Microsoft Foundry SDK for Python 2.6

Apply the Python 3 baseline first. These rules cover azure-ai-projects 2.6: project clients, connections, agents, hosted sessions, tools and toolboxes, routines, Microsoft 365 publishing, content safety, Agent Insights, and telemetry.

Client construction and authorization

• Construct AIProjectClient with a project endpoint selected from trusted configuration and an explicit production credential; never let request data, model output, or a tool argument select the endpoint, tenant, authority, credential, or token scope.
• Apply the same destination policy to get_openai_client(base_url=...) and send_request; keep API-key overrides, custom transports, and authentication policies in trusted configuration.
• Prefer a directly configured ManagedIdentityCredential or WorkloadIdentityCredential in production. If DefaultAzureCredential is unavoidable, exclude developer, interactive, and other credentials that must not satisfy the chain.
• Create separate clients, projects, identities, quotas, storage, and telemetry destinations for distinct environments and mutually untrusted tenants.
• Grant the narrowest Azure control-plane and data-plane roles at the smallest practical scope. Separate connection administration, agent authoring, publishing, monitoring, and invocation identities.
• Use the Foundry Agent Consumer role for invoke-only principals where it supplies the required data actions; do not give an application Contributor merely so it can invoke an agent.
• Treat every project, agent, deployment, connection, dataset, index, file, session, monitor, and