diff --git a/internal/editor/chunked_buffer.go b/internal/editor/chunked_buffer.go
index 4f79591..44c01fd 100644
--- a/internal/editor/chunked_buffer.go
+++ b/internal/editor/chunked_buffer.go
@@ -39,17 +39,19 @@ const (
 // therefore no stale-disk re-read (the old fixed-slot model could re-read a
 // shifted tail chunk from disk and clobber in-memory edits).
 //
-// Chunks may grow/shrink with edits and are not rebalanced; that is fine for
-// correctness (prefix sums) and keeps edits local to the affected chunk(s)
-// rather than copying the whole file.
+// Chunks grow/shrink with edits; Insert splits any chunk that grows past
+// twice the target size (see Insert), so the per-edit copy cost stays bounded
+// by O(chunkSize) even under sustained typing at one spot. Shrunken (even
+// empty) chunks are left in place: the chunk count never grows with deletes,
+// all readers walk actual lengths, and removing chunks would be pure churn.
 type ChunkedBuffer struct {
-	filename   string
-	chunkSize  int // target chunk size (e.g. 64 KB); actual chunks may vary
-	fileLen    int64     // total content length
-	chunks     [][]byte  // ordered; chunks[i] is the i-th chunk
-	dirty      bool      // true if buffer has been modified
-	FS         pool.FileSystem // filesystem for reads
-	basePath   string          // base path for file resolution
+	filename  string
+	chunkSize int             // target chunk size (e.g. 64 KB); actual chunks may vary
+	fileLen   int64           // total content length
+	chunks    [][]byte        // ordered; chunks[i] is the i-th chunk
+	dirty     bool            // true if buffer has been modified
+	FS        pool.FileSystem // filesystem for reads
+	basePath  string          // base path for file resolution
 
 	// line index is built asynchronously
 	LineIndex *types.LineIndex
@@ -320,7 +322,15 @@ func (cb *ChunkedBuffer) markDirtyChunk(idx int) {
 	}
 }
 
-// Insert inserts text at byte position pos, splicing only the affected chunk.
+// Insert inserts text at byte position pos, splicing only the affected
+// chunk. If the result grows past twice the target chunk size (sustained
+// typing at one spot, or a large paste), the chunk is split in half so that
+// chunks stay O(chunkSize) and every future edit remains a bounded
+// O(chunkSize) copy. The split cut is an arbitrary byte offset, like the
+// chunk boundaries created by SetContent: chunk boundaries may fall inside
+// multi-byte sequences, which is fine because every reader reassembles whole
+// windows from chunk bytes (windows are always line-aligned, i.e. on rune
+// boundaries).
 func (cb *ChunkedBuffer) Insert(pos int, text string) {
 	if len(text) == 0 {
 		return
@@ -334,8 +344,16 @@ func (cb *ChunkedBuffer) Insert(pos int, text string) {
 	}
 	idx, local := cb.chunkForPos(pos)
 	if idx < 0 {
-		// Empty buffer: create the first chunk.
-		cb.chunks = append(cb.chunks, []byte(text))
+		// Empty buffer: chunk the new text directly so a large first paste
+		// does not create one oversized chunk.
+		cb.chunks = nil
+		for start := 0; start < len(text); start += cb.chunkSize {
+			end := start + cb.chunkSize
+			if end > len(text) {
+				end = len(text)
+			}
+			cb.chunks = append(cb.chunks, []byte(text[start:end]))
+		}
 		cb.fileLen = int64(len(text))
 		cb.markDirtyChunk(0)
 		return
@@ -345,6 +363,16 @@ func (cb *ChunkedBuffer) Insert(pos int, text string) {
 	newChunk = append(newChunk, chunk[:local]...)
 	newChunk = append(newChunk, text...)
 	newChunk = append(newChunk, chunk[local:]...)
+	if len(newChunk) > 2*cb.chunkSize {
+		// Split in half and insert the second half after idx.
+		cut := len(newChunk) / 2
+		second := make([]byte, len(newChunk)-cut)
+		copy(second, newChunk[cut:])
+		newChunk = newChunk[:cut]
+		cb.chunks = append(cb.chunks, nil)
+		copy(cb.chunks[idx+2:], cb.chunks[idx+1:]) // overlap-safe (memmove)
+		cb.chunks[idx+1] = second
+	}
 	cb.chunks[idx] = newChunk
 	cb.fileLen += int64(len(text))
 	cb.markDirtyChunk(idx)
diff --git a/internal/editor/chunked_buffer_test.go b/internal/editor/chunked_buffer_test.go
index c1fe231..28caf92 100644
--- a/internal/editor/chunked_buffer_test.go
+++ b/internal/editor/chunked_buffer_test.go
@@ -191,3 +191,91 @@ func TestAppendAtEnd(t *testing.T) {
 		t.Fatalf("FileLen=%d, want 9", cb.FileLen())
 	}
 }
+
+// assertMaxChunkSize verifies the chunking invariant: no chunk may exceed
+// twice the target size (Insert splits oversized chunks in half).
+func assertMaxChunkSize(t *testing.T, cb *ChunkedBuffer) {
+	t.Helper()
+	for i, c := range cb.chunks {
+		if len(c) > 2*cb.chunkSize {
+			t.Fatalf("chunk %d is %d bytes, want <= %d (2 x %d)", i, len(c), 2*cb.chunkSize, cb.chunkSize)
+		}
+	}
+}
+
+// TestInsert_SustainedTyping_SplitsOversizedChunks simulates sustained typing
+// at one spot (many small inserts into the same chunk). Without the split, the
+// target chunk grows without bound and every edit copies the whole grown
+// chunk (quadratic total). With the split, chunks stay O(chunkSize).
+func TestInsert_SustainedTyping_SplitsOversizedChunks(t *testing.T) {
+	const chunkSize = 64
+	cb := NewChunkedBuffer("/test.txt", chunkSize, nil, "")
+	seed := strings.Repeat("seed line\n", chunkSize/8) // 64 bytes = 1 chunk
+	cb.SetContent([]byte(seed))
+
+	// Type 2000 characters, each inserted immediately after the previous one,
+	// starting at position 10 (inside chunk 0).
+	var typed strings.Builder
+	pos := 10
+	for i := 0; i < 2000; i++ {
+		r := rune('a' + i%26)
+		cb.Insert(pos, string(r))
+		pos++
+		typed.WriteRune(r)
+		if i%200 == 0 {
+			assertMaxChunkSize(t, cb)
+		}
+	}
+	assertMaxChunkSize(t, cb)
+	want := seed[:10] + typed.String() + seed[10:]
+	full, _ := cb.FullContent()
+	if full != want {
+		t.Fatalf("FullContent mismatch after sustained inserts (len %d vs %d)", len(full), len(want))
+	}
+	if cb.FileLen() != int64(len(want)) {
+		t.Fatalf("FileLen=%d, want %d", cb.FileLen(), len(want))
+	}
+}
+
+// TestInsert_LargePaste_EmptyBuffer verifies that a large first paste into an
+// empty buffer is chunked (not one oversized chunk) and reconstructs exactly.
+func TestInsert_LargePaste_EmptyBuffer(t *testing.T) {
+	cb := NewChunkedBuffer("/test.txt", 64*1024, nil, "")
+	paste := strings.Repeat("0123456789", 100*1024) // 1 MB
+	cb.Insert(0, paste)
+	assertMaxChunkSize(t, cb)
+	if got := len(cb.chunks); got != 16 {
+		t.Fatalf("expected 16 chunks for 1MB/64KB, got %d", got)
+	}
+	full, _ := cb.FullContent()
+	if full != paste {
+		t.Fatalf("FullContent mismatch after large paste")
+	}
+}
+
+// TestInsert_SplitContentIntegrity checks that Content stays exact across the
+// NEW chunk boundary created by a split, including spans that cross it.
+func TestInsert_SplitContentIntegrity(t *testing.T) {
+	const chunkSize = 64
+	cb := NewChunkedBuffer("/test.txt", chunkSize, nil, "")
+	cb.SetContent([]byte(strings.Repeat("x", 64)))
+	// 200 inserts of 2 chars at pos 0: the first chunk grows to 464+ bytes and
+	// must be split several times.
+	for i := 0; i < 200; i++ {
+		cb.Insert(0, "AB")
+	}
+	want := strings.Repeat("AB", 200) + strings.Repeat("x", 64)
+	full, _ := cb.FullContent()
+	if full != want {
+		t.Fatalf("FullContent mismatch: got %d bytes, want %d", len(full), len(want))
+	}
+	// Span across several chunk boundaries near the front.
+	seg := cb.Content(0, 200)
+	if seg != want[:200] {
+		t.Fatalf("Content(0,200) mismatch across split boundary")
+	}
+	seg = cb.Content(50, 300)
+	if seg != want[50:300] {
+		t.Fatalf("Content(50,300) mismatch across split boundary")
+	}
+}