Pass LanguageModelChat to tools

[roblourens]

We could finalize #244323 since it's generally useful for tools.

[roblourens]

	export interface LanguageModelToolInvocationOptions<T> {
		model?: LanguageModelChat;
	}

This is only set when the tool is called using a toolInvocationToken. The model ID is saved on the ChatRequest, we look up the model for the associated request and send that.

[wangmingliang-ms]

@roblourens
When will this API be available in VSCode Insiders? we can test it with our extension once it's ready.
Will this API be available in VSCode Stable before //Build?

[roblourens]

API sync notes:

• Would need to split the LanguageModelToolInvocationOptions interface so that model can be received but not passed in. "Resolved" options?

@wangmingliang-ms

Will this API be available in VSCode Stable before //Build?

It might be finalized, but no guarantee

we can test it with our extension once it's ready.

It's been there since March behind an api proposal. You can test it if you like but you can't publish an extension that uses it. https://github.com/microsoft/vscode/pull/244323/files#diff-0612ef0251b8125f593ece49eb7bef5154bd1b029b8d7d23de18c424078cc671R438

[isidorn]

For Build this API will only be available in Insiders. Not in Stable.

@wangmingliang-ms we always recommend that Build demos are done with Insiders. So you should be good.

[roblourens]

I'm really just hung up on what the name of the interface should be... LanguageModelToolResolvedInvocationOptions? Quite a mouthful...

• LanguageModelToolInvokeeOptions
• Just LanguageModelToolOptions for one of them but I don't know which one

[jrieken]

LanguageModelTool#invoke(LanguageModelToolInvocationInvokeOptions...)

[tlmii]

Does the current implementation have the model getting passed through to mcp tools somehow? Thats what we were looking for in the issue that was duped to this one, so just clarifying the current state.

[jrieken]

No. MCP doesn't consider this case and it would needed to be raised with them. The closet is mcp sampling which allows MCP servers to use the LM of their host. But afaik MCP servers have to ask for LMs by capabilities and not by UI state

[akaroml]

The Java Upgrade is about to hit GA by the end of July. And we need to honor user's model selection, but we could not use proposed API for a GA product. Will this be included in the June release?