Microsoft Foundry SDK for Python 2

Apply to azure-ai-projects 2.x and its AIProjectClient integration with Foundry agents and the OpenAI client. Check API availability against the installed minor; stable packages also expose preview features. Gate allow_preview=True and .beta use explicitly; .beta sub-clients do not require that flag. Apply the included Python baseline.

Identity and project boundaries

• Construct AIProjectClient(endpoint=..., credential=...) from trusted deployment configuration. Allowlist the HTTPS project endpoint, deployment and agent names; never let request content redirect an authenticated client to another account or project.
• Use an explicit production workload identity with narrowly scoped Entra roles. In deployed code pass ManagedIdentityCredential or WorkloadIdentityCredential rather than a bare DefaultAzureCredential, whose chain can pick up host environment variables or developer sign-ins; if you keep the chain, set AZURE_TOKEN_CREDENTIALS to prod or one credential name with require_envvar=True. Don't call Foundry with account keys, because key authentication bypasses role restrictions. Separate principals that create or publish agents from those that only invoke them; assign Foundry Agent Consumer at agent scope for endpoint-only access, since project scope reaches every agent endpoint in the project.
• Authorize the current application user before reading, updating, invoking or deleting agents, conversations, responses, files, datasets and connections. Bind IDs, including previous_response_id, to the current tenant and user in server-side