Remote browser capture and return

The existing desktop browser executor keeps an inactive visible tab in a hidden window with a compositor. Switching tabs preserves its WebContents, profile, DOM and consent binding. A detached native view can still answer DOM queries while Chromium screenshot delivery remains pending; visibility alone is not a capture-readiness guarantee.

Screenshot and finite PNG-frame recording use the existing CDP driver and its 10-second per-command deadline. The change adds no retries, transport, rendering flags or authority. Sensitive-page checks, media bounds and current authority checks after asynchronous CDP work remain mandatory.

The browser journey uses the production local consent shell, delegated REST workflow admission, the existing remote poller and the installed MCP SDK over local HTTP. A direct-command fixture revision is disabled after its verified receipt so later contact events cannot replay it. Same-run continuation and Stop are checked against persisted command/workflow state and an independent loopback fixture service. Late real Chromium bytes are held at the debugger boundary; local Stop must prevent both successful byte delivery and the workflow's next action. Hand back requires the same tab and server audit acknowledgement.

Qualification is limited to synthetic owned profiles and physical-device IPC fixtures. It does not certify physical OS input, an installed signed app, external-account reliability or model tool use. Full GQA-02/GQA-03 scope remains open beyond those boundaries.

Protocol references:

• https://www.electronjs.org/docs/latest/api/web-contents-view
• https://www.electronjs.org/docs/latest/api/web-contents
• https://chromedevtools.github.io/devtools-protocol/tot/Page/#method-captureScreenshot

API result storage preserves validated PNG/JPEG image bytes at supported browser screenshot, navigation and recording paths. MIME, canonical base64, size, format and bounded image decoding checks precede that exception. Prose, sibling fields and malformed media retain normal secret redaction. Preserving screenshot bytes does not remove sensitive content from pixels or embedded image metadata.

Qualification covers real Chromium pixel decoding through canonical database storage and MCP image responses, same-tab human takeover and hand-back, and Stop during capture. The test selects the original tab by its exact ID and waits for its active/controller acknowledgement before checking refusal. This is synthetic owned-fixture evidence; installed physical-device and external-account acceptance remain separate release gates.