fix(openai): emit streaming token usage from usage-only SSE chunks

Author: Tom-Ryder

Sources/AgentRunKit/LLM/OpenAIClientStreaming.swift

@@ -110,6 +110,7 @@ extension OpenAIClient {
 
     func extractDeltas(from chunk: StreamingChunk) throws -> [StreamDelta] {
         var deltas: [StreamDelta] = []
+        var emittedFinished = false
         for choice in chunk.choices ?? [] {
             if let reasoning = choice.delta.reasoning ?? choice.delta.reasoningContent, !reasoning.isEmpty {
                 deltas.append(.reasoning(reasoning))
@@ -121,8 +122,12 @@ extension OpenAIClient {
             try extractAudioDeltas(from: choice.delta, into: &deltas)
             if choice.finishReason != nil {
                 deltas.append(.finished(usage: chunk.usage.map(\.tokenUsage)))
+                emittedFinished = true
             }
         }
+        if !emittedFinished, let usage = chunk.usage {
+            deltas.append(.finished(usage: usage.tokenUsage))
+        }
         return deltas
     }
 

Tests/AgentRunKitTests/OpenAIClientResponseTests.swift

@@ -406,7 +406,7 @@ struct StreamingChunkParsingTests {
     }
 
     @Test
-    func streamingChunkWithoutChoicesDecodesCleanly() throws {
+    func streamingChunkWithUsageOnlyEmitsFinished() throws {
         let json = """
         {
             "usage": {
@@ -419,9 +419,13 @@ struct StreamingChunkParsingTests {
         let chunk = try client.parseStreamingChunk(Data(json.utf8))
         let deltas = try client.extractDeltas(from: chunk)
 
-        #expect(deltas.isEmpty)
-        #expect(chunk.usage?.promptTokens == 100)
-        #expect(chunk.usage?.completionTokens == 50)
+        #expect(deltas.count == 1)
+        if case let .finished(usage) = deltas.first {
+            #expect(usage?.input == 100)
+            #expect(usage?.output == 50)
+        } else {
+            Issue.record("Expected .finished delta with usage")
+        }
     }
 
     @Test