From c1850e3cc7472b76c7ffd4a710f7e8d40e18fdd3 Mon Sep 17 00:00:00 2001
From: Neil <4138956+nwparker@users.noreply.github.com>
Date: Fri, 4 Sep 2026 14:33:39 -0700
Subject: [PATCH] perf(claude-usage): reject non-assistant transcript lines
 before parsing them (#18640)

Only assistant records carry usage, but the parser ran JSON.parse on every line
first and checked the type after. Claude transcripts interleave user and
tool-result lines that routinely embed whole files or captured command output,
so the scanner built and discarded a full object graph for each of them.

A substring gate on the line rejects them first: median 50.5ms to 8.4ms over
20000 realistic lines.
---
 ...transcript-record-parser-prefilter.test.ts | 40 +++++++++++++++++++
 .../claude-usage/transcript-record-parser.ts  | 20 ++++++++++
 2 files changed, 60 insertions(+)
 create mode 100644 src/main/claude-usage/transcript-record-parser-prefilter.test.ts

diff --git a/src/main/claude-usage/transcript-record-parser-prefilter.test.ts b/src/main/claude-usage/transcript-record-parser-prefilter.test.ts
new file mode 100644
index 00000000000..992b4ba8fd3
--- /dev/null
+++ b/src/main/claude-usage/transcript-record-parser-prefilter.test.ts
@@ -0,0 +1,40 @@
+import { describe, expect, it } from 'vitest'
+import { parseClaudeUsageRecord } from './transcript-record-parser'
+
+function assistantLine(overrides: Record<string, unknown> = {}): string {
+  return JSON.stringify({
+    type: 'assistant',
+    sessionId: 'session-1',
+    timestamp: '2026-01-01T00:00:00.000Z',
+    message: { usage: { input_tokens: 3, output_tokens: 5 } },
+    ...overrides
+  })
+}
+
+describe('assistant-record prefilter', () => {
+  it('still parses an ordinary assistant record', () => {
+    expect(parseClaudeUsageRecord(assistantLine())?.inputTokens).toBe(3)
+  })
+
+  it('rejects a user record that never mentions assistant', () => {
+    const userLine = JSON.stringify({
+      type: 'user',
+      sessionId: 'session-1',
+      timestamp: '2026-01-01T00:00:00.000Z',
+      message: { content: 'x'.repeat(200) }
+    })
+
+    expect(parseClaudeUsageRecord(userLine)).toBeNull()
+  })
+
+  it('rejects a non-assistant record that happens to contain the word assistant', () => {
+    const userLine = JSON.stringify({
+      type: 'user',
+      sessionId: 'session-1',
+      timestamp: '2026-01-01T00:00:00.000Z',
+      message: { content: 'ask the assistant about this' }
+    })
+
+    expect(parseClaudeUsageRecord(userLine)).toBeNull()
+  })
+})
diff --git a/src/main/claude-usage/transcript-record-parser.ts b/src/main/claude-usage/transcript-record-parser.ts
index 59b0e75be39..2ad51f3ee90 100644
--- a/src/main/claude-usage/transcript-record-parser.ts
+++ b/src/main/claude-usage/transcript-record-parser.ts
@@ -84,10 +84,30 @@ function dedupeClaudeUsageTurns(
   return deduped
 }
 
+/**
+ * Necessary condition for `JSON.parse(line).type === 'assistant'`, checked before the parse.
+ *
+ * Sound for any transcript written by a standard JSON serializer: `JSON.stringify` (which writes
+ * these files) escapes only quotes, backslashes and control characters, never ASCII letters, so
+ * the decoded value can only be `assistant` if the line spells it literally. The gate over-admits
+ * freely — the `parsed.type` check below stays authoritative.
+ *
+ * A `\u`-escape fallback was measured and rejected: it costs a second full-line scan and made
+ * transcripts whose tool results contain control characters 1.43x slower overall.
+ */
+function mayEncodeAssistantType(line: string): boolean {
+  return line.includes('assistant')
+}
+
 function parseClaudeUsageSourceRecord(
   line: string,
   fallbackSessionId: string | null = null
 ): ClaudeUsageParsedSourceTurn | null {
+  // Only assistant records carry usage, but transcripts interleave user/tool-result lines that
+  // routinely embed whole files. Reject those before paying for a full parse.
+  if (!mayEncodeAssistantType(line)) {
+    return null
+  }
   let parsed: ClaudeUsageSourceRecord
   try {
     parsed = JSON.parse(line) as ClaudeUsageSourceRecord