diff --git a/src/shared/child-process/bounded-output-sink.test.ts b/src/shared/child-process/bounded-output-sink.test.ts
new file mode 100644
index 00000000000..c94458e7211
--- /dev/null
+++ b/src/shared/child-process/bounded-output-sink.test.ts
@@ -0,0 +1,34 @@
+import { describe, expect, it } from 'vitest'
+import { createOutputSink } from './bounded-output-sink'
+
+describe('bounded process output', () => {
+  it('can read empty output and continue collecting', () => {
+    const sink = createOutputSink(10)
+    expect(sink.text()).toBe('')
+    sink.write('one')
+    expect(sink.text()).toBe('one')
+    sink.write('two')
+    expect(sink.text()).toBe('onetwo')
+    expect(sink.truncated()).toBe(false)
+  })
+
+  it('decodes UTF-8 across every chunk boundary and byte limit', () => {
+    const bytes = Buffer.from('a💻é\r\nb')
+    for (let split = 0; split <= bytes.length; split += 1) {
+      for (let cap = 0; cap <= bytes.length + 1; cap += 1) {
+        const sink = createOutputSink(cap)
+        sink.write(bytes.subarray(0, split))
+        sink.write(bytes.subarray(split))
+        expect(sink.text()).toBe(bytes.subarray(0, cap).toString('utf8'))
+        expect(sink.truncated()).toBe(bytes.length > cap)
+      }
+    }
+  })
+
+  it('clips a single string chunk at the byte limit', () => {
+    const sink = createOutputSink(3)
+    sink.write('a💻')
+    expect(sink.text()).toBe('a�')
+    expect(sink.truncated()).toBe(true)
+  })
+})
diff --git a/src/shared/child-process/bounded-output-sink.ts b/src/shared/child-process/bounded-output-sink.ts
index 195e466fdf6..8e1a9309978 100644
--- a/src/shared/child-process/bounded-output-sink.ts
+++ b/src/shared/child-process/bounded-output-sink.ts
@@ -25,7 +25,10 @@ export function createOutputSink(maxBytes: number): {
       chunks.push(chunk.length > remaining ? chunk.subarray(0, remaining) : chunk)
       bytes += chunk.length
     },
-    text: () => Buffer.concat(chunks).toString('utf8'),
+    text: () =>
+      chunks.length === 0
+        ? ''
+        : (chunks.length === 1 ? chunks[0] : Buffer.concat(chunks)).toString('utf8'),
     // Why: callers that parse the output need to tell a short answer from a
     // clipped one -- truncated JSON or JSONL parses as a smaller valid result.
     truncated: () => bytes > maxBytes