From 61f630c60d42734280b5b2bc8387584e457ec00c Mon Sep 17 00:00:00 2001
From: Peter Dave Hello <3691490+PeterDaveHello@users.noreply.github.com>
Date: Fri, 11 Sep 2026 23:09:23 +0800
Subject: [PATCH] Decode SSE chunks incrementally

---
 src/utils/eventsource-parser.mjs             |  9 +--
 tests/unit/utils/eventsource-parser.test.mjs | 72 +++++++++++++++++++-
 2 files changed, 74 insertions(+), 7 deletions(-)

diff --git a/src/utils/eventsource-parser.mjs b/src/utils/eventsource-parser.mjs
index 31e019e92..84d8f6dfd 100644
--- a/src/utils/eventsource-parser.mjs
+++ b/src/utils/eventsource-parser.mjs
@@ -2,7 +2,7 @@
 
 function createParser(onParse) {
   let isFirstChunk
-  let bytes
+  let decoder
   let buffer
   let startingPosition
   let startingFieldLength
@@ -18,7 +18,7 @@ function createParser(onParse) {
   }
   function reset() {
     isFirstChunk = true
-    bytes = []
+    decoder = new TextDecoder()
     buffer = ''
     startingPosition = 0
     startingFieldLength = -1
@@ -29,8 +29,7 @@ function createParser(onParse) {
   }
 
   function feed(chunk) {
-    bytes = bytes.concat(Array.from(chunk))
-    buffer = new TextDecoder().decode(new Uint8Array(bytes))
+    buffer += decoder.decode(chunk, { stream: true })
     if (isFirstChunk && hasBom(buffer)) {
       buffer = buffer.slice(BOM.length)
     }
@@ -70,10 +69,8 @@ function createParser(onParse) {
       position += lineLength + 1
     }
     if (position === length) {
-      bytes = []
       buffer = ''
     } else if (position > 0) {
-      bytes = bytes.slice(new TextEncoder().encode(buffer.slice(0, position)).length)
       buffer = buffer.slice(position)
     }
   }
diff --git a/tests/unit/utils/eventsource-parser.test.mjs b/tests/unit/utils/eventsource-parser.test.mjs
index debe86ee0..6ca24ccfb 100644
--- a/tests/unit/utils/eventsource-parser.test.mjs
+++ b/tests/unit/utils/eventsource-parser.test.mjs
@@ -115,7 +115,22 @@ test('createParser preserves split CRLF state across empty chunks', () => {
 
 test('createParser produces the same events at every single chunk boundary', () => {
   const stream = toBytes('data: alpha\r\n\ndata: beta\ndata: gamma\r\r')
-  const expected = parseChunks(stream)
+  const expected = [
+    {
+      type: 'event',
+      id: undefined,
+      event: undefined,
+      data: 'alpha',
+      extra: undefined,
+    },
+    {
+      type: 'event',
+      id: undefined,
+      event: undefined,
+      data: 'beta\ngamma',
+      extra: undefined,
+    },
+  ]
 
   for (let split = 0; split <= stream.length; ++split) {
     const actual = parseChunks(stream.slice(0, split), stream.slice(split))
@@ -123,6 +138,61 @@ test('createParser produces the same events at every single chunk boundary', ()
   }
 })
 
+test('createParser preserves UTF-8 data at every byte boundary', () => {
+  const stream = toBytes('data: 台灣🙂 café\n\n')
+  const expected = [
+    {
+      type: 'event',
+      id: undefined,
+      event: undefined,
+      data: '台灣🙂 café',
+      extra: undefined,
+    },
+  ]
+
+  for (let split = 0; split <= stream.length; ++split) {
+    const actual = parseChunks(stream.slice(0, split), stream.slice(split))
+    assert.deepEqual(actual, expected, `split at byte ${split}`)
+  }
+})
+
+test('createParser preserves pending data after a leading UTF-8 BOM', () => {
+  const parsed = parseChunks(toBytes('\uFEFFdata: a\n\ndata:'), toBytes(' b\n\n'))
+
+  assert.deepEqual(
+    parsed.map((event) => event.data),
+    ['a', 'b'],
+  )
+})
+
+test('createParser preserves pending data after invalid UTF-8 replacement', () => {
+  const firstChunk = new Uint8Array([
+    ...toBytes('data: '),
+    0xff,
+    ...toBytes('\n\ndata:'),
+  ])
+  const parsed = parseChunks(firstChunk, toBytes(' b\n\n'))
+
+  assert.deepEqual(
+    parsed.map((event) => event.data),
+    ['�', 'b'],
+  )
+})
+
+test('createParser reset discards pending decoder bytes', () => {
+  const parsed = []
+  const parser = createParser((event) => parsed.push(event))
+
+  parser.feed(toBytes('🙂').slice(0, 2))
+  parser.reset()
+  parser.feed(toBytes('data: clean\n\n'))
+
+  assert.deepEqual(
+    parsed.map((event) => event.data),
+    ['clean'],
+  )
+})
+
 test('createParser handles \\r only line endings', () => {
   const parsed = []
   const parser = createParser((event) => parsed.push(event))