.NET: Allow custom argument marshaling for skill scripts

[SergeyMenshykh]

Summary

vLLM and some other OpenAI-compatible backends send tool-call arguments as a JSON string (e.g. "{\"query\":\"hello\"}") instead of a JSON object. Skill scripts only accept a JSON object, so these calls fail with an InvalidOperationException.

This PR lets callers plug in their own argument conversion logic.

Changes

• Adds an optional Func<JsonElement?, AIFunctionArguments>? argumentMarshaler that controls how raw JSON arguments are converted before the script delegate runs.
• It can be supplied at three levels:
  • AgentInlineSkillScript constructor - per script.
  • AgentInlineSkill constructor - applies to all scripts added via AddScript.
  • AgentClassSkill<T> constructor - applies to discovered scripts and those created via CreateScript.
• When omitted, the existing behavior is unchanged: arguments must be a JSON object (or null), otherwise an InvalidOperationException is thrown.

Closes #6020

[Copilot]

Pull request overview

This PR introduces configurable argument marshaling for .NET skill scripts so callers can handle OpenAI-compatible backends (e.g., vLLM) that emit tool-call arguments as a JSON string instead of a JSON object, while preserving the existing strict default behavior.

Changes:

• Added an optional Func<JsonElement?, AIFunctionArguments>? argumentMarshaler that converts raw JSON arguments into AIFunctionArguments before invoking the script delegate.
• Plumbed the marshaler through AgentInlineSkillScript, AgentInlineSkill (skill-level default), and AgentClassSkill<T> (applies to discovered/programmatic scripts).
• Added unit tests covering default behavior vs. custom string-parsing marshaler behavior and precedence rules.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
dotnet/tests/Microsoft.Agents.AI.UnitTests/AgentSkills/SkillScriptArgumentMarshalerTests.cs	Adds tests validating default strict object-only marshaling and custom marshaler support (including vLLM-like string-wrapped JSON).
dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentInlineSkillScript.cs	Adds a per-script marshaler hook and uses it in RunAsync prior to AIFunction invocation.
dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentInlineSkill.cs	Adds a skill-level default marshaler and wires it into scripts added via AddScript.
dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentClassSkill.cs	Adds a skill-level marshaler for class-based skills and passes it to discovered/programmatic scripts.

Comments suppressed due to low confidence (1)

dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentClassSkill.cs:124

• AgentClassSkill is a public base class; replacing the protected parameterless constructor with a single ctor that takes an (optional) parameter is a binary-breaking change for already-compiled derived skills (they will look for base .ctor() and fail at runtime). Keep a true parameterless ctor and add a new overload that accepts argumentMarshaler.

    protected AgentClassSkill(Func<JsonElement?, AIFunctionArguments>? argumentMarshaler = null)
    {
        this._argumentMarshaler = argumentMarshaler;
        this._resources = new Lazy<IReadOnlyList<AgentSkillResource>?>(this.DiscoverResources);
        this._scripts = new Lazy<IReadOnlyList<AgentSkillScript>?>(this.DiscoverScripts);
        this._content = new Lazy<string>(() => AgentInlineSkillContentBuilder.Build(
            this.Frontmatter.Name,
            this.Frontmatter.Description,
            this.Instructions,
            this.Scripts));
    }

[github-actions]

Automated Code Review

Reviewers: 5 | Confidence: 92% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Failure Modes, Design Approach

---

Automated review by SergeyMenshykh's agents

[github-actions]

Automated Code Review

Reviewers: 5 | Confidence: 91% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Failure Modes, Design Approach

---

Automated review by SergeyMenshykh's agents