Browser Automation tool (preview) in Foundry prompt agent returns empty output, no sessions in Playwright Browser activity log

Setup: Foundry project, Playwright Workspace created, access-token auth enabled, project connection created (auth type: API key), role assigned on the workspace resource to the project identity. Python, azure-ai-projects, PromptAgentDefinition with BrowserAutomationPreviewTool, project_connection_id set to the connection name (also tried the full resource ID, same result), tool_choice="required", streaming.

What happens: The stream contains a browser_automation_preview_call item (with the query), followed by browser_automation_preview_call_output with status='completed' and output=''. The model then says it could not retrieve the page. I tried several URLs. One run for example.com gave the right heading, but I can’t rule out that the model answered from memory.

Checked: The Browser activity log in the Playwright workspace shows no sessions. Foundry tracing shows the tool call but no error. Token is valid.

Region: eastUS

Question: What can cause a completed tool call with empty output and no browser session, and where can I see the actual error?