@testdriverai/mcp 7.9.115-test → 7.9.117-test

package/mcp-server/src/server.ts

@@ -12,6 +12,9 @@ process.env.TD_DEBUG = "true";
 import { registerAppResource, registerAppTool, RESOURCE_MIME_TYPE } from "@modelcontextprotocol/ext-apps/server";
 import { McpServer, ResourceTemplate } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
+import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
+import { randomUUID } from "crypto";
+import * as http from "http";
 import type { RequestHandlerExtra } from "@modelcontextprotocol/sdk/shared/protocol.js";
 import type { Variables } from "@modelcontextprotocol/sdk/shared/uriTemplate.js";
 import type { CallToolResult, ReadResourceResult, ServerNotification, ServerRequest } from "@modelcontextprotocol/sdk/types.js";
@@ -22,9 +25,10 @@ import * as path from "path";
 import { fileURLToPath, pathToFileURL } from "url";
 import { z } from "zod";
 
-import { generateActionCode } from "./codegen.js";
+import * as core from "./core/actions.js";
+import { NoActiveSessionError, type ActionResult } from "./core/actions.js";
 import { resolveE2bTemplateId, resolveOs } from "./env-utils.js";
-import { getProvisionOptions, SessionStartInputSchema, type SessionStartInput } from "./provision-types.js";
+import { SessionStartInputSchema, type SessionStartInput } from "./provision-types.js";
 import { sessionManager, type SessionState } from "./session.js";
 
 // =============================================================================
@@ -211,12 +215,6 @@ const RESOURCE_URI = "ui://testdriver/mcp-app.html";
 const SCREENSHOT_RESOURCE_BASE = "screenshot://testdriver/screenshot";
 const CROPPED_IMAGE_RESOURCE_BASE = "screenshot://testdriver/cropped";
 
-// SDK instance (will be initialized on session start)
-let sdk: any = null;
-
-// Last screenshot base64 for check comparisons
-let lastScreenshotBase64: string | null = null;
-
 // =============================================================================
 // Image Store - Stores images with unique IDs for reload persistence
 // =============================================================================
@@ -284,54 +282,46 @@ function getSessionData(session: SessionState | null) {
 }
 
 /**
- * Check if session is ready for use - returns error result if not
- * This helper provides clear, actionable error messages for the AI
- * 
- * Auto-extends the session on each successful check to prevent expiry during active use
+ * Map a {@link NoActiveSessionError} thrown by the core into the exact same
+ * "no/expired session" tool result the server's old `requireActiveSession`
+ * produced. Kept byte-identical so external behavior is unchanged.
  */
-function requireActiveSession(): { valid: true } | { valid: false; error: CallToolResult } {
-  const session = sessionManager.getCurrentSession();
-  
-  // No session ever created
-  if (!sdk || !session) {
-    return {
-      valid: false,
-      error: createToolResult(
-        false,
-        "ERROR: No active session. You must call session_start first to create a sandbox before using any other tools.",
-        { 
-          error: "NO_SESSION",
-          action: "session_start",
-          message: "No sandbox session exists. Call session_start to create one."
-        }
-      )
-    };
+function noSessionResult(err: NoActiveSessionError): CallToolResult {
+  if (err.code === "SESSION_EXPIRED") {
+    return createToolResult(
+      false,
+      "ERROR: Session has expired or timed out. The sandbox is no longer available. You must call session_start again to create a new sandbox session before continuing.",
+      {
+        error: "SESSION_EXPIRED",
+        action: "session_start",
+        message: "The previous sandbox session has expired. Call session_start to create a new one.",
+        expiredSessionId: err.expiredSessionId,
+      }
+    );
   }
-  
-  // Session exists but has expired
-  if (!sessionManager.isSessionValid(session.sessionId)) {
-    // Clear the SDK reference since the sandbox is no longer available
-    sdk = null;
-    return {
-      valid: false,
-      error: createToolResult(
-        false,
-        "ERROR: Session has expired or timed out. The sandbox is no longer available. You must call session_start again to create a new sandbox session before continuing.",
-        { 
-          error: "SESSION_EXPIRED",
-          action: "session_start",
-          message: "The previous sandbox session has expired. Call session_start to create a new one.",
-          expiredSessionId: session.sessionId
-        }
-      )
-    };
+  return createToolResult(
+    false,
+    "ERROR: No active session. You must call session_start first to create a sandbox before using any other tools.",
+    {
+      error: "NO_SESSION",
+      action: "session_start",
+      message: "No sandbox session exists. Call session_start to create one.",
+    }
+  );
+}
+
+/**
+ * Map a core {@link ActionResult} into an MCP CallToolResult, storing images as
+ * resource URIs (the core returns BARE base64, which is what `storeImage` wants).
+ */
+function resultToMcp(r: ActionResult): CallToolResult {
+  const data: Record<string, unknown> = { ...r.data };
+  for (const img of r.images ?? []) {
+    const uri = storeImage(img.base64, img.kind);
+    if (img.kind === "cropped") data.croppedImageResourceUri = uri;
+    else data.screenshotResourceUri = uri;
   }
-  
-  // Auto-extend session on each command to prevent expiry during active use
-  // This resets the expiry timer back to the original keepAlive duration
-  sessionManager.refreshSession(session.sessionId);
-  
-  return { valid: true };
+  return createToolResult(r.ok, r.text, data, r.code);
 }
 
 // =============================================================================
@@ -536,10 +526,6 @@ const server = isSentryEnabled()
       version: version,
     });
 
-// Element reference storage (for click/hover after find)
-// Stores actual Element instances - no raw coordinates as input
-const elementRefs = new Map<string, { element: any; description: string; coords: { x: number; y: number; centerX: number; centerY: number } }>();
-
 // =============================================================================
 // Register UI Resource
 // =============================================================================
@@ -681,248 +667,76 @@ Debug mode (connect to existing sandbox):
       logger.info("session_start: Using TD_E2B_TEMPLATE_ID environment variable", { e2bTemplateId: resolvedE2bTemplateId });
     }
 
-    logger.info("session_start: Starting", { 
-      type: params.type, 
+    logger.info("session_start: Starting", {
+      type: params.type,
       url: params.url,
-      os: resolvedOs, 
+      os: resolvedOs,
       reconnect: params.reconnect,
       sandboxId: params.sandboxId,
     });
 
     try {
-      // Validate required fields for specific provision types (unless connecting to existing sandbox)
-      if (!params.sandboxId) {
-        if (params.type === "installer" && !params.installerUrl) {
-          return createToolResult(false, "installer type requires 'installerUrl' parameter", { error: "Missing required parameter: installerUrl" });
-        }
-        if (params.type === "electron" && !params.appPath) {
-          return createToolResult(false, "electron type requires 'appPath' parameter", { error: "Missing required parameter: appPath" });
-        }
+      // The core owns session creation, SDK init, connect, provisioning and the
+      // initial screenshot. We keep the abort race + heartbeat scaffolding here:
+      // a heartbeat ticks while the long provisioning await is in flight, and the
+      // whole core call is raced against the client's abort signal so cancellation
+      // still returns promptly. Progress messages from the core are forwarded.
+      const stopHeartbeat = progress.heartbeat(
+        params.sandboxId
+          ? `Connecting to existing sandbox ${params.sandboxId}...`
+          : "Starting session..."
+      );
+      let result: ActionResult;
+      try {
+        result = await raceAbort(
+          extra.signal,
+          "session_start",
+          core.sessionStart(
+            params,
+            { os: resolvedOs, e2bTemplateId: resolvedE2bTemplateId },
+            { onProgress: (m) => progress.report(m) }
+          )
+        );
+      } finally {
+        stopHeartbeat();
       }
 
-      // Create new session
-      const newSession = sessionManager.createSession({
-        os: resolvedOs,
-        keepAlive: params.keepAlive,
-        testFile: params.testFile,
-      });
-      logger.debug("session_start: Session created", { sessionId: newSession.sessionId });
-
-      // Determine API root
-      const apiRoot = params.apiRoot || process.env.TD_API_ROOT || "https://api.testdriver.ai";
-      logger.debug("session_start: Using API root", { apiRoot });
+      const duration = Date.now() - startTime;
 
-      // Initialize SDK
-      logger.debug("session_start: Initializing SDK");
-      const TestDriverSDK = (await import("../../sdk.js")).default;
-      
-      // Determine preview mode from environment variable
-      // TD_PREVIEW can be "ide", "browser", or "none"
-      // Default to "ide" so the live preview shows within the IDE
-      const previewMode = process.env.TD_PREVIEW || "ide";
-      logger.debug("session_start: Preview mode", { preview: previewMode });
-      
-      // Get IP from params or environment (for self-hosted instances)
-      const instanceIp = params.ip || process.env.TD_IP;
-      
-      // Get API key - check multiple sources for GitHub Copilot coding agent compatibility
-      // 1. TD_API_KEY (standard environment variable)
-      // 2. COPILOT_MCP_TD_API_KEY (fallback for GitHub Copilot coding agent)
-      const apiKey = process.env.TD_API_KEY || process.env.COPILOT_MCP_TD_API_KEY || "";
-      
-      if (!apiKey) {
-        logger.error("session_start: No API key found", {
-          hasTD_API_KEY: !!process.env.TD_API_KEY,
-          hasCOPILOT_MCP_TD_API_KEY: !!process.env.COPILOT_MCP_TD_API_KEY,
-          availableEnvVars: Object.keys(process.env).filter(k => k.includes('TD') || k.includes('COPILOT_MCP'))
-        });
-        return createToolResult(false, "No API key found. Please set TD_API_KEY or COPILOT_MCP_TD_API_KEY environment variable.", { 
-          error: "Missing API key",
-          hint: "For GitHub Copilot coding agent, create a Copilot environment secret named COPILOT_MCP_TD_API_KEY"
-        });
+      // Set Sentry context once the session id is known.
+      if (result.ok && typeof result.data.sessionId === "string") {
+        setSessionContext(result.data.sessionId, (result.data.sandboxId as string) || undefined);
       }
-      
-      logger.debug("session_start: API key found", { 
-        source: process.env.TD_API_KEY ? "TD_API_KEY" : "COPILOT_MCP_TD_API_KEY",
-        keyPrefix: apiKey.substring(0, 7) + "..."
-      });
-      
-      sdk = new TestDriverSDK(apiKey, {
-        os: resolvedOs,
-        logging: false,
-        apiRoot,
-        preview: previewMode as "browser" | "ide" | "none",
-        ip: instanceIp,
-        e2bTemplateId: resolvedE2bTemplateId,
-      });
 
-      // Handle sandboxId mode - connect to existing sandbox (debug-on-failure mode)
-      if (params.sandboxId) {
-        logger.info("session_start: Connecting to existing sandbox (debug mode)", { sandboxId: params.sandboxId });
-        const stopHeartbeat = progress.heartbeat(`Connecting to existing sandbox ${params.sandboxId}...`);
-        try {
-          await raceAbort(extra.signal, "session_start", sdk.connect({
-            sandboxId: params.sandboxId,
-            keepAlive: params.keepAlive,
-          }));
-        } finally {
-          stopHeartbeat();
-        }
-        
-        // Get sandbox ID
-        const instance = sdk.getInstance();
-        logger.info("session_start: Connected to existing sandbox", { instanceId: instance?.instanceId });
-        sessionManager.activateSession(newSession.sessionId, instance?.instanceId || params.sandboxId);
-        
-        // Set Sentry context for error tracking
-        setSessionContext(newSession.sessionId, instance?.instanceId);
-        
-        // Capture screenshot of current state
-        logger.debug("session_start: Capturing screenshot of existing sandbox");
-        progress.report("Capturing screenshot...");
-        const screenshotBase64 = await raceAbort(extra.signal, "session_start", sdk.agent.system.captureScreenBase64(1, false, true)) as string;
-        
-        let screenshotResourceUri: string | undefined;
-        if (screenshotBase64) {
-          screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-          lastScreenshotBase64 = screenshotBase64;
-        }
+      logger.info("session_start: Completed", { duration, ok: result.ok });
 
-        const duration = Date.now() - startTime;
-        logger.info("session_start: Connected to existing sandbox", { duration, sessionId: newSession.sessionId, sandboxId: params.sandboxId });
-
-        return createToolResult(
-          true,
-          `Connected to existing sandbox (debug mode)
-Session: ${newSession.sessionId}
-Sandbox: ${params.sandboxId}
-Expires in: ${Math.round(params.keepAlive / 1000)}s
-
-You are now connected to the sandbox in its current state. Use find, click, type, etc. to interact.`,
-          { 
-            action: "session_start",
-            sessionId: newSession.sessionId, 
-            sandboxId: params.sandboxId,
-            debugMode: true,
-            screenshotResourceUri,
-            duration 
-          },
-          "// Connected to existing sandbox - no provision code needed"
-        );
+      if (!result.ok) {
+        // Validation / missing-key failures: pass through with duration added.
+        return resultToMcp({ ...result, data: { ...result.data, duration } });
       }
 
-      // Connect to sandbox
-      if (instanceIp) {
-        logger.info("session_start: Connecting to self-hosted instance...", { ip: instanceIp });
-      } else {
-        logger.info("session_start: Connecting to cloud sandbox...");
-      }
-      {
-        const stopHeartbeat = progress.heartbeat(
-          instanceIp ? `Connecting to self-hosted instance ${instanceIp}...` : "Connecting to cloud sandbox..."
-        );
-        try {
-          await raceAbort(extra.signal, "session_start", sdk.connect({
-            reconnect: params.reconnect,
-            keepAlive: params.keepAlive,
-            ip: instanceIp,
-          }));
-        } finally {
-          stopHeartbeat();
-        }
+      // Reproduce the server's exact success output (text + data) which differs
+      // from the core's neutral text. Images are mapped to resource URIs.
+      const data: Record<string, unknown> = { ...result.data, duration };
+      for (const img of result.images ?? []) {
+        const uri = storeImage(img.base64, img.kind);
+        if (img.kind === "cropped") data.croppedImageResourceUri = uri;
+        else data.screenshotResourceUri = uri;
       }
 
-      // Get sandbox ID
-      const instance = sdk.getInstance();
-      logger.info("session_start: Connected to sandbox", { instanceId: instance?.instanceId });
-      sessionManager.activateSession(newSession.sessionId, instance?.instanceId || "unknown");
-      
-      // Set Sentry context for error tracking
-      setSessionContext(newSession.sessionId, instance?.instanceId);
-
-      // Get provision-specific options
-      const provisionOptions = getProvisionOptions(params);
-      let provisionCmd = "";
-
-      // Provisioning can take tens of seconds (downloading installers, booting
-      // apps). Heartbeat so the client's idle timeout keeps resetting.
-      const stopProvisionHeartbeat = progress.heartbeat(`Provisioning ${params.type}...`);
-      try {
-      // Provision based on type
-      switch (params.type) {
-        case "chrome": {
-          const chromeOpts = provisionOptions as { url: string; maximized?: boolean; guest?: boolean };
-          logger.info("session_start: Provisioning Chrome", { url: chromeOpts.url });
-          await raceAbort(extra.signal, "session_start", sdk.provision.chrome(chromeOpts));
-          provisionCmd = "provision.chrome";
-          logger.debug("session_start: Chrome provisioned");
-          break;
-        }
-        
-        case "chromeExtension": {
-          const extOpts = provisionOptions as { extensionPath?: string; extensionId?: string; maximized?: boolean };
-          logger.info("session_start: Provisioning Chrome Extension", { extensionPath: extOpts.extensionPath, extensionId: extOpts.extensionId });
-          await raceAbort(extra.signal, "session_start", sdk.provision.chromeExtension(extOpts));
-          provisionCmd = "provision.chromeExtension";
-          logger.debug("session_start: Chrome Extension provisioned");
-          break;
-        }
-        
-        case "vscode": {
-          const vscodeOpts = provisionOptions as { workspace?: string; extensions?: string[] };
-          logger.info("session_start: Provisioning VS Code", { workspace: vscodeOpts.workspace });
-          await raceAbort(extra.signal, "session_start", sdk.provision.vscode(vscodeOpts));
-          provisionCmd = "provision.vscode";
-          logger.debug("session_start: VS Code provisioned");
-          break;
-        }
-        
-        case "installer": {
-          const installerOpts = provisionOptions as { url: string; filename?: string; appName?: string; launch?: boolean };
-          logger.info("session_start: Provisioning installer", { url: installerOpts.url });
-          await raceAbort(extra.signal, "session_start", sdk.provision.installer(installerOpts));
-          provisionCmd = "provision.installer";
-          logger.debug("session_start: Installer provisioned");
-          break;
-        }
-        
-        case "electron": {
-          const electronOpts = provisionOptions as { appPath: string; args?: string[] };
-          logger.info("session_start: Provisioning Electron", { appPath: electronOpts.appPath });
-          await raceAbort(extra.signal, "session_start", sdk.provision.electron(electronOpts));
-          provisionCmd = "provision.electron";
-          logger.debug("session_start: Electron app provisioned");
-          break;
-        }
-      }
-      } finally {
-        stopProvisionHeartbeat();
-      }
+      if (result.data.debugMode) {
+        // Debug (existing-sandbox) success text — preserve original wording.
+        const text = `Connected to existing sandbox (debug mode)
+Session: ${result.data.sessionId}
+Sandbox: ${result.data.sandboxId}
+Expires in: ${Math.round(params.keepAlive / 1000)}s
 
-      // Capture initial screenshot after provisioning
-      logger.debug("session_start: Capturing initial screenshot");
-      progress.report("Capturing screenshot...");
-      const screenshotBase64 = await raceAbort(extra.signal, "session_start", sdk.agent.system.captureScreenBase64(1, false, true)) as string;
-      
-      let screenshotResourceUri: string | undefined;
-      if (screenshotBase64) {
-        screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-        lastScreenshotBase64 = screenshotBase64;
+You are now connected to the sandbox in its current state. Use find, click, type, etc. to interact.`;
+        return createToolResult(true, text, data, result.code);
       }
 
-      const duration = Date.now() - startTime;
-      logger.info("session_start: Completed", { duration, sessionId: newSession.sessionId, selfHosted: !!instanceIp });
-
-      // Generate the code for this provision action
-      const generatedCode = generateActionCode(provisionCmd, provisionOptions);
-
-      // Build debugger URL for the session
-      const debuggerUrl = instance?.debuggerUrl || (instanceIp ? `http://${instanceIp}:9222` : null);
-
-      const connectionType = instanceIp ? `Self-hosted (${instanceIp})` : "Cloud";
-      return createToolResult(
-        true,
-        `Session started: ${newSession.sessionId}\nConnection: ${connectionType}\nType: ${params.type}\nSandbox: ${instance?.instanceId}\nExpires in: ${Math.round(params.keepAlive / 1000)}s
+      // Normal provisioning success — append the EXACT dependency guidance block.
+      const text = `${result.text}
 
 IMPORTANT - If creating a new test project, use these EXACT dependencies in package.json:
 {
@@ -934,19 +748,8 @@ IMPORTANT - If creating a new test project, use these EXACT dependencies in pack
   "scripts": {
     "test": "vitest"
   }
-}`,
-        { 
-          action: "session_start",
-          sessionId: newSession.sessionId, 
-          provisionType: params.type, 
-          selfHosted: !!instanceIp, 
-          instanceIp: instanceIp || undefined,
-          debuggerUrl,
-          screenshotResourceUri,
-          duration 
-        },
-        generatedCode
-      );
+}`;
+      return createToolResult(true, text, data, result.code);
     } catch (error) {
       // On client cancellation, tear down the half-provisioned session so we
       // don't leak a connected sandbox. The underlying SDK call may still be
@@ -954,7 +757,7 @@ IMPORTANT - If creating a new test project, use these EXACT dependencies in pack
       if (error instanceof ToolAbortError) {
         logger.info("session_start: Cancelled by client, tearing down session");
         try {
-          await sdk?.disconnect?.();
+          await core.disconnect();
         } catch (cleanupErr) {
           logger.warn("session_start: Cleanup after cancel failed", { error: String(cleanupErr) });
         }
@@ -977,27 +780,17 @@ server.registerTool(
   async (): Promise<CallToolResult> => {
     const startTime = Date.now();
     logger.info("session_status: Checking");
-    const session = sessionManager.getCurrentSession();
 
-    if (!session) {
-      logger.warn("session_status: No active session");
-      return createToolResult(false, "No active session", { error: "No active session. Call session_start first." });
-    }
-
-    const summary = sessionManager.getSessionSummary(session.sessionId);
+    const result = core.sessionStatus();
     const duration = Date.now() - startTime;
-    logger.info("session_status: Completed", { 
-      sessionId: session.sessionId, 
-      status: session.status,
-      timeRemaining: summary?.timeRemaining,
-      duration 
+    logger.info("session_status: Completed", {
+      ok: result.ok,
+      sessionId: result.data.sessionId,
+      status: result.data.status,
+      duration,
     });
 
-    return createToolResult(
-      true,
-      `Session: ${session.sessionId}\nStatus: ${session.status}\nTime remaining: ${Math.round((summary?.timeRemaining || 0) / 1000)}s`,
-      { action: "session_status", ...summary, sessionId: session.sessionId, status: session.status, duration }
-    );
+    return resultToMcp({ ...result, data: { ...result.data, duration } });
   }
 );
 
@@ -1012,22 +805,22 @@ server.registerTool(
   },
   async (params) => {
     logger.info("session_extend: Extending", { additionalMs: params.additionalMs });
-    const session = sessionManager.getCurrentSession();
 
-    if (!session) {
+    const result = core.sessionExtend(params.additionalMs);
+
+    if (!result.ok) {
       logger.warn("session_extend: No active session");
       return { content: [{ type: "text" as const, text: "No active session" }] };
     }
 
-    sessionManager.extendSession(session.sessionId, params.additionalMs);
-    const newExpiry = sessionManager.getTimeRemaining(session.sessionId);
-    logger.info("session_extend: Extended", { sessionId: session.sessionId, newExpiry });
+    logger.info("session_extend: Extended", { newExpiry: result.data.newExpiry });
 
+    // Preserve the original plain content shape (not via createToolResult).
     return {
       content: [
         {
           type: "text" as const,
-          text: `Session extended by ${params.additionalMs / 1000}s. New expiry: ${Math.round(newExpiry / 1000)}s`,
+          text: result.text,
         },
       ],
     };
@@ -1052,113 +845,25 @@ registerAppTool(
     const progress = makeProgressReporter(extra);
     logger.info("find: Starting", { description: params.description, timeout: params.timeout });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("find: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("find: Calling SDK find");
       const stopHeartbeat = progress.heartbeat(`Looking for "${params.description}"...`);
-      let element: any;
+      let result: ActionResult;
       try {
-        element = await raceAbort(
-          extra.signal,
-          "find",
-          sdk.find(params.description, params.timeout ? { timeout: params.timeout } : undefined)
-        );
+        result = await raceAbort(extra.signal, "find", core.find(params.description, params.timeout));
       } finally {
         stopHeartbeat();
       }
-      const found = element.found();
-      const coords = element.getCoordinates();
-
-      // Store element ref for later use (stores actual Element instance)
-      const elementRef = `el-${Date.now()}`;
-      if (found && coords) {
-        elementRefs.set(elementRef, {
-          element: element,  // Store the actual Element instance
-          description: params.description,
-          coords: {
-            x: coords.x,
-            y: coords.y,
-            centerX: coords.centerX,
-            centerY: coords.centerY,
-          },
-        });
-        logger.info("find: Element found", { 
-          description: params.description, 
-          coords: { x: coords.centerX, y: coords.centerY },
-          confidence: element.confidence,
-          elementRef 
-        });
-      } else {
-        logger.warn("find: Element not found", { description: params.description });
-      }
 
-      // Return raw SDK response directly
-      const rawResponse = element._response || {};
       const duration = Date.now() - startTime;
-      
-      // Store cropped image for resource serving (instead of inline data URL)
-      let croppedImageResourceUri: string | undefined;
-      let screenshotResourceUri: string | undefined;
-      const croppedImage = rawResponse.croppedImage;
-      if (croppedImage) {
-        const imageData = croppedImage.startsWith('data:') 
-          ? croppedImage.replace(/^data:image\/\w+;base64,/, '')
-          : croppedImage;
-        croppedImageResourceUri = storeImage(imageData, "cropped");
-        // Remove croppedImage from response to avoid context bloat
-        delete rawResponse.croppedImage;
-      } else if (!found) {
-        // Element not found and no cropped image - capture a fresh screenshot
-        // so the user can see what's currently visible on screen
-        try {
-          const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-          if (screenshotBase64) {
-            screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-            logger.debug("find: Captured screenshot for not-found state");
-          }
-        } catch (e) {
-          logger.warn("find: Failed to capture screenshot for not-found state", { error: String(e) });
-        }
-      }
-      
-      // Remove extractedText and pixelDiffImage from response to reduce context bloat
-      delete rawResponse.extractedText;
-      delete rawResponse.pixelDiffImage;
-
-      // Generate code for this find action
-      const generatedCode = found ? generateActionCode("find", { description: params.description }) : undefined;
-
-      // Build element info for display (cropped image is always centered on element)
-      const elementInfo = found ? {
-        description: params.description,
-        centerX: coords?.centerX,
-        centerY: coords?.centerY,
-        confidence: element.confidence,
-        ref: elementRef,
-      } : undefined;
+      logger.info("find: Completed", { description: params.description, found: result.ok, duration });
 
-      return createToolResult(
-        found,
-        found
-          ? `Found: "${params.description}" at (${rawResponse.coordinates?.x}, ${rawResponse.coordinates?.y})\nRef: ${elementRef}`
-          : `Element not found: "${params.description}"`,
-        {
-          ...rawResponse,
-          action: "find",
-          element: elementInfo,
-          ref: elementRef,
-          croppedImageResourceUri,
-          screenshotResourceUri,
-          duration,
-        },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("find: No active session");
+        return noSessionResult(error);
+      }
       const cancelled = cancelledResultOrNull(error, "find");
       if (cancelled) return cancelled;
       logger.error("find: Failed", { error: String(error), description: params.description });
@@ -1186,121 +891,25 @@ registerAppTool(
     const progress = makeProgressReporter(extra);
     logger.info("findall: Starting", { description: params.description, timeout: params.timeout });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("findall: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("findall: Calling SDK findAll");
       const stopHeartbeat = progress.heartbeat(`Looking for all "${params.description}"...`);
-      let elements: any;
+      let result: ActionResult;
       try {
-        elements = await raceAbort(
-          extra.signal,
-          "findall",
-          sdk.findAll(params.description, params.timeout ? { timeout: params.timeout } : undefined)
-        );
+        result = await raceAbort(extra.signal, "findall", core.findAll(params.description, params.timeout));
       } finally {
         stopHeartbeat();
       }
-      const count = elements.length;
 
-      // Store element refs for later use
-      const refs: string[] = [];
-      const elementInfos: Array<{ ref: string; x: number; y: number; centerX: number; centerY: number; confidence: number }> = [];
-      
-      for (let i = 0; i < elements.length; i++) {
-        const element = elements[i];
-        const coords = element.getCoordinates();
-        const elementRef = `el-${Date.now()}-${i}`;
-        
-        if (coords) {
-          elementRefs.set(elementRef, {
-            element: element,
-            description: `${params.description} [${i}]`,
-            coords: {
-              x: coords.x,
-              y: coords.y,
-              centerX: coords.centerX,
-              centerY: coords.centerY,
-            },
-          });
-          refs.push(elementRef);
-          elementInfos.push({
-            ref: elementRef,
-            x: coords.x,
-            y: coords.y,
-            centerX: coords.centerX,
-            centerY: coords.centerY,
-            confidence: element.confidence,
-          });
-        }
-      }
-
-      logger.info("findall: Elements found", { 
-        description: params.description, 
-        count,
-        refs 
-      });
-
-      // Get the first element's response for the image (shows all highlights)
-      const rawResponse = elements[0]?._response || {};
       const duration = Date.now() - startTime;
-      
-      // Store cropped image for resource serving (instead of inline data URL)
-      let croppedImageResourceUri: string | undefined;
-      let screenshotResourceUri: string | undefined;
-      const croppedImage = rawResponse.croppedImage;
-      if (croppedImage) {
-        const imageData = croppedImage.startsWith('data:') 
-          ? croppedImage.replace(/^data:image\/\w+;base64,/, '')
-          : croppedImage;
-        croppedImageResourceUri = storeImage(imageData, "cropped");
-        // Remove croppedImage from response to avoid context bloat
-        delete rawResponse.croppedImage;
-      } else if (count === 0) {
-        // No elements found and no cropped image - capture a fresh screenshot
-        // so the user can see what's currently visible on screen
-        try {
-          const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-          if (screenshotBase64) {
-            screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-            logger.debug("findall: Captured screenshot for not-found state");
-          }
-        } catch (e) {
-          logger.warn("findall: Failed to capture screenshot for not-found state", { error: String(e) });
-        }
-      }
-      
-      // Remove extractedText and pixelDiffImage from response to reduce context bloat
-      delete rawResponse.extractedText;
-      delete rawResponse.pixelDiffImage;
-
-      // Generate code for this findall action
-      const generatedCode = count > 0 ? generateActionCode("findall", { description: params.description }) : undefined;
+      logger.info("findall: Completed", { description: params.description, count: result.data.count, duration });
 
-      // Build refs list for text output
-      const refsList = refs.map((ref, i) => `  [${i}] ${ref}`).join('\n');
-
-      return createToolResult(
-        count > 0,
-        count > 0
-          ? `Found ${count} elements matching "${params.description}":\n${refsList}`
-          : `No elements found matching: "${params.description}"`,
-        {
-          ...rawResponse,
-          count,
-          refs,
-          elements: elementInfos,
-          croppedImageResourceUri,
-          screenshotResourceUri,
-          duration,
-        },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("findall: No active session");
+        return noSessionResult(error);
+      }
       const cancelled = cancelledResultOrNull(error, "findall");
       if (cancelled) return cancelled;
       logger.error("findall: Failed", { error: String(error), description: params.description });
@@ -1327,70 +936,20 @@ registerAppTool(
     const startTime = Date.now();
     logger.info("click: Starting", { elementRef: params.elementRef, action: params.action });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("click: No active session");
-      return sessionCheck.error;
-    }
-
-    // Look up the element reference
-    const ref = elementRefs.get(params.elementRef);
-    if (!ref) {
-      logger.warn("click: Element reference not found", { elementRef: params.elementRef });
-      return createToolResult(false, `Element reference "${params.elementRef}" not found. Use 'find' first to locate the element.`, { error: "Element reference not found" });
-    }
-
-    const { element, description, coords } = ref;
-
     try {
-      logger.debug("click: Executing click on element", { description, action: params.action });
-      
-      // Use the Element's click method instead of raw coordinates
-      if (params.action === "click") {
-        await element.click();
-      } else if (params.action === "double-click") {
-        await element.doubleClick();
-      } else if (params.action === "right-click") {
-        await element.rightClick();
-      }
+      logger.debug("click: Executing click on element", { elementRef: params.elementRef, action: params.action });
+      const result = await core.click(params.elementRef, params.action);
 
-      // Capture screenshot after click to show result
-      logger.debug("click: Capturing screenshot after click");
-      const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-      
-      let screenshotResourceUri: string | undefined;
-      if (screenshotBase64) {
-        screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-        lastScreenshotBase64 = screenshotBase64;
-      }
-
-      const rawResponse = element._response || {};
-      // Remove large data from response to reduce context bloat
-      delete rawResponse.croppedImage;
-      delete rawResponse.extractedText;
-      delete rawResponse.pixelDiffImage;
-      
       const duration = Date.now() - startTime;
-      logger.info("click: Completed", { description, duration });
-
-      // Generate code for this click action
-      const generatedCode = generateActionCode("click", { action: params.action });
+      logger.info("click: Completed", { elementRef: params.elementRef, duration });
 
-      return createToolResult(
-        true,
-        `Clicked on "${description}"`,
-        { 
-          ...rawResponse, 
-          action: "click", 
-          clickAction: params.action, 
-          clickPosition: coords, 
-          screenshotResourceUri,
-          duration 
-        },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
-      logger.error("click: Failed", { error: String(error), description });
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("click: No active session");
+        return noSessionResult(error);
+      }
+      logger.error("click: Failed", { error: String(error), elementRef: params.elementRef });
       captureException(error as Error, { tags: { tool: "click" }, extra: { elementRef: params.elementRef, action: params.action } });
       throw error;
     }
@@ -1413,60 +972,20 @@ registerAppTool(
     const startTime = Date.now();
     logger.info("hover: Starting", { elementRef: params.elementRef });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("hover: No active session");
-      return sessionCheck.error;
-    }
-
-    // Look up the element reference
-    const ref = elementRefs.get(params.elementRef);
-    if (!ref) {
-      logger.warn("hover: Element reference not found", { elementRef: params.elementRef });
-      return createToolResult(false, `Element reference "${params.elementRef}" not found. Use 'find' first to locate the element.`, { error: "Element reference not found" });
-    }
-
-    const { element, description, coords } = ref;
-
     try {
-      logger.debug("hover: Executing hover on element", { description });
-      await element.hover();
+      logger.debug("hover: Executing hover on element", { elementRef: params.elementRef });
+      const result = await core.hover(params.elementRef);
 
-      // Capture screenshot after hover to show result
-      logger.debug("hover: Capturing screenshot after hover");
-      const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-      
-      let screenshotResourceUri: string | undefined;
-      if (screenshotBase64) {
-        screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-        lastScreenshotBase64 = screenshotBase64;
-      }
-
-      const rawResponse = element._response || {};
-      // Remove large data from response to reduce context bloat
-      delete rawResponse.croppedImage;
-      delete rawResponse.extractedText;
-      delete rawResponse.pixelDiffImage;
-      
       const duration = Date.now() - startTime;
-      logger.info("hover: Completed", { description, duration });
+      logger.info("hover: Completed", { elementRef: params.elementRef, duration });
 
-      // Generate code for this hover action
-      const generatedCode = generateActionCode("hover", {});
-
-      return createToolResult(
-        true,
-        `Hovered over "${description}"`,
-        { 
-          ...rawResponse, 
-          action: "hover", 
-          screenshotResourceUri,
-          duration 
-        },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
-      logger.error("hover: Failed", { error: String(error), description });
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("hover: No active session");
+        return noSessionResult(error);
+      }
+      logger.error("hover: Failed", { error: String(error), elementRef: params.elementRef });
       captureException(error as Error, { tags: { tool: "hover" }, extra: { elementRef: params.elementRef } });
       throw error;
     }
@@ -1486,29 +1005,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("wait: Starting", { timeout: params.timeout });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("wait: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("wait: Waiting", { timeout: params.timeout });
-      await sdk.wait(params.timeout);
+      const result = await core.wait(params.timeout);
 
       const duration = Date.now() - startTime;
       logger.info("wait: Completed", { timeout: params.timeout, duration });
 
-      // Generate code for this wait action
-      const generatedCode = generateActionCode("wait", { timeout: params.timeout });
-
-      return createToolResult(
-        true,
-        `Waited for ${params.timeout}ms`,
-        { action: "wait", timeout: params.timeout, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("wait: No active session");
+        return noSessionResult(error);
+      }
       logger.error("wait: Failed", { error: String(error) });
       captureException(error as Error, { tags: { tool: "wait" }, extra: { timeout: params.timeout } });
       throw error;
@@ -1529,29 +1038,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("focus_application: Starting", { name: params.name });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("focus_application: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("focus_application: Focusing", { name: params.name });
-      await sdk.focusApplication(params.name);
+      const result = await core.focusApplication(params.name);
 
       const duration = Date.now() - startTime;
       logger.info("focus_application: Completed", { name: params.name, duration });
 
-      // Generate code for this focus action
-      const generatedCode = generateActionCode("focus_application", { name: params.name });
-
-      return createToolResult(
-        true,
-        `Focused application: "${params.name}"`,
-        { action: "focus", name: params.name, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("focus_application: No active session");
+        return noSessionResult(error);
+      }
       logger.error("focus_application: Failed", { error: String(error), name: params.name });
       captureException(error as Error, { tags: { tool: "focus_application" }, extra: { name: params.name } });
       throw error;
@@ -1577,146 +1076,25 @@ registerAppTool(
     const progress = makeProgressReporter(extra);
     logger.info("find_and_click: Starting", { description: params.description, action: params.action });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("find_and_click: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("find_and_click: Finding element");
       const stopHeartbeat = progress.heartbeat(`Looking for "${params.description}"...`);
-      let element: any;
+      let result: ActionResult;
       try {
-        element = await raceAbort(extra.signal, "find_and_click", sdk.find(params.description));
+        result = await raceAbort(extra.signal, "find_and_click", core.findAndClick(params.description, params.action));
       } finally {
         stopHeartbeat();
       }
-      const found = element.found();
-
-      if (!found) {
-        logger.warn("find_and_click: Element not found", { description: params.description });
-        
-        // Capture screenshot to show current state even when element not found
-        const rawResponse = element._response || {};
-        const duration = Date.now() - startTime;
-        
-        // Store cropped image (screenshot) for resource serving
-        let croppedImageResourceUri: string | undefined;
-        let screenshotResourceUri: string | undefined;
-        const croppedImage = rawResponse.croppedImage;
-        if (croppedImage) {
-          const imageData = croppedImage.startsWith('data:') 
-            ? croppedImage.replace(/^data:image\/\w+;base64,/, '')
-            : croppedImage;
-          croppedImageResourceUri = storeImage(imageData, "screenshot");
-          delete rawResponse.croppedImage;
-        } else {
-          // No cropped image - capture a fresh screenshot so the user can see
-          // what's currently visible on screen when element was not found
-          try {
-            const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-            if (screenshotBase64) {
-              screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
-              logger.debug("find_and_click: Captured screenshot for not-found state");
-            }
-          } catch (e) {
-            logger.warn("find_and_click: Failed to capture screenshot for not-found state", { error: String(e) });
-          }
-        }
-        
-        // Remove extractedText and pixelDiffImage from response to reduce context bloat
-        delete rawResponse.extractedText;
-        delete rawResponse.pixelDiffImage;
-        
-        return createToolResult(
-          false, 
-          `Element not found: "${params.description}"`, 
-          { 
-            ...rawResponse,
-            action: "find_and_click",
-            error: "Element not found", 
-            croppedImageResourceUri,
-            screenshotResourceUri,
-            duration 
-          }
-        );
-      }
-
-      const coords = element.getCoordinates();
-
-      // Store element ref for later use (in case user wants to interact again)
-      const elementRef = `el-${Date.now()}`;
-      if (coords) {
-        elementRefs.set(elementRef, {
-          element: element,
-          description: params.description,
-          coords: {
-            x: coords.x,
-            y: coords.y,
-            centerX: coords.centerX,
-            centerY: coords.centerY,
-          },
-        });
-      }
-
-      logger.debug("find_and_click: Element found, clicking", { action: params.action, elementRef });
-      if (params.action === "click") {
-        await element.click();
-      } else if (params.action === "double-click") {
-        await element.doubleClick();
-      } else if (params.action === "right-click") {
-        await element.rightClick();
-      }
 
-      // Return raw SDK response directly
-      const rawResponse = element._response || {};
       const duration = Date.now() - startTime;
-      
-      // Store cropped image for resource serving (instead of inline data URL)
-      let croppedImageResourceUri: string | undefined;
-      const croppedImage = rawResponse.croppedImage;
-      if (croppedImage) {
-        const imageData = croppedImage.startsWith('data:') 
-          ? croppedImage.replace(/^data:image\/\w+;base64,/, '')
-          : croppedImage;
-        croppedImageResourceUri = storeImage(imageData, "cropped");
-        // Remove croppedImage from response to avoid context bloat
-        delete rawResponse.croppedImage;
-      }
-      
-      // Remove extractedText and pixelDiffImage from response to reduce context bloat
-      delete rawResponse.extractedText;
-      delete rawResponse.pixelDiffImage;
-
-      // Generate code for this find_and_click action
-      const generatedCode = generateActionCode("find_and_click", { description: params.description, action: params.action });
-
-      // Build element info for display (match find action format)
-      const elementInfo = coords ? {
-        description: params.description,
-        centerX: coords.centerX,
-        centerY: coords.centerY,
-        confidence: element.confidence,
-        ref: elementRef,
-      } : undefined;
+      logger.info("find_and_click: Completed", { description: params.description, found: result.ok, duration });
 
-      return createToolResult(
-        true,
-        `Found and clicked: "${params.description}" at (${rawResponse.coordinates?.x}, ${rawResponse.coordinates?.y})\nRef: ${elementRef}`,
-        {
-          ...rawResponse,
-          action: "find_and_click",
-          element: elementInfo,
-          ref: elementRef,
-          clickAction: params.action,
-          clickPosition: coords ? { x: coords.centerX, y: coords.centerY } : undefined,
-          croppedImageResourceUri,
-          duration,
-        },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("find_and_click: No active session");
+        return noSessionResult(error);
+      }
       const cancelled = cancelledResultOrNull(error, "find_and_click");
       if (cancelled) return cancelled;
       logger.error("find_and_click: Failed", { error: String(error), description: params.description });
@@ -1741,29 +1119,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("type: Starting", { textLength: params.text.length, secret: params.secret });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("type: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("type: Typing text");
-      await sdk.type(params.text, { secret: params.secret, delay: params.delay });
+      const result = await core.type(params.text, params.secret, params.delay);
 
       const duration = Date.now() - startTime;
       logger.info("type: Completed", { duration });
 
-      // Generate code for this type action
-      const generatedCode = generateActionCode("type", { text: params.text, secret: params.secret });
-
-      return createToolResult(
-        true,
-        `Typed: ${params.secret ? "[secret text]" : `"${params.text}"`}`,
-        { action: "type", text: params.secret ? "[SECRET]" : params.text, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("type: No active session");
+        return noSessionResult(error);
+      }
       logger.error("type: Failed", { error: String(error) });
       captureException(error as Error, { tags: { tool: "type" }, extra: { textLength: params.text.length, secret: params.secret } });
       throw error;
@@ -1784,29 +1152,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("press_keys: Starting", { keys: params.keys });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("press_keys: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("press_keys: Pressing keys");
-      await sdk.pressKeys(params.keys);
+      const result = await core.pressKeys(params.keys);
 
       const duration = Date.now() - startTime;
       logger.info("press_keys: Completed", { keys: params.keys, duration });
 
-      // Generate code for this press_keys action
-      const generatedCode = generateActionCode("press_keys", { keys: params.keys });
-
-      return createToolResult(
-        true,
-        `Pressed keys: ${params.keys.join(" + ")}`,
-        { action: "press_keys", keys: params.keys, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("press_keys: No active session");
+        return noSessionResult(error);
+      }
       logger.error("press_keys: Failed", { error: String(error), keys: params.keys });
       captureException(error as Error, { tags: { tool: "press_keys" }, extra: { keys: params.keys } });
       throw error;
@@ -1828,29 +1186,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("scroll: Starting", { direction: params.direction, amount: params.amount });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("scroll: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("scroll: Scrolling");
-      await sdk.scroll(params.direction, params.amount ? { amount: params.amount } : undefined);
+      const result = await core.scroll(params.direction, params.amount);
 
       const duration = Date.now() - startTime;
       logger.info("scroll: Completed", { direction: params.direction, duration });
 
-      // Generate code for this scroll action
-      const generatedCode = generateActionCode("scroll", { direction: params.direction, amount: params.amount });
-
-      return createToolResult(
-        true,
-        `Scrolled ${params.direction}${params.amount ? ` by ${params.amount}px` : ""}`,
-        { action: "scroll", scrollDirection: params.direction, direction: params.direction, amount: params.amount, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("scroll: No active session");
+        return noSessionResult(error);
+      }
       logger.error("scroll: Failed", { error: String(error), direction: params.direction });
       captureException(error as Error, { tags: { tool: "scroll" }, extra: { direction: params.direction, amount: params.amount } });
       throw error;
@@ -1877,29 +1225,19 @@ Unlike 'check' which is for your understanding during development, 'assert' crea
     const startTime = Date.now();
     logger.info("assert: Starting", { assertion: params.assertion });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("assert: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("assert: Running assertion");
-      const result = await sdk.assert(params.assertion);
+      const result = await core.assert(params.assertion);
 
       const duration = Date.now() - startTime;
-      logger.info("assert: Completed", { assertion: params.assertion, passed: result, duration });
-
-      // Generate code for this assert action
-      const generatedCode = generateActionCode("assert", { assertion: params.assertion });
+      logger.info("assert: Completed", { assertion: params.assertion, passed: result.ok, duration });
 
-      return createToolResult(
-        result,
-        result ? `✓ Assertion passed: "${params.assertion}"` : `✗ Assertion failed: "${params.assertion}"`,
-        { action: "assert", assertion: params.assertion, passed: result, success: result, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("assert: No active session");
+        return noSessionResult(error);
+      }
       logger.error("assert: Failed", { error: String(error), assertion: params.assertion });
       captureException(error as Error, { tags: { tool: "assert" }, extra: { assertion: params.assertion } });
       throw error;
@@ -1943,122 +1281,61 @@ You can optionally provide a reference image URI to compare against a previous s
     const progress = makeProgressReporter(extra);
     logger.info("check: Starting", { task: params.task, hasReferenceImageUri: !!params.referenceImageUri });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("check: No active session");
-      return sessionCheck.error;
-    }
-
     try {
-      // Capture current screenshot
-      logger.debug("check: Capturing current screenshot");
-      progress.report("Capturing screenshot...");
-      const currentScreenshot = await sdk.agent.system.captureScreenBase64(1, false, true);
-      
-      // Use provided reference image URI, last screenshot as "before" state, or current if no previous screenshot
-      let beforeScreenshot: string;
+      // Resolve the optional reference image from the image store (a server/UI
+      // concern). The core handles "last screenshot" / current-screenshot
+      // fallback internally when no reference image is passed.
+      let referenceImage: string | undefined;
       if (params.referenceImageUri) {
         // Extract image ID from URI (e.g., "screenshot://testdriver/screenshot/screenshot-1" -> "screenshot-1")
-        const uriParts = params.referenceImageUri.split('/');
+        const uriParts = params.referenceImageUri.split("/");
         const imageId = uriParts[uriParts.length - 1];
-        
-        logger.info("check: Looking up reference image", { 
-          referenceImageUri: params.referenceImageUri, 
+
+        logger.info("check: Looking up reference image", {
+          referenceImageUri: params.referenceImageUri,
           extractedImageId: imageId,
           imageStoreSize: imageStore.size,
-          availableKeys: Array.from(imageStore.keys())
+          availableKeys: Array.from(imageStore.keys()),
         });
-        
+
         const storedImage = getStoredImage(imageId);
-        
+
         if (storedImage) {
-          logger.info("check: Found reference image", { 
-            imageId, 
+          logger.info("check: Found reference image", {
+            imageId,
             dataLength: storedImage.data?.length,
             type: storedImage.type,
-            hasData: !!storedImage.data
+            hasData: !!storedImage.data,
           });
-          beforeScreenshot = storedImage.data;
+          referenceImage = storedImage.data;
         } else {
-          logger.warn("check: Reference image NOT found in store, falling back to last screenshot", { 
-            referenceImageUri: params.referenceImageUri, 
+          logger.warn("check: Reference image NOT found in store, falling back to last screenshot", {
+            referenceImageUri: params.referenceImageUri,
             imageId,
             imageStoreSize: imageStore.size,
-            availableKeys: Array.from(imageStore.keys())
+            availableKeys: Array.from(imageStore.keys()),
           });
-          beforeScreenshot = lastScreenshotBase64 || currentScreenshot;
         }
-      } else {
-        beforeScreenshot = lastScreenshotBase64 || currentScreenshot;
       }
-      
-      // Update last screenshot for next check
-      lastScreenshotBase64 = currentScreenshot;
-      
-      // Get system state
-      const mousePosition = await sdk.agent.system.getMousePosition();
-      const activeWindow = await sdk.agent.system.activeWin();
-      
-      // Call the check endpoint
-      logger.info("check: Calling check API endpoint", { 
-        hasLastScreenshot: beforeScreenshot !== currentScreenshot,
-        usingReferenceImageUri: !!params.referenceImageUri,
-        beforeScreenshotLength: beforeScreenshot?.length || 0,
-        currentScreenshotLength: currentScreenshot?.length || 0,
-        beforeScreenshotPreview: beforeScreenshot?.substring(0, 50),
-        currentScreenshotPreview: currentScreenshot?.substring(0, 50)
-      });
+
+      progress.report("Capturing screenshot...");
       const stopHeartbeat = progress.heartbeat(`Checking: "${params.task}"...`);
-      let response: any;
+      let result: ActionResult;
       try {
-        response = await raceAbort(extra.signal, "check", sdk.agent.sdk.req("check", {
-          tasks: [params.task],
-          images: [beforeScreenshot, currentScreenshot],
-          mousePosition,
-          activeWindow,
-        }));
+        result = await raceAbort(extra.signal, "check", core.check(params.task, referenceImage));
       } finally {
         stopHeartbeat();
       }
 
-      const aiResponse = response.data;
-      
-      // Store screenshot for resource serving
-      let screenshotResourceUri: string | undefined;
-      if (currentScreenshot) {
-        screenshotResourceUri = storeImage(currentScreenshot, "screenshot");
-      }
-      
-      // Determine if the check passed based on the AI response
-      // The AI typically returns markdown with its analysis
-      // We consider it "complete" if the response doesn't contain code blocks (indicating more work needed)
-      const hasCodeBlocks = aiResponse && (
-        aiResponse.includes("```yml") || 
-        aiResponse.includes("```yaml") ||
-        aiResponse.includes("- command:")
-      );
-      const isComplete = !hasCodeBlocks;
-
       const duration = Date.now() - startTime;
-      logger.info("check: Completed", { task: params.task, complete: isComplete, duration });
+      logger.info("check: Completed", { task: params.task, complete: result.ok, duration });
 
-      // Note: check doesn't generate code - it's for AI understanding, not test recording
-      return createToolResult(
-        isComplete,
-        isComplete 
-          ? `✓ Task appears complete: "${params.task}"\n\nAI Analysis:\n${aiResponse}`
-          : `⚠ Task may not be complete: "${params.task}"\n\nAI Analysis:\n${aiResponse}`,
-        { 
-          action: "check", 
-          task: params.task, 
-          complete: isComplete, 
-          success: isComplete, 
-          aiResponse, 
-          screenshotResourceUri,
-          duration 
-        }
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("check: No active session");
+        return noSessionResult(error);
+      }
       const cancelled = cancelledResultOrNull(error, "check");
       if (cancelled) return cancelled;
       logger.error("check: Failed", { error: String(error), task: params.task });
@@ -2083,29 +1360,19 @@ server.registerTool(
     const startTime = Date.now();
     logger.info("exec: Starting", { language: params.language, codeLength: params.code.length, timeout: params.timeout });
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("exec: No active session");
-      return sessionCheck.error;
-    }
-
     try {
       logger.debug("exec: Executing code", { language: params.language });
-      const output = await sdk.exec(params.language, params.code, params.timeout);
+      const result = await core.exec(params.language, params.code, params.timeout);
 
       const duration = Date.now() - startTime;
-      logger.info("exec: Completed", { language: params.language, outputLength: output?.length || 0, duration });
+      logger.info("exec: Completed", { language: params.language, outputLength: (result.data.output as string)?.length || 0, duration });
 
-      // Generate code for this exec action
-      const generatedCode = generateActionCode("exec", { language: params.language, code: params.code, timeout: params.timeout });
-
-      return createToolResult(
-        true,
-        `Executed ${params.language} code:\n${output || "(no output)"}`,
-        { action: "exec", language: params.language, output, duration },
-        generatedCode
-      );
+      return resultToMcp({ ...result, data: { ...result.data, duration } });
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("exec: No active session");
+        return noSessionResult(error);
+      }
       logger.error("exec: Failed", { error: String(error), language: params.language });
       captureException(error as Error, { tags: { tool: "exec" }, extra: { language: params.language, codeLength: params.code.length } });
       throw error;
@@ -2485,37 +1752,27 @@ Only use 'screenshot' when you explicitly want to show something to the human us
     const startTime = Date.now();
     logger.info("screenshot: Starting");
 
-    const sessionCheck = requireActiveSession();
-    if (!sessionCheck.valid) {
-      logger.warn("screenshot: No active session");
-      return sessionCheck.error;
-    }
-
     try {
-      // Capture full screen screenshot
-      const screenshotBase64 = await sdk.agent.system.captureScreenBase64(1, false, true);
-      
-      let screenshotResourceUri: string | undefined;
-      if (screenshotBase64) {
-        // Store raw base64 for the resource blob with unique ID
-        screenshotResourceUri = storeImage(screenshotBase64, "screenshot");
+      const result = await core.screenshot();
+
+      // Store the captured image as a resource URI (kept OUT of AI context — the
+      // MCP app fetches it via resources/read). Preserve the original user-facing
+      // text rather than the core's neutral text.
+      const data: Record<string, unknown> = { action: "screenshot" };
+      for (const img of result.images ?? []) {
+        data.screenshotResourceUri = storeImage(img.base64, "screenshot");
       }
-      
+
       const duration = Date.now() - startTime;
-      logger.info("screenshot: Completed", { duration, hasImage: !!screenshotBase64 });
+      data.duration = duration;
+      logger.info("screenshot: Completed", { duration, hasImage: (result.images?.length ?? 0) > 0 });
 
-      // Only send the resource URI - the MCP app will fetch the image via resources/read
-      // This keeps the base64 image data OUT of AI context
-      return createToolResult(
-        true,
-        "Screenshot captured and displayed to user",
-        { 
-          action: "screenshot",
-          screenshotResourceUri,
-          duration 
-        }
-      );
+      return createToolResult(true, "Screenshot captured and displayed to user", data);
     } catch (error) {
+      if (error instanceof NoActiveSessionError) {
+        logger.warn("screenshot: No active session");
+        return noSessionResult(error);
+      }
       logger.error("screenshot: Failed", { error: String(error) });
       return createToolResult(false, `Screenshot failed: ${error}`, { error: String(error) });
     }
@@ -2612,27 +1869,111 @@ Learn more at https://docs.testdriver.ai/v7/getting-started/
 );
 
 
+// =============================================================================
+// HTTP transport (Streamable HTTP, no auth)
+// =============================================================================
+
+/**
+ * Serve the MCP server over Streamable HTTP with no authentication.
+ *
+ * This is what the eve agent (and any other Streamable-HTTP MCP client) connects
+ * to. The server holds single-session global state (the active `sdk`, element
+ * refs, image store), so we mirror that with a single long-lived MCP transport:
+ * the first `initialize` request mints a session id and every subsequent request
+ * reuses the same `server` + transport. Concurrent sessions are intentionally not
+ * supported — this server provisions one sandbox at a time.
+ *
+ * No auth is applied here by design (the connection is expected to be local-only
+ * or otherwise protected outside the MCP layer). Do not expose this on a public
+ * network without putting a real authenticating proxy in front of it.
+ */
+async function startHttpServer() {
+  const host = process.env.TD_MCP_HOST || "127.0.0.1";
+  const port = Number(process.env.TD_MCP_PORT || process.env.PORT || 8788);
+  const mcpPath = process.env.TD_MCP_PATH || "/mcp";
+
+  // One transport for the lifetime of the process (single-session design).
+  const transport = new StreamableHTTPServerTransport({
+    sessionIdGenerator: () => randomUUID(),
+  });
+  await server.connect(transport);
+
+  const httpServer = http.createServer((req, res) => {
+    const url = new URL(req.url || "/", `http://${req.headers.host}`);
+
+    // Lightweight health check for readiness probes / `eve dev`.
+    if (req.method === "GET" && url.pathname === "/health") {
+      res.writeHead(200, { "content-type": "application/json" });
+      res.end(JSON.stringify({ status: "ok", server: "testdriver", version }));
+      return;
+    }
+
+    if (url.pathname !== mcpPath) {
+      res.writeHead(404, { "content-type": "application/json" });
+      res.end(JSON.stringify({ error: "Not found", hint: `MCP endpoint is ${mcpPath}` }));
+      return;
+    }
+
+    // The Streamable HTTP transport handles POST (requests), GET (SSE stream),
+    // and DELETE (session teardown) on the same path. Body parsing is left to
+    // the transport so it can read the raw stream.
+    transport.handleRequest(req, res).catch((error) => {
+      logger.error("http: handleRequest failed", { error: String(error) });
+      captureException(error as Error, { tags: { phase: "http" } });
+      if (!res.headersSent) {
+        res.writeHead(500, { "content-type": "application/json" });
+        res.end(JSON.stringify({ error: "Internal server error" }));
+      }
+    });
+  });
+
+  await new Promise<void>((resolve) => httpServer.listen(port, host, resolve));
+  logger.info("TestDriver MCP Server running on Streamable HTTP", {
+    url: `http://${host}:${port}${mcpPath}`,
+  });
+
+  const shutdown = async () => {
+    logger.info("Shutting down MCP Server");
+    httpServer.close();
+    await transport.close().catch(() => {});
+    await flushSentry();
+    process.exit(0);
+  };
+  process.on("SIGINT", shutdown);
+  process.on("SIGTERM", shutdown);
+}
+
 // Start the server
 async function main() {
-  logger.info("Starting TestDriver MCP Server", { 
+  // Transport selection: TD_MCP_TRANSPORT=http serves Streamable HTTP (used by
+  // the eve agent); anything else (default) uses stdio for local CLI clients.
+  const transportMode = (process.env.TD_MCP_TRANSPORT || "stdio").toLowerCase();
+
+  logger.info("Starting TestDriver MCP Server", {
     version,
+    transport: transportMode,
     logLevel: process.env.TD_LOG_LEVEL || "INFO",
     distDir: DIST_DIR,
     sentryEnabled: isSentryEnabled(),
   });
-  
+
+  if (transportMode === "http") {
+    await startHttpServer();
+    return;
+  }
+
   const transport = new StdioServerTransport();
   await server.connect(transport);
-  
+
   logger.info("TestDriver MCP Server running on stdio");
-  
+
   // Handle graceful shutdown
   const shutdown = async () => {
     logger.info("Shutting down MCP Server");
     await flushSentry();
     process.exit(0);
   };
-  
+
   process.on("SIGINT", shutdown);
   process.on("SIGTERM", shutdown);
 }