@cutleryapp/agent 1.0.18 → 1.0.19

package/dist/mcp-executor.js

@@ -248,6 +248,57 @@ async function tryClick(page, nameRe, label) {
     // AI vision fallback
     return await aiClickFallback(page, label);
 }
+function buildAgentPrompt(stepText, round) {
+    return `You are an intelligent browser test agent with vision. Your job is to look at the current screen, understand the test step intent, and decide what actions to perform.
+
+GOAL: "${stepText}"
+ROUND: ${round + 1}
+
+## YOUR CAPABILITIES
+1. SCREEN UNDERSTANDING — Identify every visible UI element, form field, button, link, and label on screen.
+2. INTENT MAPPING — Understand what the test step MEANS even if the wording is vague or high-level (e.g. "checkout the product" = navigate to cart → fill shipping info → complete purchase).
+3. DATA SIMULATION — If a form needs data that is not specified, INVENT realistic test data:
+   - Names: "John Smith" or "Test User"
+   - Email: "testuser@example.com"
+   - Phone: "555-0100"
+   - Address: "123 Test St, Springfield"
+   - Postal/ZIP: "12345"
+   - Credit card: "4111111111111111", expiry "12/25", CVV "123"
+   - Password: "Test@1234"
+   - Any other field: invent plausible data based on the field label
+
+## RETURN FORMAT
+Return ONLY valid JSON — no markdown, no explanation:
+{
+  "reasoning": "What I see on screen and what I plan to do",
+  "done": false,
+  "actions": [
+    {"action": "click",  "selector": "SELECTOR"},
+    {"action": "fill",   "selector": "SELECTOR", "value": "VALUE"},
+    {"action": "select", "selector": "SELECTOR", "value": "OPTION_LABEL"},
+    {"action": "verify", "text": "TEXT_TO_CHECK", "not": false},
+    {"action": "wait",   "ms": 500}
+  ]
+}
+
+Set "done": true with empty "actions" array when the goal is fully accomplished.
+
+## SELECTOR RULES
+- Prefer visible text: button:has-text("Checkout"), a:has-text("Login")
+- Use id when visible: #first-name, #postal-code
+- Use data attributes: [data-test="checkout-button"], [data-testid="..."]
+- Use placeholder for inputs: input[placeholder="First Name"]
+- Use name attribute: input[name="firstName"]
+- NEVER use position or coordinates
+
+## SMART RULES
+- Look at ALL visible form fields and fill them ALL in one round
+- If you see a multi-step form, complete this step fully then click continue/next/submit
+- If the goal is "checkout": cart → fill info → continue → finish
+- If the goal is "register" or "sign up": fill all fields + submit
+- If the goal is already accomplished (correct page shown), set done: true immediately
+- Maximum 5 actions per round`;
+}
 /**
  * Universal AI agentic fallback.
  * GPT-4o sees the current page screenshot, understands the GOAL of the step,
@@ -266,41 +317,11 @@ async function aiStepFallback(page, stepText) {
         const base64 = screenshotBuffer.toString('base64');
         const response = await openai.chat.completions.create({
             model: 'gpt-4o',
-            max_tokens: 600,
+            max_tokens: 800,
             messages: [{
                     role: 'user',
                     content: [
-                        {
-                            type: 'text',
-                            text: `You are an autonomous browser test agent. Your goal is to accomplish this test step:
-
-"${stepText}"
-
-Look at the current screenshot and decide what actions are needed RIGHT NOW to make progress toward the goal.
-
-Return ONLY valid JSON (no markdown) with this structure:
-{
-  "reasoning": "brief explanation of what you see and what needs to happen",
-  "done": false,
-  "actions": [
-    {"action": "click",  "selector": "CSS_OR_TEXT_SELECTOR"},
-    {"action": "fill",   "selector": "CSS_OR_TEXT_SELECTOR", "value": "TEXT"},
-    {"action": "select", "selector": "CSS_OR_TEXT_SELECTOR", "value": "OPTION_TEXT"},
-    {"action": "verify", "text": "EXPECTED_TEXT", "not": false},
-    {"action": "wait",   "ms": 500}
-  ]
-}
-
-Set "done": true and empty "actions" array when the goal has been fully accomplished.
-
-Rules:
-- Return only the actions needed for THIS screenshot — after executing them you will see the next state
-- Use text-based selectors when possible: button:has-text("Checkout"), [data-test="..."], #id
-- For "checkout" goals: look for cart/checkout buttons and click them in sequence
-- For "login" goals: fill username, fill password, click login
-- For "add to cart" goals: find and click the Add to cart button
-- Maximum 3 actions per round to stay precise`
-                        },
+                        { type: 'text', text: buildAgentPrompt(stepText, round) },
                         { type: 'image_url', image_url: { url: `data:image/png;base64,${base64}` } }
                     ]
                 }]
@@ -343,6 +364,14 @@ Rules:
                 else if (act.action === 'wait') {
                     await page.waitForTimeout(act.ms || 1000);
                 }
+                else if (act.action === 'scroll') {
+                    if (act.selector) {
+                        await page.locator(act.selector).first().scrollIntoViewIfNeeded();
+                    }
+                    else {
+                        await page.evaluate(() => window.scrollBy(0, 400));
+                    }
+                }
             }
             catch (e) {
                 console.log(`  ⚠️ Action failed (${e.message}), continuing to next round...`);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@cutleryapp/agent",
-    "version": "1.0.18",
+    "version": "1.0.19",
     "description": "Local agent that connects your machine to the Cutlery QA platform and runs UI tests via Playwright",
     "main": "dist/cli.js",
     "bin": {