explorbot 0.1.27 → 0.1.29

package/dist/src/ai/fisherman.js

@@ -4,6 +4,7 @@ import { createDebug, tag } from "../utils/logger.js";
 const debugLog = createDebug('explorbot:fisherman');
 import { loop } from "../utils/loop.js";
 import { createFishermanTools } from "./fisherman-tools.js";
+import { dataProtectionRules } from "./rules.js";
 const MAX_ITERATIONS = 15;
 const MAX_TOOL_ROUNDTRIPS = 5;
 export class Fisherman {
@@ -65,7 +66,7 @@ export class Fisherman {
             spec: this.spec,
             baseEndpoint: this.baseEndpoint,
         });
-        const conversation = this.provider.startConversation(this.buildSystemPrompt(endpointList, scopeUrl), 'fisherman');
+        const conversation = this.provider.startConversation(this.buildSystemPrompt(endpointList, Object.keys(tools), scopeUrl), 'fisherman');
         conversation.addUserText(this.buildTaskPrompt(instructions));
         await loop(async ({ stop, iteration }) => {
             debugLog(`iteration ${iteration}`);
@@ -149,7 +150,7 @@ export class Fisherman {
         }
         return lines.join('\n');
     }
-    buildSystemPrompt(endpointList, scopeUrl) {
+    buildSystemPrompt(endpointList, toolNames, scopeUrl) {
         const scopeBlock = scopeUrl ? `\n\nSCOPE: You are operating within ${scopeUrl}.\nAll created items must belong to this scope.` : '';
         return dedent `
       You are Fisherman — a data preparation agent. You create test data by making API requests.
@@ -158,6 +159,11 @@ export class Fisherman {
       ${endpointList}
       ${scopeBlock}
 
+      AVAILABLE TOOLS:
+      ${toolNames.join(', ')}.
+      Use tool names exactly as listed. Do not invent aliases, combined names, or names with channel markers such as "commentary".
+      Match each tool input schema exactly. Do not invent parameter names or pass extra fields.
+
       WORKFLOW:
       1. Call getEndpointSpec to see the request body example for the endpoint
       2. Make requests — the response automatically extracts IDs, names, and status fields
@@ -169,6 +175,8 @@ export class Fisherman {
       - Chain requests logically — create parent resources before children
       - If a request fails, try once more with adjusted data before reporting failure
       - Use realistic but unique data for each item (vary names, titles)
+
+      ${dataProtectionRules}
     `;
     }
     buildTaskPrompt(instructions) {
@@ -177,7 +185,10 @@ export class Fisherman {
 
       ${instructions}
 
-      Execute the necessary API requests to create this data. When done, call finish with the summary.
+      ${dataProtectionRules}
+
+      If data preparation is allowed by these rules, execute the necessary API requests to create this data.
+      When done, call finish with the summary. If data preparation is forbidden, call stop with the reason.
     `;
     }
 }

package/dist/src/ai/historian/screencast.js

@@ -7,6 +7,7 @@ import { tag } from "../../utils/logger.js";
 import { relativeToCwd } from "../../utils/next-steps.js";
 import { safeFilename } from "../../utils/strings.js";
 import { debugLog } from "./mixin.js";
+const FATAL_SCREENCAST_STOP_ERRORS = /Target page, context or browser has been closed|Target closed|Session closed|Protocol error/i;
 export function WithScreencast(Base) {
     return class extends Base {
         screencastPage = null;
@@ -100,18 +101,26 @@ export function WithScreencast(Base) {
                 return;
             const path = this.screencastPath;
             const task = this.screencastTask;
+            let stopped = false;
             try {
                 await this.screencastPage.screencast.stop();
+                stopped = true;
             }
             catch (err) {
-                tag('operation').log(`Screencast stop failed: ${err.message}`);
+                const message = err.message;
+                if (FATAL_SCREENCAST_STOP_ERRORS.test(message)) {
+                    tag('operation').log('Screencast skipped: browser was closed before recording could be finalized');
+                }
+                else {
+                    tag('operation').log(`Screencast stop failed: ${message}`);
+                }
             }
             this.screencastActive = false;
             this.screencastPage = null;
             this.screencastPath = null;
             this.screencastTask = null;
             this.screencastLastChapter = null;
-            if (path) {
+            if (path && stopped) {
                 this.savedFiles.add(path);
                 task?.addArtifact?.(path);
                 tag('operation').log(`Saved screencast: ${relativeToCwd(path)}`);

package/dist/src/ai/navigator.js

@@ -121,6 +121,9 @@ class Navigator {
         return normalizeUrl(currentUrl) === normalizeUrl(expectedUrl);
     }
     async visit(url) {
+        return this.explorer.runWithBrowserRecovery('navigator.visit', () => this.visitOnce(url));
+    }
+    async visitOnce(url) {
         try {
             const action = this.explorer.createAction();
             await action.execute(`I.amOnPage('${url}')`);
@@ -150,7 +153,7 @@ class Navigator {
                     throw new Error(`Navigation to ${url} failed: ${action.lastError?.message}`);
                 }
             }
-            await action.caputrePageWithScreenshot();
+            await this.explorer.capturePageWithScreenshot();
             await this.hooksRunner.runAfterHook('navigator', url);
         }
         catch (error) {
@@ -349,7 +352,7 @@ class Navigator {
                         // URL did not transition to expectedUrl within timeout
                     }
                 }
-                const freshState = await action.capturePageState();
+                const freshState = await this.explorer.capturePageState();
                 const currentUrl = /^https?:\/\//i.test(expectedUrl) ? freshState.fullUrl || freshState.url || '' : freshState.url || '';
                 const urlMatches = this.isSameExpectedOrigin(expectedUrl, action.stateManager) && normalizeUrl(currentUrl) === normalizeUrl(expectedUrl);
                 const stateChanged = freshState.getStateHash() !== actionResult.getStateHash();

package/dist/src/ai/pilot.js

@@ -10,6 +10,7 @@ import { ErrorPageError } from "../utils/error-page.js";
 import { createDebug, tag } from "../utils/logger.js";
 const debugLog = createDebug('explorbot:pilot');
 import { truncateJson } from "../utils/strings.js";
+import { capabilityGroundingRule, dataProtectionRules } from "./rules.js";
 import { isInteractive } from "./task-agent.js";
 const CHECK_TOOLS = ['verify', 'see', 'research', 'context'];
 const META_TOOLS = ['record', 'reset', 'stop', 'finish'];
@@ -50,7 +51,7 @@ export class Pilot {
         return this.reviewDecision('finish', task, currentState, testerConversation, navigator);
     }
     async reviewCompletion(task, currentState, testerConversation, navigator) {
-        const verdictType = task.hasAchievedAny() ? 'finish' : 'stop';
+        const verdictType = this.hasCompletionEvidence(task, currentState, testerConversation) ? 'finish' : 'stop';
         return this.reviewDecision(verdictType, task, currentState, testerConversation, navigator);
     }
     async finalReview(task, currentState, testerConversation, navigator) {
@@ -67,13 +68,13 @@ export class Pilot {
         tag('substep').log(`Pilot reviewing ${type} verdict...`);
         const sessionLog = this.formatSessionLog(testerConversation);
         const stateContext = this.buildStateContext(currentState);
+        const successfulAssertions = this.formatSuccessfulAssertions(currentState, testerConversation);
         const notes = task.notesToString() || 'No notes recorded.';
         let visualAnalysis = '';
         let screenshotState = null;
-        if (this.provider.hasVision()) {
+        if (type === 'finish' && this.provider.hasVision()) {
             try {
-                const action = this.explorer.createAction();
-                screenshotState = await action.caputrePageWithScreenshot();
+                screenshotState = await this.explorer.capturePageWithScreenshot();
                 if (screenshotState.screenshot) {
                     visualAnalysis = (await this.researcher.answerQuestionAboutScreenshot(screenshotState, `Describe current page state relevant to: ${task.scenario}`)) || '';
                 }
@@ -102,6 +103,10 @@ export class Pilot {
 
       ${this.formatExpectations(task)}
 
+      <successful_assertions>
+      ${successfulAssertions || 'None'}
+      </successful_assertions>
+
       <notes>
       ${notes}
       </notes>
@@ -135,7 +140,7 @@ export class Pilot {
         try {
             const response = await this.provider.generateObject(messages, schema, this.provider.getAgenticModel('pilot'), {
                 agentName: 'pilot',
-                experimental_telemetry: { functionId: 'pilot.reviewVerdict' },
+                telemetry: { functionId: 'pilot.reviewVerdict' },
             });
             const result = response?.object;
             if (!result) {
@@ -225,7 +230,7 @@ export class Pilot {
         try {
             const response = await this.provider.generateObject(messages, schema, this.provider.getAgenticModel('pilot'), {
                 agentName: 'pilot',
-                experimental_telemetry: { functionId: 'pilot.reviewReset' },
+                telemetry: { functionId: 'pilot.reviewReset' },
             });
             const result = response?.object;
             if (!result) {
@@ -329,6 +334,8 @@ export class Pilot {
       You are Pilot — final decision maker for test pass/fail. Tester requested ${type}. Review the
       evidence and commit to a verdict; "continue" only when evidence is genuinely insufficient.
 
+      ${capabilityGroundingRule}
+
       ${this.buildSharedEvidenceRules(task)}
 
       DECISION:
@@ -337,6 +344,8 @@ export class Pilot {
         Pick assertions DOM can express; for non-DOM regions (iframes, canvas, Monaco/CodeMirror), target a
         stable landmark (container, ARIA role) instead of literal inner text. Your "pass" stands even if the
         DOM assertion can't be made.
+        Do not pass when Tester achieved only a related navigation/filter/tab/status outcome instead of the
+        requested action, workflow, or entity detail goal.
       - "fail": scenario was attempted but the goal was not achieved.
       - "skipped": scenario is irrelevant to the app, OR systematic infrastructure failures (LLM errors,
         crashes) prevented testing. NOT for "test failed to interact" — that's "fail" or "continue".
@@ -370,6 +379,10 @@ export class Pilot {
 
         FIRST: Decide if precondition() is needed.
 
+        ${capabilityGroundingRule}
+
+        ${dataProtectionRules}
+
         Call precondition() WHEN:
         - The scenario edits/deletes/modifies an item, and you want a DISPOSABLE item to act on safely
         - The scenario needs specific data clearly NOT on the current page (e.g., items with specific statuses for filtering)
@@ -494,7 +507,7 @@ export class Pilot {
             maxToolRoundtrips: opts.maxToolRoundtrips ?? 0,
             agentName: 'pilot',
             stopWhen: opts.task ? () => opts.task.hasFinished : undefined,
-            experimental_telemetry: { functionId },
+            telemetry: { functionId },
         });
         const text = result?.response?.text || '';
         const learned = (result?.toolExecutions || []).filter((e) => e.toolName === 'learnExperience' && e.output?.content).map((e) => e.output.content);
@@ -587,8 +600,7 @@ export class Pilot {
     async checkDataAvailability(task, requestedData, fishermanReason) {
         if (!this.provider.hasVision())
             return null;
-        const action = this.explorer.createAction();
-        const screenshotState = await action.caputrePageWithScreenshot().catch(() => null);
+        const screenshotState = await this.explorer.capturePageWithScreenshot().catch(() => null);
         if (!screenshotState?.screenshot)
             return null;
         const question = dedent `
@@ -791,6 +803,31 @@ export class Pilot {
         }
         return parts.join('\n\n');
     }
+    hasCompletionEvidence(task, currentState, testerConversation) {
+        if (task.hasAchievedAny())
+            return true;
+        return this.hasSuccessfulCheckEvidence(currentState, testerConversation);
+    }
+    hasSuccessfulCheckEvidence(currentState, testerConversation) {
+        if (Object.values(currentState.verifications ?? {}).some(Boolean))
+            return true;
+        return testerConversation.getToolExecutions().some((t) => CHECK_TOOLS.includes(t.toolName) && t.wasSuccessful);
+    }
+    formatSuccessfulAssertions(currentState, testerConversation) {
+        const lines = [];
+        for (const [assertion, passed] of Object.entries(currentState.verifications ?? {})) {
+            if (passed)
+                lines.push(`PASS state verification: ${assertion}`);
+        }
+        for (const exec of testerConversation.getToolExecutions()) {
+            if (!CHECK_TOOLS.includes(exec.toolName) || !exec.wasSuccessful)
+                continue;
+            const description = exec.input?.assertion || exec.input?.request || truncateJson(exec.input);
+            const result = exec.output?.message || exec.output?.analysis || exec.output?.result;
+            lines.push(`PASS ${exec.toolName}: ${description}${result ? ` -> ${result}` : ''}`);
+        }
+        return [...new Set(lines)].join('\n');
+    }
     formatActions(toolCalls) {
         return toolCalls
             .map((t) => {
@@ -902,9 +939,13 @@ export class Pilot {
 
       Tester tools: click, pressKey, form, see, verify, context, research, xpathCheck, visualClick,
       back, getVisitedStates, reset, stop, finish, record.
+      Use tool names exactly as listed. Do not invent combined names, aliases, or names with channel markers such as "commentary".
+
+      ${capabilityGroundingRule}
 
       YOUR Pilot-only tool: precondition(description) — create FRESH disposable test data via API. Never
       request users. Use when:
+
       - Scenario edits/deletes/modifies an item → create a disposable target ("1 post").
       - Scenario needs auxiliary data (labels, categories, statuses for filtering).
       - Tester failed because required data is missing (empty dropdown, empty list).
@@ -914,6 +955,8 @@ export class Pilot {
       - Current page already shows the exact data needed.
       - Scenario tests navigation, search UI, or viewing.
 
+      ${dataProtectionRules}
+
       Describe WHAT to create, not what exists. RIGHT: precondition("1 test"). WRONG:
       precondition("1 test suite named Updated Suite with existing tests"). Keep descriptions short.
 

package/dist/src/ai/planner.js

@@ -19,7 +19,7 @@ import { WithSubPages, getPlannedByStateHash, getRegisteredPlan, registerPlan }
 import { POSSIBLE_SECTIONS, Researcher } from "./researcher.js";
 import { findSimilarStateHash } from "./researcher/cache.js";
 import { hasFocusedSection } from "./researcher/focus.js";
-import { fileUploadRule, protectionRule } from "./rules.js";
+import { capabilityGroundingRule, dataProtectionRules, fileUploadRule } from "./rules.js";
 const debugLog = createDebug('explorbot:planner');
 const TasksSchema = z.object({
     planName: z.string().describe('Short descriptive name for the test plan (e.g., "User Authentication Testing", "Product Catalog Navigation", "Form Validation Tests")'),
@@ -27,7 +27,7 @@ const TasksSchema = z.object({
         .array(z.object({
         scenario: z.string().describe('A single sentence describing what to test'),
         priority: z.enum(['critical', 'important', 'high', 'normal', 'low']).describe('Priority of the task based on business importance'),
-        startUrl: z.string().nullable().describe('Start URL for the test if different from plan URL (only for tests on visited subpages)'),
+        startUrl: z.string().nullable().describe('Start URL for the test if different from plan URL. Use only stable feature/list/detail pages, not transient create/edit/modal URLs unless the scenario specifically starts inside that form.'),
         steps: z.array(z.string()).describe('List of steps to perform for this scenario. Each step should be a specific action (e.g., "Open the form", "Enter required data", "Submit the form"). Keep steps atomic and actionable.'),
         expectedOutcomes: z
             .array(z.string())
@@ -73,6 +73,9 @@ export class Planner extends PlannerBase {
         const featureDirective = feature
             ? `\n    IMPORTANT: The user requested to focus specifically on: "${feature}"\n    ALL scenarios MUST be directly related to this feature. Do not propose generic page tests unrelated to it.\n    Use the user's exact wording to guide scenario names — do not substitute different entities (e.g., do not plan "suite" actions when user said "test").`
             : '';
+        const focusExistingDataDirective = feature
+            ? '\n    If this focus asks for search, filter, tabs, sorting, or list behavior involving existing items, only use item names/values visible in the provided page research. If no concrete visible item names/values are present, do NOT propose scenarios that require an existing known item; propose no-match search, empty-state, clear-search, tab/filter empty-list, or other read-only list behavior instead.'
+            : '';
         return dedent `
     <role>
     You are ISTQB certified senior manual QA planning exploratory testing session of a web application.
@@ -96,7 +99,7 @@ export class Planner extends PlannerBase {
       Bad: "Open delete dropdown" + "Confirm deletion" — these are ONE test, not two.
       Bad: "Search for X" + "Verify search results" — searching and verifying is ONE test.
       Bad: "Leave field empty" + "Click submit" — that's one negative test, not two.
-      If two scenarios cannot run independently (one requires the other to run first), merge them into one.${featureDirective}
+      If two scenarios cannot run independently (one requires the other to run first), merge them into one.${featureDirective}${focusExistingDataDirective}
     </task>
 
     ${customPrompt || ''}
@@ -301,6 +304,11 @@ export class Planner extends PlannerBase {
       If a scenario needs existing records, recipients, results, notifications, or other target data, propose it only when that data is visible or API preconditions can create it.
       If the page appears read-only, degraded, demo-limited, maintenance-like, or lacks write controls, prefer read-only scenarios such as opening panels, inspecting visible lists, filtering, searching, or verifying current state.
       Do not assume hidden data exists just because a control is present.
+      For scenarios that act on existing items or search/filter by existing values, use only item names or values visible in research, visited pages, or prior observed flows.
+      If the list is empty or no concrete item names are visible, do not invent "known" or "existing" items. Prefer empty-state, no-match search, clear-search, or read-only list behavior scenarios.
+      Search, filter, sorting, tab, and list scenarios must start from a stable page where those controls are visible; avoid transient create/edit/new URLs unless the scenario tests that form.
+      For option values and list items, use only visible or previously observed data; do not add create/update/delete setup unless the user explicitly requests that workflow.
+      Detail-view scenarios must target visible data entities from list rows, cards, tree nodes, or detail links; do not use filter tabs, counters, status tabs, breadcrumbs, or navigation controls as detail targets.
       DO NOT propose "verification-only" tests that merely open a UI element (modal, dropdown, panel) and check it exists.
       Every test must complete a meaningful action that changes application state or produces a business outcome.
       Opening a modal is NOT a test — performing an action INSIDE the modal IS a test.
@@ -311,7 +319,8 @@ export class Planner extends PlannerBase {
       Tests that only switch views, toggle filters, or paginate are LESS valuable — propose them only after data-changing tests are covered.
       If multiple ways to create or modify data exist (different types, different forms), propose a separate test for each.
       </priority_order>
-      ${protectionRule}
+      ${capabilityGroundingRule}
+      ${dataProtectionRules}
       ${fileUploadRule}
       </rules>
 
@@ -466,7 +475,9 @@ export class Planner extends PlannerBase {
                 .join('\n')}
 
         You MAY propose tests starting from these pages if they are relevant to the plan "${this.currentPlan.title}".
-        Set startUrl for such tests. Ignore pages that belong to a different feature area.
+        Set startUrl for such tests only when the page is a stable feature/list/detail page.
+        Do not use create/edit/new/modal URLs as startUrl for scenarios that need the underlying page.
+        Ignore pages that belong to a different feature area.
         </context_from_previous_tests>
 
         Propose ONLY new scenarios that are NOT in the existing tests list.

package/dist/src/ai/provider.js

@@ -1,6 +1,7 @@
+import { OpenTelemetry } from '@ai-sdk/otel';
 import { LangfuseSpanProcessor } from '@langfuse/otel';
 import { NodeSDK } from '@opentelemetry/sdk-node';
-import { generateObject, generateText, stepCountIs } from 'ai';
+import { generateObject, generateText, isStepCount, registerTelemetry } from 'ai';
 import { clearActivity, setActivity } from "../activity.js";
 import { executionController } from "../execution-controller.js";
 import { Observability } from "../observability.js";
@@ -16,10 +17,11 @@ class AiError extends Error {
 }
 export class ContextLengthError extends Error {
 }
+let telemetryRegistered = false;
 function extractCachedTokens(usage) {
     if (!usage)
         return 0;
-    const direct = usage.cachedInputTokens ?? usage.inputTokenDetails?.cacheReadTokens;
+    const direct = usage.inputTokenDetails?.cacheReadTokens ?? usage.cachedInputTokens;
     if (typeof direct === 'number')
         return direct;
     const raw = usage.raw;
@@ -78,7 +80,7 @@ export class Provider {
             await generateText({
                 model: this.config.model,
                 prompt: 'hi',
-                maxTokens: 1,
+                maxOutputTokens: 1,
             });
         }
         catch (error) {
@@ -100,6 +102,18 @@ export class Provider {
         }
         return this.config.agenticModel || this.config.model;
     }
+    getConfiguredModels() {
+        const models = { model: this.getModelName(this.config.model) };
+        if (this.config.agenticModel)
+            models.agenticModel = this.getModelName(this.config.agenticModel);
+        if (this.config.visionModel)
+            models.visionModel = this.getModelName(this.config.visionModel);
+        for (const [agent, agentConfig] of Object.entries(this.config.agents || {})) {
+            if (agentConfig?.model)
+                models[agent] = this.getModelName(agentConfig.model);
+        }
+        return models;
+    }
     getSystemPromptForAgent(agentName, currentUrl) {
         const agentConfig = this.config.agents?.[agentName];
         const parts = [];
@@ -116,6 +130,12 @@ export class Provider {
         const agentConfig = this.config.agents?.[agentName];
         return agentConfig?.providerOptions;
     }
+    getReasoningForAgent(agentName) {
+        if (!agentName)
+            return undefined;
+        const agentConfig = this.config.agents?.[agentName];
+        return agentConfig?.reasoning;
+    }
     getRetryOptions(options = {}) {
         return {
             ...this.defaultRetryOptions,
@@ -133,6 +153,13 @@ export class Provider {
             providerOptions: { ...config.providerOptions, ...agentOptions },
         };
     }
+    finalizeConfig(config, options, telemetry) {
+        if (telemetry)
+            config.telemetry = telemetry;
+        const reasoning = this.getReasoningForAgent(options.agentName);
+        if (reasoning)
+            config.reasoning ??= reasoning;
+    }
     initLangfuse() {
         const langfuseConfig = this.config.langfuse;
         const publicKey = langfuseConfig?.publicKey || process.env.LANGFUSE_PUBLIC_KEY;
@@ -152,6 +179,10 @@ export class Provider {
             instrumentations: [],
         });
         void this.otelSdk.start();
+        if (!telemetryRegistered) {
+            registerTelemetry(new OpenTelemetry());
+            telemetryRegistered = true;
+        }
         this.telemetryEnabled = true;
     }
     getTelemetry(options) {
@@ -159,18 +190,18 @@ export class Provider {
             return undefined;
         }
         const runTelemetry = Observability.getTelemetry();
-        if (!options.experimental_telemetry) {
-            return runTelemetry || { isEnabled: true };
+        if (!options.telemetry) {
+            return runTelemetry;
         }
         if (!runTelemetry) {
-            return options.experimental_telemetry;
+            return options.telemetry;
         }
         return {
             ...runTelemetry,
-            ...options.experimental_telemetry,
+            ...options.telemetry,
             metadata: {
                 ...runTelemetry.metadata,
-                ...options.experimental_telemetry.metadata,
+                ...options.telemetry.metadata,
             },
         };
     }
@@ -210,13 +241,14 @@ export class Provider {
         promptLog(`Using model: ${modelName}`);
         const telemetry = this.getTelemetry(options);
         const config = this.mergeProviderOptions({
-            maxTokens: 16384,
+            maxOutputTokens: 16384,
+            allowSystemInMessages: true,
             ...(this.config.config || {}),
             ...options,
-            ...(telemetry ? { experimental_telemetry: telemetry } : {}),
             model,
             abortSignal: executionController.getAbortSignal(),
         }, options.agentName);
+        this.finalizeConfig(config, options, telemetry);
         promptLog(messages[messages.length - 1].content);
         try {
             const response = await withRetry(async () => {
@@ -224,7 +256,7 @@ export class Provider {
                 if (!result.text) {
                     debugLog(result);
                     if (result.finishReason === 'length') {
-                        throw new ContextLengthError('AI response empty: output truncated at maxTokens. Increase maxTokens in config or use a model with higher output capacity.');
+                        throw new ContextLengthError('AI response empty: output truncated at maxTokens. Increase maxOutputTokens in config or use a model with higher output capacity.');
                     }
                     throw new Error('No response text from AI');
                 }
@@ -277,21 +309,22 @@ export class Provider {
         const telemetry = this.getTelemetry(options);
         const maxRoundtrips = options.maxToolRoundtrips ?? 5;
         const extraStop = options.stopWhen;
-        const stopConditions = [stepCountIs(maxRoundtrips)];
+        const stopConditions = [isStepCount(maxRoundtrips)];
         if (extraStop)
             stopConditions.push(extraStop);
         const { stopWhen: _ignoredStopWhen, ...optionsWithoutStop } = options;
         const config = this.mergeProviderOptions({
             tools,
-            maxTokens: 16384,
+            maxOutputTokens: 16384,
             toolChoice: 'auto',
+            allowSystemInMessages: true,
             ...(this.config.config || {}),
             ...optionsWithoutStop,
             stopWhen: stopConditions,
-            ...(telemetry ? { experimental_telemetry: telemetry } : {}),
             model,
             abortSignal: executionController.getAbortSignal(),
         }, options.agentName);
+        this.finalizeConfig(config, options, telemetry);
         try {
             const response = await withRetry(async () => {
                 const timeout = config.timeout || 30000;
@@ -306,7 +339,7 @@ export class Provider {
                     ]));
                     const hasToolCall = (result.toolCalls?.length || 0) > 0;
                     if (!result.text && !hasToolCall && result.finishReason === 'length') {
-                        throw new ContextLengthError('AI response empty: output truncated at maxTokens. Increase maxTokens in config or use a model with higher output capacity.');
+                        throw new ContextLengthError('AI response empty: output truncated at maxTokens. Increase maxOutputTokens in config or use a model with higher output capacity.');
                     }
                     return result;
                 }
@@ -365,12 +398,13 @@ export class Provider {
         const telemetry = this.getTelemetry(options);
         const config = this.mergeProviderOptions({
             schema,
+            allowSystemInMessages: true,
             ...(this.config.config || {}),
             ...options,
-            ...(telemetry ? { experimental_telemetry: telemetry } : {}),
             model: modelToUse,
             abortSignal: executionController.getAbortSignal(),
         }, options.agentName);
+        this.finalizeConfig(config, options, telemetry);
         try {
             promptLog(messages[messages.length - 1].content);
             const response = await withRetry(async () => {
@@ -552,12 +586,13 @@ export class Provider {
         ];
         const telemetry = this.getTelemetry({});
         const config = {
-            maxTokens: 16384,
+            maxOutputTokens: 16384,
             ...(this.config.config || {}),
-            ...(telemetry ? { experimental_telemetry: telemetry } : {}),
             model: this.config.visionModel,
             abortSignal: executionController.getAbortSignal(),
         };
+        if (telemetry)
+            config.telemetry = telemetry;
         try {
             promptLog(`Processing image with prompt: ${prompt}`);
             const response = await withRetry(async () => {

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

@@ -67,7 +67,7 @@ export function WithCoordinates(Base) {
             return this._analyzeScreenshotForVisualProps();
         }
         async visuallyAnnotateElements(opts) {
-            return visuallyAnnotateContainers(this.explorer.playwrightHelper.page, opts?.containers || []);
+            return this.explorer.visuallyAnnotateElements(opts);
         }
         async _analyzeScreenshotForVisualProps() {
             const elements = new Map();
@@ -177,7 +177,6 @@ export function WithCoordinates(Base) {
             debugLog(`Merged visual props for ${merged} elements`);
         }
         async backfillCoordinates(result) {
-            const page = this.explorer.playwrightHelper.page;
             const sections = parseResearchSections(result.text);
             const eidxWithoutCoords = [];
             for (const section of sections) {
@@ -188,7 +187,7 @@ export function WithCoordinates(Base) {
             }
             if (eidxWithoutCoords.length === 0)
                 return;
-            const webElements = await WebElement.fromEidxList(page, eidxWithoutCoords);
+            const webElements = await this.explorer.runWithBrowserRecovery('backfillCoordinates', () => WebElement.fromEidxList(this.explorer.playwrightHelper.page, eidxWithoutCoords));
             if (webElements.length === 0)
                 return;
             const rectMap = new Map(webElements.map((w) => [w.eidx, w]));

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

@@ -312,10 +312,9 @@ export function WithDeepAnalysis(Base) {
                     const isCoordinateClick = el.commands[0].startsWith('I.clickXY(');
                     if (!isCoordinateClick) {
                         const hoverCmd = el.commands[0].replace('I.click(', 'I.moveCursorTo(');
-                        const hoverAction = this.explorer.createAction();
-                        await hoverAction.attempt(hoverCmd, undefined, false);
+                        await this.explorer.attemptAction(hoverCmd, undefined, false);
                         await new Promise((r) => setTimeout(r, 500));
-                        await this.explorer.createAction().capturePageState();
+                        await this.explorer.capturePageState();
                         const hoverAR = ActionResult.fromState(this.stateManager.getCurrentState());
                         const hoverDiff = await hoverAR.diff(previousState);
                         await hoverDiff.calculate();
@@ -398,7 +397,7 @@ export function WithDeepAnalysis(Base) {
         async _restorePageState(url, originalAria) {
             try {
                 await this.cancelInUi();
-                await this.explorer.createAction().capturePageState();
+                await this.explorer.capturePageState();
                 const currentAria = this.stateManager.getCurrentState()?.ariaSnapshot || '';
                 if (!diffAriaSnapshots(originalAria, currentAria))
                     return;

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

@@ -173,8 +173,7 @@ export function WithLocators(Base) {
                 }
             }
             if (needsXpath.length > 0) {
-                const page = this.explorer.playwrightHelper.page;
-                const webElements = await WebElement.fromEidxList(page, needsXpath);
+                const webElements = await this.explorer.runWithBrowserRecovery('backfillBrokenLocators', () => WebElement.fromEidxList(this.explorer.playwrightHelper.page, needsXpath));
                 const changedSections = new Set();
                 for (const w of webElements) {
                     const entry = needsXpathEls.get(w.eidx);