Add an MCP Tool

Everything a person can do in the editor, an agent must be able to do through MCP (see docs/AIControl.md and docs/Decisions/0016-mcp-server.md). Components don't need tools: the generic component tools cover every registered component. Tools are for features that aren't components - play mode, screenshots, export, multiplayer, and so on.

1. Put the behaviour in the editor context or a command

• Anything that changes a document is a Command (Source/Editor/Lodestone/Editor/Commands), run through EditorContext::Execute(), so it can be undone and the UI can use it too. It must change nothing when it fails, and give the same result - UUIDs included - when redone
• Anything else that the UI also does (opening documents, play mode) is a method on EditorContext, so the UI and MCP share it. Tools stay thin

2. Register the tool

In Source/Editor/Lodestone/Editor/Mcp/EditorTools.cpp, in the Add...Tools() function for its area:

server.AddTool({.Name = "light_bake",
    .Title = "Bake lighting",
    .Description = "Bakes the open scene's lighting. Returns how long it took",
    .InputSchema = Schema(R"({"type": "object", "properties": {
        "quality": {"type": "string", "enum": ["draft", "final"], "description": "draft by default"}}})"),
    .Handler = [&context](const Json::Value& json)
    {
        const auto quality = Arguments(json).OptionalString("quality");
        if (!quality)
            return Fail(quality.error());
        ...
        return McpToolResult::Structured({{"seconds", seconds}});
    }});

• Names are snake_case, <area>_<verb>. The schema is a JSON Schema object with a description for anything that isn't obvious; required arguments are listed in required
• Read arguments with Arguments, which reports which argument is wrong. Return Fail(error) or McpToolResult::Failure() for anything the agent can fix - never throw, and never assert on agent input
• Set ReadOnly for tools that only read, and Destructive for tools that delete or discard. Tools that replace the open document check CheckUnsaved() and take discardChanges
• Return structured content for data (McpToolResult::Structured), Run() for commands (it reports what was done), and McpToolResult::Image for images
• Paths go through EditorContext::ResolveAssetPath(), which keeps them inside the project's Assets directory
• Work on the main thread only - tool handlers already run there

3. Test it

In Tests/Editor/EditorToolsTests.cpp:

• Add the name to ToolNames (the inventory test fails until you do, and the bad-arguments test then calls the tool with junk)
• A test case per tool: success, the result's content, the effect on the editor, undo where it's a command, and every failure an agent can cause
• If the feature changes what an agent's whole workflow looks like, extend Tests/Editor/McpAgentSmokeTest.py

4. Document it

Add it to the tool table in docs/AIControl.md#tools.