free-computer-use 0.1.0 → 0.2.0

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # Changelog
 
+## 0.2.0 - 2026-09-26
+
+### Added
+
+- Deterministic, provider-free extraction of visible page text and links.
+- Bounded extraction of the first 1–1000 visible links, with exact-count
+  verification.
+
+### Changed
+
+- Zero-timeout selector checks return immediately instead of waiting through a
+  polling interval.
+
+### Fixed
+
+- Element verification evaluates all matching elements, avoiding a false
+  failure when a later match satisfies the condition.
+
 ## 0.1.0
 
 ### Added

package/README.md

@@ -3,41 +3,84 @@
 ![FreeComputerUse — a local-first browser agent that plans, acts and verifies with you in control](assets/readme/hero.svg)
 
 <p align="center">
+  <a href="https://www.npmjs.com/package/free-computer-use"><img alt="npm version" src="https://img.shields.io/npm/v/free-computer-use?style=flat-square&labelColor=18251f&color=c4e967"></a>
+  <a href="https://github.com/OthmaneBlial/FreeComputerUse/releases/latest"><img alt="Latest GitHub release" src="https://img.shields.io/github/v/release/OthmaneBlial/FreeComputerUse?style=flat-square&labelColor=18251f&color=c4e967"></a>
   <a href="LICENSE"><img alt="MIT license" src="https://img.shields.io/badge/license-MIT-c4e967?style=flat-square&labelColor=18251f"></a>
+  <img alt="Node.js 22.13 or newer" src="https://img.shields.io/badge/Node.js-22.13%2B-43853d?style=flat-square&labelColor=18251f">
   <img alt="TypeScript" src="https://img.shields.io/badge/TypeScript-3178c6?style=flat-square&labelColor=18251f">
   <img alt="Playwright" src="https://img.shields.io/badge/Playwright-45ba4b?style=flat-square&labelColor=18251f">
-  <img alt="Local-first" src="https://img.shields.io/badge/local--first-browser_runs-c4e967?style=flat-square&labelColor=18251f">
 </p>
 
 <p align="center">
   <a href="#quick-start">Quick start</a> ·
+  <a href="#no-model-key-needed">Try it without a model</a> ·
+  <a href="#watch-a-real-run">Watch the demo</a> ·
+  <a href="https://github.com/OthmaneBlial/FreeComputerUse/releases/latest">Latest release</a> ·
   <a href="https://othmaneblial.github.io/FreeComputerUse/">Project site</a> ·
-  <a href="https://othmaneblial.github.io/FreeComputerUse/lab/index.html">Try the task lab</a> ·
   <a href="SECURITY.md">Security</a>
 </p>
 
-**Your browser. Your model. Your call.** Tell it what done looks like, then stay in control of every step.
+**Your browser. Your model. Your call.** Give your browser a goal; the model proposes short action batches, and Playwright runs them locally. You approve website access and sensitive actions, then inspect what happened.
 
-FreeComputerUse is a local-first AI browser automation agent. A model plans small action batches; TypeScript and Playwright execute and verify them in your browser. Compatible workflows can be learned once and replayed later with **zero model calls**.
+Use an API or a supported model CLI for open-ended tasks. For common page-reading jobs, skip model setup entirely: FreeComputerUse can read visible text and links with deterministic local strategies.
+
+## No model key needed
+
+Install the release and start the local dashboard:
+
+```bash
+npm install --global free-computer-use@0.2.0
+FCU_BROWSER_CHANNEL=chrome agent doctor
+FCU_BROWSER_CHANNEL=chrome agent ui
+```
+
+Open **http://127.0.0.1:4318**, enter a page URL and one of these goals, then approve access to that website:
+
+| Goal | Result |
+| --- | --- |
+| `Read the page` | Extract visible page text. |
+| `Extract the links` | List visible link labels and URLs. |
+| `Extract the first 5 links` | Return a bounded list of the first five links. |
+| `Extract the table` | Try the synthetic [practice revenue table](https://othmaneblial.github.io/FreeComputerUse/lab/reports.html). |
+
+These narrow workflows use local strategies and make zero model calls. General tasks need a configured provider. The declared minimum is Node.js 22.13; the verified setup is macOS 26.6 on Apple Silicon, Node 25.9.0 and system Chrome 154.0.8037.57. Other OS/browser combinations and the declared Node minimum remain unverified. See the [support matrix](docs/SUPPORT_MATRIX.md).
 
 ## Watch a real run
 
-[![A real browser run: investigate a synthetic API incident and save a checked report](assets/readme/incident-demo.gif)](https://othmaneblial.github.io/FreeComputerUse/#watch)
+[![FreeComputerUse: a real practice-table extraction from the published package](assets/readme/product-demo-poster.jpg)](assets/readme/product-demo.mp4)
+
+[Watch the 22-second dashboard walkthrough](assets/readme/product-demo.mp4) · [Watch the short portrait cut](assets/readme/product-demo-portrait.mp4)
 
-One goal led through six pages to a verified incident brief: **21 successful actions, 10 model calls, one repaired failure**. The configured cost estimate was **$0.00650**. This is one recorded synthetic task, not a general success-rate claim. [Watch the full recording](https://othmaneblial.github.io/FreeComputerUse/lab/media/incident-demo.mp4) · [Inspect the evidence](assets/readme/incident-evidence.json).
+This recording uses the 0.1.0 package. It shows the npm install, local dashboard, explicit site approval and a verified synthetic table result: three rows, one browser action and zero model calls. The same practice workflow remains available. It is a deterministic demo, not a claim about arbitrary websites.
+
+[![Recorded browser run: investigate a synthetic API incident and save a checked report](assets/readme/incident-demo.gif)](https://othmaneblial.github.io/FreeComputerUse/#watch)
+
+One recorded task followed six pages to a checked incident brief: **21 successful actions, 10 model calls, one repaired failure**. The configured cost estimate was **$0.00650**. This is one synthetic run, not a general success-rate claim. [Watch the full recording](https://othmaneblial.github.io/FreeComputerUse/lab/media/incident-demo.mp4) · [Inspect the evidence](assets/readme/incident-evidence.json).
 
 ## How it works
 
-1. **Observe** the page’s DOM and accessible controls.
+1. **Observe** the page DOM and accessible controls.
 2. **Plan** a bounded batch with the model you choose.
 3. **Execute** browser actions locally with Playwright; direct navigation to a new origin asks for approval.
-4. **Verify** results, repair only what failed, then reuse compatible learned workflows.
+4. **Verify** results, repair only what failed, then replay compatible learned workflows later.
 
-The model does not run shell commands or arbitrary JavaScript. Browser previews stay local; the goal and selected page context are sent to your configured model provider.
+The model does not run shell commands or arbitrary JavaScript. Browser previews stay local; the task and selected page context go to your configured model provider. Screenshots are not sent.
 
 ## Quick start
 
-Requires **Node.js 22.13+**, npm and an installed browser. The verified setup is macOS 26.6 (Apple Silicon), Node 25.9.0 and system Chrome 154.0.8037.57; the complete serial validation passed on this combination. The declared Node minimum and other OS/browser combinations remain unverified. This setup selects installed Chrome and avoids a separate Playwright browser download; see the [support matrix](docs/SUPPORT_MATRIX.md) for exact coverage and limits.
+Requires Node.js 22.13+, npm and an installed browser. The verified setup selects installed Chrome and avoids a separate Playwright browser download.
+
+### Install from npm
+
+```bash
+npm install --global free-computer-use@0.2.0
+FCU_BROWSER_CHANNEL=chrome agent doctor
+FCU_BROWSER_CHANNEL=chrome agent ui
+```
+
+Open **http://127.0.0.1:4318**. Try a keyless goal above, or configure a provider for model-planned tasks.
+
+### Run from source
 
 ```bash
 git clone https://github.com/OthmaneBlial/FreeComputerUse.git
@@ -47,9 +90,7 @@ cp .env.example .env
 chmod 600 .env
 ```
 
-Set `FCU_BROWSER_CHANNEL=chrome` in `.env` to use the verified installed-Chrome path without downloading a browser binary.
-
-For model-planned tasks, configure a provider in `.env`. The browser sandbox task below works without a model key:
+Set `FCU_BROWSER_CHANNEL=chrome` in `.env` to use installed Chrome. For model-planned tasks, add the provider details to your local `.env`; for example:
 
 ```dotenv
 FCU_BROWSER_CHANNEL=chrome
@@ -64,21 +105,11 @@ npm run agent -- doctor
 npm run dev
 ```
 
-Open **http://127.0.0.1:4318**, enter a starting URL and goal, then approve site access.
-
-**Try it without an API key:** open [the practice revenue table](https://othmaneblial.github.io/FreeComputerUse/lab/reports.html) and run the goal `Extract the table`. This narrow workflow has a deterministic local strategy; general tasks need a model provider.
-
-## Troubleshooting
-
-- **The browser does not start:** install Chrome, set `FCU_BROWSER_CHANNEL=chrome` in `.env`, then run `npm run agent -- doctor`. See the [verified platform limits](docs/SUPPORT_MATRIX.md); other OS/browser combinations are not certified here.
-- **The dashboard says no model is configured:** the practice revenue-table task works without a provider. For model planning, configure one route from [Pick your model](#pick-your-model). `npm run agent -- doctor --api` makes an opt-in request to the configured provider; use it only when you want that network check.
-- **A site is blocked or asks for approval:** normal mode asks before a new origin and blocks private/reserved DNS answers. Approve only the site needed for the task. Ultra mode disables those protections; do not use it as a workaround for a blocked destination.
-- **The run finishes without proving the goal:** inspect the result checks and use a narrower goal. Browser clicks alone do not mean the requested outcome was verified.
-- **You need to find or remove local data:** see [local storage, export, retention and deletion](docs/LOCAL_DATA.md). Data is not encrypted and is not automatically expired.
+Then open **http://127.0.0.1:4318**, enter a starting URL and goal, and approve site access.
 
 ## Pick your model
 
-Use a provider API key, or sign in through the official CLI for a supported subscription. API usage and consumer subscriptions are separate billing products.
+Use a provider API key or sign in through the official CLI for a supported subscription. API usage and consumer subscriptions are separate billing products.
 
 | Provider | Configuration |
 | --- | --- |
@@ -89,51 +120,41 @@ Use a provider API key, or sign in through the official CLI for a supported subs
 | ChatGPT plan | Install Codex CLI, sign in with `codex login`, set `LLM_PROVIDER=codex-subscription` |
 | Claude Pro/Max plan | Install Claude Code 2.1.248+, sign in with `claude auth login` (not Console), set `LLM_PROVIDER=claude-subscription` |
 
-The provider API routes have not all been live-tested here; DeepSeek Flash is the measured default. Subscription modes use the local CLI sign-in, need no API key and respect plan limits. Their local agent tools and MCP servers are disabled while planning. See [provider setup and limits](docs/PROVIDERS.md), the [support matrix](docs/SUPPORT_MATRIX.md) and [.env.example](.env.example) for route-specific setup, tests and live-evidence status.
+DeepSeek Flash is the measured default. Other provider routes are not all live-tested. Subscription modes use local CLI sign-in, need no API key, respect plan limits, and disable their local agent tools and MCP servers while planning. See [provider setup and limits](docs/PROVIDERS.md), the [support matrix](docs/SUPPORT_MATRIX.md) and [.env.example](.env.example).
 
 ## Connect an MCP host
 
-Run FreeComputerUse as a local MCP server over `stdio` with `agent mcp`. It
-provides tools to start a task, inspect the observed page, follow and verify the
-result, and stop a task. Site and sensitive-action approvals still require a
-human response in an MCP host that supports form elicitation. No HTTP endpoint
-is opened. See the [MCP setup and validation limits](docs/MCP.md). The separate
-provider CLI adapters keep their own tools and MCP disabled while planning.
+Run FreeComputerUse as a local MCP server over `stdio` with `agent mcp`. It exposes bounded tools to start a task, inspect the page, follow and verify results, and stop a task. Site and sensitive-action approvals still need a human response in an MCP host that supports form elicitation. No HTTP endpoint is opened. See the [MCP setup and validation limits](docs/MCP.md).
 
-## FAQ
-
-**Can I try it without a model key?** Yes, the practice revenue-table workflow runs with a deterministic local strategy. General tasks need a configured model.
+## Security and privacy
 
-**Which providers have live evidence?** The current matrix records DeepSeek Flash and one synthetic plan through Codex CLI `0.156.1` with ChatGPT. Contract tests do not certify a live provider; see the [dated support matrix](docs/SUPPORT_MATRIX.md).
+- Normal mode asks before direct navigation to a new origin and blocks private or reserved DNS answers. A loopback-only proxy connects to the vetted numeric address for browser HTTP(S) and WebSocket traffic. Ultra mode and the low-level `allowExternal` option opt out of origin and private-address checks; Ultra mode is off by default.
+- Sensitive actions have a separate confirmation gate. Read the [security boundaries](SECURITY.md) before using personal or sensitive data.
+- Browser execution, profiles, history and downloads stay on your machine. Selected task and page context goes to the provider you choose; screenshots are not sent.
+- Planner inputs use aliases for local profile and file values, not their contents. Local data is **not encrypted** and does not expire automatically. See [storage, export, retention and deletion](docs/LOCAL_DATA.md).
+- Chromium profiles disable WebRTC UDP that the proxy cannot carry. Sites that require direct UDP for voice or video may fail.
 
-**Does the model receive screenshots or browser profiles?** No screenshots are sent. The task and selected page context go to your provider; browser execution, profiles, history and downloads stay local. Local data is not encrypted.
+## Evidence and limits
 
-**Which platform is verified?** macOS 26.6 on Apple Silicon, Node 25.9.0 and system Chrome 154. Other OS, Node and browser combinations remain unverified in the [support matrix](docs/SUPPORT_MATRIX.md).
+A focused public suite recorded **18 September 2026** passed 14/14 first-run tasks and 14/14 compatible learned repeats with independent checks. First runs used 18 model calls; repeats used zero. These single-trial sandbox results do not predict success on arbitrary websites. [Report](artifacts/benchmark-public.json) · [Method and limitations](docs/BENCHMARKS.md).
 
-## Control and privacy
+Provider, operating-system, browser and accessibility evidence stays bounded to the [support matrix](docs/SUPPORT_MATRIX.md). Contract tests do not certify a live provider; builds do not certify an untested platform.
 
-- Normal mode asks before direct navigation to a new origin and blocks hostnames resolving to private or reserved address ranges, including private IPv4 embedded in a discovered NAT64 prefix. A loopback-only proxy connects to the vetted numeric address, closing the DNS lookup-to-connection rebinding gap for browser HTTP(S) and WebSocket traffic. Explicit IP destinations require an exact origin grant. Ultra mode and the low-level `allowExternal` option opt out of origin and private-address checks. Chrome 154 tests cover redirect denial, simulated DNS rebinding, approved/blocked plain WebSockets, and approved/blocked browser-originated WSS. The WSS fixture uses a generated local certificate and ignores its certificate error only in the test; a separate opt-in smoke verifies Chrome TLS and an echo against one public WSS endpoint. Other WSS endpoints, live network-specific NAT64 discovery and other browser builds remain unverified. Sensitive actions have a separate confirmation gate. Ultra mode is explicit and off by default.
-- Browser execution, profiles, history and downloads stay on your machine. Page context needed for a plan goes to the chosen model provider; screenshots are not sent.
-- The planner receives aliases for local profile and file values, not their contents. Local storage is **not encrypted**.
-- Chromium profiles disable WebRTC UDP that the proxy cannot carry. Sites needing direct UDP for voice/video may fail; one local Chrome STUN fixture confirms no direct packet reached its receiver. Other non-HTTP traffic remains unverified.
-- Runs record actions, checks, repairs, token estimates and workflow reuse so you can inspect what happened.
+## FAQ
 
-Read [the security boundaries](SECURITY.md) before using personal or sensitive data.
-See [where local data is stored and how to inspect, export or delete it](docs/LOCAL_DATA.md).
+**Can I try it without an API key?** Yes. Use `Read the page`, `Extract the links`, `Extract the first 5 links`, or the synthetic revenue-table task. These workflows make zero model calls.
 
-## Evidence
+**Does the model receive screenshots or browser profiles?** No. The selected task and page context go to your provider; screenshots and browser profiles stay local. Local data is not encrypted.
 
-In a focused public suite recorded **18 September 2026**, 14/14 first-run results and 14/14 compatible learned repeats passed independent checks. First runs used 18 model calls; repeats used zero. These single-trial sandbox results do not predict success on arbitrary websites. [Report](artifacts/benchmark-public.json) · [Method and limitations](docs/BENCHMARKS.md).
+**Which platform is verified?** macOS 26.6 on Apple Silicon, Node 25.9.0 and system Chrome 154.0.8037.57. Check the [support matrix](docs/SUPPORT_MATRIX.md) for exact coverage and limits.
 
 ## Explore and contribute
 
 - [Project site and task library](https://othmaneblial.github.io/FreeComputerUse/)
-- [Frequently asked questions](https://othmaneblial.github.io/FreeComputerUse/#faq)
+- [Latest GitHub release](https://github.com/OthmaneBlial/FreeComputerUse/releases/latest)
 - [Useful browser-task examples](docs/USEFUL_EXAMPLES.md)
-- [Implementation notes](docs/IMPLEMENTATION.md)
-- [Architecture](docs/ARCHITECTURE.md)
-- [Local validation commands](docs/LOCAL_VALIDATION.md)
-- [Contributing](CONTRIBUTING.md) · [Changelog](CHANGELOG.md)
+- [Architecture](docs/ARCHITECTURE.md) · [implementation notes](docs/IMPLEMENTATION.md)
+- [Local validation commands](docs/LOCAL_VALIDATION.md) · [contributing](CONTRIBUTING.md) · [changelog](CHANGELOG.md)
 
 Useful contributions: reproducible browser tasks, safer permission scopes and checks that make results easier to trust. Keep shared examples synthetic or read-only; remove credentials and personal data from traces.
 

package/dist/actions/executor.d.ts

@@ -1,6 +1,6 @@
 import { type Action, type Condition } from './schema.js';
 import { type ConfirmationPolicy } from './policy.js';
-import type { Browser } from '../browser/Browser.js';
+import { type Browser } from '../browser/Browser.js';
 import type { Observer } from '../browser/Observer.js';
 import type { VariableResolver } from '../profile/VariableResolver.js';
 import type { Control } from '../agent/Control.js';
@@ -33,6 +33,7 @@ export declare class Executor {
     });
     semantic(action: Action): Action;
     semanticCondition(condition: Condition): Condition;
+    private readTextValue;
     private storeDownload;
     run(input: Action): Promise<ActionResult>;
 }

package/dist/actions/executor.js

@@ -3,6 +3,8 @@ import { randomUUID } from 'node:crypto';
 import { basename, resolve } from 'node:path';
 import { ActionSchema } from './schema.js';
 import { sensitiveReason } from './policy.js';
+import { checkedHttpURL } from '../browser/Browser.js';
+import { SecurityBoundaryError } from '../browser/SecurityBoundaryError.js';
 import { Verifier } from '../verification/Verifier.js';
 import { ActionCompiler } from './compiler.js';
 export class Executor {
@@ -34,6 +36,9 @@ export class Executor {
     semanticCondition(condition) {
         return 'target' in condition && typeof condition.target === 'string' ? { ...condition, target: this.observer.selectors.descriptor(condition.target) } : condition;
     }
+    readTextValue(locator) {
+        return locator.evaluate(el => 'value' in el ? String(el.value) : el instanceof HTMLElement && el.isContentEditable ? el.innerText : undefined);
+    }
     async storeDownload(download, requestedName) {
         const folder = resolve(this.options.downloadDir ?? '.fcu/downloads');
         await mkdir(folder, { recursive: true, mode: 0o700 });
@@ -56,7 +61,8 @@ export class Executor {
     async run(input) {
         let action = ActionSchema.parse(input);
         const startedAt = Date.now();
-        let strategy, executed = false, receiptAction = action;
+        let strategy, executed = false, navigationUncertain = false, receiptAction = action;
+        let pageBefore;
         let assertApproved, disposeApproved;
         try {
             await this.control.checkpoint();
@@ -90,6 +96,7 @@ export class Executor {
             }
             await this.control.checkpoint();
             const page = this.browser.page;
+            pageBefore = page;
             const timeout = action.timeoutMs ?? 4000;
             let data;
             const interaction = this.browser.interaction, interactionOptions = { timeout, checkpoint: () => this.control.checkpoint(), beforeEffect: assertApproved };
@@ -108,26 +115,34 @@ export class Executor {
                 case 'navigate':
                     if (action.url.includes('{{profile.') || action.url.includes('{{files.'))
                         throw new Error('Local vault values cannot be embedded in navigation URLs');
+                    checkedHttpURL(action.url, page.url());
+                    navigationUncertain = true;
                     await this.browser.navigate(action.url);
                     break;
                 case 'openTab':
                     if (action.url.includes('{{profile.') || action.url.includes('{{files.'))
                         throw new Error('Local vault values cannot be embedded in navigation URLs');
+                    checkedHttpURL(action.url, page.url());
+                    navigationUncertain = true;
                     await this.browser.openTab(action.url);
                     break;
                 case 'closeTab':
+                    navigationUncertain = this.browser.context.pages().length > 1;
                     await this.browser.closeTab();
                     break;
                 case 'switchTab':
                     await this.browser.switchTab(action.index, timeout);
                     break;
                 case 'back':
+                    navigationUncertain = true;
                     await page.goBack({ waitUntil: 'domcontentloaded' });
                     break;
                 case 'forward':
+                    navigationUncertain = true;
                     await page.goForward({ waitUntil: 'domcontentloaded' });
                     break;
                 case 'reload':
+                    navigationUncertain = true;
                     await page.reload({ waitUntil: 'domcontentloaded' });
                     break;
                 case 'click': {
@@ -162,12 +177,17 @@ export class Executor {
                     break;
                 case 'fill':
                     await interaction.enter(locator, value, true, interactionOptions);
-                    if (await locator.inputValue() !== value)
+                    if (await this.readTextValue(locator) !== value)
                         throw new Error('Fill postcondition failed');
                     break;
-                case 'type':
-                    await interaction.enter(locator, value, false, interactionOptions);
+                case 'type': {
+                    if (!await locator.isEditable())
+                        throw new Error('Type target is not editable');
+                    const { before, after, expected } = await interaction.enter(locator, value, false, interactionOptions);
+                    if (value && (expected !== undefined ? after !== expected : before !== undefined && (after ?? await this.readTextValue(locator)) === before))
+                        throw new Error('Type postcondition failed');
                     break;
+                }
                 case 'select': {
                     const options = await locator.evaluate(el => [...el.options].map(o => ({ label: o.label, value: o.value })));
                     const option = options.find(o => o.value === value) || options.find(o => o.label === value);
@@ -240,47 +260,138 @@ export class Executor {
                     interaction.cue('extract');
                     const root = locator ?? page.locator('body');
                     if (action.format === 'table')
-                        data = await root.locator('tr').filter({ visible: true }).evaluateAll(rows => rows.map(row => [...row.querySelectorAll('th,td')].map(cell => cell.textContent?.trim() ?? '')));
+                        data = await root.evaluateAll((els, { match, limit }) => {
+                            let rows = [...new Set(els.flatMap(el => el.matches('tr') ? [el] : [...el.querySelectorAll('tr')]))].filter(row => { const box = row.getBoundingClientRect(); return box.width > 0 && box.height > 0 && getComputedStyle(row).visibility === 'visible'; });
+                            if (!match && limit)
+                                rows = rows.slice(0, limit);
+                            const values = rows.map(row => [...row.querySelectorAll('th,td')].map(cell => {
+                                for (let parent = cell; parent; parent = parent.parentElement) {
+                                    const style = getComputedStyle(parent);
+                                    if (style.display === 'none' || style.visibility === 'hidden' || style.opacity === '0')
+                                        return '';
+                                }
+                                return cell.closest('[hidden],[inert],[aria-hidden="true"]') ? '' : cell.innerText?.trim() ?? '';
+                            }));
+                            return (match ? values.filter(row => JSON.stringify(row).toLowerCase().includes(match)) : values).slice(0, limit);
+                        }, { match: action.match?.toLowerCase(), limit: action.limit });
                     else if (action.format === 'links')
-                        data = await root.evaluateAll(els => els.flatMap(el => [...(el.matches('a[href]') ? [el] : el.querySelectorAll('a[href]'))].filter(a => a.getClientRects().length > 0).map(a => ({ text: a.textContent?.trim(), url: a.href }))));
+                        data = await root.evaluateAll((els, { match, limit }) => {
+                            let links = [...new Set(els.flatMap(el => el.matches('a[href]') ? [el] : [...el.querySelectorAll('a[href]')]))].filter(a => {
+                                if (!a.getClientRects().length || a.closest('[hidden],[inert],[aria-hidden="true"]'))
+                                    return false;
+                                for (let parent = a; parent; parent = parent.parentElement) {
+                                    const style = getComputedStyle(parent);
+                                    if (style.visibility !== 'visible' || style.opacity === '0')
+                                        return false;
+                                }
+                                return true;
+                            });
+                            if (!match && limit)
+                                links = links.slice(0, limit);
+                            const values = links.map(a => ({ text: a.innerText?.trim(), url: a.href }));
+                            return (match ? values.filter(link => JSON.stringify(link).toLowerCase().includes(match)) : values).slice(0, limit);
+                        }, { match: action.match?.toLowerCase(), limit: action.limit });
                     else if (action.format === 'records') {
                         if (!action.fields || !Object.keys(action.fields).length || Object.keys(action.fields).length > 20)
                             throw new Error('Records extraction requires 1..20 controlled CSS fields');
-                        data = await root.filter({ visible: true }).evaluateAll((els, fields) => els.flatMap(el => el.matches('table,tbody') ? [...el.querySelectorAll('tr')].filter(row => row.querySelector('td')) : [el]).map(el => Object.fromEntries(Object.entries(fields).map(([key, field]) => {
-                            let node = el.querySelector(field.css);
-                            let value = '';
-                            // Resolve tabular fields from actual headers, rather than guessed cell indices.
-                            if (el.tagName === 'TR' && field.attribute === 'text') {
-                                const headers = [...(el.closest('table')?.querySelectorAll('thead th,tr:first-child th') ?? [])];
-                                const [normalize] = [(s) => (s ?? '').toLowerCase().replace(/[^a-z]/g, '')];
-                                const wanted = normalize(key), indices = headers.map((h, i) => normalize(h.textContent ?? '') === wanted ? i : -1).filter(i => i >= 0);
-                                const cells = [...el.querySelectorAll(':scope > td')];
-                                if (indices.length === 1 && cells.length >= headers.length)
-                                    node = cells[indices[0] + cells.length - headers.length] ?? node;
-                            }
-                            if (node) {
-                                if (field.attribute === 'text')
-                                    value = node.innerText?.trim() ?? node.textContent?.trim() ?? '';
-                                else if (field.attribute === 'href')
-                                    value = node.href ?? '';
-                                else if (field.attribute === 'src')
-                                    value = node.src ?? '';
-                                else if (field.attribute === 'value')
-                                    value = node.matches('input[type=password],input[autocomplete=cc-number],input[autocomplete=cc-csc]') ? '[sensitive value omitted]' : node.value ?? '';
-                                else
-                                    value = node.getAttribute(field.attribute) ?? '';
-                            }
-                            return [key, value];
-                        }))), action.fields);
+                        data = await root.filter({ visible: true }).evaluateAll((els, { fields, match, limit }) => {
+                            let rows = [...new Set(els.flatMap(el => el.matches('table,tbody') ? [...el.querySelectorAll('tr')].filter(row => row.querySelector('td')) : [el]))];
+                            if (!match && limit)
+                                rows = rows.slice(0, limit);
+                            const records = rows.map(el => Object.fromEntries(Object.entries(fields).map(([key, field]) => {
+                                let node = el.querySelector(field.css);
+                                let value = '';
+                                // Resolve tabular fields from actual headers, rather than guessed cell indices.
+                                if (el.tagName === 'TR' && field.attribute === 'text') {
+                                    const headers = [...(el.closest('table')?.querySelectorAll('thead th,tr:first-child th') ?? [])];
+                                    const [normalize] = [(s) => (s ?? '').toLowerCase().replace(/[^a-z]/g, '')];
+                                    const wanted = normalize(key), indices = headers.map((h, i) => normalize(h.textContent ?? '') === wanted ? i : -1).filter(i => i >= 0);
+                                    const cells = [...el.querySelectorAll(':scope > td')];
+                                    if (indices.length === 1 && cells.length >= headers.length)
+                                        node = cells[indices[0] + cells.length - headers.length] ?? node;
+                                }
+                                if (node) {
+                                    let hidden = !!node.closest('[hidden],[inert],[aria-hidden="true"]');
+                                    for (let parent = node; parent; parent = parent.parentElement) {
+                                        const style = getComputedStyle(parent);
+                                        if (style.display === 'none' || style.visibility !== 'visible' || style.opacity === '0')
+                                            hidden = true;
+                                    }
+                                    const autocomplete = (node.getAttribute('autocomplete') ?? '').toLowerCase().split(/\s+/);
+                                    const sensitiveInput = field.attribute === 'value' && node instanceof HTMLInputElement && (node.type === 'hidden' || node.type === 'password' || autocomplete.some(token => token.startsWith('cc-') || ['current-password', 'new-password', 'one-time-code'].includes(token)));
+                                    if (hidden && !sensitiveInput)
+                                        value = '';
+                                    else if (field.attribute === 'text')
+                                        value = node.innerText?.trim() ?? node.textContent?.trim() ?? '';
+                                    else if (field.attribute === 'href')
+                                        value = node.href ?? '';
+                                    else if (field.attribute === 'src')
+                                        value = node.src ?? '';
+                                    else if (field.attribute === 'value')
+                                        value = sensitiveInput ? '[sensitive value omitted]' : node.value ?? '';
+                                    else
+                                        value = node.getAttribute(field.attribute) ?? '';
+                                }
+                                return [key, value];
+                            })));
+                            return (match ? records.filter(record => JSON.stringify(record).toLowerCase().includes(match)) : records).slice(0, limit);
+                        }, { fields: action.fields, match: action.match?.toLowerCase(), limit: action.limit });
                     }
                     else
-                        data = (await root.filter({ visible: true }).allInnerTexts()).join('\n').slice(0, 100000);
-                    if (Array.isArray(data)) {
-                        const match = action.match;
+                        data = await root.filter({ visible: true }).evaluateAll(els => els.map(el => {
+                            const modal = document.querySelector('dialog:modal,[role=dialog][aria-modal=true]'), modalVisible = !!modal && modal.getClientRects().length > 0;
+                            const [excluded] = [(node) => {
+                                    if (node.matches('[hidden],[inert],[aria-hidden="true"]') || modalVisible && modal && !modal.contains(node) && modal !== node && !node.contains(modal))
+                                        return true;
+                                    const style = getComputedStyle(node);
+                                    return style.display === 'none' || style.visibility !== 'visible' || style.opacity === '0';
+                                }];
+                            let hiddenParent = false;
+                            for (let parent = el.parentElement; parent; parent = parent.parentElement)
+                                if (excluded(parent)) {
+                                    hiddenParent = true;
+                                    break;
+                                }
+                            const omitted = new Set(), dirty = new Set();
+                            const [inspect] = [(node, hidden = false) => {
+                                    if (hidden || excluded(node)) {
+                                        omitted.add(node);
+                                        for (let parent = node; parent; parent = parent.parentElement) {
+                                            dirty.add(parent);
+                                            if (parent === el)
+                                                break;
+                                        }
+                                        return;
+                                    }
+                                    for (const child of node.children)
+                                        inspect(child);
+                                }];
+                            inspect(el, hiddenParent);
+                            const [read] = [(node) => {
+                                    if (node.nodeType === Node.TEXT_NODE)
+                                        return node.nodeValue ?? '';
+                                    if (node.nodeType !== Node.ELEMENT_NODE)
+                                        return '';
+                                    const element = node;
+                                    if (omitted.has(element))
+                                        return '';
+                                    if (!dirty.has(element))
+                                        return element.innerText ?? element.textContent ?? '';
+                                    if (element.tagName === 'BR')
+                                        return '\n';
+                                    const content = [...element.childNodes].map(read).join('');
+                                    return /^(block|flex|grid|flow-root|list-item|table)/.test(getComputedStyle(element).display) ? `\n${content}\n` : content;
+                                }];
+                            return read(el).replace(/[ \t]*\n[ \t]*/g, '\n').replace(/\n+/g, '\n').trim();
+                        }).join('\n').slice(0, 100000));
+                    const { match, limit } = action;
+                    if (typeof data === 'string' && (match || limit)) {
+                        let lines = data.split(/\r?\n/);
                         if (match)
-                            data = data.filter(item => JSON.stringify(item).toLowerCase().includes(match.toLowerCase()));
-                        if (action.limit)
-                            data = data.slice(0, action.limit);
+                            lines = lines.filter(line => line.toLowerCase().includes(match.toLowerCase()));
+                        if (limit)
+                            lines = lines.slice(0, limit);
+                        data = lines.join('\n');
                     }
                     this.browser.extractions.push({ key: action.key, value: data });
                     data = { [action.key]: data };
@@ -295,7 +406,8 @@ export class Executor {
             return { action: receiptAction, startedAt, durationMs: Date.now() - startedAt, success: true, strategy, data };
         }
         catch (error) {
-            return { action: receiptAction, startedAt, durationMs: Date.now() - startedAt, success: false, strategy, error: this.variables.redact(error instanceof Error ? error.message : 'Browser action failed'), uncertain: executed && ['click', 'doubleClick', 'press', 'submit', 'download'].includes(action.type) };
+            const blockedBeforeNavigation = error instanceof SecurityBoundaryError && ['navigate', 'openTab'].includes(action.type) && this.browser.page === pageBefore;
+            return { action: receiptAction, startedAt, durationMs: Date.now() - startedAt, success: false, strategy, error: this.variables.redact(error instanceof Error ? error.message : 'Browser action failed'), uncertain: executed && ((navigationUncertain && !blockedBeforeNavigation) || ['click', 'doubleClick', 'press', 'submit', 'download', 'type'].includes(action.type)) };
         }
         finally {
             await disposeApproved?.();