Emit a usage chunk on streamed completions

Sources/MLXServerKit/ChatCompletionsHandler.swift

@@ -77,11 +77,18 @@ enum ChatCompletionsHandler {
                         roleSent = true
                         try await writer.write(
                             SSE.event(chunk(id, created, model, delta, finishReason: nil)))
-                    case .finished(let reason, _):
+                    case .finished(let reason, let usage):
                         let delta = ChatCompletionChunk.Delta(
                             role: nil, content: nil, toolCalls: nil)
                         try await writer.write(
                             SSE.event(chunk(id, created, model, delta, finishReason: reason)))
+                        // OpenAI-style trailing usage chunk: empty choices,
+                        // populated usage. Lets clients report context-window
+                        // consumption for a streamed turn.
+                        try await writer.write(
+                            SSE.event(ChatCompletionChunk(
+                                id: id, created: created, model: model,
+                                choices: [], usage: usage)))
                     }
                 }
                 try await writer.write(SSE.done())

Sources/MLXServerKit/OpenAITypes.swift

@@ -249,6 +249,9 @@ public struct ChatCompletionChunk: Encodable, Sendable {
     public var created: Int
     public var model: String
     public var choices: [ChunkChoice]
+    /// Token usage. Populated only on the final chunk of a stream so clients
+    /// can report context-window consumption; omitted on all other chunks.
+    public var usage: Usage?
 
     public struct ChunkChoice: Encodable, Sendable {
         public var index: Int

Tests/MLXServerTests/RoutesTests.swift

@@ -97,6 +97,8 @@ struct RoutesTests {
             let text = String(buffer: response.body)
             #expect(text.contains("chat.completion.chunk"))
             #expect(text.contains("data: [DONE]"))
+            // The stream must carry a trailing usage chunk for context %.
+            #expect(text.contains("\"prompt_tokens\""))
         }
     }