@defai.digital/ax-cli 3.0.2 → 3.1.1

package/.ax-cli/memory.json

@@ -0,0 +1,55 @@
+{
+  "version": 1,
+  "created_at": "2025-11-21T09:17:20.806Z",
+  "updated_at": "2025-11-21T09:17:20.806Z",
+  "project_root": "/Users/akiralam/code/ax-cli",
+  "content_hash": "sha256:e0e29a88f3c707749dcc25f1fb3ce2428fd5a52004652e88c0968d5d5d47d920",
+  "source": {
+    "directories": [
+      {
+        "path": "src",
+        "max_depth": 3
+      },
+      {
+        "path": "packages",
+        "max_depth": 2
+      }
+    ],
+    "files": [
+      "README.md",
+      "readme.md",
+      "package.json",
+      "tsconfig.json",
+      ".ax-cli/CUSTOM.md"
+    ],
+    "ignore": [
+      "node_modules",
+      "dist",
+      "build",
+      ".git",
+      ".next",
+      ".nuxt",
+      "coverage",
+      ".cache",
+      ".turbo",
+      ".vercel",
+      "__pycache__",
+      ".pytest_cache",
+      "target",
+      "vendor",
+      "*.log",
+      "*.lock",
+      ".DS_Store"
+    ]
+  },
+  "context": {
+    "formatted": "# Project: ax-cli\n\n\n## Architecture Patterns\n- **Project Type**: cli\n- **Primary Language**: TypeScript\n- **Tech Stack**: React, Vitest, Zod, Commander, Ink, ESM, TypeScript\n- **Package Manager**: npm\n- **Entry Point**: dist/index.js\n- **Module System**: ESM\n- **Test Framework**: vitest\n- **Validation**: zod\n- **Linter**: eslint\n- **Key Directories**: source: src, tests: tests\n\n\n## Directory Structure\n```\nsrc/\n├── agent/\n│   ├── specialized/\n│   │   ├── analysis-agent.ts\n│   │   ├── debug-agent.ts\n│   │   ├── documentation-agent.ts\n│   │   ├── index.ts\n│   │   ├── performance-agent.ts\n│   │   ├── refactoring-agent.ts\n│   │   └── testing-agent.ts\n│   ├── context-manager.ts\n│   ├── context-manager.ts.bak\n│   ├── dependency-resolver.ts\n│   ├── index.ts\n│   ├── llm-agent.ts\n│   ├── llm-agent.ts.backup\n│   ├── subagent-orchestrator.ts\n│   ├── subagent-types.ts\n│   └── subagent.ts\n├── checkpoint/\n│   ├── index.ts\n│   ├── manager.ts\n│   ├── storage.ts\n│   └── types.ts\n├── commands/\n│   ├── init/\n│   │   └── wizard.ts\n│   ├── cache.ts\n│   ├── init.ts\n│   ├── mcp.ts\n│   ├── memory.ts\n│   ├── models.ts\n│   ├── plan.ts\n│   ├── rewind.ts\n│   ├── setup.ts\n│   ├── templates.ts\n│   ├── update.ts\n│   └── usage.ts\n├── hooks/\n│   ├── use-chat-reducer.ts\n│   ├── use-enhanced-input.ts\n│   ├── use-input-handler.ts\n│   └── use-input-history.ts\n├── llm/\n│   ├── client.ts\n│   ├── tools.ts\n│   └── types.ts\n├── mcp/\n│   ├── client.ts\n│   ├── config.ts\n│   └── transports.ts\n├── memory/\n│   ├── context-generator.ts\n│   ├── context-injector.ts\n│   ├── context-store.ts\n│   ├── index.ts\n│   ├── schemas.ts\n│   ├── stats-collector.ts\n│   └── types.ts\n├── planner/\n│   ├── prompts/\n│   │   └── planning-prompt.ts\n│   ├── dependency-resolver.ts\n│   ├── index.ts\n│   ├── plan-generator.ts\n│   ├── plan-storage.ts\n│   ├── task-planner.ts\n│   ├── token-estimator.ts\n│   └── types.ts\n├── schemas/\n│   ├── api-schemas.ts\n│   ├── confirmation-schemas.ts\n│   ├── index-unified.ts\n│   ├── index.ts\n│   ├── README.md\n│   ├── settings-schemas.ts\n│   ├── tool-schemas.ts\n│   └── yaml-schemas.ts\n├── tools/\n│   ├── bash-output.ts\n│   ├── bash.ts\n│   ├── confirmation-tool.ts\n│   ├── index.ts\n│   ├── search.ts\n│   ├── text-editor.ts\n│   └── todo-tool.ts\n├── types/\n│   ├── index.ts\n│   ├── project-analysis.ts\n│   └── template.ts\n├── ui/\n│   ├── components/\n│   │   ├── api-key-input.tsx\n│   │   ├── chat-history.tsx\n│   │   ├── chat-input.tsx\n│   │   ├── chat-interface.tsx\n│   │   ├── collapsible-tool-result.tsx\n│   │   ├── command-suggestions.tsx\n│   │   ├── confirmation-dialog.tsx\n│   │   ├── diff-renderer.tsx\n│   │   ├── index.ts\n│   │   ├── keyboard-hints.tsx\n│   │   ├── loading-spinner.tsx\n│   │   ├── mcp-status.tsx\n│   │   ├── model-selection.tsx\n│   │   ├── phase-progress.tsx\n│   │   ├── quick-actions.tsx\n│   │   ├── reasoning-display.tsx\n│   │   ├── status-bar.tsx\n│   │   ├── subagent-monitor.tsx\n│   │   ├── toast-notification.tsx\n│   │   ├── virtualized-chat-history.tsx\n│   │   └── welcome-panel.tsx\n│   ├── shared/\n│   │   └── max-sized-box.tsx\n│   ├── utils/\n│   │   ├── code-colorizer.tsx\n│   │   ├── colors.ts\n│   │   └── markdown-renderer.tsx\n│   └── app.tsx\n├── utils/\n│   ├── background-task-manager.ts\n│   ├── cache.ts\n│   ├── config-loader.ts\n│   ├── confirmation-service.ts\n│   ├── console-messenger.ts\n│   ├── custom-instructions.ts\n│   ├── error-handler.ts\n│   ├── error-translator.ts\n│   ├── errors.ts\n│   ├── file-cache.ts\n│   ├── history-manager.ts\n│   ├── incremental-analyzer.ts\n│   ├── index.ts\n│   ├── init-previewer.ts\n│   ├── init-validator.ts\n│   ├── instruction-generator.ts\n│   ├── json-utils.ts\n│   ├── llm-optimized-instruction-generator.ts\n│   ├── message-optimizer.ts\n│   ├── onboarding-manager.ts\n│   ├── parallel-analyzer.ts\n│   ├── path-utils.ts\n│   ├── path-validator.ts\n│   ├── performance.ts\n│   ├── progress-tracker.ts\n│   ├── project-analyzer.ts\n│   ├── prompt-builder.ts\n│   ├── settings-manager.ts\n│   ├── settings.ts\n│   ├── setup-validator.ts\n│   └── ... (7 more)\n├── constants.ts\n└── index.ts\npackages/\n└── schemas/\n    ├── __tests__/\n    ├── scripts/\n    ├── src/\n    ├── .eslintrc.js\n    ├── package-lock.json\n    ├── package.json\n    ├── README.md\n    ├── tsconfig.json\n    └── vitest.config.ts\n```\n\n\n## Key Configuration\n\n### README.md\n\n```\n# AX CLI - Enterprise-Class AI CLI\n\n[![Tests](https://img.shields.io/badge/tests-562%20passing-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli/actions/workflows/test.yml)\n[![Coverage](https://img.shields.io/badge/coverage-98.29%25-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli)\n[![npm](https://img.shields.io/npm/dt/@defai.digital/ax-cli?style=flat-square&logo=npm&label=downloads)](https://npm-stat.com/charts.html?package=%40defai.digital%2Fax-cli)\n...\n```\n\n### readme.md\n\n```\n# AX CLI - Enterprise-Class AI CLI\n\n[![Tests](https://img.shields.io/badge/tests-562%20passing-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli/actions/workflows/test.yml)\n[![Coverage](https://img.shields.io/badge/coverage-98.29%25-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli)\n[![npm](https://img.shields.io/npm/dt/@defai.digital/ax-cli?style=flat-square&logo=npm&label=downloads)](https://npm-stat.com/charts.html?package=%40defai.digital%2Fax-cli)\n...\n```\n\n### package.json\n\n- **Name**: @defai.digital/ax-cli\n- **Version**: 3.0.2\n- **Description**: Enterprise-Class AI Command Line Interface - Primary support for GLM (General Language Model) with multi-provider AI orchestration powered by AutomatosX.\n- **Type**: module\n- **CLI**: Yes\n- **Dependencies**: @ax-cli/schemas, @clack/prompts, @modelcontextprotocol/sdk, axios, cfonts, chalk, clipboardy, commander, dotenv, enquirer (+11 more)\n- **Dev Dependencies**: @types/fs-extra, @types/js-yaml, @types/marked-terminal, @types/node, @types/react, @typescript-eslint/eslint-plugin, @typescript-eslint/parser, @vitest/coverage-v8 (+5 more)\n- **Scripts**: build, build:schemas, build:bun, dev, dev:node, start\n\n### tsconfig.json\n\n- **Target**: ES2022\n- **Module**: ESNext\n- **Module Resolution**: Bundler\n- **Strict**: true\n- **Output**: ./dist\n\n### .ax-cli/CUSTOM.md\n\n# @defai.digital/ax-cli - Quick Reference\n\n**Type:** cli | **Lang:** TypeScript | **Ver:**  v2.3.1\n**Stack:** React, Vitest, Zod, Commander, Ink, ESM, TypeScript\n\n---\n\n## 🎯 Critical Rules\n\n1. **ESM Imports:** Always use `.js` extension: `import { x } from './y.js'`\n2. **Validation:** Use zod for all external inputs\n3. **Types:** Explicit return types required on all functions\n4. **Testing:** 80%+ coverage, test error paths\n5. **Modules:** Use `import/export` (not `require/module.exports`)\n6. **File Organization:** Follow standardized output paths (see below)\n\n---\n\n## 📁 Project File Organization\n\n### Standard Output Paths\n\nAll AI-generated and project artifacts must follow this structure:\n\n```\nautomatosx/\n├── PRD/              # Product Requirement Documents\n│   ├── features/     # Feature specifications\n│   ├── api/          # API documentation\n│   └── archive/      # Old/deprecated PRDs\n├── REPORT/           # Project reports and analysis\n│   ├── status/       # Status reports\n│   ├\n[...truncated]\n\n\n## README Summary\n\n# AX CLI - Enterprise-Class AI CLI\n\n[![Tests](https://img.shields.io/badge/tests-562%20passing-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli/actions/workflows/test.yml)\n[![Coverage](https://img.shields.io/badge/coverage-98.29%25-brightgreen?style=flat-square)](https://github.com/defai-digital/ax-cli)\n[![npm](https://img.shields.io/npm/dt/@defai.digital/ax-cli?style=flat-square&logo=npm&label=downloads)](https://npm-stat.com/charts.html?package=%40defai.digital%2Fax-cli)\n[![Node.js Version](https://img.shields.io/badge/node-%3E%3D24.0.0-blue?style=flat-square)](https://nodejs.org/)\n[![TypeScript](https://img.shields.io/badge/TypeScript-5.9%2B-blue?style=flat-square&logo=typescript)](https://www.typescriptlang.org/)\n[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?style=flat-square)](https://opensource.org/licenses/MIT)\n[![macOS](https://img.shields.io/badge/macOS-26-blue?style=flat-square&logo=apple)](https://www.apple.com/macos/)\n[![Windows](https://img.shields.io/badge/Windows-11-blue?style=flat-square&logo=windows)](https://www.microsoft.com/windows/)\n[![Ubuntu](https://img.shields.io/badge/Ubuntu-24.04-blue?style=flat-square&logo=ubuntu)](https://ubuntu.com/)\n\n![AX CLI Logo](.github/assets/ax-cli.png)\n\n<p align=\"center\">\n  <strong>Enterprise-Grade Architecture • 98%+ Test Coverage • TypeScript & Zod Validation</strong>\n</p>\n\n---\n\n## 🚀 Quick Start\n\n```bash\n# Install globally\nnpm install -g @defai.digital/ax-cli\n\n# Configure your API key\nax-cli setup\n\n# Initialize your project\nax-cli init\n\n# Start interactive mode\nax-cli\n```\n\n## ✨ Features\n\n- **🤖 Multi-Provider Support**: Z.AI (GLM), xAI (Grok), OpenAI, Anthropic (Claude), Ollama (local)\n- **🧠 GLM 4.6 Optimized**: Primary support for General Language Model with advanced reasoning\n  - **32K max tokens** (industry-standard, matches Claude Code CLI)\n  - 200K context window, 128K max output capability\n  - 30% more token efficient than GLM 4.5\n  - Optimized for complex code generation and refactoring\n- **🎯 Multi-Phase Task Planner** (NEW in v3.0.0): Intelligent task decomposition for complex requests\n  - Automatic complexity detection (57 keyword patterns)\n  - LLM-based plan generation with phases and dependencies\n  - Phase-by-phase execution with progress tracking\n  - File modification tracking and context pruning between phases\n  - Plan management commands: `/plans`, `/plan`, `/phases`, `/pause`, `/resume`, `/skip`, `/abandon`\n- **🔄 Session Continuity**: Directory-specific conversation history with `--continue` flag\n  - Preserve context across sessions for multi-day development\n  - Each project maintains its own independent history\n  - Seamlessly resume conversations where you left off\n- **🔌 MCP Integration**: Model Context Protocol for extensible tool support\n- **✅ Production-Ready**: 98%+ test coverage, TypeScript strict mode, Zod validation\n- **🎯 Interactive & Headless**: Chat interface or one-shot commands\n- **📝 Smart Project Init**: Automatic project analysis and custom instructions\n- **🧠 Project Memory** (NEW): Intelligent context caching for z.ai GLM-4.6\n  - Automatic project scanning and context generation\n  - z.ai implicit caching support (50% token savings on repeated context)\n  - Cache statistics tracking and efficiency monitoring\n- **🔄 Auto-Update**: Built-in update checker and installer\n\n### Max Tokens Configuration\n\nAX CLI uses **industry-standard max tokens** based on research of leading AI coding tools:\n\n| Tool | Max Tokens | Notes |\n|------|-----------|-------|\n| **Claude Code CLI** | 16k - 32k | Industry standard |\n| **GitHub Copilot** | 64k | GPT-4o default |\n| **Cursor AI** | 200k | With intelligent pruning |\n| **AX CLI (GLM 4.6)** | **32k** ✅ | Matches Claude Code upper default |\n| **AX CLI (Others)** | 8k | Appropriate for each model |\n\n**Why 32k for GLM 4.6?**\n- Competitive with industry leaders (Claude Code, GitHub Copilot)\n- GLM 4.6 supports up to 128k max output (our 32k is conservative at 25%)\n- Better for complex code generation, large file modifications, and multi-file context\n- Based on o\n\n[...truncated]",
+    "token_estimate": 3315,
+    "sections": {
+      "structure": 1252,
+      "readme": 1121,
+      "config": 835,
+      "patterns": 99
+    }
+  }
+}

package/README.md

@@ -56,6 +56,10 @@ ax-cli
 - **✅ Production-Ready**: 98%+ test coverage, TypeScript strict mode, Zod validation
 - **🎯 Interactive & Headless**: Chat interface or one-shot commands
 - **📝 Smart Project Init**: Automatic project analysis and custom instructions
+- **🧠 Project Memory** (NEW): Intelligent context caching for z.ai GLM-4.6
+  - Automatic project scanning and context generation
+  - z.ai implicit caching support (50% token savings on repeated context)
+  - Cache statistics tracking and efficiency monitoring
 - **🔄 Auto-Update**: Built-in update checker and installer
 
 ### Max Tokens Configuration
@@ -122,6 +126,7 @@ ax-cli  # Will prompt for API key on first run
 - **User Settings**: `~/.ax-cli/config.json`
 - **Project Settings**: `.ax-cli/settings.json`
 - **Custom Instructions**: `.ax-cli/CUSTOM.md`
+- **Project Memory**: `.ax-cli/memory.json` (auto-generated)
 
 [Configuration Guide →](docs/configuration.md)
 
@@ -340,6 +345,74 @@ ax-cli mcp remove linear
 
 [MCP Integration Guide →](docs/mcp.md)
 
+## 🧠 Project Memory (NEW)
+
+Project Memory enables intelligent context caching for z.ai GLM-4.6, reducing token costs and improving response consistency:
+
+```bash
+# Initialize project memory (scans codebase)
+ax-cli memory warmup
+
+# Output:
+# ✓ Project memory generated (3,305 tokens)
+#
+# 📊 Context breakdown:
+#    Structure:  1,252 tokens (38%)
+#    README:     1,111 tokens (34%)
+#    Config:       835 tokens (25%)
+#    Patterns:      99 tokens (3%)
+```
+
+### How It Works
+
+1. **Warmup**: Scans your project structure, README, configs, and detects architecture patterns
+2. **Auto-Injection**: Memory context is automatically prepended to system prompts
+3. **z.ai Caching**: Identical prompt prefixes are automatically cached by z.ai (50% token savings)
+4. **Statistics**: Track cache efficiency with `ax-cli memory cache-stats`
+
+### Memory Commands
+
+```bash
+ax-cli memory warmup        # Create project memory
+ax-cli memory refresh       # Update after changes
+ax-cli memory status        # Show memory status & token distribution
+ax-cli memory clear         # Remove project memory
+ax-cli memory cache-stats   # Show cache efficiency statistics
+
+# Options
+ax-cli memory warmup -d 5           # Custom scan depth (1-10)
+ax-cli memory warmup -m 12000       # Custom max tokens
+ax-cli memory warmup --dry-run      # Preview without saving
+ax-cli memory status --verbose      # Show full context
+ax-cli memory status --json         # JSON output
+```
+
+### Token Distribution Visualization
+
+```
+📊 Token Distribution:
+   ████████░░░░░░░░░░░░  Structure  (38%)
+   ███████░░░░░░░░░░░░░  README     (34%)
+   █████░░░░░░░░░░░░░░░  Config     (25%)
+   █░░░░░░░░░░░░░░░░░░░  Patterns   (3%)
+```
+
+### Recommended Workflow
+
+```bash
+# 1. Initialize project (if not done)
+ax-cli init
+
+# 2. Create project memory
+ax-cli memory warmup
+
+# 3. Use ax-cli normally - memory is auto-injected
+ax-cli -p "refactor authentication module"
+
+# 4. After major changes, refresh memory
+ax-cli memory refresh
+```
+
 ## 🎯 Multi-Phase Task Planner (v3.0.0)
 
 AX CLI now includes an intelligent multi-phase task planner that automatically decomposes complex requests:
@@ -404,6 +477,7 @@ AX CLI implements enterprise-grade architecture with:
 - [Usage](docs/usage.md) - Comprehensive usage guide
 - [CLI Reference](docs/cli-reference.md) - Command-line interface reference
 - [MCP Integration](docs/mcp.md) - Model Context Protocol guide
+- [Project Memory](automatosx/prd/project-memory-prd.md) - Project memory feature specification
 - [Architecture](docs/architecture.md) - Technical architecture details
 - [Development](docs/development.md) - Development and contribution guide
 - [Troubleshooting](docs/troubleshooting.md) - Common issues and solutions

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

@@ -1,4 +1,5 @@
 import { LLMToolCall } from "../llm/client.js";
+import type { SamplingConfig } from "../llm/types.js";
 import { ToolResult } from "../types/index.js";
 import { EventEmitter } from "events";
 import { CheckpointManager } from "../checkpoint/index.js";
@@ -52,10 +53,30 @@ export declare class LLMAgent extends EventEmitter {
     private taskPlanner;
     private currentPlan;
     private planningEnabled;
+    /** Sampling configuration for deterministic/reproducible mode */
+    private samplingConfig;
     constructor(apiKey: string, baseURL?: string, model?: string, maxToolRounds?: number);
     private initializeCheckpointManager;
     private initializeMCP;
     private isGrokModel;
+    /**
+     * Build chat options with sampling configuration included
+     * Merges provided options with the agent's sampling config
+     */
+    private buildChatOptions;
+    /**
+     * Set sampling configuration for this agent session
+     * @param config Sampling configuration to apply
+     */
+    setSamplingConfig(config: SamplingConfig | undefined): void;
+    /**
+     * Get current sampling configuration
+     */
+    getSamplingConfig(): SamplingConfig | undefined;
+    /**
+     * Check if agent is running in deterministic mode
+     */
+    isDeterministicMode(): boolean;
     /**
      * Detect if a tool call is repetitive (likely causing a loop)
      * Returns true if the same tool with similar arguments was called multiple times recently

package/dist/agent/llm-agent.js

@@ -35,6 +35,8 @@ export class LLMAgent extends EventEmitter {
     taskPlanner;
     currentPlan = null;
     planningEnabled = PLANNER_CONFIG.ENABLED;
+    /** Sampling configuration for deterministic/reproducible mode */
+    samplingConfig;
     constructor(apiKey, baseURL, model, maxToolRounds) {
         super();
         const manager = getSettingsManager();
@@ -55,6 +57,8 @@ export class LLMAgent extends EventEmitter {
         this.checkpointManager = getCheckpointManager();
         this.subagentOrchestrator = new SubagentOrchestrator({ maxConcurrentAgents: 5 });
         this.taskPlanner = getTaskPlanner();
+        // Load sampling configuration from settings (supports env vars, project, and user settings)
+        this.samplingConfig = manager.getSamplingSettings();
         // Wire up checkpoint callback for automatic checkpoint creation
         this.textEditor.setCheckpointCallback(async (files, description) => {
             await this.checkpointManager.createCheckpoint({
@@ -77,9 +81,18 @@ export class LLMAgent extends EventEmitter {
             customInstructions: customInstructions || undefined,
         });
         // Initialize with system message
+        // OPTIMIZATION: Keep static system prompt separate from dynamic context
+        // This maximizes cache hit rates on the xAI API (cached tokens = 50% cost savings)
+        // The API automatically caches identical content across requests
         this.messages.push({
             role: "system",
-            content: `${systemPrompt}\n\nCurrent working directory: ${process.cwd()}`,
+            content: systemPrompt,
+        });
+        // Add dynamic context as a separate system message
+        // This allows the main system prompt to be cached while context varies
+        this.messages.push({
+            role: "system",
+            content: `Current working directory: ${process.cwd()}\nTimestamp: ${new Date().toISOString().split('T')[0]}`,
         });
     }
     initializeCheckpointManager() {
@@ -125,6 +138,37 @@ export class LLMAgent extends EventEmitter {
         const currentModel = this.llmClient.getCurrentModel();
         return currentModel.toLowerCase().includes("grok");
     }
+    /**
+     * Build chat options with sampling configuration included
+     * Merges provided options with the agent's sampling config
+     */
+    buildChatOptions(options) {
+        const result = { ...options };
+        // Include sampling configuration if set and not overridden
+        if (this.samplingConfig && !result.sampling) {
+            result.sampling = this.samplingConfig;
+        }
+        return result;
+    }
+    /**
+     * Set sampling configuration for this agent session
+     * @param config Sampling configuration to apply
+     */
+    setSamplingConfig(config) {
+        this.samplingConfig = config;
+    }
+    /**
+     * Get current sampling configuration
+     */
+    getSamplingConfig() {
+        return this.samplingConfig;
+    }
+    /**
+     * Check if agent is running in deterministic mode
+     */
+    isDeterministicMode() {
+        return this.samplingConfig?.doSample === false;
+    }
     /**
      * Detect if a tool call is repetitive (likely causing a loop)
      * Returns true if the same tool with similar arguments was called multiple times recently
@@ -289,7 +333,7 @@ export class LLMAgent extends EventEmitter {
             let toolRounds = 0;
             const maxPhaseRounds = Math.min(this.maxToolRounds, 50); // Limit per phase
             while (toolRounds < maxPhaseRounds) {
-                const response = await this.llmClient.chat(this.messages, tools);
+                const response = await this.llmClient.chat(this.messages, tools, this.buildChatOptions());
                 const assistantMessage = response.choices[0]?.message;
                 if (!assistantMessage)
                     break;
@@ -431,7 +475,7 @@ export class LLMAgent extends EventEmitter {
                     { role: "system", content: systemPrompt },
                     { role: "user", content: userPrompt },
                 ];
-                const response = await this.llmClient.chat(planMessages, []);
+                const response = await this.llmClient.chat(planMessages, [], this.buildChatOptions());
                 return response.choices[0]?.message?.content || "";
             }, {
                 projectType: "typescript", // Could be detected
@@ -658,11 +702,11 @@ export class LLMAgent extends EventEmitter {
                 // Load tools safely
                 const tools = await this.loadToolsSafely();
                 // Create chat stream
-                const stream = this.llmClient.chatStream(this.messages, tools, {
+                const stream = this.llmClient.chatStream(this.messages, tools, this.buildChatOptions({
                     searchOptions: this.isGrokModel() && this.shouldUseSearchFor(message)
                         ? { search_parameters: { mode: "auto" } }
                         : { search_parameters: { mode: "off" } }
-                });
+                }));
                 // Process streaming chunks
                 const chunkGen = this.processStreamingChunks(stream, inputTokensRef.value, lastTokenUpdateRef, totalOutputTokensRef);
                 let streamResult;
@@ -765,11 +809,11 @@ export class LLMAgent extends EventEmitter {
         let toolRounds = 0;
         try {
             const tools = await getAllGrokTools();
-            let currentResponse = await this.llmClient.chat(this.messages, tools, {
+            let currentResponse = await this.llmClient.chat(this.messages, tools, this.buildChatOptions({
                 searchOptions: this.isGrokModel() && this.shouldUseSearchFor(message)
                     ? { search_parameters: { mode: "auto" } }
                     : { search_parameters: { mode: "off" } }
-            });
+            }));
             // Agent loop - continue until no more tool calls or max rounds reached
             while (toolRounds < maxToolRounds) {
                 const assistantMessage = currentResponse.choices[0]?.message;
@@ -864,11 +908,11 @@ export class LLMAgent extends EventEmitter {
                         this.messages = this.contextManager.pruneMessages(this.messages, this.tokenCounter);
                     }
                     // Get next response - this might contain more tool calls
-                    currentResponse = await this.llmClient.chat(this.messages, tools, {
+                    currentResponse = await this.llmClient.chat(this.messages, tools, this.buildChatOptions({
                         searchOptions: this.isGrokModel() && this.shouldUseSearchFor(message)
                             ? { search_parameters: { mode: "auto" } }
                             : { search_parameters: { mode: "off" } }
-                    });
+                    }));
                 }
                 else {
                     // No more tool calls, add final response
@@ -1122,17 +1166,21 @@ export class LLMAgent extends EventEmitter {
      * Add assistant message to history and conversation
      */
     addAssistantMessage(accumulatedMessage) {
+        // Safely extract tool_calls with proper validation
+        const toolCalls = Array.isArray(accumulatedMessage.tool_calls)
+            ? accumulatedMessage.tool_calls
+            : undefined;
         const assistantEntry = {
             type: "assistant",
             content: accumulatedMessage.content || "Using tools to help you...",
             timestamp: new Date(),
-            toolCalls: accumulatedMessage.tool_calls || undefined,
+            toolCalls,
         };
         this.chatHistory.push(assistantEntry);
         this.messages.push({
             role: "assistant",
             content: accumulatedMessage.content || "",
-            tool_calls: accumulatedMessage.tool_calls,
+            tool_calls: toolCalls,
         });
         // Apply context pruning after adding message to prevent overflow
         // Critical for long assistant responses and tool results
@@ -1240,11 +1288,11 @@ export class LLMAgent extends EventEmitter {
                     };
                 }
                 // Create chat stream
-                const stream = this.llmClient.chatStream(this.messages, tools, {
+                const stream = this.llmClient.chatStream(this.messages, tools, this.buildChatOptions({
                     searchOptions: this.isGrokModel() && this.shouldUseSearchFor(message)
                         ? { search_parameters: { mode: "auto" } }
                         : { search_parameters: { mode: "off" } }
-                });
+                }));
                 // Process streaming chunks
                 const chunkGen = this.processStreamingChunks(stream, inputTokensRef.value, lastTokenUpdateRef, totalOutputTokensRef);
                 let streamResult;
@@ -1342,14 +1390,15 @@ export class LLMAgent extends EventEmitter {
      * @returns Parsed arguments or error result
      */
     parseToolArguments(toolCall, toolType = 'Tool') {
-        if (!toolCall.function.arguments || toolCall.function.arguments.trim() === '') {
+        const argsString = toolCall.function.arguments;
+        if (!argsString || typeof argsString !== 'string' || argsString.trim() === '') {
             return {
                 success: false,
                 error: `${toolType} ${toolCall.function.name} called with empty arguments`,
             };
         }
         try {
-            const args = JSON.parse(toolCall.function.arguments);
+            const args = JSON.parse(argsString);
             // Validate that args is an object (not null, array, or primitive)
             if (typeof args !== 'object' || args === null || Array.isArray(args)) {
                 return {