cactus-react-native 0.2.5 → 0.2.7

package/lib/typescript/lm.d.ts

@@ -8,9 +8,13 @@ interface CactusLMReturn {
 export declare class CactusLM {
     protected context: LlamaContext;
     protected conversationHistoryManager: ConversationHistoryManager;
+    private initParams;
     private static _initCache;
     private static getCacheKey;
-    protected constructor(context: LlamaContext);
+    protected constructor(context: LlamaContext, initParams: ContextParams);
+    private static isContextNotFoundError;
+    private reinit;
+    private run;
     static init(params: ContextParams, onProgress?: (progress: number) => void, cactusToken?: string, retryOptions?: {
         maxRetries?: number;
         delayMs?: number;

package/lib/typescript/lm.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"lm.d.ts","sourceRoot":"","sources":["../../src/lm.ts"],"names":[],"mappings":"AAAA,OAAO,EAAa,YAAY,EAAE,MAAM,SAAS,CAAA;AAGjD,OAAO,KAAK,EACV,aAAa,EACb,gBAAgB,EAChB,0BAA0B,EAC1B,sBAAsB,EACtB,eAAe,EACf,qBAAqB,EACtB,MAAM,SAAS,CAAA;AAIhB,OAAO,EAAE,0BAA0B,EAAE,MAAM,QAAQ,CAAA;AAEnD,UAAU,cAAc;IACtB,EAAE,EAAE,QAAQ,GAAG,IAAI,CAAA;IACnB,KAAK,EAAE,KAAK,GAAG,IAAI,CAAA;CACpB;AAED,qBAAa,QAAQ;IACnB,SAAS,CAAC,OAAO,EAAE,YAAY,CAAA;IAC/B,SAAS,CAAC,0BAA0B,EAAE,0BAA0B,CAAA;IAIhE,OAAO,CAAC,MAAM,CAAC,UAAU,CAAmD;IAE5E,OAAO,CAAC,MAAM,CAAC,WAAW;IAI1B,SAAS,aAAa,OAAO,EAAE,YAAY;WAK9B,IAAI,CACf,MAAM,EAAE,aAAa,EACrB,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,EACvC,WAAW,CAAC,EAAE,MAAM,EACpB,YAAY,CAAC,EAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,GACvD,OAAO,CAAC,cAAc,CAAC;IA8E1B,UAAU,GACR,UAAU,0BAA0B,EAAE,EACtC,SAAQ,gBAAqB,EAC7B,WAAW,CAAC,IAAI,EAAE,GAAG,KAAK,IAAI,KAC7B,OAAO,CAAC,sBAAsB,CAAC,CAqBjC;IAEK,SAAS,CACb,IAAI,EAAE,MAAM,EACZ,MAAM,CAAC,EAAE,eAAe,EACxB,IAAI,GAAE,MAAgB,GACrB,OAAO,CAAC,qBAAqB,CAAC;cAoCjB,qBAAqB,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,qBAAqB,CAAC;cAI7F,sBAAsB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,CAAC;IAOpF,MAAM,QAAa,OAAO,CAAC,IAAI,CAAC,CAE/B;IAEK,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IAIxB,cAAc,IAAI,OAAO,CAAC,IAAI,CAAC;CAItC"}
+{"version":3,"file":"lm.d.ts","sourceRoot":"","sources":["../../src/lm.ts"],"names":[],"mappings":"AAAA,OAAO,EAAa,YAAY,EAAE,MAAM,SAAS,CAAA;AAGjD,OAAO,KAAK,EACV,aAAa,EACb,gBAAgB,EAChB,0BAA0B,EAC1B,sBAAsB,EACtB,eAAe,EACf,qBAAqB,EACtB,MAAM,SAAS,CAAA;AAIhB,OAAO,EAAE,0BAA0B,EAAE,MAAM,QAAQ,CAAA;AAEnD,UAAU,cAAc;IACtB,EAAE,EAAE,QAAQ,GAAG,IAAI,CAAA;IACnB,KAAK,EAAE,KAAK,GAAG,IAAI,CAAA;CACpB;AAED,qBAAa,QAAQ;IACnB,SAAS,CAAC,OAAO,EAAE,YAAY,CAAA;IAC/B,SAAS,CAAC,0BAA0B,EAAE,0BAA0B,CAAA;IAChE,OAAO,CAAC,UAAU,CAAe;IACjC,OAAO,CAAC,MAAM,CAAC,UAAU,CAAmD;IAE5E,OAAO,CAAC,MAAM,CAAC,WAAW;IAI1B,SAAS,aAAa,OAAO,EAAE,YAAY,EAAE,UAAU,EAAE,aAAa;IAMtE,OAAO,CAAC,MAAM,CAAC,sBAAsB;YAKvB,MAAM;YAMN,GAAG;WAUJ,IAAI,CACf,MAAM,EAAE,aAAa,EACrB,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,EACvC,WAAW,CAAC,EAAE,MAAM,EACpB,YAAY,CAAC,EAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,GACvD,OAAO,CAAC,cAAc,CAAC;IA4E1B,UAAU,GACR,UAAU,0BAA0B,EAAE,EACtC,SAAQ,gBAAqB,EAC7B,WAAW,CAAC,IAAI,EAAE,GAAG,KAAK,IAAI,KAC7B,OAAO,CAAC,sBAAsB,CAAC,CAuBjC;IAEK,SAAS,CACb,IAAI,EAAE,MAAM,EACZ,MAAM,CAAC,EAAE,eAAe,EACxB,IAAI,GAAE,MAAgB,GACrB,OAAO,CAAC,qBAAqB,CAAC;cAoCjB,qBAAqB,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,qBAAqB,CAAC;cAI7F,sBAAsB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,CAAC;IAOpF,MAAM,QAAa,OAAO,CAAC,IAAI,CAAC,CAE/B;IAEK,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IAUxB,cAAc,IAAI,OAAO,CAAC,IAAI,CAAC;CAItC"}

package/lib/typescript/telemetry.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"telemetry.d.ts","sourceRoot":"","sources":["../../src/telemetry.ts"],"names":[],"mappings":"AAKA,MAAM,WAAW,eAAe;IAC9B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAA;IAC3B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;CACrB;AAmBD,UAAU,eAAe;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,qBAAa,SAAS;IACpB,OAAO,CAAC,MAAM,CAAC,QAAQ,CAA0B;IACjD,OAAO,CAAC,MAAM,CAA4B;IAE1C,OAAO;IAOP,OAAO,CAAC,MAAM,CAAC,WAAW;IAQ1B,MAAM,CAAC,QAAQ,IAAI,IAAI;IASvB,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI;IAM1C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,EAAE,OAAO,EAAE,eAAe,EAAE,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,GAAG,IAAI;IAOhH,MAAM,CAAC,KAAK,CAAC,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,eAAe,GAAG,IAAI;IAO1D,OAAO,CAAC,aAAa;IAoBrB,OAAO,CAAC,aAAa;YAuBP,UAAU;CAYzB"}
+{"version":3,"file":"telemetry.d.ts","sourceRoot":"","sources":["../../src/telemetry.ts"],"names":[],"mappings":"AAIA,MAAM,WAAW,eAAe;IAC9B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAA;IAC3B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;CACrB;AAmBD,UAAU,eAAe;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,qBAAa,SAAS;IACpB,OAAO,CAAC,MAAM,CAAC,QAAQ,CAA0B;IACjD,OAAO,CAAC,MAAM,CAA4B;IAE1C,OAAO;IAOP,OAAO,CAAC,MAAM,CAAC,WAAW;IAQ1B,MAAM,CAAC,QAAQ,IAAI,IAAI;IASvB,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI;IAM1C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,EAAE,OAAO,EAAE,eAAe,EAAE,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,GAAG,IAAI;IAOhH,MAAM,CAAC,KAAK,CAAC,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,eAAe,GAAG,IAAI;IAO1D,OAAO,CAAC,aAAa;IAoBrB,OAAO,CAAC,aAAa;YAuBP,UAAU;CAYzB"}

package/lib/typescript/vlm.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"vlm.d.ts","sourceRoot":"","sources":["../../src/vlm.ts"],"names":[],"mappings":"AAAA,OAAO,EAKL,KAAK,aAAa,EAClB,KAAK,gBAAgB,EACrB,KAAK,0BAA0B,EAC/B,KAAK,sBAAsB,EAC5B,MAAM,SAAS,CAAA;AAIhB,OAAO,EAAE,0BAA0B,EAAE,MAAM,QAAQ,CAAA;AAEnD,UAAU,eAAe;IACvB,GAAG,EAAE,SAAS,GAAG,IAAI,CAAA;IACrB,KAAK,EAAE,KAAK,GAAG,IAAI,CAAA;CACpB;AAED,MAAM,MAAM,gBAAgB,GAAG,aAAa,GAAG;IAC7C,MAAM,EAAE,MAAM,CAAA;CACf,CAAA;AAED,MAAM,MAAM,mBAAmB,GAAG,IAAI,CAAC,gBAAgB,EAAE,QAAQ,CAAC,GAAG;IACnE,MAAM,CAAC,EAAE,MAAM,EAAE,CAAA;IACjB,IAAI,CAAC,EAAE,MAAM,CAAA;CACd,CAAA;AAED,qBAAa,SAAS;IACpB,OAAO,CAAC,OAAO,CAAc;IAC7B,SAAS,CAAC,0BAA0B,EAAE,0BAA0B,CAAA;IAGhE,OAAO,CAAC,MAAM,CAAC,UAAU,CAAoD;IAE7E,OAAO,CAAC,MAAM,CAAC,WAAW;IAI1B,OAAO;WAKM,IAAI,CACf,MAAM,EAAE,gBAAgB,EACxB,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,EACvC,WAAW,CAAC,EAAE,MAAM,EACpB,YAAY,CAAC,EAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,GACvD,OAAO,CAAC,eAAe,CAAC;IA8ErB,UAAU,CACd,QAAQ,EAAE,0BAA0B,EAAE,EACtC,MAAM,GAAE,mBAAwB,EAChC,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,GAAG,KAAK,IAAI,GAC7B,OAAO,CAAC,sBAAsB,CAAC;IAuClC,OAAO,CAAC,sBAAsB,CAyC7B;YAEa,uBAAuB;IA+C/B,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IAIxB,cAAc,IAAI,OAAO,CAAC,IAAI,CAAC;CAGtC"}
+{"version":3,"file":"vlm.d.ts","sourceRoot":"","sources":["../../src/vlm.ts"],"names":[],"mappings":"AAAA,OAAO,EAKL,KAAK,aAAa,EAClB,KAAK,gBAAgB,EACrB,KAAK,0BAA0B,EAC/B,KAAK,sBAAsB,EAC5B,MAAM,SAAS,CAAA;AAIhB,OAAO,EAAE,0BAA0B,EAAE,MAAM,QAAQ,CAAA;AAEnD,UAAU,eAAe;IACvB,GAAG,EAAE,SAAS,GAAG,IAAI,CAAA;IACrB,KAAK,EAAE,KAAK,GAAG,IAAI,CAAA;CACpB;AAED,MAAM,MAAM,gBAAgB,GAAG,aAAa,GAAG;IAC7C,MAAM,EAAE,MAAM,CAAA;CACf,CAAA;AAED,MAAM,MAAM,mBAAmB,GAAG,IAAI,CAAC,gBAAgB,EAAE,QAAQ,CAAC,GAAG;IACnE,MAAM,CAAC,EAAE,MAAM,EAAE,CAAA;IACjB,IAAI,CAAC,EAAE,MAAM,CAAA;CACd,CAAA;AAED,qBAAa,SAAS;IACpB,OAAO,CAAC,OAAO,CAAc;IAC7B,SAAS,CAAC,0BAA0B,EAAE,0BAA0B,CAAA;IAEhE,OAAO,CAAC,MAAM,CAAC,UAAU,CAAoD;IAE7E,OAAO,CAAC,MAAM,CAAC,WAAW;IAI1B,OAAO;WAKM,IAAI,CACf,MAAM,EAAE,gBAAgB,EACxB,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,EACvC,WAAW,CAAC,EAAE,MAAM,EACpB,YAAY,CAAC,EAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,GACvD,OAAO,CAAC,eAAe,CAAC;IA8ErB,UAAU,CACd,QAAQ,EAAE,0BAA0B,EAAE,EACtC,MAAM,GAAE,mBAAwB,EAChC,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,GAAG,KAAK,IAAI,GAC7B,OAAO,CAAC,sBAAsB,CAAC;IAuClC,OAAO,CAAC,sBAAsB,CAyC7B;YAEa,uBAAuB;IA+C/B,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;IAIxB,cAAc,IAAI,OAAO,CAAC,IAAI,CAAC;CAGtC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cactus-react-native",
-  "version": "0.2.5",
+  "version": "0.2.7",
   "description": "Run AI models locally on mobile devices",
   "main": "lib/commonjs/index",
   "module": "lib/module/index",

package/src/agent.ts

@@ -0,0 +1,109 @@
+import { CactusLM } from './lm'
+
+import type {
+  ContextParams,
+  CompletionParams,
+  CactusOAICompatibleMessage,
+  NativeCompletionResult,
+  TokenData,
+} from './index'
+
+import { Tools } from './tools'
+
+interface Parameter {
+  type: string
+  description: string
+  required?: boolean
+}
+
+interface CactusAgentReturn {
+  agent: CactusAgent | null
+  error: Error | null
+  lm: CactusAgent | null
+}
+
+export interface AgentCompletionParams extends CompletionParams {
+  tools?: Tools
+}
+
+export class CactusAgent extends CactusLM {
+  private tools!: Tools
+
+  static async init(
+    params: ContextParams,
+    onProgress?: (progress: number) => void,
+    cactusToken?: string,
+    retryOptions?: { maxRetries?: number; delayMs?: number },
+  ): Promise<CactusAgentReturn> {
+    const result = await CactusLM.init(params, onProgress, cactusToken, retryOptions)
+    
+    if (result.error || !result.lm) {
+      return { agent: null, error: result.error, lm: null }
+    }
+
+    const agent = Object.setPrototypeOf(result.lm, CactusAgent.prototype) as CactusAgent
+    agent.tools = new Tools()
+    return { agent, error: null, lm: agent }
+  }
+
+  addTool(
+    func: Function,
+    description: string,
+    parameters: { [key: string]: Parameter }
+  ): Function {
+    return this.tools.add(func, description, parameters)
+  }
+
+  getTools(): Tools {
+    return this.tools
+  }
+
+  async completionWithTools(
+    messages: CactusOAICompatibleMessage[],
+    params: AgentCompletionParams = {},
+    callback?: (data: TokenData) => void,
+    recursionCount: number = 0,
+    recursionLimit: number = 3
+  ): Promise<NativeCompletionResult> {
+    const tools = params.tools || this.tools
+    
+    if (!messages?.length) {
+      return this.completion([], params, callback)
+    }
+    if (!tools || tools.getSchemas().length === 0) {
+      return this.completion(messages, params, callback)
+    }
+    if (recursionCount >= recursionLimit) {
+      return this.completion(messages, {
+        ...params,
+        jinja: true,
+        tools: tools.getSchemas()
+      }, callback)
+    }
+
+    const { newMessages, requiresReset } =
+      this.conversationHistoryManager.processNewMessages(messages)
+
+    if (requiresReset) {
+      this.context?.rewind()
+      this.conversationHistoryManager.reset()
+    }
+
+    const result = await this.context.completionWithTools({
+      ...params,
+      messages: newMessages.length > 0 ? newMessages : messages,
+      tools: tools
+    }, callback, recursionCount, recursionLimit)
+
+    this.conversationHistoryManager.update(
+      newMessages.length > 0 ? newMessages : messages,
+      {
+        role: 'assistant',
+        content: result.content,
+        tool_calls: result.tool_calls
+      }
+    )
+
+    return result
+  }
+}

package/src/chat.ts

@@ -6,7 +6,9 @@ export type CactusMessagePart = {
 
 export type CactusOAICompatibleMessage = {
   role: string
-  content?: string | CactusMessagePart[] | any 
+  content?: string | CactusMessagePart[] | any
+  tool_calls?: any[]
+  tool_call_id?: string
 }
 
 export function formatChat(

package/src/index.ts

@@ -27,7 +27,6 @@ import type { CactusMessagePart, CactusOAICompatibleMessage } from './chat'
 import { formatChat } from './chat'
 import { Tools, parseAndExecuteTool } from './tools'
 import { Telemetry, type TelemetryParams } from './telemetry'
-
 export type {
   NativeContextParams,
   NativeLlamaContext,
@@ -45,9 +44,11 @@ export type {
   JinjaFormattedChatResult,
   NativeAudioDecodeResult,
 }
-
-export {Tools }
 export * from './remote'
+export {
+  Tools, 
+  parseAndExecuteTool
+} from './tools'
 
 const EVENT_ON_INIT_CONTEXT_PROGRESS = '@Cactus_onInitContextProgress'
 const EVENT_ON_TOKEN = '@Cactus_onToken'
@@ -72,7 +73,6 @@ if (EventEmitter) {
       logListeners.forEach((listener) => listener(evt.level, evt.text))
     },
   )
-  // Trigger unset to use default log callback
   Cactus?.toggleNativeLog?.(false)?.catch?.(() => {})
 }
 
@@ -181,18 +181,12 @@ export class LlamaContext {
     this.model = model
   }
 
-  /**
-   * Load cached prompt & completion state from a file.
-   */
   async loadSession(filepath: string): Promise<NativeSessionLoadResult> {
     let path = filepath
     if (path.startsWith('file://')) path = path.slice(7)
     return Cactus.loadSession(this.id, path)
   }
 
-  /**
-   * Save current cached prompt & completion state to a file.
-   */
   async saveSession(
     filepath: string,
     options?: { tokenSize: number },
@@ -242,10 +236,10 @@ export class LlamaContext {
     recursionCount: number = 0,
     recursionLimit: number = 3
 ): Promise<NativeCompletionResult> {
-    if (!params.messages) { // tool calling only works with messages
+    if (!params.messages) { 
         return this.completion(params, callback);
     }
-    if (!params.tools) { // no tools => default completion
+    if (!params.tools) { 
         return this.completion(params, callback);
     }
     if (recursionCount >= recursionLimit) {
@@ -256,7 +250,7 @@ export class LlamaContext {
         }, callback);
     }
 
-    const messages = [...params.messages]; // avoid mutating the original messages
+    const messages = [...params.messages];
 
     const result = await this.completion({
         ...params, 
@@ -307,7 +301,7 @@ export class LlamaContext {
       emit_partial_completion: !!callback,
     }
     if (params.messages) {
-      // messages always win
+
       const formattedResult = await this.getFormattedChat(
         params.messages,
         params.chat_template || params.chatTemplate,
@@ -502,7 +496,6 @@ export async function loadLlamaModelInfo(model: string): Promise<Object> {
 }
 
 const poolTypeMap = {
-  // -1 is unspecified as undefined
   none: 0,
   mean: 1,
   cls: 2,
@@ -665,4 +658,5 @@ export const getDeviceInfo = async (contextId: number): Promise<NativeDeviceInfo
 
 export { CactusLM } from './lm';
 export { CactusVLM } from './vlm';
-export { CactusTTS } from './tts';
+export { CactusTTS } from './tts';
+export { CactusAgent } from './agent';

package/src/lm.ts

@@ -22,20 +22,40 @@ interface CactusLMReturn {
 export class CactusLM {
   protected context: LlamaContext
   protected conversationHistoryManager: ConversationHistoryManager
-
-  // the initPromise enables a "async singleton" initialization pattern which
-  // protects against a race condition in the event of multiple init attempts
+  private initParams: ContextParams
   private static _initCache: Map<string, Promise<CactusLMReturn>> = new Map();
 
   private static getCacheKey(params: ContextParams, cactusToken?: string, retryOptions?: { maxRetries?: number; delayMs?: number }): string {
     return JSON.stringify({ params, cactusToken, retryOptions });
   }
 
-  protected constructor(context: LlamaContext) {
+  protected constructor(context: LlamaContext, initParams: ContextParams) {
     this.context = context
+    this.initParams = initParams
     this.conversationHistoryManager = new ConversationHistoryManager()
   }
 
+  private static isContextNotFoundError(e: unknown): boolean {
+    const message = String((e as any)?.message ?? e ?? '')
+    return /context not found/i.test(message)
+  }
+
+  private async reinit(): Promise<void> {
+    const newContext = await initLlama(this.initParams)
+    this.context = newContext
+    this.conversationHistoryManager.reset()
+  }
+
+  private async run<T>(op: () => Promise<T>): Promise<T> {
+    try {
+      return await op()
+    } catch (e) {
+      if (!CactusLM.isContextNotFoundError(e)) throw e
+      await this.reinit()
+      return await op()
+    }
+  }
+
   static async init(
     params: ContextParams,
     onProgress?: (progress: number) => void,
@@ -49,7 +69,6 @@ export class CactusLM {
 
     const key = CactusLM.getCacheKey(params, cactusToken, retryOptions);
     if (CactusLM._initCache.has(key)) {
-      // concurrent initialization calls with the same params all get the same cached Promise
       return CactusLM._initCache.get(key)!;
     }
 
@@ -82,7 +101,7 @@ export class CactusLM {
         for (let attempt = 1; attempt <= maxRetries; attempt++) {
           try {
             const context = await initLlama(config, onProgress);
-            return { lm: new CactusLM(context), error: null };
+            return { lm: new CactusLM(context, config), error: null };
           } catch (e) {
             lastError = e as Error;
             const isLastConfig = configs.indexOf(config) === configs.length - 1;
@@ -113,9 +132,8 @@ export class CactusLM {
     CactusLM._initCache.set(key, initPromise);
 
     const result = await initPromise;
-    if (result.error) {
-      CactusLM._initCache.delete(key); // Reset on failure to allow retries
-    }
+    // Cache only while in-flight; never cache resolved instances
+    CactusLM._initCache.delete(key);
     return result;
   }
 
@@ -128,7 +146,7 @@ export class CactusLM {
       this.conversationHistoryManager.processNewMessages(messages);
 
     if (requiresReset) {
-      this.context?.rewind();
+      await this.run(() => this.context.rewind())
       this.conversationHistoryManager.reset();
     }
 
@@ -136,7 +154,9 @@ export class CactusLM {
       console.warn('No messages to complete!');
     }
 
-    const result = await this.context.completion({ messages: newMessages, ...params }, callback);
+    const result = await this.run(() =>
+      this.context.completion({ messages: newMessages, ...params }, callback),
+    )
 
     this.conversationHistoryManager.update(newMessages, {
       role: 'assistant',
@@ -187,7 +207,7 @@ export class CactusLM {
   }
 
   protected async _handleLocalEmbedding(text: string, params?: EmbeddingParams): Promise<NativeEmbeddingResult> {
-    return this.context.embedding(text, params);
+    return this.run(() => this.context.embedding(text, params))
   }
 
   protected async _handleRemoteEmbedding(text: string): Promise<NativeEmbeddingResult> {
@@ -198,15 +218,21 @@ export class CactusLM {
   }
 
   rewind = async (): Promise<void> => {
-    return this.context?.rewind()
+    return this.run(() => this.context.rewind())
   }
 
   async release(): Promise<void> {
-    return this.context.release()
+    try {
+      return await this.context.release()
+    } catch (e) {
+      // Treat missing context as already released
+      if (CactusLM.isContextNotFoundError(e)) return
+      throw e
+    }
   }
 
   async stopCompletion(): Promise<void> {
-    return await this.context.stopCompletion()
+    return await this.run(() => this.context.stopCompletion())
   }
 
 } 

package/src/telemetry.ts

@@ -1,5 +1,4 @@
 import { Platform } from 'react-native'
-// Import package.json to get version
 const packageJson = require('../package.json');
 import { PROJECT_ID } from './projectId';
 

package/src/tools.ts

@@ -3,7 +3,7 @@ import type { NativeCompletionResult } from "./NativeCactus";
 interface Parameter {
   type: string,
   description: string,
-  required?: boolean // parameter is optional if not specified
+  required?: boolean 
 }
 
 interface Tool {

package/src/vlm.ts

@@ -31,7 +31,6 @@ export class CactusVLM {
   private context: LlamaContext
   protected conversationHistoryManager: ConversationHistoryManager
 
-  // see CactusLM for detailed docs
   private static _initCache: Map<string, Promise<CactusVLMReturn>> = new Map();
 
   private static getCacheKey(params: VLMContextParams, cactusToken?: string, retryOptions?: { maxRetries?: number; delayMs?: number }): string {