Fix SDK 0.16 stream completion at DONE (LAN-873)

.github/workflows/main.yml

@@ -5,7 +5,7 @@ on:
     tags:
       - "*"
   pull_request:
-    branches: [main]
+    branches: [main, "v0.16.x"]
 
 jobs:
   build:

CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## 0.16.22
+
+- Finish Vercel AI SDK streams when a complete `[DONE]` event arrives, preserving tool calls and usage without waiting for the HTTP connection to close.
+- Handle split SSE records and a leading UTF-8 BOM; cancel the response body without delaying completion when cancellation stalls or fails.
+
 ## 0.16.21
 
 - Add Meta's `meta-responses-v1` to the `reasoning_details` format union. It was missing, and since `ReasoningDetailArraySchema` drops entries it cannot parse, Meta reasoning details (e.g. from Muse Spark 1.2) were silently discarded instead of erroring — losing reasoning continuity across turns.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "langtail",
-  "version": "0.16.21",
+  "version": "0.16.22",
   "description": "",
   "main": "./Langtail.js",
   "packageManager": "pnpm@8.15.6",
@@ -41,7 +41,7 @@
     "copyfiles": "^2.4.1",
     "fs-extra": "^11.2.0",
     "jsdom": "^24.1.0",
-    "nock": "14.0.0-beta.5",
+    "nock": "14.0.17",
     "prettier": "^3.2.5",
     "tsup": "^8.0.2",
     "typescript": "^5.7.3",

src/LangtailPrompts.spec.ts

@@ -75,6 +75,11 @@ describe.skipIf(!liveTesting)(
         const completion = await lt.invoke({
           prompt,
           environment: "staging",
+          // Override the deployed fixture's retired model for live inference.
+          model: "gpt-6-luna",
+          temperature: 1,
+          reasoning_effort: "low",
+          max_tokens: 512,
           variables: {
             about: "cowboy Bebop",
           },
@@ -91,6 +96,11 @@ describe.skipIf(!liveTesting)(
         const completion = await lt.invoke({
           prompt: "short-story-teller",
           environment: "staging",
+          // Override the deployed fixture's retired model for live inference.
+          model: "gpt-6-luna",
+          temperature: 1,
+          reasoning_effort: "low",
+          max_tokens: 512,
           variables: {
             about: "Aragorn",
           },
@@ -107,6 +117,11 @@ describe.skipIf(!liveTesting)(
         const proxyCompletion = await lt.invoke({
           prompt,
           environment: "staging",
+          // Override the deployed fixture's retired model for live inference.
+          model: "gpt-6-luna",
+          temperature: 1,
+          reasoning_effort: "low",
+          max_tokens: 512,
           variables: {
             about: "napoleon",
           },

src/openai.spec.ts

@@ -5,41 +5,45 @@ import nock from "nock"
 import { ChatCompletionChunkSchema } from "./dataSchema"
 import OpenAI from "openai"
 
-
-const lt = createOpenAIProxy(new OpenAI({ apiKey: process.env.LANGTAIL_API_KEY }))
+const lt = createOpenAIProxy(
+  new OpenAI({ apiKey: process.env.LANGTAIL_API_KEY || "local-test" }),
+)
 
 describe("Langtail Proxy", () => {
-  it("should support streaming", async () => {
-    const proxyCompletion = await lt.chat.completions.create({
-      messages: [
-        { role: "system", content: "You are a helpful assistant." },
-        { role: "user", content: "What is the weather like?" },
-      ],
-      stream: true,
-      model: "gpt-3.5-turbo",
-      doNotRecord: false,
-      metadata: {
-        "custom-field": "1",
-      },
-    })
-    let partCount = 0
-    for await (const part of proxyCompletion) {
-      partCount++
+  it.skipIf(process.env.TEST_LIVE !== "true")(
+    "should support streaming",
+    async () => {
+      const proxyCompletion = await lt.chat.completions.create({
+        messages: [
+          { role: "system", content: "You are a helpful assistant." },
+          { role: "user", content: "What is the weather like?" },
+        ],
+        stream: true,
+        model: "gpt-6-luna",
+        reasoning_effort: "low",
+        max_tokens: 512,
+        doNotRecord: false,
+        metadata: {
+          "custom-field": "1",
+        },
+      })
+      let partCount = 0
+      for await (const part of proxyCompletion) {
+        partCount++
 
-      ChatCompletionChunkSchema.parse(part)
-    }
+        ChatCompletionChunkSchema.parse(part)
+      }
 
-    expect(partCount > 1).toBe(true)
-  })
+      expect(partCount > 1).toBe(true)
+    },
+  )
 
   it("should not record", async (t) => {
     nock(baseURL) // nock works by intercepting requests at the network level, if open AI switches to undici we will need to intercept differently
       .post("/chat/completions")
       .reply(200, function (uri, req) {
-        expect(this.req.headers["x-langtail-do-not-record"][0]).toBe("true")
-        expect(this.req.headers["x-langtail-metadata-custom-field"][0]).toBe(
-          "1",
-        )
+        expect(this.req.headers["x-langtail-do-not-record"]).toBe("true")
+        expect(this.req.headers["x-langtail-metadata-custom-field"]).toBe("1")
 
         // check that body has no metadata or doNotRecord
         expect(req).toMatchInlineSnapshot(`