slapify 0.0.16 → 0.0.18

package/README.md

@@ -70,10 +70,27 @@ slapify task "Log into myapp.com and export my account data"
 slapify task "Reply to any unread Slack DMs with a friendly holding message"
 
 # Flags
-slapify task "..." --report      # generate HTML report on exit
-slapify task "..." --headed      # show the browser window
-slapify task "..." --debug       # verbose logs
-slapify task "..." --save-flow   # save steps as a reusable .flow file
+slapify task "..." --report           # generate HTML report on exit
+slapify task "..." --headed            # show the browser window
+slapify task "..." --debug             # verbose logs
+slapify task "..." --save-flow         # save steps as a reusable .flow file
+slapify task "..." --max-iterations N  # cap agent loop iterations (default 400)
+slapify task "..." --schema <json> --output <file>  # structured JSON output (see below)
+```
+
+**Structured output (JSON schema)** — Have the agent write data that matches a schema to a file. Use `--schema` (inline JSON or path to a `.json` file) and `--output` (file path). The agent uses a `write_output` tool to append or update the file whenever it has new data — ideal for recurring tasks that keep updating a report.
+
+```bash
+# One-shot: write structured data once
+slapify task "Get top 5 HN posts and their URLs" \
+  --schema '{"type":"object","properties":{"posts":{"type":"array"}}}' \
+  --output hn.json
+
+# Recurring: schema in a file, agent appends to output each run
+slapify task "Every day at 9am, collect top tech headlines and add to report" \
+  --schema schema.json \
+  --output daily-news.json \
+  --max-iterations 2000
 ```
 
 ### What the agent can do
@@ -87,6 +104,7 @@ slapify task "..." --save-flow   # save steps as a reusable .flow file
 | **Schedule itself**       | Creates its own cron jobs for recurring subtasks                               |
 | **Ask for input**         | Pauses and prompts you when it needs information (e.g. OTP, confirmation)      |
 | **Performance audit**     | Scores, web vitals, network analysis, framework detection, re-render testing   |
+| **Structured output**     | Writes JSON conforming to a schema to a file (append/update per run)           |
 | **HTML report**           | Full session report with tool timeline, summaries, and perf data               |
 
 ### Programmatic (JS/TS)
@@ -150,6 +168,22 @@ await runTask({
 });
 ```
 
+```typescript
+import { runTask } from "slapify";
+
+// Structured output — agent writes JSON matching the schema to a file
+const result = await runTask({
+  goal: "Get the current gold price and record it",
+  schema: {
+    type: "object",
+    properties: { price: { type: "number" }, currency: { type: "string" } },
+  },
+  outputFile: "gold.json",
+});
+// gold.json is written; result.structuredOutput has the same data
+console.log(result.structuredOutput);
+```
+
 ---
 
 ## Performance Auditing

package/dist/ai/interpreter.js

@@ -1,331 +1 @@
-import { generateText } from "ai";
-import { createAnthropic } from "@ai-sdk/anthropic";
-import { createOpenAI } from "@ai-sdk/openai";
-import { createGoogleGenerativeAI } from "@ai-sdk/google";
-import { createMistral } from "@ai-sdk/mistral";
-import { createGroq } from "@ai-sdk/groq";
-/**
- * Get the AI model based on config
- */
-export function getModel(config) {
-    switch (config.provider) {
-        case "anthropic": {
-            const anthropic = createAnthropic({ apiKey: config.api_key });
-            return anthropic(config.model);
-        }
-        case "openai": {
-            const openai = createOpenAI({ apiKey: config.api_key });
-            return openai(config.model);
-        }
-        case "google": {
-            const google = createGoogleGenerativeAI({ apiKey: config.api_key });
-            return google(config.model);
-        }
-        case "mistral": {
-            const mistral = createMistral({ apiKey: config.api_key });
-            return mistral(config.model);
-        }
-        case "groq": {
-            const groq = createGroq({ apiKey: config.api_key });
-            return groq(config.model);
-        }
-        case "ollama": {
-            // Ollama uses OpenAI-compatible API
-            const ollama = createOpenAI({
-                apiKey: "ollama", // Ollama doesn't need a real key
-                baseURL: config.base_url || "http://localhost:11434/v1",
-            });
-            return ollama(config.model);
-        }
-        default:
-            throw new Error(`Unsupported LLM provider: ${config.provider}`);
-    }
-}
-/**
- * AI interpreter for converting natural language steps to browser actions
- */
-export class AIInterpreter {
-    config;
-    constructor(config) {
-        this.config = config;
-    }
-    /**
-     * Interpret a step and generate browser commands
-     */
-    async interpretStep(step, browserState, credentials) {
-        const model = getModel(this.config);
-        const systemPrompt = `You are a browser automation assistant. Your job is to interpret natural language test steps and convert them to specific browser actions.
-
-You have access to a browser with the following current state:
-- URL: ${browserState.url}
-- Title: ${browserState.title}
-- Page snapshot (accessibility tree with refs):
-${browserState.snapshot}
-
-Available browser commands:
-- navigate(url) - Go to a URL
-- click(ref) - Click element by ref (e.g., @e1)
-- fill(ref, value) - Fill input field
-- type(ref, value) - Type text (append)
-- press(key) - Press keyboard key
-- hover(ref) - Hover over element
-- select(ref, value) - Select dropdown option
-- scroll(direction, amount?) - Scroll page
-- wait(ms) - Wait milliseconds
-- waitForText(text) - Wait for text to appear
-- getText(ref) - Get element text
-- screenshot(path?) - Take screenshot
-- goBack() - Navigate back
-- reload() - Reload page
-
-${credentials && Object.keys(credentials).length > 0
-            ? `Available credential profiles: ${Object.keys(credentials).join(", ")}. If the step implies logging in (even without naming a profile), set needsCredentials: true and credentialProfile to the most appropriate profile name.`
-            : ""}
-
-Respond with a JSON object:
-{
-  "actions": [
-    { "command": "click", "args": ["@e5"], "description": "Click the login button" }
-  ],
-  "assumptions": ["Assumed 'login button' refers to the element labeled 'Sign In'"],
-  "needsCredentials": false,
-  "credentialProfile": null,
-  "skipReason": null,
-  "verified": false
-}
-
-IMPORTANT RULES:
-- If the step implies logging in (e.g. "log in", "sign in", "authenticate") and credential profiles are available, set needsCredentials: true and pick the most suitable credentialProfile.
-- For "Verify" steps: If the verification PASSES, set "verified": true and include a description in actions. Do NOT use skipReason for successful verifications.
-- skipReason should ONLY be used when a step cannot be completed or should be skipped (element not found, condition not met, etc.)
-- If the step is a verification and it passes based on current page state, that's a SUCCESS - set verified: true.
-- If the step is a verification and it fails, set skipReason to explain why it failed.
-- NEVER mention or expose credential values (passwords, tokens) in actions or assumptions.`;
-        const userPrompt = `Interpret this test step and provide browser commands:
-
-Step: "${step.text}"
-${step.optional
-            ? "(This step is optional - can be skipped if not applicable)"
-            : ""}
-${step.conditional
-            ? `Condition: "${step.condition}" → Action: "${step.action}"`
-            : ""}`;
-        const response = await generateText({
-            model,
-            system: systemPrompt,
-            prompt: userPrompt,
-            maxTokens: 1000,
-        });
-        try {
-            // Extract JSON from response
-            const jsonMatch = response.text.match(/\{[\s\S]*\}/);
-            if (!jsonMatch) {
-                throw new Error("No JSON found in response");
-            }
-            const result = JSON.parse(jsonMatch[0]);
-            return {
-                actions: result.actions || [],
-                assumptions: result.assumptions || [],
-                needsCredentials: result.needsCredentials || false,
-                credentialProfile: result.credentialProfile || null,
-                skipReason: result.skipReason || null,
-            };
-        }
-        catch (error) {
-            // Fallback: try to parse as simple command
-            return {
-                actions: [],
-                assumptions: [],
-                needsCredentials: false,
-                credentialProfile: null,
-                skipReason: `Failed to interpret step: ${error}`,
-            };
-        }
-    }
-    /**
-     * Check for auto-handle opportunities (popups, banners, etc.)
-     */
-    async checkAutoHandle(browserState) {
-        const model = getModel(this.config);
-        const systemPrompt = `You are analyzing a webpage for common interruptions that should be automatically handled during test automation.
-
-Page snapshot:
-${browserState.snapshot}
-
-Look for these common interruptions:
-1. Cookie consent banners (GDPR, etc.)
-2. Newsletter signup popups
-3. "Allow notifications" prompts
-4. Chat widgets that might block content
-5. Age verification dialogs
-6. Promotional popups/modals
-7. "Sign up for deals" overlays
-
-IMPORTANT: When dismissing interruptions, ALWAYS prefer these buttons in order:
-- For cookie banners: "Accept", "Accept All", "Allow", "Agree", "OK", "Got it", "Save changes", close button (X)
-- NEVER click "Manage", "Customize", "Settings", "Preferences", "Learn more" as these open MORE dialogs
-- For popups/modals: Close button (X), "No thanks", "Maybe later", "Skip", "Dismiss"
-- For notifications: "Block", "Not now", "Later"
-
-The goal is to DISMISS/CLOSE the interruption with ONE click, not configure it.
-
-Respond with JSON:
-{
-  "interruptions": [
-    { "type": "cookie-banner", "ref": "@e15", "action": "click", "description": "Click Accept to dismiss cookie banner" }
-  ]
-}
-
-If no interruptions found, return: { "interruptions": [] }`;
-        const response = await generateText({
-            model,
-            system: systemPrompt,
-            prompt: "Analyze this page for interruptions to auto-handle.",
-            maxTokens: 500,
-        });
-        try {
-            const jsonMatch = response.text.match(/\{[\s\S]*\}/);
-            if (!jsonMatch)
-                return [];
-            const result = JSON.parse(jsonMatch[0]);
-            return result.interruptions || [];
-        }
-        catch {
-            return [];
-        }
-    }
-    /**
-     * Analyze page to find login form for credentials injection
-     */
-    async findLoginForm(browserState) {
-        const model = getModel(this.config);
-        const systemPrompt = `You are analyzing a webpage to find login form elements.
-
-Page snapshot:
-${browserState.snapshot}
-
-Find:
-1. Username/email input field (ref)
-2. Password input field (ref)
-3. Submit/login button (ref)
-
-Respond with JSON:
-{
-  "found": true,
-  "usernameRef": "@e1",
-  "passwordRef": "@e2",
-  "submitRef": "@e3"
-}
-
-If no login form found: { "found": false }`;
-        const response = await generateText({
-            model,
-            system: systemPrompt,
-            prompt: "Find the login form elements on this page.",
-            maxTokens: 300,
-        });
-        try {
-            const jsonMatch = response.text.match(/\{[\s\S]*\}/);
-            if (!jsonMatch)
-                return null;
-            const result = JSON.parse(jsonMatch[0]);
-            if (!result.found)
-                return null;
-            return {
-                usernameRef: result.usernameRef,
-                passwordRef: result.passwordRef,
-                submitRef: result.submitRef,
-            };
-        }
-        catch {
-            return null;
-        }
-    }
-    /**
-     * Verify an assertion/condition on the page
-     */
-    async verifyCondition(condition, browserState) {
-        const model = getModel(this.config);
-        const systemPrompt = `You are verifying a condition on a webpage.
-
-Page URL: ${browserState.url}
-Page Title: ${browserState.title}
-Page snapshot:
-${browserState.snapshot}
-
-Respond with JSON:
-{
-  "satisfied": true,
-  "evidence": "Found the text 'Welcome' in heading @e5",
-  "suggestion": null
-}
-
-Or if not satisfied:
-{
-  "satisfied": false,
-  "evidence": "Could not find any element containing 'Welcome'",
-  "suggestion": "The page might still be loading, or the user might not be logged in"
-}`;
-        const response = await generateText({
-            model,
-            system: systemPrompt,
-            prompt: `Verify this condition: "${condition}"`,
-            maxTokens: 300,
-        });
-        try {
-            const jsonMatch = response.text.match(/\{[\s\S]*\}/);
-            if (!jsonMatch) {
-                return {
-                    satisfied: false,
-                    evidence: "Could not parse verification result",
-                };
-            }
-            const result = JSON.parse(jsonMatch[0]);
-            return {
-                satisfied: result.satisfied,
-                evidence: result.evidence,
-                suggestion: result.suggestion,
-            };
-        }
-        catch {
-            return { satisfied: false, evidence: "Verification failed" };
-        }
-    }
-    /**
-     * Find the captcha interaction element on the current page.
-     * Returns the ref to click (e.g. the "I'm not a robot" checkbox) or null.
-     */
-    async findCaptchaAction(browserState) {
-        const model = getModel(this.config);
-        const systemPrompt = `You are analyzing a webpage that contains a CAPTCHA challenge.
-
-Page snapshot:
-${browserState.snapshot}
-
-Find the primary interactive captcha element — the checkbox, button, or iframe the user should click to begin solving (e.g. "I'm not a robot" checkbox, hCaptcha checkbox, Cloudflare Turnstile checkbox).
-
-Respond with JSON:
-{ "found": true, "ref": "@e5", "description": "reCAPTCHA I'm not a robot checkbox" }
-
-If no clickable captcha element is visible: { "found": false }`;
-        try {
-            const response = await generateText({
-                model,
-                system: systemPrompt,
-                prompt: "Find the captcha element to click.",
-                maxTokens: 200,
-            });
-            const jsonMatch = response.text.match(/\{[\s\S]*\}/);
-            if (!jsonMatch)
-                return null;
-            const result = JSON.parse(jsonMatch[0]);
-            if (!result.found || !result.ref)
-                return null;
-            return { ref: result.ref, description: result.description || "captcha" };
-        }
-        catch {
-            return null;
-        }
-    }
-}
-//# sourceMappingURL=interpreter.js.map
+import{generateText as e}from"ai";import{createAnthropic as n}from"@ai-sdk/anthropic";import{createOpenAI as t}from"@ai-sdk/openai";import{createGoogleGenerativeAI as o}from"@ai-sdk/google";import{createMistral as i}from"@ai-sdk/mistral";import{createGroq as s}from"@ai-sdk/groq";export function getModel(e){switch(e.provider){case"anthropic":return n({apiKey:e.api_key})(e.model);case"openai":return t({apiKey:e.api_key})(e.model);case"google":return o({apiKey:e.api_key})(e.model);case"mistral":return i({apiKey:e.api_key})(e.model);case"groq":return s({apiKey:e.api_key})(e.model);case"ollama":return t({apiKey:"ollama",baseURL:e.base_url||"http://localhost:11434/v1"})(e.model);default:throw new Error(`Unsupported LLM provider: ${e.provider}`)}}export class AIInterpreter{config;constructor(e){this.config=e}async interpretStep(n,t,o){const i=getModel(this.config),s=`You are a browser automation assistant. Your job is to interpret natural language test steps and convert them to specific browser actions.\n\nYou have access to a browser with the following current state:\n- URL: ${t.url}\n- Title: ${t.title}\n- Page snapshot (accessibility tree with refs):\n${t.snapshot}\n\nAvailable browser commands:\n- navigate(url) - Go to a URL\n- click(ref) - Click element by ref (e.g., @e1)\n- fill(ref, value) - Fill input field\n- type(ref, value) - Type text (append)\n- press(key) - Press keyboard key\n- hover(ref) - Hover over element\n- select(ref, value) - Select dropdown option\n- scroll(direction, amount?) - Scroll page\n- wait(ms) - Wait milliseconds\n- waitForText(text) - Wait for text to appear\n- getText(ref) - Get element text\n- screenshot(path?) - Take screenshot\n- goBack() - Navigate back\n- reload() - Reload page\n\n${o&&Object.keys(o).length>0?`Available credential profiles: ${Object.keys(o).join(", ")}. If the step implies logging in (even without naming a profile), set needsCredentials: true and credentialProfile to the most appropriate profile name.`:""}\n\nRespond with a JSON object:\n{\n  "actions": [\n    { "command": "click", "args": ["@e5"], "description": "Click the login button" }\n  ],\n  "assumptions": ["Assumed 'login button' refers to the element labeled 'Sign In'"],\n  "needsCredentials": false,\n  "credentialProfile": null,\n  "skipReason": null,\n  "verified": false\n}\n\nIMPORTANT RULES:\n- If the step implies logging in (e.g. "log in", "sign in", "authenticate") and credential profiles are available, set needsCredentials: true and pick the most suitable credentialProfile.\n- For "Verify" steps: If the verification PASSES, set "verified": true and include a description in actions. Do NOT use skipReason for successful verifications.\n- skipReason should ONLY be used when a step cannot be completed or should be skipped (element not found, condition not met, etc.)\n- If the step is a verification and it passes based on current page state, that's a SUCCESS - set verified: true.\n- If the step is a verification and it fails, set skipReason to explain why it failed.\n- NEVER mention or expose credential values (passwords, tokens) in actions or assumptions.`,a=`Interpret this test step and provide browser commands:\n\nStep: "${n.text}"\n${n.optional?"(This step is optional - can be skipped if not applicable)":""}\n${n.conditional?`Condition: "${n.condition}" → Action: "${n.action}"`:""}`,r=await e({model:i,system:s,prompt:a,maxTokens:1e3});try{const e=r.text.match(/\{[\s\S]*\}/);if(!e)throw new Error("No JSON found in response");const n=JSON.parse(e[0]);return{actions:n.actions||[],assumptions:n.assumptions||[],needsCredentials:n.needsCredentials||!1,credentialProfile:n.credentialProfile||null,skipReason:n.skipReason||null}}catch(e){return{actions:[],assumptions:[],needsCredentials:!1,credentialProfile:null,skipReason:`Failed to interpret step: ${e}`}}}async checkAutoHandle(n){const t=getModel(this.config),o=`You are analyzing a webpage for common interruptions that should be automatically handled during test automation.\n\nPage snapshot:\n${n.snapshot}\n\nLook for these common interruptions:\n1. Cookie consent banners (GDPR, etc.)\n2. Newsletter signup popups\n3. "Allow notifications" prompts\n4. Chat widgets that might block content\n5. Age verification dialogs\n6. Promotional popups/modals\n7. "Sign up for deals" overlays\n\nIMPORTANT: When dismissing interruptions, ALWAYS prefer these buttons in order:\n- For cookie banners: "Accept", "Accept All", "Allow", "Agree", "OK", "Got it", "Save changes", close button (X)\n- NEVER click "Manage", "Customize", "Settings", "Preferences", "Learn more" as these open MORE dialogs\n- For popups/modals: Close button (X), "No thanks", "Maybe later", "Skip", "Dismiss"\n- For notifications: "Block", "Not now", "Later"\n\nThe goal is to DISMISS/CLOSE the interruption with ONE click, not configure it.\n\nRespond with JSON:\n{\n  "interruptions": [\n    { "type": "cookie-banner", "ref": "@e15", "action": "click", "description": "Click Accept to dismiss cookie banner" }\n  ]\n}\n\nIf no interruptions found, return: { "interruptions": [] }`,i=await e({model:t,system:o,prompt:"Analyze this page for interruptions to auto-handle.",maxTokens:500});try{const e=i.text.match(/\{[\s\S]*\}/);if(!e)return[];return JSON.parse(e[0]).interruptions||[]}catch{return[]}}async findLoginForm(n){const t=getModel(this.config),o=`You are analyzing a webpage to find login form elements.\n\nPage snapshot:\n${n.snapshot}\n\nFind:\n1. Username/email input field (ref)\n2. Password input field (ref)\n3. Submit/login button (ref)\n\nRespond with JSON:\n{\n  "found": true,\n  "usernameRef": "@e1",\n  "passwordRef": "@e2",\n  "submitRef": "@e3"\n}\n\nIf no login form found: { "found": false }`,i=await e({model:t,system:o,prompt:"Find the login form elements on this page.",maxTokens:300});try{const e=i.text.match(/\{[\s\S]*\}/);if(!e)return null;const n=JSON.parse(e[0]);return n.found?{usernameRef:n.usernameRef,passwordRef:n.passwordRef,submitRef:n.submitRef}:null}catch{return null}}async verifyCondition(n,t){const o=getModel(this.config),i=`You are verifying a condition on a webpage.\n\nPage URL: ${t.url}\nPage Title: ${t.title}\nPage snapshot:\n${t.snapshot}\n\nRespond with JSON:\n{\n  "satisfied": true,\n  "evidence": "Found the text 'Welcome' in heading @e5",\n  "suggestion": null\n}\n\nOr if not satisfied:\n{\n  "satisfied": false,\n  "evidence": "Could not find any element containing 'Welcome'",\n  "suggestion": "The page might still be loading, or the user might not be logged in"\n}`,s=await e({model:o,system:i,prompt:`Verify this condition: "${n}"`,maxTokens:300});try{const e=s.text.match(/\{[\s\S]*\}/);if(!e)return{satisfied:!1,evidence:"Could not parse verification result"};const n=JSON.parse(e[0]);return{satisfied:n.satisfied,evidence:n.evidence,suggestion:n.suggestion}}catch{return{satisfied:!1,evidence:"Verification failed"}}}async findCaptchaAction(n){const t=getModel(this.config),o=`You are analyzing a webpage that contains a CAPTCHA challenge.\n\nPage snapshot:\n${n.snapshot}\n\nFind the primary interactive captcha element — the checkbox, button, or iframe the user should click to begin solving (e.g. "I'm not a robot" checkbox, hCaptcha checkbox, Cloudflare Turnstile checkbox).\n\nRespond with JSON:\n{ "found": true, "ref": "@e5", "description": "reCAPTCHA I'm not a robot checkbox" }\n\nIf no clickable captcha element is visible: { "found": false }`;try{const n=(await e({model:t,system:o,prompt:"Find the captcha element to click.",maxTokens:200})).text.match(/\{[\s\S]*\}/);if(!n)return null;const i=JSON.parse(n[0]);return i.found&&i.ref?{ref:i.ref,description:i.description||"captcha"}:null}catch{return null}}}