From afe024e8c7990e9ced295f544c3cc5502c7136d4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Yi=C4=9Fit=20ERDO=C4=9EAN?= <yigiterdogan023@gmail.com>
Date: Sat, 1 Aug 2026 12:23:08 +0300
Subject: [PATCH] fix(ai): assistant text blocks are run together on replay
 (#115743)

Chat Completions replay flattened every assistant text block with an empty
separator, so two distinct blocks came back as one word-joined sentence. The
same message shape survives distinctly on the Anthropic, Responses and Mistral
lanes, and the string-content flattener for strict OpenAI-compatible servers
already joins with a newline.

Two blocks arise routinely: streaming opens a new text block after a tool call,
and cross-model replay converts a thinking block into a text block.
---
 .../src/openai-completions-messages.test.ts   | 52 +++++++++++++++++++
 .../ai/src/openai-completions-messages.ts     |  4 +-
 2 files changed, 55 insertions(+), 1 deletion(-)
 create mode 100644 packages/ai/src/openai-completions-messages.test.ts

diff --git a/packages/ai/src/openai-completions-messages.test.ts b/packages/ai/src/openai-completions-messages.test.ts
new file mode 100644
index 000000000000..42e552d7c7ce
--- /dev/null
+++ b/packages/ai/src/openai-completions-messages.test.ts
@@ -0,0 +1,52 @@
+import { describe, expect, it } from "vitest";
+import { convertMessages } from "./openai-completions-messages.js";
+import { resolveOpenAICompletionsCompat } from "./transports/openai-completions-compat.js";
+import type { AssistantMessage, Context, Model } from "./types.js";
+
+const model: Model<"openai-completions"> = {
+  id: "test-model",
+  name: "Test model",
+  api: "openai-completions",
+  provider: "custom-openai-compatible",
+  baseUrl: "https://proxy.example/v1",
+  reasoning: false,
+  input: ["text"],
+  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+  contextWindow: 128_000,
+  maxTokens: 4_096,
+};
+
+const emptyUsage = {
+  input: 0,
+  output: 0,
+  cacheRead: 0,
+  cacheWrite: 0,
+  totalTokens: 0,
+  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+};
+
+describe("convertMessages assistant text replay", () => {
+  it("keeps separate assistant text blocks apart", () => {
+    const assistant: AssistantMessage = {
+      role: "assistant",
+      api: model.api,
+      provider: model.provider,
+      model: model.id,
+      content: [
+        { type: "text", text: "Let me check the file." },
+        { type: "text", text: "The file contains X." },
+      ],
+      usage: emptyUsage,
+      stopReason: "stop",
+      timestamp: 2,
+    };
+    const context: Context = {
+      messages: [{ role: "user", content: "hello", timestamp: 1 }, assistant],
+    };
+
+    const converted = convertMessages(model, context, resolveOpenAICompletionsCompat(model));
+
+    const replayed = converted.find((message) => message.role === "assistant");
+    expect(replayed?.content).toBe("Let me check the file.\nThe file contains X.");
+  });
+});
diff --git a/packages/ai/src/openai-completions-messages.ts b/packages/ai/src/openai-completions-messages.ts
index 3aa348a55f5e..5a693a4278ce 100644
--- a/packages/ai/src/openai-completions-messages.ts
+++ b/packages/ai/src/openai-completions-messages.ts
@@ -141,7 +141,9 @@ export function convertMessages(
               text: sanitizeSurrogates(block.text),
             }) satisfies ChatCompletionContentPartText,
         );
-      const assistantText = assistantTextParts.map((part) => part.text).join("");
+      // Separate content blocks are distinct utterances, so replay them the way
+      // the string-content flattener does rather than running them together.
+      const assistantText = assistantTextParts.map((part) => part.text).join("\n");
 
       const nonEmptyThinkingBlocks = msg.content
         .filter(isThinkingContentBlock)