fix(tui): drop duplicated in-transcript thinking placeholder

internal/tui/model_smoke_test.go

@@ -74,6 +74,35 @@ func TestRenderStreamingTurn(t *testing.T) {
 	}
 }
 
+// TestEmptyStreamingTurnHidden verifies that while a streamed assistant turn
+// has no content yet, the transcript shows neither a "thinking…" placeholder
+// nor a bare odek block — the top-bar spinner is the only progress signal.
+// The turn block appears atomically once the first real content arrives.
+func TestEmptyStreamingTurnHidden(t *testing.T) {
+	m := newTestModel()
+	m.msgs = append(m.msgs,
+		message{role: roleUser, content: "hi"},
+		message{role: roleAsst, streaming: true},
+	)
+	m.curIdx = 1
+	m.busy = true
+
+	out := plain(m.conversation())
+	if strings.Contains(out, "thinking…") {
+		t.Errorf("transcript shows duplicated thinking placeholder:\n%s", out)
+	}
+	if strings.Contains(out, "⬡ odek") {
+		t.Errorf("empty streaming turn renders a bare odek block:\n%s", out)
+	}
+
+	// First reasoning chunk makes the turn block visible.
+	m.handleEvent(client.Event{Type: "thinking", Content: "hmm"})
+	out = plain(m.conversation())
+	if !strings.Contains(out, "⬡ odek") || !strings.Contains(out, "hmm") {
+		t.Errorf("turn block did not appear after first content:\n%s", out)
+	}
+}
+
 // TestApprovalAndAutocompleteRender ensures the approval panel and the @-popup
 // render at full and narrow widths without panicking.
 func TestApprovalAndAutocompleteRender(t *testing.T) {

internal/tui/view.go

@@ -283,6 +283,11 @@ func (m *Model) conversation() string {
 		blocks = append(blocks, m.convPrefix)
 	}
 	for i := tail; i < len(m.msgs); i++ {
+		if emptyStreamingTurn(m.msgs[i]) {
+			// Nothing to show yet: the top-bar spinner is the sole
+			// progress signal — no bare odek block, no placeholder.
+			continue
+		}
 		s, r := m.renderMessage(m.msgs[i], i, lineOffset)
 		blocks = append(blocks, s)
 		refs = append(refs, r...)
@@ -297,6 +302,23 @@ func (m *Model) conversation() string {
 	return strings.Join(blocks, "\n\n")
 }
 
+// emptyStreamingTurn reports whether a message is an in-flight assistant
+// turn without any visible content yet — no tokens, reasoning, or steps.
+func emptyStreamingTurn(msg message) bool {
+	if msg.role != roleAsst || !msg.streaming {
+		return false
+	}
+	if strings.TrimSpace(msg.content) != "" || strings.TrimSpace(msg.thinking) != "" || len(msg.steps) > 0 {
+		return false
+	}
+	for _, it := range msg.items {
+		if !it.thinking || strings.TrimSpace(it.text) != "" {
+			return false
+		}
+	}
+	return true
+}
+
 func (m *Model) renderMessage(msg message, msgIdx, lineOffset int) (string, []stepRef) {
 	th := m.th
 	// Pre-styled cards (e.g. /stats) render verbatim — no label, no bar, and
@@ -320,9 +342,6 @@ func (m *Model) renderMessage(msg message, msgIdx, lineOffset int) (string, []st
 		if !msg.streaming && msg.rendered != "" {
 			content = msg.rendered
 		}
-		if strings.TrimSpace(content) == "" && msg.streaming {
-			content = th.thinkStyle.Render(m.sp.View() + " thinking…")
-		}
 		// Compose the turn body from the chronological timeline: reasoning
 		// blocks and tool steps interleaved in arrival order.
 		items := msg.items