.NET: [Bug]: code definition skills,agent returning an empty response

[dalangtk]

Description

I am currently testing code definition skills ，now using deepseek-v4-flash. Why the agent returning an empty response?

Code Sample

Error Messages / Stack Traces

Package Versions

Microsoft.Agents.AI Version=1.17.0, Microsoft.Agents.AI.OpenAI Version=1.17.0

.NET Version

.NET 10

Additional Context

test code github address
https://github.com/dalangtk/agenttest

[github-actions]

👋 Hi dalang (@dalangtk) — thanks for the report!

Your report mentions empty responses when using deepseek-v4-flash but provides no error messages, stack traces, or inline code sample. Looking at the framework's skill injection code, tools (load_skill, read_skill_resource, run_skill_script) and instructions are correctly merged into the chat context. An empty response.Text typically means the model returned no text content — for example, the model may have issued only tool calls without a final text reply, or may not fully support the OpenAI function-calling format that the SDK sends. To help investigate further, it would be useful to see: (1) the full console output including any exceptions, (2) whether the same code works with a model known to support function calling (e.g., gpt-4o), and (3) the response.Messages contents (not just response.Text) to see if the model returned tool-call messages instead of text.

Once you can share the details above, we'll take another look. 🙏

[dalangtk]

👋 Hi dalang (@dalangtk) — thanks for the report!

Your report mentions empty responses when using deepseek-v4-flash but provides no error messages, stack traces, or inline code sample. Looking at the framework's skill injection code, tools (load_skill, read_skill_resource, run_skill_script) and instructions are correctly merged into the chat context. An empty response.Text typically means the model returned no text content — for example, the model may have issued only tool calls without a final text reply, or may not fully support the OpenAI function-calling format that the SDK sends. To help investigate further, it would be useful to see: (1) the full console output including any exceptions, (2) whether the same code works with a model known to support function calling (e.g., gpt-4o), and (3) the response.Messages contents (not just response.Text) to see if the model returned tool-call messages instead of text.

Once you can share the details above, we'll take another look. 🙏

If I comment out the line of code AIContextProviders = [skillsProvider], the agent can return messages normally.

[rogeriorfp]

Starting with version 1.11.1, all AgentSkillsProvider tools require approval by default. Please try this:

 AIAgent agent = chatClientWrapper.AsAIAgent(new ChatClientAgentOptions
   {
       Name = "UnitConverterAgent",
       ChatOptions = new()
       {
           Instructions = "You are a helpful assistant",
       },
       AIContextProviders = [skillsProvider],
   })
   .AsBuilder()
   .UseToolApproval(new ToolApprovalAgentOptions
   {
       // NOTE: Auto-approving all skill tools is done here for simplicity in
       // this demonstration. In production, you should prompt the user before
       // allowing script execution. See Agent_Step07_SkillsAutoApproval for a
       // walkthrough of the full approval flow.
       AutoApprovalRules = [AgentSkillsProvider.AllToolsAutoApprovalRule],
   })
   .Build();