explorbot 0.1.32-beta.1 → 0.2.0

package/dist/src/ai/researcher/deep-analysis.js

@@ -74,7 +74,6 @@ export function WithDeepAnalysis(Base) {
                 return null;
             }
             const diff = await current.diff(previous);
-            await diff.calculate();
             if (!diff.ariaChanged && diff.htmlParts.length === 0) {
                 debugLog(`No diff between current and previous state for overlay "${focusArea.name}"`);
                 return null;
@@ -312,12 +311,11 @@ export function WithDeepAnalysis(Base) {
                     const isCoordinateClick = el.commands[0].startsWith('I.clickXY(');
                     if (!isCoordinateClick) {
                         const hoverCmd = el.commands[0].replace('I.click(', 'I.moveCursorTo(');
-                        await this.explorer.attemptAction(hoverCmd, undefined, false);
+                        await this.explorer.attemptAction(hoverCmd, undefined);
                         await new Promise((r) => setTimeout(r, 500));
                         await this.explorer.capturePageState();
                         const hoverAR = ActionResult.fromState(this.stateManager.getCurrentState());
                         const hoverDiff = await hoverAR.diff(previousState);
-                        await hoverDiff.calculate();
                         const hoverHtmlSize = hoverDiff.htmlParts.reduce((sum, p) => sum + p.subtree.length, 0);
                         const hoverRevealed = hoverDiff.ariaChanged && hoverHtmlSize > 500;
                         if (hoverRevealed) {
@@ -355,7 +353,7 @@ export function WithDeepAnalysis(Base) {
             let clickCode = null;
             const action = this.explorer.createAction();
             for (const cmd of commands) {
-                if (await action.attempt(cmd, undefined, false)) {
+                if (await action.attempt(cmd, undefined)) {
                     clickCode = cmd;
                     break;
                 }
@@ -370,7 +368,6 @@ export function WithDeepAnalysis(Base) {
                 await this.explorer.createAction().capturePageState();
                 const currAR = ActionResult.fromState(this.stateManager.getCurrentState());
                 diff = await currAR.diff(previousState);
-                await diff.calculate();
             }
             catch (err) {
                 tag('warning').log(`State capture failed after click: ${err instanceof Error ? err.message : err}`);
@@ -399,7 +396,7 @@ export function WithDeepAnalysis(Base) {
                 await this.cancelInUi();
                 await this.explorer.capturePageState();
                 const currentAria = this.stateManager.getCurrentState()?.ariaSnapshot || '';
-                if (!diffAriaSnapshots(originalAria, currentAria))
+                if (!diffAriaSnapshots(originalAria, currentAria).text)
                     return;
             }
             catch (err) {

package/dist/src/ai/researcher/parser.js

@@ -69,7 +69,7 @@ export function extractContainerFromBlockquote(sectionMarkdown) {
     return css;
 }
 export function parseResearchSections(markdown) {
-    const hasExtendedResearch = markdown.includes('\n# Extended Research') || markdown.startsWith('# Extended Research');
+    const hasExtendedResearch = mdq(markdown).query('section1(~"Extended Research")').count() > 0;
     return parseSections(markdown)
         .filter((s) => !SKIP_SECTIONS.has(s.name.toLowerCase()) && !s.name.toLowerCase().includes('data:'))
         .map((section) => {

package/dist/src/ai/researcher/research-result.d.ts

@@ -9,7 +9,6 @@ export declare class ResearchResult {
     parseLocators(): void;
     get containers(): string[];
     get containerLocators(): Locator[];
-    updateSection(sectionName: string, locators: Locator[]): void;
     rebuildSectionInText(section: ResearchSection): void;
     cleanup(): void;
     reorderColumns(columns: string[]): string[];

package/dist/src/ai/researcher/research-result.js

@@ -42,25 +42,6 @@ export class ResearchResult {
             pwLocator: null,
         }));
     }
-    updateSection(sectionName, locators) {
-        const sections = parseResearchSections(this.text);
-        const section = sections.find((s) => s.name === sectionName);
-        if (!section)
-            return;
-        for (const el of section.elements) {
-            const elLocators = locators.filter((l) => l.element === el.name);
-            for (const loc of elLocators) {
-                const value = loc.valid === false ? null : loc.locator || null;
-                if (loc.type === 'css')
-                    el.css = value;
-                if (loc.type === 'xpath')
-                    el.xpath = value;
-                if (loc.type === 'aria')
-                    el.aria = value ? parseAriaLocator(value) : null;
-            }
-        }
-        this.rebuildSectionInText(section);
-    }
     rebuildSectionInText(section) {
         if (section.elements.length === 0)
             return;

package/dist/src/ai/researcher/sections.js

@@ -89,9 +89,13 @@ export function WithSections(Base) {
         </section_format>
 
         <rules>
-        - Every element with eidx MUST appear in the table.
+        - List only elements physically inside this section's declared container.
+        - Do not copy global toolbar, navigation, list, or detail elements into this section unless they are descendants of this section container.
+        - Every element with eidx inside this section's container MUST appear in the table.
         - Every row needs CSS; ARIA may be "-" for icon-only buttons.
         - ARIA locator JSON uses keys "role" and "text" (NOT "name").
+        - Elements marked data-explorbot-hit="covered" or "offscreen" are not directly actionable; describe the covering or focused UI first.
+        - In split-pane pages, entity detail panels are active detail context; include close/back/pin controls in the detail panel section when present.
         </rules>
 
         ${generalLocatorRuleText}

package/dist/src/ai/researcher.js

@@ -9,7 +9,7 @@ import { diffAriaSnapshots } from "../utils/aria.js";
 import { ErrorPageError, detectPageCondition } from "../utils/error-page.js";
 import { HooksRunner } from "../utils/hooks-runner.js";
 import { isBodyEmpty } from "../utils/html.js";
-import { createDebug, pluralize, tag } from '../utils/logger.js';
+import { createDebug, tag } from '../utils/logger.js';
 import { mdq } from "../utils/markdown-query.js";
 import { RulesLoader } from "../utils/rules-loader.js";
 import { ContextLengthError } from './provider.js';
@@ -261,12 +261,10 @@ export class Researcher extends ResearcherBase {
             if (stateHash) {
                 researchFile = saveResearch(stateHash, result.text, combinedHtml);
             }
-            const summaryMatch = result.text.match(/## Summary\s*\n+([\s\S]*?)(?=\n##|$)/i);
-            if (summaryMatch) {
-                const summaryLine = summaryMatch[1].trim().split('\n')[0].trim().slice(0, 200);
-                if (summaryLine)
-                    this.experienceTracker.updateSummary(this.actionResult, summaryLine);
-            }
+            const summaryText = mdq(result.text).query('section2(/^summary/)').query('paragraph[0]').text().trim();
+            const summaryLine = summaryText.split('\n')[0]?.trim().slice(0, 200);
+            if (summaryLine)
+                this.experienceTracker.updateSummary(this.actionResult, summaryLine);
             tag('multiline').log(formatResearchSummary(result.text, { visionUsed: this.hasScreenshotToAnalyze }));
             tag('success').log('Research complete');
             if (researchFile)
@@ -365,7 +363,10 @@ export class Researcher extends ResearcherBase {
       - Research all menus and navigational areas.
       - Ignore decorative sidebars, footer-only links, and external links.
       - Detect layout patterns: list/detail split, 2-pane, or 3-pane layouts.
-      - Every element with an eidx attribute MUST appear in the UI map — describe icon-only buttons by their visual role.
+      - In split-pane layouts, keep global toolbar, list pane, and detail pane as separate sections. Do not copy toolbar/list elements into the detail section unless they are physically inside that detail container.
+      - If the URL or page state opens an entity detail panel, describe that panel as the active detail context and include its close/back/pin controls when present.
+      - If an element has data-explorbot-hit="covered" or "offscreen", do not present it as directly actionable. Prefer the overlay, drawer, dialog, or focused section covering it, and mention what must be dismissed or revealed first.
+      - Every element with an eidx attribute MUST appear in exactly one matching UI map section — describe icon-only buttons by their visual role.
       - Every UI map row needs a CSS selector; ARIA may be "-" for icon-only buttons, CSS must never be "-".
       - ARIA locator JSON uses keys "role" and "text" (NOT "name").
       - Mark elements with likely hover interactions (title, aria-describedby, menu items with submenus) as "(hover)".
@@ -382,6 +383,7 @@ export class Researcher extends ResearcherBase {
             .join('\n')}
 
       - Sections can overlap; prefer more detailed sections over broader ones.
+      - Section tables must list only elements physically inside the declared container. If a control belongs to a global toolbar, list it only in the toolbar section.
       - Never name a section "Focus" or "Focused" — use what it contains (Detail, Modal, Form, Content, List).
       - Omit sections that are not present or not relevant.
       - Each section needs a container CSS locator; UI map CSS locators are relative to it.
@@ -402,21 +404,7 @@ export class Researcher extends ResearcherBase {
         if (!this.actionResult)
             throw new Error('actionResult is not set');
         const html = await this.actionResult.combinedHtml();
-        const knowledgeFiles = this.stateManager.getRelevantKnowledge();
-        let knowledge = '';
-        if (knowledgeFiles.length > 0) {
-            const knowledgeContent = knowledgeFiles
-                .map((k) => k.content)
-                .filter((k) => !!k)
-                .join('\n\n');
-            tag('operation').log(`Found ${knowledgeFiles.length} relevant knowledge ${pluralize(knowledgeFiles.length, 'file')} for: ${this.actionResult.url}`);
-            knowledge = `
-        <hint>
-        Here is relevant knowledge for this page:
-
-        ${knowledgeContent}
-        </hint>`;
-        }
+        const knowledge = this.explorer.getKnowledgeTracker().renderRelevantKnowledge(this.actionResult);
         const ariaSnapshot = this.actionResult.getCompactARIA();
         return dedent `
       Analyze this web page and provide a comprehensive research report in markdown format.
@@ -471,35 +459,6 @@ export class Researcher extends ResearcherBase {
 
     `;
     }
-    async textContent(state) {
-        const actionResult = ActionResult.fromState(state);
-        const html = await actionResult.combinedHtml();
-        const prompt = dedent `
-      Transform into markdown.
-      Identify headers, footers, asides, special application parts and main contant.
-      Content should be in markdown format. If it is content: tables must be tables, lists must be lists.
-      Navigation elements should be represented as standalone blocks after the content.
-      Do not summarize content, just transform it into markdown.
-      It is important to list all the content text
-      If it is link it must be linked
-      You can summarize footers/navigation/aside elements.
-      But main conteint should be kept as text and formatted as markdown based on its current markup.
-      Links to external web sites should be avoided in output.
-
-      Break down into sections:
-
-      ## Content Area
-
-      ## Navigation Area
-
-      <page_html>
-      ${html}
-      </page_html>
-    `;
-        const model = this.provider.getModelForAgent('researcher');
-        const r = await this.provider.chat([{ role: 'user', content: prompt }], model, { agentName: 'researcher', telemetryFunctionId: 'researcher.textContent' });
-        return r.text;
-    }
     getScreenshotFromState(state) {
         const actionResult = ActionResult.fromState(state);
         const image = actionResult.screenshot;
@@ -693,7 +652,7 @@ export class Researcher extends ResearcherBase {
     async cancelInUi() {
         const beforeAria = this.stateManager.getCurrentState()?.ariaSnapshot || null;
         await this.explorer.executeAction('I.clickXY(0, 0)');
-        if (diffAriaSnapshots(beforeAria, this.stateManager.getCurrentState()?.ariaSnapshot || null))
+        if (diffAriaSnapshots(beforeAria, this.stateManager.getCurrentState()?.ariaSnapshot || null).text)
             return;
         await this.explorer.executeAction(`I.pressKey('Escape')`);
     }

package/dist/src/ai/rules.d.ts

@@ -1,5 +1,6 @@
 export declare const recommendedCodeceptCommands: readonly ["I.click", "I.type", "I.fillField", "I.see", "I.seeElement"];
 export declare const locatorRule: string;
+export declare const drillLocatorRule: string;
 export declare const fileUploadRule: string;
 export declare const formRequirementsRule: string;
 export declare const protectionRule: string;

package/dist/src/ai/rules.js

@@ -1,6 +1,6 @@
 import dedent from 'dedent';
 export const recommendedCodeceptCommands = ['I.click', 'I.type', 'I.fillField', 'I.see', 'I.seeElement'];
-export const locatorRule = dedent `
+const locatorPriorityRule = dedent `
   <locator_priority>
   Use the following priority when selecting locators:
 
@@ -17,10 +17,11 @@ export const locatorRule = dedent `
      Example: '#login-btn', '[data-testid="submit"]', 'form#login input[name="email"]'
 
   4. XPath (last resort) - for complex hierarchy or when CSS can't express the path
-     Always start with //, never use positional indices like [1], [2]
+     Always start with //. Avoid positional indices like [1], [2] except as a last-resort disambiguator
      Example: '//form[@id="login"]//input[@name="email"]'
   </locator_priority>
-
+`;
+const contextSimplificationRule = dedent `
   <context_simplification>
   When container is available from UI map sections:
   - Text + container is simplest and PREFERRED: I.click('Save', '.modal')
@@ -28,7 +29,8 @@ export const locatorRule = dedent `
   - ALWAYS use context parameter unless locator is XPath or unique ID
   - No need for complex ARIA when container narrows scope sufficiently
   </context_simplification>
-
+`;
+const locatorStrategyRule = dedent `
   <disambiguation>
   When multiple elements could match the request, select based on intent:
   1. Match the context of recent actions - if filling a form, use elements in that same form
@@ -46,8 +48,9 @@ export const locatorRule = dedent `
   - Use aria-label value if present: { "role": "button", "text": "Close" } (from aria-label="Close")
   - Use title attribute if present: { "role": "button", "text": "Settings" } (from title="Settings")
   - If no accessible name exists, mark ARIA as "-" and use CSS/XPath:
-    * CSS: use partial href a[href*="settings"] or SVG icon class a:has(svg.md-icon-cog)
-    * XPath: use contains(@href,"settings") or SVG class //a[.//svg[contains(@class,"md-icon-cog")]]
+    * CSS: use partial href a[href*="settings"] or SVG icon class a:has(svg.icon-settings)
+    * XPath: use contains(@href,"settings") or SVG class //a[.//svg[contains(@class,"icon-settings")]]
+  - In inline create/edit rows, confirmation can be an icon-only control near the edited field instead of a text Save button. Anchor the locator to the same row/form as the field and target the adjacent confirm icon/control.
   - NEVER use empty text: { "role": "button", "text": "" } is INVALID and useless
 
   <good_aria_locator_example>
@@ -78,9 +81,9 @@ export const locatorRule = dedent `
   - Vue: data-v-* attributes
   Avoid locators that seem to have generated ids or class names (long random numbers, uuids, hashes, etc)
   Prefer text or ARIA locators over href-based ones. But for icon-only links with no accessible name, use:
-  - Partial href match: a[href*="settings"], a[href*="requirements"] (use path segments, not full URLs)
-  - SVG icon class: a:has(svg.md-icon-cog), button:has(svg.md-icon-plus) (target the SVG class inside the link/button)
-  Avoid full absolute href like a[href="/projects/imr_manual12/settings"] — use generic path segments instead
+  - Partial href match: a[href*="settings"], a[href*="reports"] (use path segments, not full URLs)
+  - SVG icon class: a:has(svg.icon-settings), button:has(svg.icon-add) (target the SVG class inside the link/button)
+  Avoid full absolute href like a[href="/items/12345/settings"] — use generic path segments instead
   Avoid CSS framework utility classes as containers (Tailwind: flex, grid, space-x-*, justify-*, items-*, w-*, h-*, p-*, m-*, etc; Bootstrap: col-*, row, d-flex, etc)
   Prefer semantic class names, roles, data attributes, or element hierarchy for containers
 
@@ -92,7 +95,7 @@ export const locatorRule = dedent `
 
   <xpath_rules>
   XPath locators must start with //.
-  XPath should use positional indices [1], [2], [3] and contains(., "text") for disambiguation.
+  XPath may use positional indices [1], [2] and contains(., "text") as a last-resort disambiguator when attribute/text strategies are exhausted.
   XPath should rely less on class names — prefer element hierarchy, position, and text content.
   XPath and CSS MUST provide different strategies for finding the same element.
   </xpath_rules>
@@ -104,21 +107,23 @@ export const locatorRule = dedent `
     '#content-top #user_name'
     '#content-top form input[name="name"]'
     'a.nav-item[href*="settings"]' // icon-only link matched by partial href
-    'a.nav-item:has(svg.md-icon-cog)' // icon-only link matched by SVG icon class
+    'a.nav-item:has(svg.icon-settings)' // icon-only link matched by SVG icon class
     '//nav//a[contains(@href,"settings")]' // XPath for icon-only nav link
   </good locator example>
 
   <bad locator example>
-    'a.filter-tab:nth-of-type(1)' // WRONG: positional in CSS, use :has-text("Manual") instead
-    '//a[contains(@class,"filter-tab") and contains(@class,"active")]' // WRONG: XPath repeats CSS approach, use positional //a[contains(@class,"filter-tab")][1]
+    'a.filter-tab:nth-of-type(1)' // WRONG: positional in CSS, use :has-text("Active") instead
+    '//a[contains(@class,"filter-tab") and contains(@class,"active")]' // WRONG: XPath repeats CSS approach, provide a different strategy e.g. //a[contains(@class,"filter-tab")][1]
     '//table//tbody/tr[1]//button[contains(@onclick,'fn()')]' // onclick is not semantic attribute
     '//html/body/vue-button-123' // vue-framework specific locator
-    'link "New Template"'  // WRONG: malformed string, use {"role":"link","text":"New Template"}
-    'a[href="/projects/imr_manual12/settings"]' // WRONG: full absolute href, use a[href*="settings"] instead
+    'link "New Item"'  // WRONG: malformed string, use {"role":"link","text":"New Item"}
+    'a[href="/items/12345/settings"]' // WRONG: full absolute href, use a[href*="settings"] instead
   </bad locator example>
 
   HTML locators must be valid JS strings
 `;
+export const locatorRule = [locatorPriorityRule, contextSimplificationRule, locatorStrategyRule].join('\n\n');
+export const drillLocatorRule = [locatorPriorityRule, locatorStrategyRule].join('\n\n');
 export const fileUploadRule = dedent `
   <file_upload>
   Explorbot CAN upload files using I.attachFile() via form() tool.
@@ -294,6 +299,7 @@ export const actionRule = dedent `
   Use context parameter (second argument) to narrow click area when:
   - The same text/button appears multiple times on page
   - You need to click inside a specific form, modal, or section
+  - You need an icon-only confirm/save control next to a field in an inline create/edit row
   Context should be a CSS selector pointing to a unique container.
 
   <example>
@@ -306,6 +312,7 @@ export const actionRule = dedent `
   </example>
 
   Prefer text/ARIA locators with context over complex CSS/XPath selectors.
+  For inline create/edit flows, after filling a field verify it contains the value, then confirm using the nearest explicit button/link, an adjacent icon-only confirm control in the same row/form, or Enter if the field remains focused.
   If locator doesn't work, try CSS or XPath locators.
   If nothing works, use I.clickXY(x, y) as last resort.
 

package/dist/src/ai/session-analyst.js

@@ -42,6 +42,9 @@ export class SessionAnalyst {
       ## Severity (defects only)
       [High] blocks a core flow · [Medium] degrades a flow but workaround exists · [Low] cosmetic / edge case
 
+      ## Naming tests
+      Reference every test by its full label wrapped in backticks: \`\`\`ET-<number> <test title>\`\`\` — the \`ref\` and \`scenario\` from its \`<test>\` block. Backticks only, never bold. Never write a bare \`#2\`.
+
       ## Format
 
       # Session Analysis
@@ -53,22 +56,22 @@ export class SessionAnalyst {
       - Features: <capabilities>
 
       ## What works
-      - **<feature>** — #2, #7, #8
+      - **<feature>** — \`ET-2 <title>\`, \`ET-7 <title>\`
 
       ## Defects
 
       ### [Medium] <plain-English bug title>
-      Affects: #3, #5
+      Affects: \`ET-3 <title>\`, \`ET-5 <title>\`
       Reproduce:
         1. <concrete UI step>
         2. <next>
       Evidence: <one short observation>
 
       ## UX issues
-      - **<feature>** — <what's confusing> (#7)
+      - **<feature>** — <what's confusing> (\`ET-7 <title>\`)
 
       ## Execution Issues
-      - **#2 <scenario>** — <≤10 words, what was unreliable>
+      - \`ET-2 <title>\` — <≤10 words, what was unreliable>
 
       ## Brevity rules
 
@@ -127,7 +130,7 @@ export class SessionAnalyst {
         `
             : 'verification_status: none';
         return dedent `
-      <test ref="#${ref}">
+      <test ref="ET-${ref}">
       url: ${test.startUrl || '/'}
       scenario: ${test.scenario}
       result: ${test.result || 'unknown'}

package/dist/src/ai/task-agent.d.ts

@@ -1,5 +1,5 @@
 import type { ActionResult } from '../action-result.js';
-import { type ExperienceTracker } from '../experience-tracker.js';
+import type { ExperienceTracker } from '../experience-tracker.js';
 import type { KnowledgeTracker } from '../knowledge-tracker.js';
 import { Historian } from './historian.js';
 import type { Navigator } from './navigator.js';

package/dist/src/ai/task-agent.js

@@ -1,7 +1,3 @@
-import dedent from 'dedent';
-import { renderExperienceToc } from '../experience-tracker.js';
-import { createDebug, pluralize, tag } from '../utils/logger.js';
-const debugLog = createDebug('explorbot:task-agent');
 export function isInteractive() {
     return process.env.INK_RUNNING === 'true';
 }
@@ -18,31 +14,10 @@ export class TaskAgent {
     _historian = null;
     _quartermaster = null;
     getKnowledge(actionResult) {
-        const knowledgeFiles = this.getKnowledgeTracker().getRelevantKnowledge(actionResult);
-        if (knowledgeFiles.length === 0)
-            return '';
-        const knowledgeContent = knowledgeFiles
-            .map((k) => k.content)
-            .filter((k) => !!k)
-            .join('\n\n');
-        tag('operation').log(`Found ${knowledgeFiles.length} relevant knowledge ${pluralize(knowledgeFiles.length, 'file')}`);
-        return dedent `
-      <knowledge>
-      Here is relevant knowledge for this page:
-
-      ${knowledgeContent}
-      </knowledge>
-    `;
+        return this.getKnowledgeTracker().renderRelevantKnowledge(actionResult);
     }
     getExperience(actionResult) {
-        const tracker = this.getExperienceTracker();
-        const toc = tracker.getExperienceTableOfContents(actionResult);
-        if (toc.length === 0)
-            return '';
-        const totalSections = toc.reduce((sum, entry) => sum + entry.sections.length, 0);
-        debugLog(`injecting experience TOC (${toc.length} files, ${totalSections} sections)`);
-        tag('operation').log(`Found ${toc.length} experience ${pluralize(toc.length, 'file')} (${totalSections} sections)`);
-        return renderExperienceToc(toc);
+        return this.getExperienceTracker().renderExperienceTocFor(actionResult);
     }
     setHistorian(historian) {
         this._historian = historian;

package/dist/src/ai/tester.d.ts

@@ -26,7 +26,6 @@ export declare class Tester extends TaskAgent implements Agent {
     researcher: Researcher;
     navigator: Navigator;
     agentTools: any;
-    executionLogFile: string | null;
     previousUrl: string | null;
     previousStateHash: string | null;
     pageStateHash: string | null;
@@ -56,7 +55,6 @@ export declare class Tester extends TaskAgent implements Agent {
     shouldStopForStalledExecution(task: Test, previousState: ActionResult, toolExecutions: any[]): boolean;
     prepareInstructionsForNextStep(task: Test): Promise<string>;
     reinjectContextIfNeeded(iteration: number, currentState: ActionResult): Promise<string>;
-    promptLogStep(task: Test): Promise<string>;
     finishTest(task: Test): void;
     abortStartedTestOnErrorPage(task: Test, actionResult: ActionResult): Promise<{
         success: boolean;

package/dist/src/ai/tester.js

@@ -5,7 +5,6 @@ import dedent from 'dedent';
 import { z } from 'zod';
 import { ActionResult } from "../action-result.js";
 import { clearActivity, setActivity } from "../activity.js";
-import { ConfigParser } from "../config.js";
 import { Observability } from "../observability.js";
 import { Stats } from "../stats.js";
 import { TestResult } from "../test-plan.js";
@@ -16,7 +15,7 @@ import { createDebug, tag } from "../utils/logger.js";
 import { loop } from "../utils/loop.js";
 import { actionRule, capabilityGroundingRule, dataProtectionRules, focusedElementRule, formRequirementsRule, locatorRule, multipleTabsRule, sectionContextRule } from "./rules.js";
 import { TaskAgent } from "./task-agent.js";
-import { createCodeceptJSTools, createSpecialContextTools } from "./tools.js";
+import { createCodeceptJSTools, createIframeTools } from "./tools.js";
 const debugLog = createDebug('explorbot:tester');
 const SAMPLE_FILES_DIR = resolve(dirname(fileURLToPath(import.meta.url)), '../../assets/sample-files');
 const SAMPLE_FILES = {
@@ -43,7 +42,6 @@ export class Tester extends TaskAgent {
     researcher;
     navigator;
     agentTools;
-    executionLogFile = null;
     previousUrl = null;
     previousStateHash = null;
     pageStateHash = null;
@@ -120,9 +118,6 @@ export class Tester extends TaskAgent {
         const conversation = this.provider.startConversation(this.getSystemMessage(), 'tester');
         conversation.markLastMessageCacheable();
         this.currentConversation = conversation;
-        const outputDir = ConfigParser.getInstance().getOutputDir();
-        this.executionLogFile = join(outputDir, `tester_${task.sessionName}.md`);
-        // Note: Markdown saving functionality removed from Conversation class
         const scenarioBlock = this.buildScenarioBlock(task, initialState);
         conversation.addUserText(scenarioBlock);
         conversation.markLastMessageCacheable();
@@ -217,10 +212,12 @@ export class Tester extends TaskAgent {
                 const currentState = this.getCurrentState();
                 const tools = {
                     ...codeceptjsTools,
-                    ...(currentState.isInsideIframe ? createSpecialContextTools(this.explorer, 'iframe') : {}),
                     ...this.createTestFlowTools(task, currentState, conversation),
                     ...this.agentTools,
                 };
+                if (currentState.isInsideIframe) {
+                    Object.assign(tools, createIframeTools(this.explorer));
+                }
                 debugLog(`Test ${task.scenario} iteration ${iteration}`);
                 if (this.explorer.getStateManager().isInDeadLoop()) {
                     task.addNote('Dead loop detected. Stopped');
@@ -456,7 +453,6 @@ export class Tester extends TaskAgent {
         const currentUrl = currentState.url;
         const currentStateHash = currentState.hash;
         const isNewUrl = this.previousUrl !== currentUrl;
-        const isStateChanged = !isNewUrl && this.previousStateHash !== currentStateHash;
         this.previousUrl = currentUrl;
         this.previousStateHash = currentStateHash;
         let context = '';
@@ -564,23 +560,6 @@ export class Tester extends TaskAgent {
         `;
             }
         }
-        // if (isStateChanged) {
-        //   const combinedHtml = await currentState.combinedHtml();
-        //   context += dedent`
-        //     Context (state changed):
-        //     <page>
-        //     CURRENT URL: ${currentState.url}
-        //     CURRENT TITLE: ${currentState.title}
-        //     </page>
-        //     <page_html>
-        //     ${combinedHtml}
-        //     </page_html>
-        //     <page_aria>
-        //     ${currentState.ariaSnapshot}
-        //     </page_aria>
-        //   `;
-        //   return context;
-        // }
         if (context)
             return context;
         if (iteration % 5)
@@ -598,34 +577,6 @@ export class Tester extends TaskAgent {
       </page_aria>
     `;
     }
-    async promptLogStep(task) {
-        let logPrompt = dedent `
-      <task>
-        Add a note explaining what you achieved with previous action.
-        Use tools to interact with the page to achieve the scenario goal or expected outcomes.
-        Call record tool to explain the last action
-        Format: record([<action performed>, <what has changed>, <what you expect to do next>]) 
-      </task>
-    `;
-        if (task.getPrintableNotes()) {
-            logPrompt = dedent `
-        Your interaction log notes:
-        <notes>
-        ${task.getPrintableNotes()}
-        </notes>
-
-        <rules>
-        Use your previous interaction notes to guide your next actions.
-        Do not perform the same checks.
-        </rules>
-      `;
-        }
-        const remaining = task.getRemainingExpectations();
-        if (remaining.length > 0) {
-            logPrompt += `\nExpected steps to check: ${remaining.join(', ')}`;
-        }
-        return logPrompt;
-    }
     finishTest(task) {
         if (!task.hasFinished) {
             task.finish(TestResult.FAILED);
@@ -725,11 +676,13 @@ export class Tester extends TaskAgent {
     - Before retrying your actions check maybe they already achived expected results. Use see() tool for that
     - If the current URL is already a create/edit/new form and the scenario is about creating/editing that entity, fill and submit that form. Do not click the list-page "New" button again from inside the form.
     - If the scenario is about search/filter/sort/tabs/list inspection and the current URL is a create/edit/new form, go back or reset to the stable list page before interacting with list controls.
-    - When selecting related entities from a list, do not choose rows/options/cards marked as "0 items", "0 tests", or otherwise empty if the scenario requires selecting real content.
-    - In selection pickers, counters such as "Selected 0", "Matched tests 0", or disabled Save/Apply mean the selection did not register. Choose a non-empty item or change filters before submitting.
+    - When selecting related entities from a list, do not choose rows/options/cards marked as "0 items", "0 results", or otherwise empty if the scenario requires selecting real content.
+    - In selection pickers, counters such as "Selected 0", "Matched 0", or disabled Save/Apply mean the selection did not register. Choose a non-empty item or change filters before submitting.
     - A passed form/click command only means the command executed. If a required field remains empty, submit stays disabled, or the expected text is not visible, treat the action as not completed and correct the missing field/state.
-    - For filter/tab scenarios, success requires BOTH: the requested filter/tab is visibly active/selected AND the list content matches that filter. Do not finish from only one of these signals.
-    - Empty-state text such as "No matched items" only proves a filter when the requested filter/tab is active and the empty state belongs to the filtered list.
+    - For filter/tab scenarios, success requires BOTH: the requested state is evidenced by a selected control, URL/query, or another explicit state indicator AND the list content matches that state. Do not finish from only one of these signals.
+    - Once the requested control state and matching content are both visible, finish the scenario instead of repeating the interaction. Do not require an aggregate count change unless a baseline was observed immediately before the action.
+    - Empty-state text such as "No matched items" only proves a filter when the requested filter state is explicit and the empty state belongs to the filtered list.
+    - Associate validation feedback with a field only through explicit evidence such as the field label in the message, an accessibility relationship, focus on the invalid control, or visual confirmation. Do not infer the affected field from DOM order or proximity alone.
     - When filling complex form with lot of actions performed, use see() to look which fields were filled and which are not
     - When verify() fails, use see() to visually confirm the result — visual confirmation is equally valid evidence
     - For visual state verification (active tabs, selected items, counts, colors), prefer see() over DOM-based verify()
@@ -788,10 +741,6 @@ export class Tester extends TaskAgent {
       When creating or editing items via form() or type() you should include ${task.sessionName} in the value (if it is not restricted by the application logic)
       Initial page URL: ${actionResult.url}
 
-      ${capabilityGroundingRule}
-
-      ${dataProtectionRules}
-
       ${this.buildDeletionScope(task)}
 
       ${this.buildAvailableFiles()}

package/dist/src/ai/tools.d.ts

@@ -1,4 +1,4 @@
-import { ActionResult } from '../action-result.js';
+import { ActionResult, type PageDiff } from '../action-result.js';
 import type { ExperienceTracker } from '../experience-tracker.js';
 import type Explorer from '../explorer.js';
 import { type Task } from '../test-plan.js';
@@ -24,18 +24,29 @@ export declare function createCodeceptJSTools(explorer: Explorer, task: Task): {
         explanation: any;
     }, Record<string, any>, import("@ai-sdk/provider-utils").Context>>;
 };
-export declare function createSpecialContextTools(explorer: Explorer, context: 'iframe'): {
-    exitIframe?: undefined;
-} | {
+export declare function createIframeTools(explorer: Explorer): {
     exitIframe: import("@ai-sdk/provider-utils").ExecutableTool<import("ai").Tool<{
         reason?: string;
     }, Record<string, any>, import("@ai-sdk/provider-utils").Context>>;
 };
-export declare function createAgentTools({ explorer, researcher, navigator, experienceTracker, getState, supervisor, }: {
+export declare function createLearnExperienceTool({ getExperienceTracker, getState }: {
+    getExperienceTracker: () => ExperienceTracker;
+    getState: () => ActionResult | null;
+}): import("@ai-sdk/provider-utils").ExecutableTool<import("ai").Tool<{
+    fileTag: any;
+    sectionIndex: any;
+}, {
+    title: string;
+    url: string;
+    content: string;
+} | {
+    error: string;
+}, import("@ai-sdk/provider-utils").Context>>;
+export declare function createAgentTools({ explorer, researcher, navigator, supervisor, withExperience, }: {
     explorer: Explorer;
     researcher: Researcher;
     navigator: Navigator;
-    experienceTracker?: ExperienceTracker;
-    getState?: () => ActionResult | null;
     supervisor?: boolean;
+    withExperience?: boolean;
 }): any;
+export declare function isMajorPageChange(pageDiff: PageDiff): boolean;