Foundry Responses API returns function tool <server>___<tool> as <server>.<tool> on claude-sonnet-4-6

[Eduardo Nicastro (edumuellerFSL)]

Technical Feedback

On a claude-sonnet-4-6 deployment, the Foundry Responses API returns function calls with a dot where the tool name has a triple underscore. A toolbox publishes MCP tools as <server>___<tool>, and the function call comes back as <server>.<tool>.

Sending a function tool named docs___search straight to {project}/openai/v1/responses with tool_choice "required" returns docs.search for claude-sonnet-4-6 and docs___search for gpt-5-mini. Replaying the conversation with the original ___ name is accepted.

A hosted agent using the toolbox (AddFoundryToolboxes, Responses protocol 2.0.0) can't call toolbox tools because of this. FunctionInvokingChatClient reports Requested function "<server>.<tool>" not found., and the next model call fails with HTTP 400:

Invalid 'input[4].name': string does not match pattern. Expected a string that matches the pattern '^[a-zA-Z0-9_-]+$'.

Packages: Microsoft.Agents.AI.Foundry.Hosting 1.21.0-preview.260911.1, Azure.AI.Projects 3.0.0-beta.2, Microsoft.Extensions.AI.OpenAI 10.10.0, .NET 10.0.

Desired Outcome

The function call name matches the name the toolbox published (<server>___<tool>), as it already does for gpt-5-mini.

Current Workaround

A DelegatingChatClient below the function invoker maps <server>.<tool> back to <server>___<tool> when that tool is registered, and drops the content's RawRepresentation, because OpenAIResponsesChatClient would otherwise replay the raw item with the dotted name.