MCP act focus by ref fails unless the DOM document was requested first

[avinashgola]

Problem

The Rust MCP contract suite currently documents a live limitation:

act with kind: "focus" and a DOM ref fails instead of focusing the referenced element.

The contract case is named act: focus by ref reports the current DOM-domain limitation in packages/browseros-agent/contracts/claw-mcp/tests/cases-act.ts.

While debugging PR #2583, the same underlying path surfaced in CI as:

act failed: CDP error: Document needs to be requested first

That error came from the core focus path:

• packages/browseros-agent/crates/browseros-core/src/input/geometry.rs
• focus_element() calls DOM.pushNodesByBackendIdsToFrontend
• then calls DOM.focus

In the BrowserOS CI browser, DOM.pushNodesByBackendIdsToFrontend can fail unless the DOM document has first been requested for that session.

Impact

Agents can click/fill/type referenced elements, but cannot reliably call the explicit focus action by ref. This also makes future ref-targeted keyboard behavior easy to regress, because focusing via the DOM domain is not usable in every session state.

Expected behavior

act { kind: "focus", ref } should focus the referenced element successfully, or core should use a focus implementation that does not require prior DOM.getDocument state for the session.

Possible fix

Use the runtime/object path already used by other semantic ref actions:

1. Resolve the backend node with DOM.resolveNode.
2. Call Runtime.callFunctionOn with this.focus().
3. Keep the existing DOM-domain DOM.focus path only if there is a reason it is preferred after DOM.getDocument.

PR #2583 added a runtime focus path for press(ref) because it hit this limitation; the standalone focus action still uses the brittle DOM-domain path.