joonecli 0.1.0 → 0.2.0

package/src/test_cache.ts

@@ -1,72 +0,0 @@
-import "dotenv/config";
-import { ChatAnthropic } from "@langchain/anthropic";
-import { ExecutionHarness } from "./core/agentLoop.js";
-import { ContextState } from "./core/promptBuilder.js";
-import { HumanMessage } from "@langchain/core/messages";
-
-async function runCacheTest() {
-    console.log("=== Starting Prompt Caching Test ===\n");
-    
-    // We need Anthropic API Key for this to actually hit the network and measure cache
-    if (!process.env.ANTHROPIC_API_KEY) {
-        console.warn("WARNING: ANTHROPIC_API_KEY is not set in .env file. The LLM call will fail.");
-    }
-
-    const model = new ChatAnthropic({
-        modelName: "claude-3-5-sonnet-20241022",
-        temperature: 0,
-        maxTokens: 4096,
-    }); // No tools needed for this basic string test
-
-    const harness = new ExecutionHarness(model, []);
-
-    // 1. Create a massive static prefix (simulating a lot of project context)
-    // We repeat a string many times to ensure we pass the 1024 token minimum for Anthropic caching.
-    const massiveProjectMemory = Array(500).fill("Project Rule: Always write clean, modular TypeScript code with strict typings. ").join("\n");
-
-    const state: ContextState = {
-        globalSystemInstructions: "You are a helpful coding assistant. You remember rules carefully.",
-        projectMemory: massiveProjectMemory,
-        sessionContext: "User OS: Windows 11",
-        conversationHistory: []
-    };
-
-    console.log("Turn 1: Initial Query (Should create cache)");
-    state.conversationHistory.push(new HumanMessage("Hello! What is one of the project rules?"));
-    
-    const response1 = await harness.step(state);
-    state.conversationHistory.push(response1);
-    
-    // In @langchain/anthropic, the actual usage stats (including cache hits/misses) 
-    // are stored in the response_metadata of the AIMessage.
-    console.log(`Response 1: ${response1.content}`);
-    console.log(`Token Usage 1: ${JSON.stringify(response1.response_metadata?.usage, null, 2)}\n`);
-
-    // --- TURN 2 ---
-    console.log("Turn 2: Follow-up Query (Should hit cache)");
-    // We DO NOT change the globalSystemInstructions, projectMemory, or sessionContext.
-    // We only append to the conversation history. This preserves the prefix!
-    state.conversationHistory.push(new HumanMessage("Could you summarize the rule again briefly?"));
-    
-    const response2 = await harness.step(state);
-    state.conversationHistory.push(response2);
-    
-    console.log(`Response 2: ${response2.content}`);
-    console.log(`Token Usage 2: ${JSON.stringify(response2.response_metadata?.usage, null, 2)}\n`);
-
-    // --- TURN 3: The System Reminder Pattern ---
-    console.log("Turn 3: Using <system-reminder> to simulate environment change without breaking cache");
-    // If a file changed, we DON'T update `state.projectMemory`. We inject a reminder.
-    state.conversationHistory.push(new HumanMessage(
-        "<system-reminder>\nThe file 'auth.ts' has just been deleted by the user.\n</system-reminder>\nWhat should we do if we need auth now?"
-    ));
-
-    const response3 = await harness.step(state);
-    state.conversationHistory.push(response3);
-    
-    console.log(`Response 3: ${response3.content}`);
-    console.log(`Token Usage 3: ${JSON.stringify(response3.response_metadata?.usage, null, 2)}\n`);}
-
-if (require.main === module) {
-    runCacheTest().catch(console.error);
-}

package/src/test_google.js

@@ -1,40 +0,0 @@
-import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
-import { HumanMessage, SystemMessage } from "@langchain/core/messages";
-import * as dotenv from "dotenv";
-
-dotenv.config();
-
-async function main() {
-  const model = new ChatGoogleGenerativeAI({
-    model: "gemini-2.5-flash",
-  });
-
-  const messagesGrouped = [
-    new SystemMessage("You are a helpful assistant."),
-    new SystemMessage("Also, be concise."),
-    new HumanMessage("Hello!"),
-  ];
-
-  try {
-    await model.invoke(messagesGrouped);
-    console.log("Success: Grouped SystemMessages at the start work perfectly.");
-  } catch (e) {
-    console.error("Grouped systems failed:", e.message);
-  }
-
-  const messagesMidstream = [
-    new SystemMessage("You are a helpful assistant."),
-    new HumanMessage("Hello!"),
-    new SystemMessage("System recovery hint here."),
-    new HumanMessage("What did I say?"),
-  ];
-
-  try {
-    await model.invoke(messagesMidstream);
-    console.log("Success: Mid-stream SystemMessages work.");
-  } catch (e) {
-    console.error("Midstream systems failed:", e.message);
-  }
-}
-
-main();

package/src/test_google.ts

@@ -1,40 +0,0 @@
-import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
-import { HumanMessage, SystemMessage } from "@langchain/core/messages";
-import * as dotenv from "dotenv";
-
-dotenv.config();
-
-async function main() {
-  const model = new ChatGoogleGenerativeAI({
-    model: "gemini-2.5-flash",
-  });
-
-  const messagesGrouped = [
-    new SystemMessage("You are a helpful assistant."),
-    new SystemMessage("Also, be concise."),
-    new HumanMessage("Hello!"),
-  ];
-
-  try {
-    await model.invoke(messagesGrouped);
-    console.log("Success: Grouped SystemMessages at the start work perfectly.");
-  } catch (e: any) {
-    console.error("Grouped systems failed:", e.message);
-  }
-
-  const messagesMidstream = [
-    new SystemMessage("You are a helpful assistant."),
-    new HumanMessage("Hello!"),
-    new SystemMessage("System recovery hint here."),
-    new HumanMessage("What did I say?"),
-  ];
-
-  try {
-    await model.invoke(messagesMidstream);
-    console.log("Success: Mid-stream SystemMessages work.");
-  } catch (e: any) {
-    console.error("Midstream systems failed:", e.message);
-  }
-}
-
-main();

package/src/tools/askUser.ts

@@ -1,47 +0,0 @@
-import { HITLBridge } from "../hitl/bridge.js";
-import { DynamicToolInterface, ToolResult } from "./index.js";
-
-/**
- * AskUserQuestionTool — allows the agent to ask the user a clarifying question mid-turn.
- *
- * Use cases:
- * - Resolving ambiguous requirements before coding.
- * - Getting user preferences (framework choice, styling, naming).
- * - Requesting approval of an implementation plan before proceeding.
- */
-export const AskUserQuestionTool: DynamicToolInterface = {
-    name: "ask_user_question",
-    description:
-        "Ask the user a question and wait for their response. " +
-        "Use this when you need clarification on the task, user preferences, " +
-        "or approval before proceeding with a significant change. " +
-        "You may optionally provide a list of answer choices.",
-    schema: {
-        type: "object" as const,
-        properties: {
-            question: {
-                type: "string",
-                description: "The question to ask the user.",
-            },
-            options: {
-                type: "array",
-                items: { type: "string" },
-                description: "Optional list of predefined answer choices.",
-            },
-        },
-        required: ["question"],
-    },
-    async execute(args: Record<string, unknown>): Promise<ToolResult> {
-        const question = args.question as string;
-        const options = args.options as string[] | undefined;
-
-        if (!question || question.trim() === "") {
-            return { content: "Error: You must provide a non-empty question.", isError: true };
-        }
-
-        const bridge = HITLBridge.getInstance();
-        const answer = await bridge.askUser(question, options);
-
-        return { content: answer };
-    },
-};

package/src/tools/browser.ts

@@ -1,137 +0,0 @@
-import { SandboxManager } from "../sandbox/manager.js";
-import { LazyInstaller } from "../sandbox/bootstrap.js";
-import { DynamicToolInterface, ToolResult } from "./index.js";
-
-// ─── Sandbox + Installer references ─────────────────────────────────────────
-
-let _sandboxManager: SandboxManager | null = null;
-let _installer: LazyInstaller | null = null;
-
-export function bindBrowserSandbox(
-  sandbox: SandboxManager,
-  installer: LazyInstaller
-): void {
-  _sandboxManager = sandbox;
-  _installer = installer;
-}
-
-// ─── Helpers ────────────────────────────────────────────────────────────────────
-
-/**
- * Escapes a string so it can be safely used as an argument in a Bash shell command.
- * It wraps the string in single quotes and safely escapes internal single quotes.
- */
-function escapeBashArg(arg: string): string {
-  return `'${arg.replace(/'/g, "'\\''")}'`;
-}
-
-// ─── BrowserTool ────────────────────────────────────────────────────────────────
-
-/**
- * Web Browser Tool — wraps Vercel Labs' `agent-browser` CLI.
- *
- * Provides compact accessibility-tree output optimized for LLMs
- * (low token usage vs raw HTML). Runs inside the E2B sandbox.
- *
- * Supported actions:
- * - navigate: Go to a URL
- * - snapshot: Get the accessibility tree (compact text representation)
- * - click: Click an element by ref
- * - type: Type text into a form field by ref
- * - screenshot: Capture a screenshot
- * - scroll: Scroll the page up or down
- */
-export const BrowserTool: DynamicToolInterface = {
-  name: "browser",
-  description:
-    "Interact with web pages using a headless browser. Actions: navigate, snapshot, click, type, screenshot, scroll. " +
-    "Returns compact accessibility-tree text output optimized for AI consumption.",
-  schema: {
-    type: "object",
-    properties: {
-      action: {
-        type: "string",
-        enum: ["navigate", "snapshot", "click", "type", "screenshot", "scroll"],
-        description: "The browser action to perform",
-      },
-      url: {
-        type: "string",
-        description: "URL to navigate to (required for 'navigate')",
-      },
-      ref: {
-        type: "string",
-        description:
-          "Element reference from the accessibility tree (required for 'click' and 'type')",
-      },
-      text: {
-        type: "string",
-        description: "Text to type (required for 'type')",
-      },
-      direction: {
-        type: "string",
-        enum: ["up", "down"],
-        description: "Scroll direction for 'scroll' action (optional, defaults to 'down')",
-      },    },
-    required: ["action"],
-  },
-  execute: async (args: {
-    action: string;
-    url?: string;
-    ref?: string;
-    text?: string;
-    direction?: string;
-  }): Promise<ToolResult> => {
-    if (!_sandboxManager || !_sandboxManager.isActive()) {
-      return { content: "Sandbox is not active. Cannot use browser tool.", isError: true };
-    }
-
-    // Build the CLI command
-    let command: string;
-
-    switch (args.action) {
-      case "navigate":
-        if (!args.url) return { content: "Error: 'url' is required for navigate action.", isError: true };
-        command = `agent-browser navigate ${escapeBashArg(args.url)} 2>&1`;
-        break;
-
-      case "snapshot":
-        command = "agent-browser snapshot 2>&1";
-        break;
-
-      case "click":
-        if (!args.ref) return { content: "Error: 'ref' is required for click action.", isError: true };
-        command = `agent-browser click ${escapeBashArg(args.ref)} 2>&1`;
-        break;
-
-      case "type":
-        if (!args.ref) return { content: "Error: 'ref' is required for type action.", isError: true };
-        if (!args.text) return { content: "Error: 'text' is required for type action.", isError: true };
-        command = `agent-browser type ${escapeBashArg(args.ref)} ${escapeBashArg(args.text)} 2>&1`;
-        break;
-
-      case "screenshot":
-        command = "agent-browser screenshot 2>&1";
-        break;
-
-      case "scroll":
-        const dir = args.direction || "down";
-        command = `agent-browser scroll ${escapeBashArg(dir)} 2>&1`;
-        break;
-
-      default:
-        return { content: `Error: Unknown action "${args.action}". Use: navigate, snapshot, click, type, screenshot, scroll.`, isError: true };
-    }
-
-    const result = await _sandboxManager.exec(command);
-
-    if (result.exitCode !== 0) {
-      return {
-        content: `Browser action failed (exit code ${result.exitCode}):\n${result.stdout}\n${result.stderr}`,
-        metadata: { exitCode: result.exitCode },
-        isError: true
-      };
-    }
-
-    return { content: result.stdout || "(no output)", metadata: { exitCode: result.exitCode }, isError: false };
-  },
-};

package/src/tools/index.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["index.ts"],"names":[],"mappings":"AAAA,MAAM,WAAW,oBAAoB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,CAAC;IACpB,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;IAC5B,OAAO,EAAE,CAAC,IAAI,EAAE,GAAG,KAAK,OAAO,CAAC,MAAM,CAAC,GAAG,MAAM,CAAC;CACpD;AAED;;;GAGG;AACH,eAAO,MAAM,QAAQ,EAAE,oBActB,CAAC;AAEF;;;;GAIG;AACH,eAAO,MAAM,YAAY,EAAE,oBAc1B,CAAC;AAEF,eAAO,MAAM,UAAU,wBAA2B,CAAC"}

package/src/tools/index.ts

@@ -1,237 +0,0 @@
-import * as fs from "node:fs";
-import * as path from "node:path";
-import { SandboxManager } from "../sandbox/manager.js";
-import { FileSync } from "../sandbox/sync.js";
-
-export interface ToolResult {
-    content: string;
-    metadata?: Record<string, any>;
-    isError?: boolean;
-}
-
-export interface DynamicToolInterface {
-    name: string;
-    description: string;
-    schema: Record<string, any>;
-    execute: (args: any) => Promise<ToolResult> | ToolResult;
-}
-
-// ─── Configuration ──────────────────────────────────────────────────────────────
-
-/** Maximum file size (in bytes) the agent is allowed to read into context. */
-const MAX_FILE_SIZE_BYTES = 512 * 1024; // 512 KB
-
-/** Maximum number of lines to return if file exceeds line limit. */
-const MAX_FILE_LINES = 2000;
-
-// ─── Sandbox reference (set at session start) ───────────────────────────────────
-
-let _sandboxManager: SandboxManager | null = null;
-let _fileSync: FileSync | null = null;
-
-/**
- * Binds the tools to a SandboxManager and FileSync instance.
- * Must be called at session start before any tool executions.
- */
-export function bindSandbox(sandbox: SandboxManager, fileSync: FileSync): void {
-    _sandboxManager = sandbox;
-    _fileSync = fileSync;
-}
-
-/**
- * Security: Validates that a resolved path is strictly inside the given workspace dir.
- * Prevents directory traversal attacks from accessing sensitive host files.
- */
-export function isPathInsideWorkspace(
-    resolvedPath: string,
-    workspaceDir = process.cwd()
-): boolean {
-    const normalizedWorkspace = path.resolve(workspaceDir);
-    // path.relative returns a path relative to the first argument.
-    // If it starts with '..' or is absolute, it means resolvedPath has escaped normalizedWorkspace.
-    const relative = path.relative(normalizedWorkspace, resolvedPath);
-    return relative !== "" && !relative.startsWith("..") && !path.isAbsolute(relative);
-}
-
-// ─── BashTool ───────────────────────────────────────────────────────────────────
-// Executes shell commands inside the E2B sandbox.
-// The ToolRouter routes this to SANDBOX — the host machine is never exposed.
-
-export const BashTool: DynamicToolInterface = {
-    name: "bash",
-    description:
-        "Runs a shell command inside an isolated sandbox. Use for tests, scripts, or installing dependencies. The host machine is never exposed.",
-    schema: {
-        type: "object",
-        properties: {
-            command: {
-                type: "string",
-                description: "The shell command to execute",
-            },
-        },
-        required: ["command"],
-    },
-    execute: async (args: { command: string }): Promise<ToolResult> => {
-        if (!_sandboxManager || !_sandboxManager.isActive()) {
-            throw new Error(
-                "Sandbox is not active. Cannot execute bash commands without an active sandbox session."
-            );
-        }
-
-        // Sync any dirty files from host → sandbox before executing
-        if (_fileSync && _fileSync.pendingCount() > 0) {
-            await _fileSync.syncToSandbox(_sandboxManager);
-        }
-
-        const result = await _sandboxManager.exec(args.command);
-
-        if (result.exitCode !== 0) {
-            return {
-                content: `Command failed (exit code ${result.exitCode}):\nSTDOUT:\n${result.stdout}\nSTDERR:\n${result.stderr}`,
-                metadata: { exitCode: result.exitCode },
-                isError: true
-            };
-        }
-
-        return {
-            content: result.stdout || "(no output)",
-            metadata: { exitCode: result.exitCode },
-            isError: false
-        };
-    },
-};
-
-// ─── ReadFileTool ───────────────────────────────────────────────────────────────
-// Reads files from the HOST filesystem (so the user's real project is visible).
-// Includes a built-in file size guardrail to prevent sending huge files to the LLM.
-
-export const ReadFileTool: DynamicToolInterface = {
-    name: "read_file",
-    description:
-        "Reads a file from the host filesystem. Includes a file size guardrail — files over 512KB are truncated to prevent context overflow.",
-    schema: {
-        type: "object",
-        properties: {
-            path: {
-                type: "string",
-                description: "Absolute or relative path to the file",
-            },
-            startLine: {
-                type: "number",
-                description: "Optional 1-indexed start line for partial reads",
-            },
-            endLine: {
-                type: "number",
-                description: "Optional 1-indexed end line for partial reads",
-            },
-        },
-        required: ["path"],
-    },
-    execute: async (args: { path: string; startLine?: number; endLine?: number }): Promise<ToolResult> => {
-        const filePath = path.resolve(args.path);
-
-        // ── Security Guardrail ──
-        if (!isPathInsideWorkspace(filePath)) {
-            return {
-                content: `Security Error: Access Denied. Cannot read files outside the current project workspace (${process.cwd()}).`,
-                isError: true
-            };
-        }
-
-        // ── Check existence ──
-        if (!fs.existsSync(filePath)) {
-            return {
-                content: `Error: File not found — ${filePath}`,
-                isError: true
-            };
-        }
-
-        // ── File Size Guardrail ──
-        const stat = fs.statSync(filePath);
-        if (stat.size > MAX_FILE_SIZE_BYTES) {
-            return {
-                content: `Error: File too large (${Math.round(stat.size / 1024)}KB). Maximum size is 512KB. Use line ranges or grep_search.`,
-                isError: true
-            };
-        }
-
-        const fileContent = fs.readFileSync(filePath, "utf-8");
-        const lines = fileContent.split("\n");
-
-        if (args.startLine !== undefined || args.endLine !== undefined) {
-            const start = Math.max(1, args.startLine ?? 1) - 1;
-            const end = Math.min(lines.length, args.endLine ?? lines.length);
-            const sliced = lines.slice(start, end);
-            return { content: sliced.map((line, i) => `${start + i + 1}: ${line}`).join("\n") };
-        }
-
-        // ── Line count guardrail ──
-        if (lines.length > MAX_FILE_LINES) {
-            const truncated = lines.slice(0, MAX_FILE_LINES);
-            return {
-                content: truncated.map((line, i) => `${i + 1}: ${line}`).join("\n") +
-                `\n\n--- Truncated at ${MAX_FILE_LINES} lines (total: ${lines.length}) ---`
-            };
-        }
-
-        return { content: fileContent };
-    },
-};
-
-// ─── WriteFileTool ──────────────────────────────────────────────────────────────
-// Writes files to the HOST filesystem (so the user sees changes in their IDE).
-// Marks written files as dirty so FileSync uploads them before sandbox execution.
-
-export const WriteFileTool: DynamicToolInterface = {
-    name: "write_file",
-    description:
-        "Writes content to a file on the host filesystem. The user will see changes in their IDE immediately. The file is automatically synced to the sandbox before the next command execution.",
-    schema: {
-        type: "object",
-        properties: {
-            path: {
-                type: "string",
-                description: "Absolute or relative path to the file",
-            },
-            content: {
-                type: "string",
-                description: "The full file content to write",
-            },
-        },
-        required: ["path", "content"],
-    },
-    execute: async (args: { path: string; content: string }): Promise<ToolResult> => {
-        const filePath = path.resolve(args.path);
-
-        // ── Security Guardrail ──
-        if (!isPathInsideWorkspace(filePath)) {
-            return {
-                content: `Security Error: Access Denied. Cannot write files outside the current project workspace (${process.cwd()}).`,
-                isError: true
-            };
-        }
-
-        // Create parent directories if needed
-        const dir = path.dirname(filePath);
-        if (!fs.existsSync(dir)) {
-            fs.mkdirSync(dir, { recursive: true });
-        }
-
-        fs.writeFileSync(filePath, args.content, "utf-8");
-
-        // Mark file as dirty for next sandbox sync
-        if (_fileSync) {
-            _fileSync.markDirty(filePath);
-        }
-
-        return { content: `File written: ${filePath}` };
-    },
-};
-
-// ─── Core Tool Set ──────────────────────────────────────────────────────────────
-
-export const CORE_TOOLS: DynamicToolInterface[] = [
-    BashTool,
-    ReadFileTool,
-    WriteFileTool,
-];