@mlx-node/lm 0.0.8 → 0.0.10

package/dist/models/model-loader.d.ts

@@ -112,6 +112,7 @@ declare const MODEL_FAMILY_REGISTRY: readonly [{
     };
     readonly load: (modelPath: string) => Promise<import("../stream.js").StreamingInstance<typeof NativeLfm2Model, {
         readonly recordModelPath: true;
+        readonly replayAssistantRawText: true;
     }>>;
     readonly nativeModelClass: typeof NativeLfm2Model;
 }, {
@@ -122,6 +123,7 @@ declare const MODEL_FAMILY_REGISTRY: readonly [{
     };
     readonly load: (modelPath: string) => Promise<import("../stream.js").StreamingInstance<typeof NativeLfm2Model, {
         readonly recordModelPath: true;
+        readonly replayAssistantRawText: true;
     }>>;
     readonly nativeModelClass: typeof NativeLfm2Model;
 }, {

package/dist/models/model-loader.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-loader.d.ts","sourceRoot":"","sources":["../../src/models/model-loader.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAKH,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,YAAY,EACZ,SAAS,IAAI,eAAe,EAC5B,eAAe,EACf,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,WAAW,EAAE,KAAK,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAG3E,qEAAqE;AACrE,MAAM,WAAW,gBAAgB;IAC/B;;;;;;;;;;;;;OAaG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAID,UAAU,qBAAqB;IAC7B,QAAQ,CAAC,oBAAoB,EAAE,OAAO,CAAC;IACvC,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,SAAS,CAAC;IAC1C,QAAQ,CAAC,iBAAiB,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,aAAa,EAAE,WAAW,CAAC,MAAM,CAAC,CAAC;CAC7C;AAED,UAAU,uBAAwB,SAAQ,qBAAqB;IAC7D,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,SAAS,CAAC;CACxC;AAgCD;;;;;;;;;;;;;;;;;;GAkBG;AACH,QAAA,MAAM,qBAAqB;wBAEZ,QAAQ;mBACb,UAAU;;;;;+BAKE,MAAM,YAAY,gBAAgB;;;;;;wBASzC,SAAS;mBACd,WAAW;;;;;+BAMC,MAAM;;;wBAIb,OAAO;mBACZ,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;;;wBAKb,SAAS;mBACd,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;wBAIb,aAAa;mBAClB,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;wBAIb,MAAM;mBACX,UAAU;;iBACP,aAAa;;+BACJ,MAAM;;;;;wBAIb,UAAU;mBACf,UAAU;;iBACP,aAAa;;+BACJ,MAAM;;;;;wBAIb,eAAe;mBACpB,KAAK;;iBACF,aAAa;;+BACJ,MAAM;;;wBAIb,aAAa;mBAClB,KAAK;;iBACF,aAAa;;+BACJ,MAAM;;EAGyB,CAAC;AAEtD,MAAM,MAAM,SAAS,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAC,WAAW,CAAC,CAAC;AAE5E,KAAK,qBAAqB,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAC;AACpE,KAAK,yBAAyB,GAAG,OAAO,CAAC,qBAAqB,EAAE;IAAE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAA;CAAE,CAAC,CAAC;AAEhG;;;;;;;GAOG;AACH,MAAM,MAAM,aAAa,GAAG,YAAY,CAAC,qBAAqB,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAEpF;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,OAAO,CAAC,UAAU,CAAC,yBAAyB,CAAC,MAAM,CAAC,CAAC,CAAC,GACtD,YAAY,CAAC,yBAAyB,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAsIhE;;;;;;;;;;;GAWG;AACH,wBAAsB,SAAS,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,aAAa,CAAC,CAGrG;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAsB,WAAW,CAC/B,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE,gBAAgB,GACzB,OAAO,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC,CAa3C;AAED,wBAAsB,eAAe,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,CAAC,CAkB3E"}
+{"version":3,"file":"model-loader.d.ts","sourceRoot":"","sources":["../../src/models/model-loader.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAKH,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,YAAY,EACZ,SAAS,IAAI,eAAe,EAC5B,eAAe,EACf,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,WAAW,EAAE,KAAK,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAG3E,qEAAqE;AACrE,MAAM,WAAW,gBAAgB;IAC/B;;;;;;;;;;;;;OAaG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAID,UAAU,qBAAqB;IAC7B,QAAQ,CAAC,oBAAoB,EAAE,OAAO,CAAC;IACvC,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,SAAS,CAAC;IAC1C,QAAQ,CAAC,iBAAiB,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,aAAa,EAAE,WAAW,CAAC,MAAM,CAAC,CAAC;CAC7C;AAED,UAAU,uBAAwB,SAAQ,qBAAqB;IAC7D,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,SAAS,CAAC;CACxC;AAgCD;;;;;;;;;;;;;;;;;;GAkBG;AACH,QAAA,MAAM,qBAAqB;wBAEZ,QAAQ;mBACb,UAAU;;;;;+BAKE,MAAM,YAAY,gBAAgB;;;;;;wBASzC,SAAS;mBACd,WAAW;;;;;+BAMC,MAAM;;;wBAIb,OAAO;mBACZ,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;;;wBAKb,SAAS;mBACd,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;wBAIb,aAAa;mBAClB,WAAW;;iBACR,aAAa;;+BACJ,MAAM;;;;;wBAIb,MAAM;mBACX,UAAU;;iBACP,aAAa;;+BACJ,MAAM;;;;;;wBAIb,UAAU;mBACf,UAAU;;iBACP,aAAa;;+BACJ,MAAM;;;;;;wBAIb,eAAe;mBACpB,KAAK;;iBACF,aAAa;;+BACJ,MAAM;;;wBAIb,aAAa;mBAClB,KAAK;;iBACF,aAAa;;+BACJ,MAAM;;EAGyB,CAAC;AAEtD,MAAM,MAAM,SAAS,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAC,WAAW,CAAC,CAAC;AAE5E,KAAK,qBAAqB,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAC;AACpE,KAAK,yBAAyB,GAAG,OAAO,CAAC,qBAAqB,EAAE;IAAE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAA;CAAE,CAAC,CAAC;AAEhG;;;;;;;GAOG;AACH,MAAM,MAAM,aAAa,GAAG,YAAY,CAAC,qBAAqB,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAEpF;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,OAAO,CAAC,UAAU,CAAC,yBAAyB,CAAC,MAAM,CAAC,CAAC,CAAC,GACtD,YAAY,CAAC,yBAAyB,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAsIhE;;;;;;;;;;;GAWG;AACH,wBAAsB,SAAS,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,aAAa,CAAC,CAGrG;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAsB,WAAW,CAC/B,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE,gBAAgB,GACzB,OAAO,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC,CAa3C;AAED,wBAAsB,eAAe,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,CAAC,CAkB3E"}

package/dist/models/paged-config-override.d.ts

@@ -48,8 +48,20 @@ export declare class PagedConfigOverrideManager {
      * A caller-supplied canonical family takes precedence over the raw config
      * type (for example, `gemma4` for a `gemma4_unified` checkpoint).
      * Unmanaged, unreadable, or malformed checkpoints pass through unchanged.
+     *
+     * `persistPagedCache` is a tri-state cold-tier directive. `undefined` leaves
+     * the field untouched (families with no cold-tier opt-in). A boolean is
+     * AUTHORITATIVE: it writes `persist_paged_cache: <value>` into the cloned
+     * config, overriding whatever the source config.json carries (either alias),
+     * and forces a clone whenever the source's value disagrees so the directive
+     * actually reaches the loader. Callers gate the boolean to the families whose
+     * paged cold restore is sound — the allowlist in
+     * `packages/agent/src/cold-tier.ts`, mirrored by `COLD_RESTORE_FAMILIES` in
+     * `crates/mlx-core/src/cold_tier.rs`. It is a SET, not one family, and it is
+     * deliberately not repeated here: `MlxModelHost` calls this for every
+     * allowlisted family, and a second copy of the list would drift.
      */
-    resolve(modelPath: string, canonicalModelType?: string): Promise<string>;
+    resolve(modelPath: string, canonicalModelType?: string, persistPagedCache?: boolean): Promise<string>;
     private resolveInternal;
     /** Remove this manager's temporary root without affecting other managers. */
     cleanup(): Promise<void>;

package/dist/models/paged-config-override.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"paged-config-override.d.ts","sourceRoot":"","sources":["../../src/models/paged-config-override.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAMH,2EAA2E;AAC3E,eAAO,MAAM,uBAAuB,YAAI,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,QAAQ,EAAE,MAAM,EAAE,UAAU,CAAU,CAAC;AAElH,iEAAiE;AACjE,eAAO,MAAM,wBAAwB,YAAI,SAAS,EAAE,aAAa,CAAU,CAAC;AAO5E,MAAM,WAAW,iCAAiC;IAChD,qFAAqF;IACrF,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC/B,0EAA0E;IAC1E,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACtC;AAED;;;;;;;GAOG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,0BAA0B,CAAU;IACrD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAsC;IAChE,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA8B;IAC7D,OAAO,CAAC,WAAW,CAA8B;IACjD,OAAO,CAAC,cAAc,CAA4B;IAClD,OAAO,CAAC,QAAQ,CAAS;IAEzB,YAAY,OAAO,GAAE,iCAAsC,EAI1D;IAED;;;;;OAKG;IACG,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,kBAAkB,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAY7E;YAEa,eAAe;IAgD7B,6EAA6E;IAC7E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAKvB;YAEa,cAAc;YAad,cAAc;IA8C5B,OAAO,CAAC,OAAO;CAIhB"}
+{"version":3,"file":"paged-config-override.d.ts","sourceRoot":"","sources":["../../src/models/paged-config-override.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAMH,2EAA2E;AAC3E,eAAO,MAAM,uBAAuB,YAAI,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,QAAQ,EAAE,MAAM,EAAE,UAAU,CAAU,CAAC;AAElH,iEAAiE;AACjE,eAAO,MAAM,wBAAwB,YAAI,SAAS,EAAE,aAAa,CAAU,CAAC;AAO5E,MAAM,WAAW,iCAAiC;IAChD,qFAAqF;IACrF,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC/B,0EAA0E;IAC1E,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACtC;AAED;;;;;;;GAOG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,0BAA0B,CAAU;IACrD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAsC;IAChE,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA8B;IAC7D,OAAO,CAAC,WAAW,CAA8B;IACjD,OAAO,CAAC,cAAc,CAA4B;IAClD,OAAO,CAAC,QAAQ,CAAS;IAEzB,YAAY,OAAO,GAAE,iCAAsC,EAI1D;IAED;;;;;;;;;;;;;;;;;OAiBG;IACG,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,kBAAkB,CAAC,EAAE,MAAM,EAAE,iBAAiB,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC,CAY1G;YAEa,eAAe;IAuE7B,6EAA6E;IAC7E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAKvB;YAEa,cAAc;YAad,cAAc;IAsD5B,OAAO,CAAC,OAAO;CAIhB"}

package/dist/models/paged-config-override.js

@@ -47,12 +47,24 @@ export class PagedConfigOverrideManager {
      * A caller-supplied canonical family takes precedence over the raw config
      * type (for example, `gemma4` for a `gemma4_unified` checkpoint).
      * Unmanaged, unreadable, or malformed checkpoints pass through unchanged.
+     *
+     * `persistPagedCache` is a tri-state cold-tier directive. `undefined` leaves
+     * the field untouched (families with no cold-tier opt-in). A boolean is
+     * AUTHORITATIVE: it writes `persist_paged_cache: <value>` into the cloned
+     * config, overriding whatever the source config.json carries (either alias),
+     * and forces a clone whenever the source's value disagrees so the directive
+     * actually reaches the loader. Callers gate the boolean to the families whose
+     * paged cold restore is sound — the allowlist in
+     * `packages/agent/src/cold-tier.ts`, mirrored by `COLD_RESTORE_FAMILIES` in
+     * `crates/mlx-core/src/cold_tier.rs`. It is a SET, not one family, and it is
+     * deliberately not repeated here: `MlxModelHost` calls this for every
+     * allowlisted family, and a second copy of the list would drift.
      */
-    async resolve(modelPath, canonicalModelType) {
+    async resolve(modelPath, canonicalModelType, persistPagedCache) {
         if (this.disposed) {
             throw new Error('PagedConfigOverrideManager: resolve() called after cleanup()');
         }
-        const operation = this.resolveInternal(modelPath, canonicalModelType);
+        const operation = this.resolveInternal(modelPath, canonicalModelType, persistPagedCache);
         this.activeResolves.add(operation);
         try {
             return await operation;
@@ -61,7 +73,7 @@ export class PagedConfigOverrideManager {
             this.activeResolves.delete(operation);
         }
     }
-    async resolveInternal(modelPath, canonicalModelType) {
+    async resolveInternal(modelPath, canonicalModelType, persistPagedCache) {
         const sourcePath = isAbsolute(modelPath) ? modelPath : resolve(modelPath);
         let config;
         try {
@@ -75,6 +87,15 @@ export class PagedConfigOverrideManager {
         if (modelType === null || !this.modelTypes.has(modelType)) {
             return modelPath;
         }
+        // A boolean persist directive is authoritative: force a clone whenever the
+        // source config's EFFECTIVE value disagrees, so the directive actually reaches
+        // the loader (both `--no-persist-cache` off AND default-on). Mirror the native
+        // parser's snake-first precedence — `persist_paged_cache` wins when present,
+        // else the camelCase alias — so a config with snake=false AND camel=true does
+        // not read as `true` here (an OR) while native reads snake=false, silently
+        // dropping an authoritative `true` request.
+        const sourcePersist = typeof config.persist_paged_cache === 'boolean' ? config.persist_paged_cache : config.persistPagedCache === true;
+        const persistOverrideNeeded = persistPagedCache !== undefined && persistPagedCache !== sourcePersist;
         // Gemma4's DSpark / assistant speculative executor currently owns flat KV
         // caches. Preserve native draft discovery only for explicit callers; the
         // default paged clone intentionally omits subdirectories, hiding `draft/`
@@ -89,20 +110,28 @@ export class PagedConfigOverrideManager {
         const memorySatisfied = cacheFloorMb === undefined || (configuredMemoryMb ?? 0) >= cacheFloorMb;
         // Even an already-paged Gemma config must be cloned when `draft/` exists:
         // returning the source would expose the draft to native auto-discovery and
-        // trigger the flat-speculation/paged-cache conflict.
-        if (pagedEnabled && memorySatisfied && !hasEmbeddedGemmaDraft) {
+        // trigger the flat-speculation/paged-cache conflict. An authoritative
+        // persist directive that disagrees with the source likewise blocks the
+        // pass-through so the resolved value reaches the loader.
+        if (pagedEnabled && memorySatisfied && !hasEmbeddedGemmaDraft && !persistOverrideNeeded) {
             return modelPath;
         }
-        const existing = this.overrides.get(sourcePath);
+        // Memoize per (source, resolved family, persist directive): the same
+        // checkpoint resolved with a different persist tri-state must yield a distinct
+        // clone, not the first one cached under the bare path. `cleanup()` iterates
+        // every value, so multiple entries per path are all still disposed.
+        const persistKey = persistPagedCache === undefined ? 'u' : persistPagedCache ? 't' : 'f';
+        const cacheKey = `${sourcePath}\0${modelType}\0${persistKey}`;
+        const existing = this.overrides.get(cacheKey);
         if (existing !== undefined)
             return existing;
-        const pending = this.createOverride(sourcePath, config, cacheFloorMb, modelType);
-        this.overrides.set(sourcePath, pending);
+        const pending = this.createOverride(sourcePath, config, cacheFloorMb, modelType, persistPagedCache);
+        this.overrides.set(cacheKey, pending);
         try {
             return await pending;
         }
         catch (error) {
-            this.overrides.delete(sourcePath);
+            this.overrides.delete(cacheKey);
             throw error;
         }
     }
@@ -126,13 +155,21 @@ export class PagedConfigOverrideManager {
             await rm(root, { recursive: true, force: true }).catch(() => undefined);
         }
     }
-    async createOverride(sourcePath, sourceConfig, cacheFloorMb, modelType) {
+    async createOverride(sourcePath, sourceConfig, cacheFloorMb, modelType, persistPagedCache) {
         const root = await this.getRoot();
         const overrideDir = await mkdtemp(join(root, 'model-'));
         const config = {
             ...sourceConfig,
             use_block_paged_cache: true,
         };
+        if (persistPagedCache !== undefined) {
+            // Authoritative: the loader reads snake_case, so that spelling is the one
+            // that decides persistence. Reconcile a stray camelCase alias spread from
+            // the source config so it can never contradict the authoritative value.
+            config.persist_paged_cache = persistPagedCache;
+            if ('persistPagedCache' in config)
+                config.persistPagedCache = persistPagedCache;
+        }
         if (cacheFloorMb !== undefined) {
             config.paged_cache_memory_mb = Math.max(positiveNumber(sourceConfig.paged_cache_memory_mb) ?? 0, cacheFloorMb);
         }

package/dist/stream.d.ts

@@ -1,6 +1,6 @@
-import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative } from '@mlx-node/core';
-import type { ChatStreamChunk, ChatStreamHandle, PerformanceMetrics, ToolCallResult } from '@mlx-node/core';
-import type { SessionCapableModel } from './chat-session.js';
+import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative } from "@mlx-node/core";
+import type { ChatStreamChunk, ChatStreamHandle, PerformanceMetrics, ToolCallResult } from "@mlx-node/core";
+import type { SessionCapableModel } from "./chat-session.js";
 export interface ChatStreamDelta {
     text: string;
     done: false;
@@ -12,10 +12,23 @@ export interface ChatStreamFinal {
     finishReason: string;
     toolCalls: ToolCallResult[];
     thinking: string | null;
+    /** Effective `enable_thinking` value passed to the model chat template. */
+    thinkingEnabled: boolean;
     numTokens: number;
     promptTokens: number;
     reasoningTokens: number;
     rawText: string;
+    /**
+     * Native token-aware reasoning-redacted raw output. ChatSession uses this
+     * when it captures full reasoning internally for deterministic replay while
+     * keeping `includeReasoning: false` private to the caller.
+     */
+    publicRawText?: string;
+    /**
+     * Whether terminal `text` is the complete parsed assistant content.
+     * Gemma emits visible content exclusively as deltas and sets false.
+     */
+    textAuthoritative?: boolean;
     /**
      * Number of prompt tokens served from the reused KV-cache prefix on
      * this turn. Mirrors the `cachedTokens` field on the non-streaming
@@ -42,6 +55,15 @@ export interface ChatStreamFinal {
     performance?: PerformanceMetrics;
 }
 export type ChatStreamEvent = ChatStreamDelta | ChatStreamFinal;
+type TemplateContentOrder = "textThenMedia" | "imagesThenText";
+interface TemplateContentPolicy {
+    order: TemplateContentOrder;
+    /**
+     * When sanitized text already contains this model-owned placeholder, keep
+     * the message structured but do not synthesize additional image parts.
+     */
+    existingImagePlaceholder?: string;
+}
 /**
  * Shared AsyncGenerator adapter for callback-based native streaming methods.
  *
@@ -130,6 +152,18 @@ interface StreamingModelOptions {
      * only the factory's path-backed replacement.
      */
     applyTemplate?: boolean;
+    /**
+     * Model-specific ordering for structured multimodal content parts. The
+     * tokenizer applies this policy after sanitization while the checkpoint
+     * Jinja template continues to own all role and wire-format tokens.
+     */
+    templateContentPolicy?: TemplateContentPolicy;
+    /**
+     * Preserve the native raw assistant bytes in session history. LFM2's
+     * checkpoint template consumes reasoning inside `message.content` and does
+     * not read the structured `reasoning_content` field used by Qwen/Gemma.
+     */
+    replayAssistantRawText?: boolean;
 }
 /**
  * Shared base type produced by the factory: a `SessionCapableModel`
@@ -147,10 +181,10 @@ export type StreamingModel = SessionCapableModel;
  */
 type ResolvedApplyTemplate<O extends StreamingModelOptions> = O extends {
     applyTemplate: boolean;
-} ? O['applyTemplate'] : O['recordModelPath'];
+} ? O["applyTemplate"] : O["recordModelPath"];
 /** @internal Method names whose callback ABI is replaced by the generator wrapper. */
 export type NativeStreamingMethod = keyof NativeStreamingInstance;
-type StreamingReplacementMethod<O extends StreamingModelOptions> = NativeStreamingMethod | (ResolvedApplyTemplate<O> extends true ? 'applyChatTemplate' : never);
+type StreamingReplacementMethod<O extends StreamingModelOptions> = NativeStreamingMethod | (ResolvedApplyTemplate<O> extends true ? "applyChatTemplate" : never);
 /**
  * Instance surface of a generated streaming wrapper. Only methods replaced at
  * runtime are removed from the native instance: the three callback streaming
@@ -165,7 +199,7 @@ type StreamingReplacementMethod<O extends StreamingModelOptions> = NativeStreami
  *
  * @internal Concrete instance type returned by {@link makeStreamingModel}.
  */
-export type StreamingInstance<C extends NativeStreamingCtor, O extends StreamingModelOptions> = Omit<InstanceType<C>, StreamingReplacementMethod<O>> & SessionCapableModel & (ResolvedApplyTemplate<O> extends true ? Required<Pick<SessionCapableModel, 'applyChatTemplate'>> : object);
+export type StreamingInstance<C extends NativeStreamingCtor, O extends StreamingModelOptions> = Omit<InstanceType<C>, StreamingReplacementMethod<O>> & SessionCapableModel & (ResolvedApplyTemplate<O> extends true ? Required<Pick<SessionCapableModel, "applyChatTemplate">> : object);
 /**
  * Build the streaming-model subclass for a native chat model class.
  *
@@ -185,7 +219,7 @@ export type StreamingInstance<C extends NativeStreamingCtor, O extends Streaming
  */
 export declare function makeStreamingModel<C extends NativeStreamingCtor, const O extends StreamingModelOptions>(NativeClass: C, opts: O): {
     new (...args: ConstructorParameters<C>): StreamingInstance<C, O>;
-    load(...args: Parameters<C['load']>): Promise<StreamingInstance<C, O>>;
+    load(...args: Parameters<C["load"]>): Promise<StreamingInstance<C, O>>;
 };
 declare const Qwen35Model_base: {
     new (): StreamingInstance<typeof Qwen35ModelNative, {
@@ -219,9 +253,11 @@ export declare class Qwen35MoeModel extends Qwen35MoeModel_base {
 declare const Lfm2Model_base: {
     new (): StreamingInstance<typeof Lfm2ModelNative, {
         readonly recordModelPath: true;
+        readonly replayAssistantRawText: true;
     }>;
     load(modelPath: string): Promise<StreamingInstance<typeof Lfm2ModelNative, {
         readonly recordModelPath: true;
+        readonly replayAssistantRawText: true;
     }>>;
 };
 /** LFM2 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */

package/dist/stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAE5B,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAGV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAE7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAiChE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CAAC,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAAK,OAAO,CAAC,gBAAgB,CAAC,EACvG,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CA6IjC;AAcD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChF;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;CACzB;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,iBAAiB,CAAC,CAAC,SAAS,mBAAmB,EAAE,CAAC,SAAS,qBAAqB,IAAI,IAAI,CAClG,YAAY,CAAC,CAAC,CAAC,EACf,0BAA0B,CAAC,CAAC,CAAC,CAC9B,GACC,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;AAE9G;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,SAAS,mBAAmB,EAAE,KAAK,CAAC,CAAC,SAAS,qBAAqB,EACrG,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAcD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CA8GA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAAgE;CAAG;;;;;;;;;AAEpG,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAAmE;CAAG;;;;;;;;;AAE1G,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAA8D;CAAG;;;;;;;;;AAEhG,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAAgE;CAAG;;;;;;;;;;;AAEpG;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}
+{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAE5B,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAGV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAE7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,2EAA2E;IAC3E,eAAe,EAAE,OAAO,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAKhE,KAAK,oBAAoB,GAAG,eAAe,GAAG,gBAAgB,CAAC;AAE/D,UAAU,qBAAqB;IAC7B,KAAK,EAAE,oBAAoB,CAAC;IAC5B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC;AA6DD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CACT,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAC1D,OAAO,CAAC,gBAAgB,CAAC,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CA+JjC;AAcD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAC7B,GAAG,IAAI,EAAE,KAAK,EAAE,KACb,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChC;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;OAIG;IACH,qBAAqB,CAAC,EAAE,qBAAqB,CAAC;IAC9C;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAC;CAClC;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,iBAAiB,CAC3B,CAAC,SAAS,mBAAmB,EAC7B,CAAC,SAAS,qBAAqB,IAC7B,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,0BAA0B,CAAC,CAAC,CAAC,CAAC,GACtD,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAClC,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GACxD,MAAM,CAAC,CAAC;AAEd;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAChC,CAAC,SAAS,mBAAmB,EAC7B,KAAK,CAAC,CAAC,SAAS,qBAAqB,EAErC,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAcD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CAqIA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;;;AAEL,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAG7B;CAAG;;;;;;;;;AAEL,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;;;AAEL;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}

package/dist/stream.js

@@ -1,25 +1,28 @@
-import { join } from 'node:path';
-import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Tokenizer, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative, } from '@mlx-node/core';
+import { join } from "node:path";
+import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Tokenizer, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative, } from "@mlx-node/core";
 const modelPathsForTokenizers = new WeakMap();
 const tokenizerPromises = new WeakMap();
 function getNativeIsReasoning(chunk) {
-    return typeof chunk.isReasoning === 'boolean' ? chunk.isReasoning : undefined;
+    return typeof chunk.isReasoning === "boolean" ? chunk.isReasoning : undefined;
 }
 function rememberModelPath(model, modelPath) {
     modelPathsForTokenizers.set(model, modelPath);
 }
-async function applyChatTemplateFromModelPath(model, messages, addGenerationPrompt, tools, enableThinking) {
+async function applyChatTemplateFromModelPath(model, messages, addGenerationPrompt, tools, enableThinking, contentPolicy) {
     const modelPath = modelPathsForTokenizers.get(model);
     if (modelPath == null) {
-        throw new Error('applyChatTemplate unavailable: model path was not recorded when this model was loaded');
+        throw new Error("applyChatTemplate unavailable: model path was not recorded when this model was loaded");
     }
     let tokenizerPromise = tokenizerPromises.get(model);
     if (tokenizerPromise == null) {
-        tokenizerPromise = Qwen3Tokenizer.fromPretrained(join(modelPath, 'tokenizer.json'));
+        tokenizerPromise = Qwen3Tokenizer.fromPretrained(join(modelPath, "tokenizer.json"));
         tokenizerPromises.set(model, tokenizerPromise);
     }
     const tokenizer = await tokenizerPromise;
-    return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking);
+    if (contentPolicy == null) {
+        return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking);
+    }
+    return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking, contentPolicy.order, contentPolicy.existingImagePlaceholder);
 }
 /**
  * Shared AsyncGenerator adapter for callback-based native streaming methods.
@@ -127,7 +130,7 @@ export async function* _runChatStream(startCall, signal) {
         }
         else {
             onAbort = triggerAbort;
-            signal.addEventListener('abort', onAbort, { once: true });
+            signal.addEventListener("abort", onAbort, { once: true });
         }
     }
     try {
@@ -149,6 +152,9 @@ export async function* _runChatStream(startCall, signal) {
                     throw item.error;
                 const chunk = item.chunk;
                 if (chunk.done) {
+                    if (typeof chunk.thinkingEnabled !== "boolean") {
+                        throw new Error("Native terminal chat stream chunk is missing thinkingEnabled");
+                    }
                     // The native `ChatStreamChunk` carries `cachedTokens` on the
                     // terminal (`done == true`) chunk for every streaming entry
                     // point. Emit it on the final event verbatim — undefined means
@@ -164,15 +170,22 @@ export async function* _runChatStream(startCall, signal) {
                         finishReason: chunk.finishReason,
                         toolCalls: chunk.toolCalls ?? [],
                         thinking: chunk.thinking ?? null,
+                        thinkingEnabled: chunk.thinkingEnabled,
                         numTokens: chunk.numTokens,
                         promptTokens: chunk.promptTokens ?? 0,
                         reasoningTokens: chunk.reasoningTokens ?? 0,
                         rawText: chunk.rawText,
                         performance: chunk.performance ?? undefined,
                     };
-                    if (typeof chunkWithCached.cachedTokens === 'number') {
+                    if (typeof chunkWithCached.cachedTokens === "number") {
                         finalEvent.cachedTokens = chunkWithCached.cachedTokens;
                     }
+                    if (typeof chunkWithCached.publicRawText === "string") {
+                        finalEvent.publicRawText = chunkWithCached.publicRawText;
+                    }
+                    if (typeof chunkWithCached.textAuthoritative === "boolean") {
+                        finalEvent.textAuthoritative = chunkWithCached.textAuthoritative;
+                    }
                     yield finalEvent;
                     return;
                 }
@@ -188,7 +201,7 @@ export async function* _runChatStream(startCall, signal) {
     finally {
         if (signal != null && onAbort != null) {
             try {
-                signal.removeEventListener('abort', onAbort);
+                signal.removeEventListener("abort", onAbort);
             }
             catch {
                 // removeEventListener shouldn't throw, but stay defensive —
@@ -218,6 +231,8 @@ export async function* _runChatStream(startCall, signal) {
 export function makeStreamingModel(NativeClass, opts) {
     const recordPath = opts.recordModelPath;
     const applyTemplate = opts.applyTemplate ?? recordPath;
+    const templateContentPolicy = opts.templateContentPolicy;
+    const replayAssistantRawText = opts.replayAssistantRawText ?? false;
     // Capture the native callback-based methods before the subclass
     // overrides below shadow them on the prototype.
     const nativeStart = NativeClass.prototype.chatStreamSessionStart;
@@ -228,6 +243,12 @@ export function makeStreamingModel(NativeClass, opts) {
     // unchanged — we extend the real native class.
     const Base = NativeClass;
     class StreamingModelImpl extends Base {
+        supportsReplayReasoningCapture() {
+            return true;
+        }
+        replaysAssistantRawText() {
+            return replayAssistantRawText;
+        }
         static async load(modelPath, ...rest) {
             // Forward any trailing family-specific load options verbatim (e.g.
             // Gemma4's `Gemma4LoadOptions` with `draftModelPath`); families whose
@@ -251,19 +272,19 @@ export function makeStreamingModel(NativeClass, opts) {
         async *chatStreamSessionStart(messages, config, signal) {
             yield* _runChatStream((callback) => nativeStart.call(this, messages, (config ?? null), callback), signal);
         }
-        async *chatStreamSessionContinue(userMessage, images, audio, config, signal) {
-            yield* _runChatStream((callback) => nativeContinue.call(this, userMessage, images, audio, (config ?? null), callback), signal);
+        async *chatStreamSessionContinue(messages, config, signal) {
+            yield* _runChatStream((callback) => nativeContinue.call(this, messages, (config ?? null), callback), signal);
         }
-        async *chatStreamSessionContinueTool(toolCallId, content, config, signal, isError) {
-            yield* _runChatStream((callback) => nativeContinueTool.call(this, toolCallId, content, (config ?? null), callback, (isError ?? null)), signal);
+        async *chatStreamSessionContinueTool(messages, config, signal) {
+            yield* _runChatStream((callback) => nativeContinueTool.call(this, messages, (config ?? null), callback), signal);
         }
     }
     if (applyTemplate) {
-        Object.defineProperty(StreamingModelImpl.prototype, 'applyChatTemplate', {
+        Object.defineProperty(StreamingModelImpl.prototype, "applyChatTemplate", {
             configurable: true,
             writable: true,
             value(messages, addGenerationPrompt, tools, enableThinking) {
-                return applyChatTemplateFromModelPath(this, messages, addGenerationPrompt, tools, enableThinking);
+                return applyChatTemplateFromModelPath(this, messages, addGenerationPrompt, tools, enableThinking, templateContentPolicy);
             },
         });
     }
@@ -277,16 +298,25 @@ export function makeStreamingModel(NativeClass, opts) {
  * `.name === 'Qwen35Model'` and a working `instanceof`. Records its
  * model path so `applyChatTemplate` can serve a lazily built tokenizer.
  */
-export class Qwen35Model extends makeStreamingModel(Qwen35ModelNative, { recordModelPath: true }) {
+export class Qwen35Model extends makeStreamingModel(Qwen35ModelNative, {
+    recordModelPath: true,
+}) {
 }
 /** Qwen3.5 MoE model — see {@link Qwen35Model} for the wrapper shape. */
-export class Qwen35MoeModel extends makeStreamingModel(Qwen35MoeModelNative, { recordModelPath: true }) {
+export class Qwen35MoeModel extends makeStreamingModel(Qwen35MoeModelNative, {
+    recordModelPath: true,
+}) {
 }
 /** LFM2 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */
-export class Lfm2Model extends makeStreamingModel(Lfm2ModelNative, { recordModelPath: true }) {
+export class Lfm2Model extends makeStreamingModel(Lfm2ModelNative, {
+    recordModelPath: true,
+    replayAssistantRawText: true,
+}) {
 }
 /** Gemma4 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */
-export class Gemma4Model extends makeStreamingModel(Gemma4ModelNative, { recordModelPath: true }) {
+export class Gemma4Model extends makeStreamingModel(Gemma4ModelNative, {
+    recordModelPath: true,
+}) {
 }
 /**
  * Qwen3 (first-gen, text-only) model.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mlx-node/lm",
-  "version": "0.0.8",
+  "version": "0.0.10",
   "homepage": "https://github.com/mlx-node/mlx-node",
   "bugs": {
     "url": "https://github.com/mlx-node/mlx-node/issues"
@@ -28,7 +28,7 @@
     "test": "vite test run"
   },
   "dependencies": {
-    "@mlx-node/core": "0.0.8"
+    "@mlx-node/core": "0.0.10"
   },
   "devDependencies": {
     "@types/node": "^26.0.0"