diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
index 4268fb6..117a03c 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
@@ -291,12 +291,40 @@ class ParsedReplies {
 
     private val chunks = ConcurrentHashMap<String, List<String>>()
 
-    fun blocksOf(text: String): List<String> = blocks.computeIfAbsent(text) { markdownBlocks(it) }
+    private val ready = ConcurrentHashMap.newKeySet<String>()
+
+    fun blocksOf(text: String): List<String> =
+        blocks.computeIfAbsent(text) {
+            DebugStats.timed("markdown split into blocks") { markdownBlocks(it) }
+        }
 
     /** How a long user message divides into slices; cached for the same reason as [blocksOf]. */
-    fun chunksOf(text: String): List<String> = chunks.computeIfAbsent(text) { userChunks(it) }
+    fun chunksOf(text: String): List<String> =
+        chunks.computeIfAbsent(text) {
+            DebugStats.timed("user message cut into slices") { userChunks(it) }
+        }
 
-    fun partsOf(text: String): List<MessagePart> = parts.computeIfAbsent(text) { messageParts(it) }
+    /**
+     * Whether [warm] has made everything drawing [text] as blocks will look up.
+     *
+     * What the flatten asks before drawing a reply that way. Splitting costs a parse of the whole
+     * message and the flatten runs on the composing thread -- so a reply not marked yet stays
+     * whole, drawing the parse it already has, until the screen has warmed it and re-flattens. An
+     * explicit mark rather than a peek into [blocksOf]'s cache, because a message with memory notes
+     * is warmed as its *parts*: nothing ever splits its full text, and inferring readiness from the
+     * cache left exactly that message unsplittable forever, re-warmed on every fold.
+     */
+    fun splitReady(text: String): Boolean = text in ready
+
+    /** The other half of [splitReady]; [warm] calls it once a message's parses exist. */
+    fun markSplitReady(text: String) {
+        ready.add(text)
+    }
+
+    fun partsOf(text: String): List<MessagePart> =
+        parts.computeIfAbsent(text) {
+            DebugStats.timed("message cut into parts") { messageParts(it) }
+        }
 
     /** The parse of [text] -- the one made ahead, or one made now. */
     fun of(text: String): State =
@@ -326,5 +354,6 @@ class ParsedReplies {
         blocks.clear()
         parts.clear()
         chunks.clear()
+        ready.clear()
     }
 }
diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/MessageBlocks.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/MessageBlocks.kt
index d65bb10..f638cde 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/MessageBlocks.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/MessageBlocks.kt
@@ -4,6 +4,8 @@ import androidx.compose.foundation.layout.Arrangement
 import androidx.compose.foundation.layout.Column
 import androidx.compose.foundation.layout.fillMaxWidth
 import androidx.compose.runtime.Composable
+import androidx.compose.runtime.LaunchedEffect
+import androidx.compose.runtime.mutableStateOf
 import androidx.compose.runtime.remember
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.draw.drawWithContent
@@ -11,6 +13,8 @@ import androidx.compose.ui.graphics.graphicsLayer
 import androidx.compose.ui.unit.dp
 import com.mikepenz.markdown.model.State
 import com.mikepenz.markdown.model.parseMarkdown
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.withContext
 
 /**
  * A message's top-level markdown blocks, cut where the parser says the blocks are.
@@ -70,7 +74,23 @@ fun BlockedMarkdown(
     modifier: Modifier = Modifier,
     live: Boolean = false,
 ) {
-    val blocks = remember(text) { replies.blocksOf(text) }
+    // The first split is inline for the reason [parsedMarkdown]'s first parse is: the row must
+    // have its real height in its first frame. Every split after that is a delta landing, and it
+    // runs off the composing thread with the message drawing the split it already has until the
+    // new one arrives -- when this recomputed wherever composition ran, one streamed reply cost
+    // 815 whole-message parses and 2.9 seconds of them, a few milliseconds per delta, on the
+    // thread that draws. Not through [replies]: a reply mid-stream is a different text per
+    // delta, and each would leave a cache entry nothing reads again.
+    val split = remember { mutableStateOf(text to markdownBlocks(text)) }
+    LaunchedEffect(text) {
+        if (split.value.first == text) return@LaunchedEffect
+        split.value =
+            text to
+                withContext(Dispatchers.Default) {
+                    DebugStats.timed("blocks split while streaming") { markdownBlocks(text) }
+                }
+    }
+    val blocks = split.value.second
     if (blocks.size == 1) {
         MarkdownText(blocks.first(), replies, modifier)
         return
diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt
index 7551fb0..d5c4a3c 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt
@@ -343,7 +343,23 @@ fun SessionScreen(settings: ServerSettings, summary: SessionSummary, onBack: ()
     // What is actually drawn: the transcript with runs of adjacent tool
     // calls folded into one row each, flattened into the list's units.
     val rows = remember(items) { groupToolRuns(items) }
-    val units = remember(rows, expandedNotes) { transcriptUnits(rows, replies, expandedNotes) }
+    // Bumped when a cold reply's parses become ready, so the flatten runs again and can split
+    // it; see [unwarmedReplies].
+    var warmedTick by remember { mutableIntStateOf(0) }
+    val units =
+        remember(rows, expandedNotes, warmedTick) { transcriptUnits(rows, replies, expandedNotes) }
+    // The reply that just finished streaming is the one row whose parses nobody has made: pages
+    // warm before their fold lands, but nothing warms live deltas. Off the composing thread,
+    // then the tick re-flattens -- so settling never costs a whole-message parse in a frame.
+    // Re-launched per fold and almost always finds nothing; during streaming the last row is
+    // unsettled and not wanted.
+    LaunchedEffect(rows) {
+        val cold = unwarmedReplies(rows, replies)
+        if (cold.isNotEmpty()) {
+            warm(replies, cold)
+            warmedTick++
+        }
+    }
     // The same list, readable from effects launched before this composition: an effect's closure
     // keeps the values of the composition that launched it, and both the anchor saver and the
     // restore need the units as they are *now*.
diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptItems.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptItems.kt
index a04bfd5..bec357a 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptItems.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptItems.kt
@@ -576,5 +576,8 @@ suspend fun warm(replies: ParsedReplies, rows: List<TranscriptItem>) {
             }
         }
         if (texts.isNotEmpty()) replies.warm(texts)
+        // After the parses exist, not before: [ParsedReplies.splitReady] is the flatten's
+        // licence to draw these as blocks on the composing thread.
+        rows.forEach { if (it is TranscriptItem.AssistantMsg) replies.markSplitReady(it.text) }
     }
 }
diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptUnits.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptUnits.kt
index 35187ba..a4da80a 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptUnits.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/TranscriptUnits.kt
@@ -164,6 +164,7 @@ fun transcriptUnits(
     replies: ParsedReplies,
     openNotes: Set<Long>,
 ): List<TranscriptUnit> {
+    val started = System.nanoTime()
     val units = ArrayList<TranscriptUnit>(rows.size)
     rows.forEachIndexed { index, row ->
         val rowGap = if (index == 0) 0.dp else TRANSCRIPT_SPACING
@@ -204,7 +205,9 @@ fun transcriptUnits(
                     )
             }
         } else if (
-            item is TranscriptItem.AssistantMsg && (item.settled || index != rows.lastIndex)
+            item is TranscriptItem.AssistantMsg &&
+                splitWanted(item, index, rows.lastIndex) &&
+                replies.splitReady(item.text)
         ) {
             var ordinal = 0
             fun gap() = if (ordinal == 0) rowGap else BLOCK_SPACING
@@ -227,9 +230,37 @@ fun transcriptUnits(
     }
     units.reverse()
     reportDuplicateKeys(units)
+    // Timed because this runs per fold on the composing thread: "loading messages feels bumpy"
+    // is this number growing, and it was invisible until it was written down.
+    DebugStats.record("units flattened", System.nanoTime() - started)
     return units
 }
 
+/**
+ * Whether this reply should be drawn as blocks: settled, or anywhere but the newest row.
+ *
+ * Wanting is not being ready -- the flatten also asks [ParsedReplies.splitReady], and the two
+ * questions are separate because they are answered by different things: this one by the fold, the
+ * other by whether [warm] has run for the text. [unwarmedReplies] is the gap between them.
+ */
+private fun splitWanted(item: TranscriptItem.AssistantMsg, index: Int, lastIndex: Int) =
+    item.settled || index != lastIndex
+
+/**
+ * The replies among [rows] that should draw as blocks but whose parses are not made yet.
+ *
+ * Normally empty: every page's rows are warmed before the fold lands. The one row that can be cold
+ * is the reply that just finished streaming -- nothing warms live deltas, so at the moment its turn
+ * ends its split would cost a whole-message parse on the composing thread. The session screen warms
+ * what this returns off-thread and re-flattens, so the whole-to-blocks swap always composes against
+ * ready parses.
+ */
+fun unwarmedReplies(rows: List<TranscriptRow>, replies: ParsedReplies): List<TranscriptItem> =
+    rows.mapIndexedNotNull { index, row ->
+        val item = (row as? TranscriptRow.Single)?.item as? TranscriptItem.AssistantMsg
+        item?.takeIf { splitWanted(it, index, rows.lastIndex) && !replies.splitReady(it.text) }
+    }
+
 /**
  * Above this many characters, a user message is drawn in slices rather than as one bubble.
  *
@@ -240,8 +271,14 @@ fun transcriptUnits(
  */
 const val USER_SPLIT_CHARS = 4000
 
-/** Roughly how much text one slice holds -- bounded, like a markdown block, is the whole point. */
-private const val USER_CHUNK_CHARS = 2500
+/**
+ * Roughly how much text one slice holds -- bounded, like a markdown block, is the whole point.
+ *
+ * About one viewport of wrapped text: a slice is composed whole in the frame it scrolls into, so
+ * its size is a frame-budget decision, and one screenful keeps that to a few milliseconds on the
+ * phone. Smaller buys nothing -- the seams are free -- but the units multiply.
+ */
+private const val USER_CHUNK_CHARS = 1000
 
 /**
  * A long user message cut at line starts into slices of roughly [USER_CHUNK_CHARS].