From 87deee7fbdc6a0d3ef672866c90b75090689c9fd Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 10 Jun 2026 02:35:27 -0700
Subject: [PATCH] Studio: faithful conversation export and import round trips
 (ShareGPT system role, CSV quoted newlines) (#6131)

* fix(studio): preserve system role in ShareGPT exports

System messages in ShareGPT conversation exports were serialized as
gpt turns, which changes the semantics of exported training data.
Map role system to from system in both the single-thread and bulk
export paths, matching the importer (sharegptToRecords), which
already maps from system back to a system role.

Extracted from #5606 by @LeoBorcherding (commit 7b277913).

* fix(studio): parse quoted newlines when importing conversation CSV

csvToRecords split the file on raw newlines before parsing quotes, so
any exported message containing a newline broke on re-import (the
record was cut mid-field and remainder lines were dropped). The module
already ships an RFC 4180 parser (parseCsv) used by the prompt and
list importers; use it for conversation CSV too.

Multi-line content, embedded quotes and commas, and CRLF files now
round-trip. Unquoted commas in hand-made CSV keep the previous
behavior (rest of line is the content). Flagged in #5606 review as
'Preserve quoted newlines when importing CSV'.
---
 .../prompt-storage/prompt-storage-dialog.tsx  | 34 ++++++-------------
 1 file changed, 10 insertions(+), 24 deletions(-)

diff --git a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
index 80c2de54a2..3908c313b6 100644
--- a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
+++ b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
@@ -331,7 +331,7 @@ function messageToOpenAI(msg: { role: unknown; content: unknown; attachments?: u
   return contentParts.length > 0 ? [{ role: role as "user" | "system", content: contentParts }] : [];
 }
 
-// ShareGPT training JSONL (human/gpt turns).
+// ShareGPT training JSONL (human/system/gpt turns).
 export async function exportConversationShareGPT(threadId: string): Promise<void> {
   const messages = await loadConversationMessages(threadId);
   if (!messages) return;
@@ -339,7 +339,7 @@ export async function exportConversationShareGPT(threadId: string): Promise<void
   const conversations: Array<{ from: string; value: string }> = [];
   for (const msg of messages) {
     const role = msg.role as string;
-    const from = role === "user" ? "human" : "gpt";
+    const from = role === "user" ? "human" : role === "system" ? "system" : "gpt";
     const value = messageToText(msg);
     if (value.trim()) conversations.push({ from, value });
   }
@@ -413,7 +413,7 @@ async function buildThreadContent(
     for (const msg of messages) {
       const role = msg.role as string;
       const value = messageToText(msg);
-      if (value.trim()) conversations.push({ from: role === "user" ? "human" : "gpt", value });
+      if (value.trim()) conversations.push({ from: role === "user" ? "human" : role === "system" ? "system" : "gpt", value });
     }
     if (conversations.length === 0) return null;
     return JSON.stringify({ conversations });
@@ -626,30 +626,16 @@ function sharegptToRecords(
 }
 
 function csvToRecords(csvText: string, threadId: string, baseTs: number): MessageRecord[] {
-  const lines = csvText.split(/\r?\n/);
+  // parseCsv handles quoted newlines, so multi-line message content
+  // round-trips; a naive per-line split would break those records.
+  const rows = parseCsv(csvText).slice(1);
   const records: MessageRecord[] = [];
   let prevId: string | null = null;
   let idx = 0;
-  for (let i = 1; i < lines.length; i++) {
-    const line = lines[i].trim();
-    if (!line) continue;
-    let role = "";
-    let content = "";
-    if (line.startsWith('"')) {
-      const m = line.match(/^"((?:[^"]|"")*)"\s*,\s*([\s\S]*)$/);
-      if (!m) continue;
-      role = m[1].replace(/""/g, '"');
-      content = m[2].startsWith('"')
-        ? m[2].slice(1, m[2].endsWith('"') ? -1 : undefined).replace(/""/g, '"')
-        : m[2];
-    } else {
-      const comma = line.indexOf(",");
-      if (comma === -1) continue;
-      role = line.slice(0, comma);
-      content = line.slice(comma + 1).startsWith('"')
-        ? line.slice(comma + 2, line.endsWith('"') ? -1 : undefined).replace(/""/g, '"')
-        : line.slice(comma + 1);
-    }
+  for (const row of rows) {
+    if (row.length < 2) continue;
+    const role = row[0].trim();
+    const content = row.slice(1).join(",");
     if (!content.trim()) continue;
     const validRole = role === "user" || role === "assistant" || role === "system" ? role : "user";
     const id = crypto.randomUUID();