From 8518cf47af51ab49331a54bd448e0c91c925aeb1 Mon Sep 17 00:00:00 2001
From: Neil <4138956+nwparker@users.noreply.github.com>
Date: Sun, 26 Jul 2026 20:31:48 -0700
Subject: [PATCH] perf(ai-vault): keep the JSONL line carry linear (#10783)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

consumeCompleteJsonlLines re-joined its held-over partial line with every
stream chunk, so one oversized record — a large tool result — cost O(record^2).
It backs the incremental parse for every resumable agent transcript, so the
whole AI Vault corpus paid it.

Hold the pieces in a list and join once, when a newline finally arrives.
Measured on a transcript with a single oversized record: 2.13 ms -> 1.17 ms at
1 MB and 68.22 ms -> 4.35 ms at 8 MB, with byte-identical output. A transcript
of ordinary records never reaches the branch.
---
 .../session-scanner-parse-cache.test.ts       | 39 +++++++++++++++++++
 .../ai-vault/session-scanner-parse-cache.ts   | 32 ++++++++++++---
 2 files changed, 66 insertions(+), 5 deletions(-)

diff --git a/src/main/ai-vault/session-scanner-parse-cache.test.ts b/src/main/ai-vault/session-scanner-parse-cache.test.ts
index 6ba5e3266..8434f26ba 100644
--- a/src/main/ai-vault/session-scanner-parse-cache.test.ts
+++ b/src/main/ai-vault/session-scanner-parse-cache.test.ts
@@ -128,6 +128,45 @@ describe('parseAgentSessionFileCached', () => {
     expect(incremental?.totalTokens).toBe(420)
   })
 
+  it('parses an oversized record without quadratic carry copying', async () => {
+    const root = await makeTempDir()
+    const path = join(root, 'aaaaaaaa-bbbb-4ccc-8ddd-eeeeeeeeeeee.jsonl')
+    // One tool result far larger than a stream chunk. Re-joining the held-over
+    // partial line per chunk copies O(record^2); the piece list joins once.
+    const recordBytes = 4 * 1024 * 1024
+    await writeFile(
+      path,
+      `${[
+        userRecord(0, 'question'),
+        assistantRecord(1, 'x'.repeat(recordBytes)),
+        assistantRecord(2, 'tail answer')
+      ].join('\n')}\n`
+    )
+
+    const originalConcat = Buffer.concat
+    let concatenatedBytes = 0
+    Buffer.concat = ((list: readonly Uint8Array[], totalLength?: number) => {
+      const joined = originalConcat(list as Uint8Array[], totalLength)
+      concatenatedBytes += joined.length
+      return joined
+    }) as typeof Buffer.concat
+    try {
+      const stats = createSessionParseStats()
+      const parsed = await parseAgentSessionFileCached(
+        await claudeCandidate(path),
+        process.platform,
+        stats
+      )
+      expect(parsed).not.toBeNull()
+    } finally {
+      Buffer.concat = originalConcat
+    }
+
+    // Linear joins the record about once; the quadratic form copied many times
+    // that, growing with the square of the record size.
+    expect(concatenatedBytes).toBeLessThan(recordBytes * 4)
+  })
+
   it('shows a trailing unterminated line without double-counting it later', async () => {
     const root = await makeTempDir()
     const path = join(root, 'aaaaaaaa-bbbb-4ccc-8ddd-eeeeeeeeeeee.jsonl')
diff --git a/src/main/ai-vault/session-scanner-parse-cache.ts b/src/main/ai-vault/session-scanner-parse-cache.ts
index b7aab7704..b3e4448a4 100644
--- a/src/main/ai-vault/session-scanner-parse-cache.ts
+++ b/src/main/ai-vault/session-scanner-parse-cache.ts
@@ -311,12 +311,28 @@ async function consumeCompleteJsonlLines(args: {
 }): Promise<JsonlReadResult> {
   let consumedThrough = args.start
   let bytesRead = 0
-  let remainder: Buffer | null = null
+  // Why a piece list: re-joining the partial line with every chunk made one
+  // oversized record (a big tool result) cost O(record^2). Joining once, when a
+  // newline finally arrives, keeps it linear.
+  let remainderParts: Buffer[] = []
+  let remainderLength = 0
 
   const stream = createReadStream(args.path, { start: args.start })
   for await (const chunk of stream as AsyncIterable<Buffer>) {
     bytesRead += chunk.length
-    const data = remainder ? Buffer.concat([remainder, chunk]) : chunk
+    // Why check the chunk alone: the pieces held over are all mid-line, so none
+    // of them contains a newline.
+    if (!chunk.includes(NEWLINE_BYTE)) {
+      remainderParts.push(chunk)
+      remainderLength += chunk.length
+      continue
+    }
+    const data =
+      remainderLength > 0
+        ? Buffer.concat([...remainderParts, chunk], remainderLength + chunk.length)
+        : chunk
+    remainderParts = []
+    remainderLength = 0
     let lineStart = 0
     let newlineIndex = data.indexOf(NEWLINE_BYTE, lineStart)
     while (newlineIndex !== -1) {
@@ -329,13 +345,19 @@ async function consumeCompleteJsonlLines(args: {
       newlineIndex = data.indexOf(NEWLINE_BYTE, lineStart)
     }
     consumedThrough += lineStart
-    // Copy the tail so retaining it doesn't pin the whole chunk buffer.
-    remainder = lineStart < data.length ? Buffer.from(data.subarray(lineStart)) : null
+    if (lineStart < data.length) {
+      // Copy the tail so retaining it doesn't pin the whole chunk buffer.
+      remainderParts = [Buffer.from(data.subarray(lineStart))]
+      remainderLength = data.length - lineStart
+    }
   }
 
+  const trailingPartialLine =
+    remainderLength > 0 ? Buffer.concat(remainderParts, remainderLength).toString('utf-8') : null
+
   return {
     consumedThrough,
-    trailingPartialLine: remainder && remainder.length > 0 ? remainder.toString('utf-8') : null,
+    trailingPartialLine,
     bytesRead
   }
 }