Add additionalContext support to MLXLanguageModel

[noorbhatia]

Add CustomGenerationOptions struct to MLXLanguageModel with an additionalContext: [String: JSONValue]? property for injecting variables into Jinja2 chat templates

Usage

var options = GenerationOptions()
options[custom: MLXLanguageModel.self] = .init(
    additionalContext: ["enable_thinking": .bool(true)]
)
let response = try await session.respond(to: "Hello", options: options)

[noorbhatia]

Hi @mattt , can you take a look at this. Thanks!

[mattt]

Sorry for not reviewing sooner. #147 also introduces MLXLanguageModel.CustomGenerationOptions, so we'll have to do some traffic control here. Should have both of these merged today.

[mattt]

This logic for converting from MLXLMCommon.JSONValue to Any values for UserInput is repeated a few times, so it'd probably make sense to move it into CustomGenerationOptions itself.

[mattt]

Reminder to myself that we should add a parameter for UserInput.Processing in a follow-up.

[Copilot]

Pull request overview

Adds support for passing an additionalContext dictionary through GenerationOptions into MLXLMCommon’s chat-template rendering context, enabling callers to inject custom variables for Jinja2-based templates.

Changes:

• Extends MLXLanguageModel.CustomGenerationOptions with additionalContext and threads it into MLXLMCommon.UserInput construction paths.
• Adds a basic MLX test that exercises setting additionalContext via GenerationOptions.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
Sources/AnyLanguageModel/Models/MLXLanguageModel.swift	Introduces additionalContext custom option and forwards it into MLX user input creation for standard, streaming, and structured generation flows.
Tests/AnyLanguageModelTests/MLXLanguageModelTests.swift	Adds a regression test ensuring additionalContext can be set and a response is produced.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

additionalContext is exposed as [String: MLXLMCommon.JSONValue]?, which leaks the MLXLMCommon dependency into AnyLanguageModel’s public API. Other models’ custom options use the package-wide JSONValue (e.g., Anthropic’s extraBody: [String: JSONValue]?), so consider changing this to [String: JSONValue]? (or a typealias) and converting internally before building MLXLMCommon.UserInput.

[Copilot]

MLXLMCommon.JSONValue.toSendable() maps .null to NSNull(), while elsewhere in this file nulls are represented as MLXLMCommon.JSONValue.null when converting JSON for Sendable contexts. Mixing null representations can lead to inconsistent template rendering and may also interact poorly with strict Sendable checking; prefer a single null representation (ideally MLXLMCommon.JSONValue.null if additionalContext ultimately expects Sendable values).

Suggested change

	case .null: return NSNull()
	case .null: return MLXLMCommon.JSONValue.null

[Copilot]

makeUserInput(...) now appears unused, and MLXLMCommon.UserInput construction (including processing defaults) is duplicated in respond, streamResponse, and structured generation. To avoid drift and keep future changes (like additionalContext) consistent, consider updating makeUserInput to accept additionalContext and reusing it in these call sites, or removing the helper if you prefer inline construction.