@illuma-ai/agents 1.0.89 → 1.0.93

package/dist/types/tools/handlers.d.ts

@@ -1,5 +1,5 @@
 import type { ToolCall, ToolCallChunk } from '@langchain/core/messages/tool';
-import type { MultiAgentGraph, StandardGraph } from '@/graphs';
+import type { Graph, MultiAgentGraph, StandardGraph } from '@/graphs';
 import type { AgentContext } from '@/agents/AgentContext';
 import type * as t from '@/types';
 export declare function handleToolCallChunks({ graph, stepKey, toolCallChunks, metadata, }: {
@@ -8,7 +8,7 @@ export declare function handleToolCallChunks({ graph, stepKey, toolCallChunks, m
     toolCallChunks: ToolCallChunk[];
     metadata?: Record<string, unknown>;
 }): Promise<void>;
-export declare const handleToolCalls: (toolCalls?: ToolCall[], metadata?: Record<string, unknown>, graph?: StandardGraph | MultiAgentGraph) => Promise<void>;
+export declare const handleToolCalls: (toolCalls?: ToolCall[], metadata?: Record<string, unknown>, graph?: Graph | StandardGraph | MultiAgentGraph) => Promise<void>;
 export declare const toolResultTypes: Set<string>;
 /**
  * Handles the result of a server tool call; in other words, a provider's built-in tool.

package/dist/types/types/graph.d.ts

@@ -220,8 +220,8 @@ export type GraphEdge = {
     description?: string;
     /** Can return boolean or specific destination(s) */
     condition?: (state: BaseGraphState) => boolean | string | string[];
-    /** 'handoff' creates tools for dynamic routing, 'direct' creates direct edges, which also allow parallel execution */
-    edgeType?: 'handoff' | 'direct';
+    /** EdgeType.HANDOFF creates tools for dynamic routing, EdgeType.DIRECT creates direct edges with parallel execution */
+    edgeType?: import('@/common').EdgeType;
     /**
      * For direct edges: Optional prompt to add when transitioning through this edge.
      * String prompts can include variables like {results} which will be replaced with
@@ -405,4 +405,11 @@ export interface AgentInputs {
      * using a cheap LLM call, and the summary is prepended to the context.
      */
     summarizeCallback?: (messagesToRefine: import('@langchain/core/messages').BaseMessage[]) => Promise<string | undefined>;
+    /**
+     * Pre-existing summary text loaded from persistent storage (MongoDB/Redis).
+     * When provided, this summary is injected into the initial message context
+     * so the agent has prior conversation history even on new turns.
+     * Set by Ranger's SummaryStore when resuming a conversation.
+     */
+    persistedSummary?: string;
 }

package/dist/types/types/llm.d.ts

@@ -2,7 +2,6 @@ import { ChatMistralAI } from '@langchain/mistralai';
 import type { BindToolsInput, BaseChatModelParams } from '@langchain/core/language_models/chat_models';
 import type { OpenAIChatInput, ChatOpenAIFields, AzureOpenAIInput, ClientOptions as OAIClientOptions } from '@langchain/openai';
 import type { GoogleGenerativeAIChatInput } from '@langchain/google-genai';
-import type { GeminiGenerationConfig } from '@langchain/google-common';
 import type { ChatVertexAIInput } from '@langchain/google-vertexai';
 import type { ChatDeepSeekCallOptions } from '@langchain/deepseek';
 import type { ChatOpenRouterCallOptions } from '@/llm/openrouter';
@@ -36,6 +35,11 @@ export type AnthropicReasoning = {
     thinking?: ThinkingConfig | boolean;
     thinkingBudget?: number;
 };
+export type GoogleThinkingConfig = {
+    thinkingBudget?: number;
+    includeThoughts?: boolean;
+    thinkingLevel?: string;
+};
 export type OpenAIClientOptions = ChatOpenAIFields;
 export type AnthropicClientOptions = AnthropicInput & {
     promptCache?: boolean;
@@ -43,6 +47,7 @@ export type AnthropicClientOptions = AnthropicInput & {
 export type MistralAIClientOptions = ChatMistralAIInput;
 export type VertexAIClientOptions = ChatVertexAIInput & {
     includeThoughts?: boolean;
+    thinkingConfig?: GoogleThinkingConfig;
 };
 export type BedrockAnthropicInput = ChatBedrockConverseInput & {
     additionalModelRequestFields?: ChatBedrockConverseInput['additionalModelRequestFields'] & AnthropicReasoning;
@@ -52,7 +57,7 @@ export type BedrockConverseClientOptions = ChatBedrockConverseInput;
 export type BedrockAnthropicClientOptions = BedrockAnthropicInput;
 export type GoogleClientOptions = GoogleGenerativeAIChatInput & {
     customHeaders?: RequestOptions['customHeaders'];
-    thinkingConfig?: GeminiGenerationConfig['thinkingConfig'];
+    thinkingConfig?: GoogleThinkingConfig;
 };
 export type DeepSeekClientOptions = ChatDeepSeekCallOptions;
 export type XAIClientOptions = ChatXAIInput;
@@ -92,7 +97,7 @@ export type ChatModelMap = {
     [Providers.MISTRALAI]: ChatMistralAI;
     [Providers.MISTRAL]: ChatMistralAI;
     [Providers.OPENROUTER]: ChatOpenRouter;
-    [Providers.BEDROCK]: CustomChatBedrockConverse;
+    [Providers.BEDROCK]: InstanceType<typeof CustomChatBedrockConverse>;
     [Providers.GOOGLE]: CustomChatGoogleGenerativeAI;
     [Providers.MOONSHOT]: ChatMoonshot;
 };

package/dist/types/types/run.d.ts

@@ -103,6 +103,8 @@ export type RunConfig = {
     returnContent?: boolean;
     tokenCounter?: TokenCounter;
     indexTokenCountMap?: Record<string, number>;
+    /** Skip post-stream cleanup (clearHeavyState) — useful for tests that inspect graph state after processStream */
+    skipCleanup?: boolean;
 };
 export type ProvidedCallbacks = (BaseCallbackHandler | CallbackHandlerMethods)[] | undefined;
 export type TokenCounter = (message: BaseMessage) => number;

package/dist/types/types/tools.d.ts

@@ -37,6 +37,14 @@ export type ToolNodeOptions = {
     toolDefinitions?: Map<string, LCTool>;
     /** Agent ID for event-driven mode (used to identify which agent's context to use) */
     agentId?: string;
+    /** Tool names that must be executed directly (via runTool) even in event-driven mode (e.g., graph-managed handoff tools) */
+    directToolNames?: Set<string>;
+    /**
+     * StreamingToolCallBuffer for recovering truncated tool call arguments.
+     * When parsePartialJson loses content due to max_tokens truncation,
+     * the ToolNode can fall back to extracting raw field values from the buffer.
+     */
+    streamingToolCallBuffer?: import('@/tools/StreamingToolCallBuffer').StreamingToolCallBuffer;
     /**
      * HITL tool approval configuration.
      * When provided, tools matching the policy will call interrupt() before execution.
@@ -51,6 +59,7 @@ export type ToolEndEvent = {
     tool_call: ToolCall;
     /** The content index of the tool call */
     index: number;
+    type?: 'tool_call';
 };
 export type CodeEnvFile = {
     id: string;
@@ -126,6 +135,11 @@ export type ToolCallRequest = {
     stepId?: string;
     /** Usage turn count for this tool */
     turn?: number;
+    /** Code execution session context for session continuity in event-driven mode */
+    codeSessionContext?: {
+        session_id: string;
+        files?: CodeEnvFile[];
+    };
 };
 /** Batch request containing ALL tool calls for a graph step */
 export type ToolExecuteBatchRequest = {
@@ -168,7 +182,7 @@ export type ProgrammaticCache = {
  * - 'interactive': User is present in the chat, approval prompts are shown.
  * - 'scheduled': Automated/scheduled execution, approval is auto-granted.
  */
-export type ExecutionContext = 'interactive' | 'scheduled';
+export type ExecutionContext = 'interactive' | 'scheduled' | 'browser';
 /**
  * Policy for when a tool requires human approval.
  * - 'always': Always require approval (default for interactive context).
@@ -247,6 +261,8 @@ export type ToolApprovalEvent = ToolApprovalRequest & {
 };
 /** Search mode: code_interpreter uses external sandbox, local uses safe substring matching */
 export type ToolSearchMode = 'code_interpreter' | 'local';
+/** Format for MCP tool names in search results */
+export type McpNameFormat = 'full' | 'base';
 /** Parameters for creating a Tool Search tool */
 export type ToolSearchParams = {
     apiKey?: string;
@@ -257,6 +273,8 @@ export type ToolSearchParams = {
     mode?: ToolSearchMode;
     /** Filter tools to only those from specific MCP server(s). Can be a single name or array of names. */
     mcpServer?: string | string[];
+    /** Format for MCP tool names: 'full' (tool_mcp_server) or 'base' (tool only). Default: 'full' */
+    mcpNameFormat?: McpNameFormat;
     [key: string]: unknown;
 };
 /** Simplified tool metadata for search purposes */
@@ -359,7 +377,7 @@ export type CodeSessionContext = {
     /** Session ID from the code execution environment */
     session_id: string;
     /** Files generated in this session (for context/tracking) */
-    files: FileRefs;
+    files?: FileRefs;
     /** Timestamp of last update */
     lastUpdated: number;
 };

package/dist/types/utils/contextAnalytics.d.ts

@@ -6,9 +6,7 @@
  */
 import type { BaseMessage } from '@langchain/core/messages';
 import type { TokenCounter } from '@/types/run';
-/**
- * Context analytics data for traces
- */
+/** Context analytics data for traces */
 export interface ContextAnalytics {
     /** Total messages in context */
     messageCount: number;
@@ -27,7 +25,10 @@ export interface ContextAnalytics {
     }>;
 }
 /**
- * Build context analytics for traces (no logging)
+ * Build context analytics for traces (no logging).
+ *
+ * @param messages - Array of BaseMessage (may originate from different @langchain/core installations)
+ * @param options  - Token counting and context budget info
  */
 export declare function buildContextAnalytics(messages: BaseMessage[], options: {
     tokenCounter?: TokenCounter;

package/dist/types/utils/index.d.ts

@@ -7,3 +7,4 @@ export * from './tokens';
 export * from './toonFormat';
 export * from './contextAnalytics';
 export * from './schema';
+export * from './toolCallContinuation';

package/dist/types/utils/tokens.d.ts

@@ -1,28 +1,15 @@
 import type { BaseMessage } from '@langchain/core/messages';
+export type EncodingName = 'o200k_base' | 'claude';
+export declare function encodingForModel(model: string): EncodingName;
 export declare function getTokenCountForMessage(message: BaseMessage, getTokenCount: (text: string) => number): number;
 /**
- * Creates a singleton token counter function that reuses the same encoder instance.
- * Falls back to character-based estimation (~4 chars per token) if encoder fails.
- * This ensures token counting never blocks LLM calls.
- */
-export declare const createTokenCounter: () => Promise<(message: BaseMessage) => number>;
-/**
- * Utility to manage the token encoder lifecycle explicitly.
- * Useful for applications that need fine-grained control over resource management.
+ * Creates a token counter function using the specified encoding.
+ * Lazily loads the encoding data on first use via dynamic import.
  */
+export declare const createTokenCounter: (encoding?: EncodingName) => Promise<(message: BaseMessage) => number>;
+/** Utility to manage the token encoder lifecycle explicitly. */
 export declare const TokenEncoderManager: {
-    /**
-     * Pre-initializes the encoder. This can be called during app startup
-     * to avoid lazy loading delays later.
-     */
     initialize(): Promise<void>;
-    /**
-     * Clears the cached encoder and token counter.
-     * Useful for testing or when you need to force a fresh reload.
-     */
     reset(): void;
-    /**
-     * Checks if the encoder has been initialized.
-     */
     isInitialized(): boolean;
 };

package/dist/types/utils/toolCallContinuation.d.ts

@@ -0,0 +1,30 @@
+/**
+ * toolCallContinuation — Utilities for detecting and classifying LLM finish
+ * reasons across providers. Used to detect max_tokens truncation and normalize
+ * stop reason extraction from provider-specific metadata formats.
+ *
+ * @module
+ */
+import type { AIMessageChunk } from '@langchain/core/messages';
+/**
+ * Extract the finish/stop reason from an AIMessageChunk's response_metadata.
+ * Centralizes provider-specific metadata access patterns.
+ *
+ * Provider formats:
+ * - OpenAI/Azure: `response_metadata.finish_reason`
+ * - Anthropic direct: `response_metadata.stop_reason`
+ * - Bedrock invoke: `response_metadata.stopReason`
+ * - Bedrock streaming: `response_metadata.messageStop.stopReason`
+ * - VertexAI/Google: `response_metadata.finishReason`
+ *
+ * @param chunk - The final accumulated AIMessageChunk
+ * @returns The finish reason string, or undefined if not found
+ */
+export declare function extractFinishReason(chunk: AIMessageChunk | undefined): string | undefined;
+/**
+ * Check if a finish reason indicates the response was truncated by token limits.
+ *
+ * @param reason - The finish reason from extractFinishReason()
+ * @returns true if the response was cut short by max_tokens
+ */
+export declare function isMaxTokensFinish(reason: string | undefined): boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@illuma-ai/agents",
-  "version": "1.0.89",
+  "version": "1.0.93",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",
@@ -86,6 +86,7 @@
     "test-parallel-agent-labeling": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-parallel-agent-labeling.ts",
     "test-thinking-handoff": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-thinking-handoff.ts",
     "test-thinking-handoff-bedrock": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-thinking-handoff-bedrock.ts",
+    "test-thinking-to-thinking-handoff-bedrock": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-thinking-to-thinking-handoff-bedrock.ts",
     "multi-agent-hybrid-flow": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/multi-agent-hybrid-flow.ts",
     "test-handoff-input": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-handoff-input.ts",
     "test-custom-prompt-key": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/test-custom-prompt-key.ts",
@@ -106,17 +107,21 @@
     "test:all": "npm test -- --testPathIgnorePatterns=title.memory-leak.test.ts && npm run test:memory",
     "reinstall": "npm run clean && npm ci && rm -rf ./dist && npm run build",
     "re": "bun run clean && bun install && rm -rf ./dist && turbo build",
-    "lint": "eslint \"{,!(node_modules|venv)/**/}*.{js,jsx,ts,tsx}\" --fix",
+    "lint": "eslint src --fix",
     "format": "prettier --write ."
   },
   "overrides": {
     "@langchain/openai": "0.5.18",
+    "@anthropic-ai/sdk": "$@anthropic-ai/sdk",
     "@browserbasehq/stagehand": {
       "openai": "$openai"
     },
-    "fast-xml-parser": "5.3.4"
+    "fast-xml-parser": "5.3.8",
+    "ajv": "6.14.0",
+    "minimatch": "3.1.4"
   },
   "dependencies": {
+    "@anthropic-ai/sdk": "^0.73.0",
     "@aws-sdk/client-bedrock-runtime": "^3.980.0",
     "@langchain/anthropic": "^0.3.26",
     "@langchain/aws": "^0.1.15",
@@ -129,12 +134,13 @@
     "@langchain/openai": "0.5.18",
     "@langchain/textsplitters": "^0.1.0",
     "@langchain/xai": "^0.0.3",
-    "@langfuse/langchain": "^4.3.0",
-    "@langfuse/otel": "^4.3.0",
-    "@langfuse/tracing": "^4.3.0",
+    "@illuma-ai/observability-langchain": "^0.2.1",
+    "@illuma-ai/observability-otel": "^0.2.1",
     "@opentelemetry/sdk-node": "^0.207.0",
+    "@scarf/scarf": "^1.4.0",
     "@toon-format/toon": "^2.1.0",
-    "axios": "^1.12.1",
+    "ai-tokenizer": "^1.0.6",
+    "axios": "^1.13.5",
     "cheerio": "^1.0.0",
     "commander": "^12.1.0",
     "dotenv": "^16.4.7",
@@ -165,6 +171,7 @@
     "@typescript-eslint/parser": "^8.24.0",
     "cross-env": "^10.1.0",
     "eslint": "^9.39.1",
+    "eslint-config-prettier": "^10.1.8",
     "eslint-import-resolver-typescript": "^3.7.0",
     "eslint-plugin-import": "^2.31.0",
     "husky": "^9.1.7",
@@ -172,7 +179,7 @@
     "jest-util": "^30.2.0",
     "lint-staged": "^15.2.7",
     "prettier": "^3.6.2",
-    "rollup": "^4.34.6",
+    "rollup": "^4.59.0",
     "rollup-plugin-cleandir": "^2.0.0",
     "ts-jest": "^29.4.5",
     "ts-node": "^10.9.2",