@zds-ai/cli 0.1.7 → 0.1.10

package/dist/agent/llm-agent.d.ts

@@ -1,7 +1,13 @@
-import { LLMMessage, LLMToolCall } from "../grok/client.js";
+import { LLMToolCall } from "../grok/client.js";
 import type { ChatCompletionContentPart } from "openai/resources/chat/completions.js";
 import { ToolResult } from "../types/index.js";
 import { EventEmitter } from "events";
+import { SessionState } from "../utils/chat-history-manager.js";
+/**
+ * Represents a single entry in the conversation history.
+ * Supports various message types including user messages, assistant responses,
+ * tool calls, tool results, and system messages.
+ */
 export interface ChatEntry {
     type: "user" | "assistant" | "tool_result" | "tool_call" | "system";
     content?: string | ChatCompletionContentPart[];
@@ -22,6 +28,10 @@ export interface ChatEntry {
         [key: string]: any;
     };
 }
+/**
+ * Represents a chunk of data in the streaming response.
+ * Used for real-time communication between the agent and UI components.
+ */
 export interface StreamingChunk {
     type: "content" | "tool_calls" | "tool_result" | "done" | "token_count" | "user_message";
     content?: string;
@@ -32,6 +42,60 @@ export interface StreamingChunk {
     userEntry?: ChatEntry;
     systemMessages?: ChatEntry[];
 }
+/**
+ * Main LLM Agent class that orchestrates AI conversations with tool execution capabilities.
+ *
+ * ## Architecture Overview
+ *
+ * The LLMAgent serves as the central coordinator for AI-powered conversations, managing:
+ * - **Conversation Flow**: Handles user messages, AI responses, and multi-turn conversations
+ * - **Tool Execution**: Coordinates with various tools (file editing, shell commands, web search, etc.)
+ * - **Context Management**: Tracks conversation history and manages token limits
+ * - **Session State**: Maintains persona, mood, active tasks, and other session data
+ * - **Streaming Support**: Provides real-time response streaming for better UX
+ *
+ * ## Delegation Architecture
+ *
+ * The agent delegates specialized functionality to focused manager classes:
+ * - **ToolExecutor**: Handles all tool execution, validation, and approval workflows
+ * - **HookManager**: Manages persona/mood/task hooks and backend testing
+ * - **SessionManager**: Handles session persistence and state restoration
+ * - **MessageProcessor**: Processes user input, handles rephrasing, and XML parsing
+ * - **ContextManager**: Manages context warnings, compaction, and token tracking
+ *
+ * ## Key Features
+ *
+ * - **Multi-Model Support**: Works with various LLM backends (Grok, OpenAI, etc.)
+ * - **Tool Integration**: Seamlessly integrates with 15+ built-in tools
+ * - **MCP Support**: Extends capabilities via Model Context Protocol servers
+ * - **Vision Support**: Handles image inputs for vision-capable models
+ * - **Streaming Responses**: Real-time response generation with token counting
+ * - **Context Awareness**: Intelligent context management and automatic compaction
+ * - **Hook System**: Extensible hook system for custom behaviors
+ * - **Session Persistence**: Maintains conversation state across restarts
+ *
+ * ## Usage Patterns
+ *
+ * ```typescript
+ * // Initialize agent
+ * const agent = new LLMAgent(apiKey, baseURL, model);
+ * await agent.initialize();
+ *
+ * // Process messages (non-streaming)
+ * const entries = await agent.processUserMessage("Hello, world!");
+ *
+ * // Process messages (streaming)
+ * for await (const chunk of agent.processUserMessageStream("Write a file")) {
+ *   console.log(chunk);
+ * }
+ *
+ * // Manage session state
+ * await agent.setPersona("helpful assistant");
+ * await agent.startActiveTask("coding", "writing tests");
+ * ```
+ *
+ * @extends EventEmitter Emits 'contextChange' events for token usage updates
+ */
 export declare class LLMAgent extends EventEmitter {
     private llmClient;
     private textEditor;
@@ -48,6 +112,7 @@ export declare class LLMAgent extends EventEmitter {
     private imageTool;
     private fileConversionTool;
     private restartTool;
+    private audioTool;
     private chatHistory;
     private messages;
     private tokenCounter;
@@ -57,8 +122,6 @@ export declare class LLMAgent extends EventEmitter {
     private temperature;
     private maxTokens;
     private firstMessageProcessed;
-    private contextWarningAt80;
-    private contextWarningAt90;
     private persona;
     private personaColor;
     private mood;
@@ -69,72 +132,314 @@ export declare class LLMAgent extends EventEmitter {
     private apiKeyEnvVar;
     private pendingContextEditSession;
     private rephraseState;
-    private hookPrefillText;
+    private toolExecutor;
+    private hookManager;
+    private sessionManager;
+    private messageProcessor;
+    private contextManager;
+    private maxContextSize;
+    /**
+     * Cleans up incomplete tool calls in the message history.
+     * Ensures all tool calls have corresponding tool results to prevent API errors.
+     *
+     * This method scans the last assistant message for tool calls and adds
+     * "[Cancelled by user]" results for any tool calls that don't have results.
+     *
+     * @private
+     */
+    private cleanupIncompleteToolCalls;
+    /**
+     * Executes the instance hook if it hasn't been run yet.
+     *
+     * The instance hook runs once per agent session and can:
+     * - Set prompt variables
+     * - Add system messages
+     * - Provide prefill text for responses
+     *
+     * @private
+     */
+    private executeInstanceHookIfNeeded;
+    /**
+     * Creates a new LLMAgent instance.
+     *
+     * @param apiKey - API key for the LLM service
+     * @param baseURL - Optional base URL for the API endpoint
+     * @param model - Optional model name (defaults to saved model or "grok-code-fast-1")
+     * @param maxToolRounds - Maximum number of tool execution rounds (default: 400)
+     * @param debugLogFile - Optional path for MCP debug logging
+     * @param startupHookOutput - Optional output from startup hook execution
+     * @param temperature - Optional temperature for API requests (0.0-2.0)
+     * @param maxTokens - Optional maximum tokens for API responses
+     */
     constructor(apiKey: string, baseURL?: string, model?: string, maxToolRounds?: number, debugLogFile?: string, startupHookOutput?: string, temperature?: number, maxTokens?: number);
     private startupHookOutput?;
     private systemPrompt;
     private hasRunInstanceHook;
     /**
-     * Initialize the agent with dynamic system prompt
-     * Must be called after construction
+     * Initialize the agent with dynamic system prompt.
+     *
+     * This method must be called after construction to:
+     * - Build the system message with current tool availability
+     * - Set up the initial conversation context
+     * - Execute the instance hook if configured
+     *
+     * @throws {Error} If system message generation fails
      */
     initialize(): Promise<void>;
     /**
-     * Build/rebuild the system message with current tool availability
-     * Updates this.systemPrompt which is always used for messages[0]
+     * Build/rebuild the system message with current tool availability.
+     *
+     * This method:
+     * - Generates a dynamic tool list using the introspect tool
+     * - Sets the APP:TOOLS variable for template rendering
+     * - Renders the full SYSTEM template with all variables
+     * - Updates messages[0] with the new system prompt
+     *
+     * The system prompt is always at messages[0] and contains the core
+     * instructions, tool descriptions, and current context information.
      */
     buildSystemMessage(): Promise<void>;
+    /**
+     * Render system message with current variable state
+     * Called before LLM API calls and task processing to ensure fresh content
+     */
+    renderSystemMessage(): void;
+    /**
+     * Load initial conversation history from persistence.
+     *
+     * This method:
+     * - Loads the chat history (excluding system messages)
+     * - Sets or generates the system prompt
+     * - Converts history to API message format
+     * - Handles tool call/result matching
+     * - Updates token counts
+     *
+     * @param history - Array of chat entries to load
+     * @param systemPrompt - Optional system prompt (will generate if not provided)
+     */
     loadInitialHistory(history: ChatEntry[], systemPrompt?: string): Promise<void>;
+    /**
+     * Initialize Model Context Protocol (MCP) servers in the background.
+     *
+     * This method loads MCP configuration and initializes any configured
+     * servers without blocking agent construction. Errors are logged but
+     * don't prevent agent operation.
+     *
+     * @param debugLogFile - Optional path for MCP debug output
+     * @private
+     */
     private initializeMCP;
+    /**
+     * Checks if the current model is a Grok model.
+     * Used to enable Grok-specific features like web search.
+     *
+     * @returns True if the current model name contains "grok"
+     * @private
+     */
     private isGrokModel;
+    /**
+     * Heuristic to determine if web search should be enabled for a message.
+     *
+     * Analyzes the message content for keywords that suggest the user is
+     * asking for current information, news, or time-sensitive data.
+     *
+     * @param message - The user message to analyze
+     * @returns True if web search should be enabled
+     * @private
+     */
     private shouldUseSearchFor;
+    /**
+     * Process a user message and return all conversation entries generated.
+     *
+     * This is the main non-streaming message processing method that:
+     * - Handles rephrase commands and message preprocessing
+     * - Manages the agent loop with tool execution
+     * - Processes multiple rounds of AI responses and tool calls
+     * - Handles errors and context management
+     * - Returns all new conversation entries
+     *
+     * ## Processing Flow
+     *
+     * 1. **Setup**: Parse rephrase commands, clean incomplete tool calls
+     * 2. **Message Processing**: Parse images, assemble content, add to history
+     * 3. **Agent Loop**: Continue until no more tool calls or max rounds reached
+     *    - Get AI response
+     *    - Execute any tool calls
+     *    - Add results to conversation
+     *    - Get next response if needed
+     * 4. **Cleanup**: Handle errors, update context, return entries
+     *
+     * @param message - The user message to process
+     * @returns Promise resolving to array of new conversation entries
+     * @throws {Error} If message processing fails critically
+     */
     processUserMessage(message: string): Promise<ChatEntry[]>;
     /**
-     * Parse XML-formatted tool calls from message content (x.ai format)
-     * Converts <xai:function_call> elements to standard LLMToolCall format
+     * Process a user message with real-time streaming response.
+     *
+     * This is the main streaming message processing method that yields
+     * chunks of data as the conversation progresses. Provides real-time
+     * updates for:
+     * - User message processing
+     * - AI response streaming (content as it's generated)
+     * - Tool execution progress
+     * - Token count updates
+     * - System messages from hooks
+     *
+     * ## Streaming Flow
+     *
+     * 1. **Setup**: Process user message, yield user entry
+     * 2. **Agent Loop**: Stream AI responses and execute tools
+     *    - Stream AI response content in real-time
+     *    - Yield tool calls when detected
+     *    - Execute tools and yield results
+     *    - Continue until completion
+     * 3. **Completion**: Yield final token counts and done signal
+     *
+     * ## Chunk Types
+     *
+     * - `user_message`: Initial user message entry
+     * - `content`: Streaming AI response content
+     * - `tool_calls`: Tool calls detected in AI response
+     * - `tool_result`: Results from tool execution
+     * - `token_count`: Updated token usage
+     * - `done`: Processing complete
+     *
+     * @param message - The user message to process
+     * @yields StreamingChunk objects with real-time updates
+     * @throws {Error} If streaming fails critically
      */
-    private parseXMLToolCalls;
-    private messageReducer;
     processUserMessageStream(message: string): AsyncGenerator<StreamingChunk, void, unknown>;
     /**
      * Apply default parameter values for tools
      * This ensures the approval hook sees the same parameters that will be used during execution
      */
-    private applyToolParameterDefaults;
     /**
      * Validate tool arguments against the tool's schema
      * Returns null if valid, or an error message if invalid
      */
-    private validateToolArguments;
-    private executeTool;
-    private executeMCPTool;
+    /**
+     * Get a copy of the current chat history.
+     * @returns Array of chat entries (defensive copy)
+     */
     getChatHistory(): ChatEntry[];
+    /**
+     * Set the chat history to a new array of entries.
+     * @param history - New chat history entries
+     */
     setChatHistory(history: ChatEntry[]): void;
+    /**
+     * Get the current system prompt.
+     * @returns The system prompt string
+     */
     getSystemPrompt(): string;
+    /**
+     * Set a new system prompt and update the first message.
+     * @param prompt - Ignored (deprecated) - system prompt is always rendered from variables
+     */
     setSystemPrompt(prompt: string): void;
+    /**
+     * Get a copy of the current API messages array.
+     * @returns Array of LLM messages (defensive copy)
+     */
     getMessages(): any[];
+    /**
+     * Get the current token count for the conversation.
+     * @returns Number of tokens in the current message context
+     */
     getCurrentTokenCount(): number;
+    /**
+     * Get the maximum context size for the current model.
+     * @returns Maximum number of tokens supported
+     * @todo Make this model-specific for different context windows
+     */
     getMaxContextSize(): number;
+    /**
+     * Get the current context usage as a percentage.
+     * @returns Percentage of context window used (0-100)
+     */
     getContextUsagePercent(): number;
     /**
-     * Convert context messages to markdown format for viewing
+     * Convert the conversation context to markdown format for viewing.
+     *
+     * Creates a human-readable markdown representation of the conversation
+     * including:
+     * - Header with context file path and token usage
+     * - Numbered messages with timestamps
+     * - Formatted tool calls and results
+     * - Proper attribution (User/Assistant/System)
+     *
      * Format: (N) Name (role) - timestamp
+     *
+     * @returns Promise resolving to markdown-formatted conversation
      */
     convertContextToMarkdown(): Promise<string>;
+    /**
+     * Get the current persona setting.
+     * @returns Current persona string
+     */
     getPersona(): string;
+    /**
+     * Get the current persona display color.
+     * @returns Color name for persona display
+     */
     getPersonaColor(): string;
+    /**
+     * Get the current mood setting.
+     * @returns Current mood string
+     */
     getMood(): string;
+    /**
+     * Get the current mood display color.
+     * @returns Color name for mood display
+     */
     getMoodColor(): string;
+    /**
+     * Get the current active task.
+     * @returns Current active task string
+     */
     getActiveTask(): string;
+    /**
+     * Get the current active task action.
+     * @returns Current task action string
+     */
     getActiveTaskAction(): string;
+    /**
+     * Get the current active task display color.
+     * @returns Color name for task display
+     */
     getActiveTaskColor(): string;
+    /**
+     * Set a pending context edit session for file-based context editing.
+     * @param tmpJsonPath - Path to temporary JSON file
+     * @param contextFilePath - Path to actual context file
+     */
     setPendingContextEditSession(tmpJsonPath: string, contextFilePath: string): void;
+    /**
+     * Get the current pending context edit session.
+     * @returns Edit session info or null if none pending
+     */
     getPendingContextEditSession(): {
         tmpJsonPath: string;
         contextFilePath: string;
     } | null;
+    /**
+     * Clear the pending context edit session.
+     */
     clearPendingContextEditSession(): void;
+    /**
+     * Set the rephrase state for message editing operations.
+     * @param originalAssistantMessageIndex - Index of original assistant message
+     * @param rephraseRequestIndex - Index of rephrase request
+     * @param newResponseIndex - Index of new response (-1 if not yet created)
+     * @param messageType - Type of message being rephrased
+     * @param prefillText - Optional prefill text for the response
+     */
     setRephraseState(originalAssistantMessageIndex: number, rephraseRequestIndex: number, newResponseIndex: number, messageType: "user" | "system", prefillText?: string): void;
+    /**
+     * Get the current rephrase state.
+     * @returns Rephrase state info or null if none active
+     */
     getRephraseState(): {
         originalAssistantMessageIndex: number;
         rephraseRequestIndex: number;
@@ -142,135 +447,200 @@ export declare class LLMAgent extends EventEmitter {
         messageType: "user" | "system";
         prefillText?: string;
     } | null;
+    /**
+     * Clear the current rephrase state.
+     */
     clearRephraseState(): void;
+    /**
+     * Set the agent's persona with optional color.
+     *
+     * Executes the persona hook if configured and updates the agent's
+     * persona state on success.
+     *
+     * @param persona - The persona description
+     * @param color - Optional display color (defaults to "white")
+     * @returns Promise resolving to success/error result
+     */
     setPersona(persona: string, color?: string): Promise<{
         success: boolean;
         error?: string;
     }>;
+    /**
+     * Set the agent's mood with optional color.
+     *
+     * Executes the mood hook if configured and updates the agent's
+     * mood state on success.
+     *
+     * @param mood - The mood description
+     * @param color - Optional display color (defaults to "white")
+     * @returns Promise resolving to success/error result
+     */
     setMood(mood: string, color?: string): Promise<{
         success: boolean;
         error?: string;
     }>;
+    /**
+     * Start an active task with specified action and color.
+     *
+     * Executes the task start hook if configured and updates the agent's
+     * task state on success.
+     *
+     * @param activeTask - The task description
+     * @param action - The current action within the task
+     * @param color - Optional display color (defaults to "white")
+     * @returns Promise resolving to success/error result
+     */
     startActiveTask(activeTask: string, action: string, color?: string): Promise<{
         success: boolean;
         error?: string;
     }>;
+    /**
+     * Transition the active task to a new action/status.
+     *
+     * Updates the current task action without changing the task itself.
+     *
+     * @param action - The new action description
+     * @param color - Optional display color (defaults to current color)
+     * @returns Promise resolving to success/error result
+     */
     transitionActiveTaskStatus(action: string, color?: string): Promise<{
         success: boolean;
         error?: string;
     }>;
+    /**
+     * Stop the current active task with reason and documentation.
+     *
+     * Executes the task stop hook if configured and clears the agent's
+     * task state on success.
+     *
+     * @param reason - Reason for stopping the task
+     * @param documentationFile - Path to documentation file
+     * @param color - Optional display color (defaults to "white")
+     * @returns Promise resolving to success/error result
+     */
     stopActiveTask(reason: string, documentationFile: string, color?: string): Promise<{
         success: boolean;
         error?: string;
     }>;
-    private emitContextChange;
-    private addContextWarningIfNeeded;
-    executeCommand(command: string, skipConfirmation?: boolean): Promise<ToolResult>;
-    getCurrentModel(): string;
-    setModel(model: string): void;
-    /**
-     * Strip in-progress tool calls from messages for backend/model testing
-     * Removes tool_calls from the last assistant message and any corresponding tool results
-     * @returns Cleaned copy of messages array, or original if no stripping needed
-     */
-    static stripInProgressToolCalls(messages: LLMMessage[]): LLMMessage[];
     /**
-     * Test a model change by making a test API call with current conversation context
-     * Rolls back to previous model if test fails
-     * @param newModel Model to test
-     * @returns Promise with success status and optional error message
+     * Delegation method for hook processing (used by ToolExecutor).
+     *
+     * Processes hook results through the HookManager to handle commands,
+     * variable transformations, and other hook-specific logic.
+     *
+     * @param hookResult - Result object from hook execution
+     * @param envKey - Optional environment key for variable transformation
+     * @returns Promise resolving to processing result
      */
-    testModel(newModel: string): Promise<{
+    processHookResult(hookResult: {
+        commands?: any[];
+    }, envKey?: string): Promise<{
         success: boolean;
-        error?: string;
+        transformedValue?: string;
     }>;
     /**
-     * Test backend/baseUrl/model changes by making a test API call with current conversation context
-     * Rolls back all changes if test fails
-     * @param backend Backend display name
-     * @param baseUrl Base URL for API calls
-     * @param apiKeyEnvVar Name of environment variable containing API key
-     * @param model Model to use (optional, uses current model if not specified)
-     * @returns Promise with success status and optional error message
+     * Execute a shell command through the ZSH tool.
+     *
+     * @param command - Shell command to execute
+     * @param skipConfirmation - Whether to skip confirmation prompts
+     * @returns Promise resolving to tool execution result
      */
-    testBackendModelChange(backend: string, baseUrl: string, apiKeyEnvVar: string, model?: string): Promise<{
-        success: boolean;
-        error?: string;
-    }>;
+    executeCommand(command: string, skipConfirmation?: boolean): Promise<ToolResult>;
     /**
-     * Process hook result including commands and transformations
-     * Handles ENV transformations, model/backend testing, and error messaging
-     * @param hookResult Hook execution result
-     * @param envKey Optional ENV key to check for transformation (e.g., ZDS_AI_AGENT_PERSONA)
-     * @returns Object with success status and transformed value (if any)
+     * Get the current LLM model name.
+     * @returns Current model identifier
      */
-    private processHookResult;
+    getCurrentModel(): string;
     /**
-     * Process hook commands (MODEL, BACKEND, BASE_URL, SYSTEM, ENV)
-     * Handles model/backend testing and error messaging
-     * @param commands Hook commands from applyHookCommands()
+     * Set a new LLM model and update related components.
+     *
+     * This method:
+     * - Updates the LLM client model
+     * - Resets vision support flag
+     * - Updates the token counter for the new model
+     * - Handles model name suffixes (e.g., :nothinking)
+     *
+     * @param model - New model identifier
+     */
+    setModel(model: string): void;
+    /**
+     * Get the backend name (e.g., "grok", "openai").
+     * @returns Backend identifier string
      */
-    private processHookCommands;
     getBackend(): string;
+    /**
+     * Abort the current operation if one is in progress.
+     *
+     * This will cancel streaming responses and tool execution.
+     */
     abortCurrentOperation(): void;
+    /**
+     * Clear the conversation cache and reinitialize the agent.
+     *
+     * This method:
+     * - Backs up current conversation to timestamped files
+     * - Clears chat history and messages
+     * - Resets context warnings and processing flags
+     * - Re-executes startup and instance hooks
+     * - Saves the cleared state
+     * - Emits context change events
+     *
+     * Used when context becomes too large or user requests a fresh start.
+     */
     clearCache(): Promise<void>;
     /**
-     * Get current session state for persistence
-     */
-    getSessionState(): {
-        session: string;
-        persona: string;
-        personaColor: string;
-        mood: string;
-        moodColor: string;
-        activeTask: string;
-        activeTaskAction: string;
-        activeTaskColor: string;
-        cwd: string;
-        contextCurrent: number;
-        contextMax: number;
-        backend: string;
-        baseUrl: string;
-        apiKeyEnvVar: string;
-        model: string;
-        supportsVision: boolean;
-    };
+     * Get current session state for persistence.
+     *
+     * Collects all session-related state including:
+     * - Model and backend configuration
+     * - Persona, mood, and task settings
+     * - Context usage statistics
+     * - API key environment variable
+     *
+     * @returns Complete session state object
+     */
+    getSessionState(): SessionState;
+    /**
+     * Restore session state from persistence.
+     *
+     * Restores all session-related state including:
+     * - Model and backend configuration
+     * - Persona, mood, and task settings
+     * - Token counter and API client setup
+     *
+     * @param state - Session state to restore
+     */
+    restoreSessionState(state: SessionState): Promise<void>;
     /**
-     * Restore session state from persistence
-     */
-    restoreSessionState(state: {
-        session?: string;
-        persona: string;
-        personaColor: string;
-        mood: string;
-        moodColor: string;
-        activeTask: string;
-        activeTaskAction: string;
-        activeTaskColor: string;
-        cwd: string;
-        contextCurrent?: number;
-        contextMax?: number;
-        backend?: string;
-        baseUrl?: string;
-        apiKeyEnvVar?: string;
-        model?: string;
-        supportsVision?: boolean;
-    }): Promise<void>;
-    /**
-     * Compact conversation context by keeping system prompt and last N messages
-     * Reduces context size when it grows too large for backend to handle
+     * Compact conversation context by keeping system prompt and last N messages.
+     *
+     * Reduces context size when it grows too large for the backend to handle.
+     * Removes older messages while preserving the system prompt and recent context.
+     *
+     * @param keepLastMessages - Number of recent messages to keep (default: 20)
      * @returns Number of messages removed
      */
     compactContext(keepLastMessages?: number): number;
     /**
-     * Get all tool instances and their class names for display purposes
+     * Get all tool instances and their class names for display purposes.
+     *
+     * Uses reflection to find all tool instances and extract their
+     * class names and handled method names for introspection.
+     *
+     * @returns Array of tool info objects with class names and methods
      */
     getToolClassInfo(): Array<{
         className: string;
         methods: string[];
     }>;
     /**
-     * Get all tool instances via reflection
+     * Get all tool instances via reflection.
+     *
+     * Scans all properties of the agent instance to find objects that
+     * implement the tool interface (have getHandledToolNames method).
+     *
+     * @returns Array of tool instances with their class names
+     * @private
      */
     private getToolInstances;
 }