Extend V1/V2 per-audience token support to all SDK extensions

- Pass authorization/auth_handler_name/turn_context to list_tool_servers
  in OpenAI and Semantic Kernel extensions so _attach_per_audience_tokens
  runs for all frameworks (Google ADK already staged in prior commit)
- Replace shared single-token header injection with per-server header
  merge ({**base_headers, **server.headers}) in OpenAI and SK extensions
- Use explicit scope for V2 token resolution: {audience}/{scope} when
  scope is present, {audience}/.default as pre-consented fallback
- Fix api:// V2 audience handling — only ATG AppId in api:// URI form
  is treated as V1; all other api:// audiences are correctly V2
- Add comprehensive tests for resolve_token_scope_for_server covering
  all V1/V2/null/api:// scenarios including test env audience handling
- Add CHANGELOG entry for microsoft-agents-a365-tooling package

Backward compatible: V1 agents (null/ATG audience) continue using the
shared ATG token unchanged. V2 blueprint + old SDK is by design not
supported — upgrade to this SDK is the migration path.

Author: biswapm

libraries/microsoft-agents-a365-tooling-extensions-openai/microsoft_agents_a365/tooling/extensions/openai/mcp_tool_registration_service.py

@@ -161,7 +161,10 @@ async def add_tool_servers_to_agent(
                         )
                     }
                     server_headers = dict(si.headers) if si.headers else {}
-                    headers = {**base_headers, **server_headers}  # per-audience token takes precedence
+                    headers = {
+                        **base_headers,
+                        **server_headers,
+                    }  # per-audience token takes precedence
 
                     # Create MCPServerStreamableHttpParams with proper configuration
                     params = MCPServerStreamableHttpParams(url=si.url, headers=headers)

libraries/microsoft-agents-a365-tooling-extensions-semantickernel/microsoft_agents_a365/tooling/extensions/semantickernel/services/mcp_tool_registration_service.py

@@ -119,7 +119,9 @@ async def add_tool_servers_to_agent(
         # Get and process servers
         options = ToolOptions(orchestrator_name=self._orchestrator_name)
         servers = await self._mcp_server_configuration_service.list_tool_servers(
-            agentic_app_id, auth_token, options,
+            agentic_app_id,
+            auth_token,
+            options,
             authorization=auth,
             auth_handler_name=auth_handler_name,
             turn_context=context,

libraries/microsoft-agents-a365-tooling/CHANGELOG.md

@@ -9,6 +9,22 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- Added V1/V2 per-audience token acquisition support in `McpToolServerConfigurationService.list_tool_servers()`. When `authorization`, `auth_handler_name`, and `turn_context` are provided, each MCP server receives its own OAuth token scoped to its audience — V1 servers (no audience, or shared ATG AppId) share a single ATG-scoped token; V2 servers (unique non-ATG audience GUID or `api://` URI) each receive a token scoped to `{audience}/{scope}` (or `{audience}/.default` when scope is absent and pre-consented)
+- Added `_attach_per_audience_tokens()` private method to `McpToolServerConfigurationService` — acquires one token per unique scope, caches within the call to avoid redundant exchanges, and attaches `Authorization: Bearer` headers to each server config
+- Added `resolve_token_scope_for_server()` utility function to derive the correct OAuth scope for a given `MCPServerConfig` based on its `audience` and `scope` fields
+- Added `audience`, `scope`, `publisher`, and `headers` fields to `MCPServerConfig`
+- Gateway discovery endpoint bumped to `/agents/v2/{id}/mcpServers`
+- `_parse_gateway_server_config()` and `_parse_manifest_server_config()` now map `audience`, `scope`, and `publisher` fields from gateway/manifest responses into `MCPServerConfig`
+
+### Changed
+
+- OpenAI, Semantic Kernel, and Google ADK extensions now pass auth context to `list_tool_servers()` and merge per-server headers (`{**base_headers, **server.headers}`) instead of injecting a single shared ATG token for all servers — fully backward compatible, V1 agents continue to receive the same shared ATG token
+
+### Notes
+
+- **Backward compatible**: agents with V1 manifests (null audience or shared ATG AppId) work identically with the new SDK — no token exchange behaviour changes
+- **Migration required for V2**: agents upgraded to V2 blueprint permissions (per-audience MCP servers) require this SDK version. Running a V2 blueprint with the old SDK will result in MCP tool auth failures (401/403)
+
 - Added `send_chat_history` method to `McpToolServerConfigurationService` for sending chat conversation history to the MCP platform for real-time threat protection analysis
 - Added `ChatHistoryMessage` Pydantic model for representing individual messages in chat history
 - Added `ChatMessageRequest` Pydantic model for the chat history API request payload

tests/tooling/test_mcp_server_configuration.py

@@ -323,7 +323,10 @@ def test_v2_api_uri_audience_with_explicit_scope(self):
         server = self._make_server(
             audience="api://mcp-calendartools", scope="McpServers.Calendar.All"
         )
-        assert resolve_token_scope_for_server(server) == "api://mcp-calendartools/McpServers.Calendar.All"
+        assert (
+            resolve_token_scope_for_server(server)
+            == "api://mcp-calendartools/McpServers.Calendar.All"
+        )
 
     def test_v2_guid_audience_null_scope_falls_back_to_default(self):
         """V2: unique GUID audience + null scope → <guid>/.default (pre-consented)."""
@@ -420,11 +423,11 @@ async def test_mixed_v1_v2_servers_deduplicate_by_scope(self, service):
         guid1 = "aaaaaaaa-0000-0000-0000-000000000001"
         guid2 = "bbbbbbbb-0000-0000-0000-000000000002"
         servers = [
-            self._make_server("mail"),        # V1
-            self._make_server("cal1", guid1), # V2 guid1
-            self._make_server("cal2", guid2), # V2 guid2
-            self._make_server("files"),       # V1 (same scope as mail → no 2nd exchange)
-            self._make_server("cal3", guid1), # V2 guid1 again → no 2nd exchange
+            self._make_server("mail"),  # V1
+            self._make_server("cal1", guid1),  # V2 guid1
+            self._make_server("cal2", guid2),  # V2 guid2
+            self._make_server("files"),  # V1 (same scope as mail → no 2nd exchange)
+            self._make_server("cal3", guid1),  # V2 guid1 again → no 2nd exchange
         ]
 
         authorization = MagicMock()