fix(desktop): do not sandwich structured mid-turn rows with inflight dump

Skip pure-text inflight.assistant projections when the transcript already
has reasoning/tool-call structure, and only overlay journal answer text
on strict extension.

Fixes #76444

apps/desktop/src/app/session/hooks/use-session-actions/utils.test.ts

@@ -1120,4 +1120,41 @@ describe('appendLiveSessionProjection', () => {
 
     expect(appendLiveSessionProjection(stored, { session_id: 'runtime-1' })).toBe(stored)
   })
+
+  it('does not sandwich a structured mid-turn row with the inflight flat dump (#76444)', () => {
+    const stored: ChatMessage[] = [
+      msg('stored-user', 'user', 'do the work'),
+      {
+        id: 'live-assistant',
+        role: 'assistant',
+        pending: true,
+        parts: [
+          { type: 'reasoning', text: 'thinking about tools' },
+          { type: 'tool-call', toolCallId: 'c1', toolName: 'terminal', args: {} },
+          { type: 'text', text: 'partial' }
+        ]
+      }
+    ]
+
+    const restored = appendLiveSessionProjection(stored, {
+      session_id: 'runtime-1',
+      inflight: {
+        user: 'do the work',
+        // Flat dump includes thinking chatter + tool narration — longer than
+        // the answer text alone, which is how the sandwich used to grow.
+        assistant: 'thinking about tools\nRan terminal\npartial and more dump',
+        streaming: true
+      }
+    })
+
+    const assistants = restored.filter(message => message.role === 'assistant')
+    expect(assistants).toHaveLength(1)
+    expect(assistants[0].id).toBe('live-assistant')
+    expect(assistants[0].parts.some(part => part.type === 'reasoning')).toBe(true)
+    expect(assistants[0].parts.some(part => part.type === 'tool-call')).toBe(true)
+    // Answer text stays the structured row's text, not the dump.
+    expect(assistants[0].parts.filter(part => part.type === 'text').map(part => ('text' in part ? part.text : ''))).toEqual([
+      'partial'
+    ])
+  })
 })

apps/desktop/src/app/session/hooks/use-session-actions/utils.ts

@@ -47,6 +47,27 @@ function withAppendedText(message: ChatMessage, suffix: string): ChatMessage {
   return appended ? { ...message, parts } : message
 }
 
+/** Reasoning / tool-call parts that the gateway inflight dump cannot express. */
+function hasStructuralParts(message: ChatMessage): boolean {
+  return message.parts.some(part => part.type === 'reasoning' || part.type === 'tool-call')
+}
+
+/**
+ * True when `next` is a pure forward extension of the previous *answer* text.
+ * Empty previous answer never accepts a dump as an extension — that is how the
+ * mid-turn inflight flat dump used to sandwich structured rows (#76444).
+ */
+export function isStrictAnswerTextExtension(next: string, previous: string): boolean {
+  const n = next.trim()
+  const p = previous.trim()
+
+  if (!p || !n) {
+    return false
+  }
+
+  return n.startsWith(p)
+}
+
 /**
  * Carry structural parts an authoritative row cannot express.
  *
@@ -273,11 +294,33 @@ export function reconcileResumeMessages(nextMessages: ChatMessage[], previousMes
     // for structural carry-over. Attachment refs and image re-appending stay on
     // the strict equality path — they reconcile a SETTLED row, and a growing
     // row is by definition not settled.
+    //
+    // Structured local assistant + text-only live projection (gateway inflight
+    // dump) at the same ordinal is still the same turn even when answer text is
+    // still empty — without this the dump would drop structure and reappear as
+    // a plain-text sandwich (#76444).
     const sameTurn =
-      sameText || (nextText.length > 0 && previousTrimmed.length > 0 && nextText.startsWith(previousTrimmed))
+      sameText ||
+      (nextText.length > 0 && previousTrimmed.length > 0 && isStrictAnswerTextExtension(nextText, previousTrimmed)) ||
+      (message.role === 'assistant' &&
+        previous.role === 'assistant' &&
+        hasStructuralParts(previous) &&
+        !hasStructuralParts(message))
 
     if (sameTurn) {
       preserved = preserveStructuralParts(preserved, previous)
+
+      // Never replace structured answer text with a non-extending flat dump.
+      if (
+        message.role === 'assistant' &&
+        hasStructuralParts(previous) &&
+        !hasStructuralParts(message) &&
+        !isStrictAnswerTextExtension(nextText, previousVisibleText)
+      ) {
+        const nonText = preserved.parts.filter(part => part.type !== 'text')
+        const priorAnswer = previous.parts.filter(part => part.type === 'text')
+        preserved = { ...preserved, parts: [...nonText, ...priorAnswer] }
+      }
     }
 
     if (
@@ -646,14 +689,27 @@ export function appendLiveSessionProjection(
 
   // Keep a pending assistant boundary even before the first delta when a
   // queued user turn follows it. This preserves the two distinct turns.
+  //
+  // When the transcript already holds a structured mid-turn assistant row
+  // (reasoning / tool-call parts from the live stream or journal), do NOT
+  // append a pure-text projection of `inflight.assistant` — that flat dump
+  // re-renders thinking as answer text and sandwiches the structured parts
+  // (#76444). Errors still force a projection so the failure is visible.
+  const lastAssistant = [...messages].reverse().find(message => message.role === 'assistant')
+  const turnAlreadyStructured = Boolean(lastAssistant && hasStructuralParts(lastAssistant))
+
   if (inflightAssistant || inflightStreaming || inflightError || (inflightUser && queuedUser)) {
-    projected.push({
-      id: `assistant-stream-${sessionId}`,
-      role: 'assistant',
-      parts: inflightAssistant ? [assistantTextPart(inflightAssistant)] : [],
-      pending: inflightStreaming,
-      ...(inflightError ? { error: inflightError } : {})
-    })
+    if (turnAlreadyStructured && !inflightError) {
+      // Structure is authoritative; skip the text-only dump row.
+    } else {
+      projected.push({
+        id: `assistant-stream-${sessionId}`,
+        role: 'assistant',
+        parts: inflightAssistant ? [assistantTextPart(inflightAssistant)] : [],
+        pending: inflightStreaming,
+        ...(inflightError ? { error: inflightError } : {})
+      })
+    }
   }
 
   if (queuedUser) {

apps/desktop/src/lib/inflight-turn-journal.ts

@@ -254,6 +254,10 @@ function assistantTextLength(message: ChatMessage): number {
  * BASE row's id so live deltas keep appending to the row the stream handler
  * already targets.
  */
+function hasStructuralParts(message: ChatMessage): boolean {
+  return message.parts.some(part => part.type === 'reasoning' || part.type === 'tool-call')
+}
+
 function overlayProjectionRow(projection: ChatMessage, journalRow: ChatMessage): ChatMessage {
   // A projected error (retained failed turn) must survive the overlay.
   const error = journalRow.error ?? projection.error
@@ -271,7 +275,20 @@ function overlayProjectionRow(projection: ChatMessage, journalRow: ChatMessage):
 
   // Backend text is newer than the journal's last throttled write — swap it
   // into the journal's first text part, keeping tool calls and reasoning.
+  // When the journal already carries structure, only accept a *strict*
+  // extension of the answer text. A longer flat dump that starts with
+  // thinking chatter must not overwrite / insert as answer text (#76444).
   const projectionText = chatMessageText(projection)
+  const journalText = chatMessageText(journalRow).trim()
+
+  if (hasStructuralParts(journalRow)) {
+    const next = projectionText.trim()
+
+    if (!journalText || !next.startsWith(journalText)) {
+      return merged
+    }
+  }
+
   const parts: ChatMessagePart[] = []
   let textReplaced = false