initialising models: ".concat(modelId, " + ").concat(LOCAL_MLC_EMBEDDING_MODEL_ID, "\u2026"));
-            if ('gpu' in navigator) {
+            gpu = navigator.gpu;
+            if (gpu) {
               _context2.next = 1;
               break;
             }
-            throw new Error('WebGPU not supported');
+            return _context2.abrupt("return", {
+              available: false
+            });
           case 1:
+            _context2.prev = 1;
             _context2.next = 2;
+            return gpu.requestAdapter();
+          case 2:
+            adapter = _context2.sent;
+            if (adapter) {
+              _context2.next = 3;
+              break;
+            }
+            return _context2.abrupt("return", {
+              available: true,
+              adapterAvailable: false
+            });
+          case 3:
+            _context2.prev = 3;
+            if (!((_adapter$info = adapter.info) !== null && _adapter$info !== void 0)) {
+              _context2.next = 4;
+              break;
+            }
+            _t = _adapter$info;
+            _context2.next = 6;
+            break;
+          case 4:
+            _context2.next = 5;
+            return (_adapter$requestAdapt = adapter.requestAdapterInfo) === null || _adapter$requestAdapt === void 0 ? void 0 : _adapter$requestAdapt.call(adapter);
+          case 5:
+            _t = _context2.sent;
+          case 6:
+            info = _t;
+            vendor = (info === null || info === void 0 ? void 0 : info.vendor) || undefined;
+            architecture = (info === null || info === void 0 ? void 0 : info.architecture) || undefined;
+            _context2.next = 8;
+            break;
+          case 7:
+            _context2.prev = 7;
+            _t2 = _context2["catch"](3);
+          case 8:
+            return _context2.abrupt("return", {
+              available: true,
+              adapterAvailable: true,
+              shaderF16Supported: adapter.features.has('shader-f16'),
+              maxBufferSizeMB: bytesToMB((_adapter$limits = adapter.limits) === null || _adapter$limits === void 0 ? void 0 : _adapter$limits.maxBufferSize),
+              maxStorageBufferBindingSizeMB: bytesToMB((_adapter$limits2 = adapter.limits) === null || _adapter$limits2 === void 0 ? void 0 : _adapter$limits2.maxStorageBufferBindingSize),
+              vendor: vendor,
+              architecture: architecture
+            });
+          case 9:
+            _context2.prev = 9;
+            _t3 = _context2["catch"](1);
+            return _context2.abrupt("return", {
+              available: true,
+              adapterAvailable: false
+            });
+          case 10:
+          case "end":
+            return _context2.stop();
+        }
+      }, _callee2, null, [[1, 9], [3, 7]]);
+    }));
+    return function probeWebGpuCapabilities() {
+      return _ref8.apply(this, arguments);
+    };
+  }();
+
+  /** Map an MLC/WebLLM engine-creation error message to a coarse reason. */
+  var classifyEngineError = function classifyEngineError(message) {
+    var lower = message.toLowerCase();
+    if (lower.includes('loading chunk') || lower.includes('dynamically imported module') || lower.includes('dynamic import')) {
+      return 'module_load_failed';
+    }
+    if (lower.includes('out of memory') || /\boom\b/.test(lower) || lower.includes('allocation') || lower.includes('exceeds') || lower.includes('buffer size') || lower.includes('not enough memory')) {
+      return 'insufficient_memory';
+    }
+    // Pre-flight already returns missing_shader_f16 when the feature is absent,
+    // so only match the exact feature token here — not bare 'shader' (compile
+    // errors) or bare 'f16' (present in model ids like q0f16-MLC).
+    if (lower.includes('shader-f16') || lower.includes('shader_f16')) {
+      return 'missing_shader_f16';
+    }
+    if (lower.includes('fetch') || lower.includes('network') || lower.includes('download') || lower.includes('http') || lower.includes('cache')) {
+      return 'model_download_failed';
+    }
+    return 'init_failed';
+  };
+
+  // Canonical, controlled failure descriptions. We never emit the raw engine
+  // error into analytics — it can embed customer-context URLs/paths (HOT-120175)
+  // — so the analytics `message` is always one of these fixed strings.
+  var LOAD_FAILURE_MESSAGE = {
+    webgpu_unavailable: 'WebGPU is not available in this browser',
+    webgpu_no_adapter: 'No compatible WebGPU adapter found',
+    missing_shader_f16: 'GPU does not support the shader-f16 feature',
+    insufficient_memory: 'Insufficient GPU memory to load the model',
+    model_download_failed: 'Failed to download model assets',
+    module_load_failed: 'Failed to load the web-llm runtime module',
+    init_failed: 'Model engine failed to initialise'
+  };
+  var handleLoadFailure = function handleLoadFailure(reason, capabilities, debugDetail) {
+    ready = false;
+    var message = LOAD_FAILURE_MESSAGE[reason];
+    if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
+      // eslint-disable-next-line no-console
+      console.log("[LocalSlowLane] Engine initialisation failed (".concat(reason, "): ").concat(debugDetail !== null && debugDetail !== void 0 ? debugDetail : message));
+    }
+    onStatus === null || onStatus === void 0 || onStatus("Engine initialisation failed: ".concat(message));
+    onLoadError === null || onLoadError === void 0 || onLoadError({
+      reason: reason,
+      message: message,
+      modelId: modelId,
+      embeddingModelId: LOCAL_MLC_EMBEDDING_MODEL_ID,
+      capabilities: capabilities
+    });
+    engineInitPromise = null;
+    initFailed = true;
+  };
+  var initEngine = /*#__PURE__*/function () {
+    var _ref9 = (0, _asyncToGenerator2.default)( /*#__PURE__*/_regenerator.default.mark(function _callee3() {
+      var capabilities, startTime, _yield$Promise$all3, _yield$Promise$all4, _yield$Promise$all4$, MLCEngineCtor, prebuiltAppConfig, customModelRecord, appConfig, newEngine, loadDurationMs, errorMsg, _t4;
+      return _regenerator.default.wrap(function (_context3) {
+        while (1) switch (_context3.prev = _context3.next) {
+          case 0:
+            _context3.next = 1;
+            return probeWebGpuCapabilities();
+          case 1:
+            capabilities = _context3.sent;
+            if (capabilities.available) {
+              _context3.next = 2;
+              break;
+            }
+            handleLoadFailure('webgpu_unavailable', capabilities);
+            return _context3.abrupt("return");
+          case 2:
+            if (!(capabilities.adapterAvailable === false)) {
+              _context3.next = 3;
+              break;
+            }
+            handleLoadFailure('webgpu_no_adapter', capabilities);
+            return _context3.abrupt("return");
+          case 3:
+            if (!(capabilities.shaderF16Supported === false)) {
+              _context3.next = 4;
+              break;
+            }
+            handleLoadFailure('missing_shader_f16', capabilities);
+            return _context3.abrupt("return");
+          case 4:
+            startTime = performance.now();
+            _context3.prev = 5;
+            if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
+              // eslint-disable-next-line no-console
+              console.log("%c[LocalSlowLane] %c\uD83D\uDE80 Initialising MLC engine with models: ".concat(modelId, " (LM) + ").concat(LOCAL_MLC_EMBEDDING_MODEL_ID, " (embedder)"), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
+            }
+            onStatus === null || onStatus === void 0 || onStatus("Initialising models: ".concat(modelId, " + ").concat(LOCAL_MLC_EMBEDDING_MODEL_ID, "\u2026"));
+
+            // Fetch the web-llm runtime and the BE-parity lookup tables in parallel;
+            // both are dynamically imported so they stay out of the main editor chunk.
+            _context3.next = 6;
             return Promise.all([Promise.resolve().then(function () {
               return _interopRequireWildcard(require( /* webpackChunkName: "@atlaskit-internal_editor-plugin-autocomplete-mlc-web-llm" */'@mlc-ai/web-llm'));
             }), loadBePayloadData()]);
-          case 2:
-            _yield$Promise$all3 = _context2.sent;
+          case 6:
+            _yield$Promise$all3 = _context3.sent;
             _yield$Promise$all4 = (0, _slicedToArray2.default)(_yield$Promise$all3, 1);
             _yield$Promise$all4$ = _yield$Promise$all4[0];
             MLCEngineCtor = _yield$Promise$all4$.MLCEngine;
@@ -708,19 +884,20 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               initProgressCallback: initProgressCallback,
               logitProcessorRegistry: new Map([[modelId, lmLogitsCapture]])
             });
-            _context2.next = 3;
+            _context3.next = 7;
             return newEngine.reload([modelId, LOCAL_MLC_EMBEDDING_MODEL_ID]);
-          case 3:
+          case 7:
             if (!destroyed) {
-              _context2.next = 4;
+              _context3.next = 8;
               break;
             }
             // destroy() was called while we were loading — clean up
             unloadEngine(newEngine);
-            return _context2.abrupt("return");
-          case 4:
+            return _context3.abrupt("return");
+          case 8:
             engine = newEngine;
             ready = true;
+            loadDurationMs = Math.round(performance.now() - startTime);
             if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
               // eslint-disable-next-line no-console
               console.log('%c[LocalSlowLane] %c✅ Both models loaded and ready', 'color: #9c27b0; font-weight: bold;', 'color: #4caf50;');
@@ -732,28 +909,27 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               console.log('%c[LocalSlowLane] %c🔢 Embedder     →', 'color: #9c27b0; font-weight: bold;', 'color: #009688; font-weight: bold;', LOCAL_MLC_EMBEDDING_MODEL_ID);
             }
             onStatus === null || onStatus === void 0 || onStatus('Model loaded and ready.');
-            _context2.next = 6;
+            onLoadSuccess === null || onLoadSuccess === void 0 || onLoadSuccess({
+              modelId: modelId,
+              embeddingModelId: LOCAL_MLC_EMBEDDING_MODEL_ID,
+              loadDurationMs: loadDurationMs,
+              capabilities: capabilities
+            });
+            _context3.next = 10;
             break;
-          case 5:
-            _context2.prev = 5;
-            _t = _context2["catch"](0);
-            errorMsg = _t instanceof Error ? _t.message : String(_t);
-            ready = false;
-            if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
-              // eslint-disable-next-line no-console
-              console.log("[LocalSlowLane] Engine initialisation failed: ".concat(errorMsg));
-            }
-            onStatus === null || onStatus === void 0 || onStatus("Engine initialisation failed: ".concat(errorMsg));
-            engineInitPromise = null;
-            initFailed = true;
-          case 6:
+          case 9:
+            _context3.prev = 9;
+            _t4 = _context3["catch"](5);
+            errorMsg = _t4 instanceof Error ? _t4.message : String(_t4);
+            handleLoadFailure(classifyEngineError(errorMsg), capabilities, errorMsg);
+          case 10:
           case "end":
-            return _context2.stop();
+            return _context3.stop();
         }
-      }, _callee2, null, [[0, 5]]);
+      }, _callee3, null, [[5, 9]]);
     }));
     return function initEngine() {
-      return _ref8.apply(this, arguments);
+      return _ref9.apply(this, arguments);
     };
   }();
   var ensureEngineInitialized = function ensureEngineInitialized() {
@@ -780,16 +956,16 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
    *     384-d semantic vector (passage-encoded; see `wrapForArctic`).
    */
   var runInference = /*#__PURE__*/function () {
-    var _ref9 = (0, _asyncToGenerator2.default)( /*#__PURE__*/_regenerator.default.mark(function _callee3(text, requestId) {
-      var lmText, semanticText, arcticInput, captureCompletionTime, _data, tStart, tLmDone, tEmbDone, _yield$Promise$all5, _yield$Promise$all6, embeddingResponse, rawLogits, payload, embedding, sumSq, i, topTokens, errorMsg, _t2;
-      return _regenerator.default.wrap(function (_context3) {
-        while (1) switch (_context3.prev = _context3.next) {
+    var _ref0 = (0, _asyncToGenerator2.default)( /*#__PURE__*/_regenerator.default.mark(function _callee4(text, requestId) {
+      var lmText, semanticText, arcticInput, captureCompletionTime, _data, tStart, tLmDone, tEmbDone, _yield$Promise$all5, _yield$Promise$all6, embeddingResponse, rawLogits, payload, embedding, sumSq, i, topTokens, errorMsg, _t5;
+      return _regenerator.default.wrap(function (_context4) {
+        while (1) switch (_context4.prev = _context4.next) {
           case 0:
             if (!(!engine || destroyed)) {
-              _context3.next = 1;
+              _context4.next = 1;
               break;
             }
-            return _context3.abrupt("return");
+            return _context4.abrupt("return");
           case 1:
             // Clear the capture buffer so we read only this pass's logits. The engine
             // serialises per-model requests and updateContext is debounced, so the
@@ -816,11 +992,11 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               // eslint-disable-next-line no-console
               console.log("%c[LocalSlowLane] %c\uD83E\uDDE0 LM input (".concat(lmText.length, " chars, ").concat(splitOnWhitespace(lmText).length, " words): \"").concat(lmText.length > 100 ? "".concat(lmText.slice(0, 100), "\u2026") : lmText, "\""), 'color: #9c27b0; font-weight: bold;', 'color: #2196f3;');
             }
-            _context3.prev = 2;
+            _context4.prev = 2;
             tStart = performance.now();
             tLmDone = 0;
             tEmbDone = 0;
-            _context3.next = 3;
+            _context4.next = 3;
             return Promise.all([captureCompletionTime(engine.completions.create({
               model: modelId,
               prompt: lmText,
@@ -836,7 +1012,7 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               tEmbDone = resolvedAt;
             })]);
           case 3:
-            _yield$Promise$all5 = _context3.sent;
+            _yield$Promise$all5 = _context4.sent;
             _yield$Promise$all6 = (0, _slicedToArray2.default)(_yield$Promise$all5, 2);
             embeddingResponse = _yield$Promise$all6[1];
             if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
@@ -846,10 +1022,10 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
 
             // Discard stale results
             if (!(requestId < latestRequestId || destroyed)) {
-              _context3.next = 4;
+              _context4.next = 4;
               break;
             }
-            return _context3.abrupt("return");
+            return _context4.abrupt("return");
           case 4:
             // ── LM logits: whole-word BE-parity payload ──────────────────
             rawLogits = lmLogitsCapture.captured;
@@ -883,17 +1059,17 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               // eslint-disable-next-line no-console
               console.log(storedLmLogits ? "\u2705 lm_logits: ".concat(Object.keys(storedLmLogits).length, " words") : '❌ No lm_logits');
               if (storedLmLogits) {
-                topTokens = Object.entries(storedLmLogits).sort(function (_ref0, _ref1) {
-                  var _ref10 = (0, _slicedToArray2.default)(_ref0, 2),
-                    a = _ref10[1];
+                topTokens = Object.entries(storedLmLogits).sort(function (_ref1, _ref10) {
                   var _ref11 = (0, _slicedToArray2.default)(_ref1, 2),
-                    b = _ref11[1];
+                    a = _ref11[1];
+                  var _ref12 = (0, _slicedToArray2.default)(_ref10, 2),
+                    b = _ref12[1];
                   return b - a;
                 }).slice(0, 10); // eslint-disable-next-line no-console
-                console.log('Top 10 predictions:', topTokens.map(function (_ref12) {
-                  var _ref13 = (0, _slicedToArray2.default)(_ref12, 2),
-                    t = _ref13[0],
-                    p = _ref13[1];
+                console.log('Top 10 predictions:', topTokens.map(function (_ref13) {
+                  var _ref14 = (0, _slicedToArray2.default)(_ref13, 2),
+                    t = _ref14[0],
+                    p = _ref14[1];
                   return "".concat(t, ": ").concat((p * 100).toFixed(1), "%");
                 }).join(', '));
               }
@@ -905,16 +1081,16 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               hasVector: storedContextVector !== null,
               hasLmLogits: storedLmLogits !== null
             });
-            _context3.next = 7;
+            _context4.next = 7;
             break;
           case 5:
-            _context3.prev = 5;
-            _t2 = _context3["catch"](2);
+            _context4.prev = 5;
+            _t5 = _context4["catch"](2);
             if (!(requestId < latestRequestId || destroyed)) {
-              _context3.next = 6;
+              _context4.next = 6;
               break;
             }
-            return _context3.abrupt("return");
+            return _context4.abrupt("return");
           case 6:
             storedContextVector = null;
             storedLmLogits = null;
@@ -923,19 +1099,19 @@ var createLocalSlowLaneClient = exports.createLocalSlowLaneClient = function cre
               hasVector: false,
               hasLmLogits: false
             });
-            errorMsg = _t2 instanceof Error ? _t2.message : String(_t2);
+            errorMsg = _t5 instanceof Error ? _t5.message : String(_t5);
             if ((0, _debugMode.isAutocompleteDebugEnabled)()) {
               // eslint-disable-next-line no-console
               console.log("%c[LocalSlowLane] %c\u274C Inference error (request #".concat(requestId, "): ").concat(errorMsg), 'color: #9c27b0; font-weight: bold;', 'color: #f44336;');
             }
           case 7:
           case "end":
-            return _context3.stop();
+            return _context4.stop();
         }
-      }, _callee3, null, [[2, 5]]);
+      }, _callee4, null, [[2, 5]]);
     }));
     return function runInference(_x, _x2) {
-      return _ref9.apply(this, arguments);
+      return _ref0.apply(this, arguments);
     };
   }();