portal: add a permissioned browser node for agent-assisted UI testing

[guicybercode]

Phase

Post-MVP. Do not begin until the task model, event stream, canvas, and security boundaries are stable.

Context

A browser node connected to a task agent can close the loop for frontend implementation and verification, especially against localhost applications.

Goal

Add a Portal node with explicit user-approved browser automation capabilities.

Scope

• Embedded browser navigation with an origin allowlist.
• Typed actions for navigate, click, type, scroll, screenshot, inspect DOM, and read console messages.
• Connect a Portal to specific task agents through capability grants.
• Capture screenshots and findings as task artifacts.
• Add timeouts, cancellation, action limits, and visible automation state.
• Keep cookies, storage, and credentials isolated from the user's normal browser profile.

Acceptance criteria

• Automation is disabled until the user grants a task-scoped origin.
• Navigation outside approved origins pauses and requests confirmation.
• Arbitrary JavaScript execution is disabled by default and requires a separate explicit capability.
• Password fields, cookies, tokens, and browser storage never enter logs or diagnostics.
• Every agent action is visible, cancellable, and recorded with sanitized metadata.
• Closing the node cancels automation without stopping unrelated agent sessions.
• Deterministic tests cover localhost navigation, denied origins, timeout, console errors, and screenshots.
• Linux and macOS WebView differences are documented and tested.

Out of scope

Mobile-device control, remote browsers, CAPTCHA bypass, and production credential automation.