diff --git a/src/contexts/conversation-runtime-context.test.tsx b/src/contexts/conversation-runtime-context.test.tsx
index 2c635ba5c5..602c9e324d 100644
--- a/src/contexts/conversation-runtime-context.test.tsx
+++ b/src/contexts/conversation-runtime-context.test.tsx
@@ -709,6 +709,14 @@ describe("ConversationRuntimeProvider delegation kickoff projection", () => {
     mockGetFolderConversation.mockImplementation(() => new Promise(() => {}))
   })
 
+  /** The live message the strip stands in for: one that IS showing the reply. */
+  const streamingReply: LiveMessage = {
+    id: "lm-streaming",
+    role: "assistant",
+    content: [{ type: "text", text: "working on it" }],
+    startedAt: 0,
+  }
+
   it("synthesizes the kickoff user turn (and strips the persisted reply) while the transcript has no user turn yet", async () => {
     // DB lags: only a partial assistant turn is persisted, no user turn.
     mockGetFolderConversation.mockResolvedValueOnce(
@@ -721,7 +729,7 @@ describe("ConversationRuntimeProvider delegation kickoff projection", () => {
       api().setLiveOwnsActiveTurn(99, true, "do the thing")
     })
     act(() => {
-      api().setLiveMessage(99, LIVE_MSG, true)
+      api().setLiveMessage(99, streamingReply, true)
     })
     await act(async () => {
       api().refetchDetail(99, { preserveLive: true })
@@ -744,6 +752,35 @@ describe("ConversationRuntimeProvider delegation kickoff projection", () => {
     ).toBe(false)
   })
 
+  it("keeps the persisted reply while the live message is showing nothing", async () => {
+    // The child's next turn has begun: `status_changed → prompting` put a fresh
+    // `content: []` live message on the connection and the viewer bridged it,
+    // but no chunk has arrived. Stripping the reply then leaves the dialog
+    // showing a prompt with nothing under it.
+    mockGetFolderConversation.mockResolvedValueOnce(
+      detailWithTurns([userTurn("u1"), assistantTurn("a1")])
+    )
+    renderProvider(<RuntimeCapture />)
+    const api = () => runtimeHolder.current!
+
+    act(() => {
+      api().setLiveOwnsActiveTurn(99, true, "do the thing")
+    })
+    act(() => {
+      api().setLiveMessage(99, LIVE_MSG, true)
+    })
+    await act(async () => {
+      api().refetchDetail(99, { preserveLive: true })
+      await Promise.resolve()
+    })
+
+    expect(
+      api()
+        .getTimelineTurns(99)
+        .map((t) => t.turn.id)
+    ).toEqual(["u1", "a1"])
+  })
+
   it("uses the real persisted user turn instead of synthesizing once it has landed", async () => {
     mockGetFolderConversation.mockResolvedValueOnce(
       detailWithTurns([userTurn("u1"), assistantTurn("a1")])
diff --git a/src/stores/conversation-runtime-store.ts b/src/stores/conversation-runtime-store.ts
index 64fa185355..22eb517231 100644
--- a/src/stores/conversation-runtime-store.ts
+++ b/src/stores/conversation-runtime-store.ts
@@ -2780,7 +2780,7 @@ interface TimelinePrefixDeps {
   optimisticTurns: MessageTurn[]
   liveOwnsActiveTurn: boolean
   delegationKickoffText: string | null
-  hasLiveMessage: boolean
+  liveShowsReply: boolean
   liveStartedAt: number | null
 }
 interface TimelinePrefixEntry {
@@ -3029,7 +3029,7 @@ function timelinePrefixDepsEqual(
     a.optimisticTurns === b.optimisticTurns &&
     a.liveOwnsActiveTurn === b.liveOwnsActiveTurn &&
     a.delegationKickoffText === b.delegationKickoffText &&
-    a.hasLiveMessage === b.hasLiveMessage &&
+    a.liveShowsReply === b.liveShowsReply &&
     a.liveStartedAt === b.liveStartedAt
   )
 }
@@ -3078,9 +3078,22 @@ function collectInFlightPersistedToolCalls(
   return out
 }
 
+/**
+ * @param liveShowsReply whether the live message this session holds actually
+ * produced an assistant turn — see [`computeTimeline`], which derives it from
+ * the same build the streaming tail is made of. Both suppressions below hide a
+ * persisted assistant turn *because the live stream is showing that reply*, so
+ * they must key off what the live message RENDERS, never off the existence of a
+ * live message object. Those two differ: a live message with nothing renderable
+ * in it is an ordinary state (`STATUS_CHANGED` → `prompting` installs
+ * `content: []` at the start of every turn, and the runtime mirror never writes
+ * a null back over it), and keying off the object hid a reply with nothing put
+ * in its place — a blank agent turn.
+ */
 function computeTimelinePrefix(
   session: ConversationRuntimeSession,
-  conversationId: number
+  conversationId: number,
+  liveShowsReply: boolean
 ): TimelinePrefixEntry {
   const detail = session.detail
   // Everything Phases 1–3 read, snapshotted for the `===` validity check.
@@ -3094,7 +3107,7 @@ function computeTimelinePrefix(
     optimisticTurns: session.optimisticTurns,
     liveOwnsActiveTurn: session.liveOwnsActiveTurn,
     delegationKickoffText: session.delegationKickoffText,
-    hasLiveMessage: session.liveMessage !== null,
+    liveShowsReply,
     liveStartedAt: session.liveMessage?.startedAt ?? null,
   }
   if (detail) {
@@ -3121,7 +3134,7 @@ function computeTimelinePrefix(
   const rawPersistedTurns = session.detail?.turns ?? []
   const hasLiveOrLocalReply =
     session.liveOwnsActiveTurn &&
-    (session.liveMessage !== null || session.localTurns.length > 0)
+    (liveShowsReply || session.localTurns.length > 0)
   let stripFrom = -1
   if (hasLiveOrLocalReply) {
     let lastUserIdx = -1
@@ -3153,13 +3166,12 @@ function computeTimelinePrefix(
   // into `detail` it sits beside the live reply (a separate assistant turn
   // under a `live-…` id), and `mergeConsecutiveAssistantTurns` concatenates
   // the two — so the already-persisted head (e.g. the first reasoning block)
-  // renders twice. Hide that persisted partial, but ONLY while `liveMessage`
-  // is in hand: the live stream carries the full reply (the attach snapshot is
-  // built atomically and includes it), so this only ever hides from render
-  // what the live stream is concurrently showing — never dropping a reply we
-  // can't re-show. The moment the turn ends, `liveMessage` clears and the
-  // persisted copy (now complete) renders normally; the brief promote→refetch
-  // grace window can show a transient visible duplicate, never a hidden turn.
+  // renders twice. Hide that persisted partial, but ONLY while the live message
+  // is actually SHOWING a reply (`liveShowsReply`): that is what makes this a
+  // choice between two renderings of one reply rather than a deletion. The
+  // moment the turn ends, `liveMessage` clears and the persisted copy (now
+  // complete) renders normally; the brief promote→refetch grace window can show
+  // a transient visible duplicate, never a hidden turn.
   //
   // The in-flight prompt is identified authoritatively by the backend, which
   // reports the id of the persisted user turn it stamped as the in-flight one
@@ -3171,9 +3183,7 @@ function computeTimelinePrefix(
   // id, so an earlier completed round's reply is never mistaken for a partial.
   const inFlightPromptId = session.detail?.in_flight_user_turn_id ?? null
   const inFlightPromptIdx =
-    !hasLiveOrLocalReply &&
-    session.liveMessage !== null &&
-    inFlightPromptId !== null
+    !hasLiveOrLocalReply && liveShowsReply && inFlightPromptId !== null
       ? persistedTurns.findIndex(
           (t) => t.role === "user" && t.id === inFlightPromptId
         )
@@ -3415,14 +3425,28 @@ function computeTimeline(
   const cached = timelineCache.get(session)
   if (cached) return cached
 
-  // Phases 1–3 (already deduped), reused across streaming batches.
-  const { prefix, prefixKeys } = computeTimelinePrefix(session, conversationId)
-
-  // Phase 4: Streaming turns (live agent response, split into rounds)
+  // Phase 4 first: Phases 1–3 hide the persisted copy of the reply this build
+  // is showing, so they need its verdict, and deriving that from the same build
+  // is what keeps the two from disagreeing. A live message can hold nothing
+  // renderable — `content: []` from the turn's own `prompting` transition, or
+  // only blocks this build drops — and a check for the message OBJECT then hid
+  // a persisted reply that nothing replaced.
   const streamingMessage = session.liveMessage
   const built = streamingMessage
     ? buildStreamingTurnsFromLiveMessage(conversationId, streamingMessage)
     : null
+  // A `user` turn here is a message the user sent mid-turn (native steering),
+  // not a rendering of the reply — a live message that produced only those is
+  // showing no reply and must suppress nothing.
+  const liveShowsReply =
+    built?.turns.some((turn) => turn.role === "assistant") ?? false
+
+  // Phases 1–3 (already deduped), reused across streaming batches.
+  const { prefix, prefixKeys } = computeTimelinePrefix(
+    session,
+    conversationId,
+    liveShowsReply
+  )
 
   let deduped: ConversationTimelineTurn[]
   if (!built || built.turns.length === 0) {
diff --git a/src/stores/runtime-empty-live-message.test.ts b/src/stores/runtime-empty-live-message.test.ts
new file mode 100644
index 0000000000..175b6a26c2
--- /dev/null
+++ b/src/stores/runtime-empty-live-message.test.ts
@@ -0,0 +1,206 @@
+/**
+ * Two timeline rules hide a persisted assistant turn while a reply streams: the
+ * `liveOwnsActiveTurn` tail strip (delegation-child dialog) and the
+ * `in_flight_user_turn_id` partial suppression (cross-client viewer). Both are
+ * only sound because the live stream is showing that same reply — so both have
+ * to key off what the live message RENDERS, not off a live message existing.
+ *
+ * Those differ, and routinely. `status_changed → prompting` installs a fresh
+ * `content: []` live message at the start of every turn and mirrors it into
+ * this store (see "fires with isLive=true and a fresh non-null liveMessage when
+ * a turn starts" in acp-connections-context.test.tsx); the mirror never writes a
+ * null back over it, so the same object stays in hand for any part of a turn
+ * that produces nothing this build renders. Keyed on the object, the persisted
+ * reply was hidden with nothing put in its place: a blank agent turn.
+ */
+
+import { afterEach, describe, expect, it } from "vitest"
+
+import type { LiveMessage } from "@/contexts/acp-connections-context"
+import type { DbConversationDetail, MessageTurn, TurnRole } from "@/lib/types"
+import {
+  getTimelineTurns,
+  resetConversationRuntimeStore,
+  useConversationRuntimeStore,
+} from "@/stores/conversation-runtime-store"
+
+const CID = 77
+const TS = "2026-09-06T00:00:00.000Z"
+
+function turn(id: string, role: TurnRole): MessageTurn {
+  return { id, role, blocks: [{ type: "text", text: id }], timestamp: TS }
+}
+
+/** What the turn's own `prompting` transition installs, before any content. */
+const promptingLiveMessage: LiveMessage = {
+  id: "m1",
+  role: "assistant",
+  content: [],
+  startedAt: Date.parse(TS),
+}
+
+/** A block Phase 2 drops, so this message renders exactly as much as `[]`. */
+const emptyTextLiveMessage: LiveMessage = {
+  ...promptingLiveMessage,
+  content: [{ type: "text", text: "" }],
+}
+
+const replyLiveMessage: LiveMessage = {
+  ...promptingLiveMessage,
+  content: [{ type: "text", text: "streaming…" }],
+}
+
+/** A message the user sent mid-turn, with no reply to it yet. */
+const steeringOnlyLiveMessage: LiveMessage = {
+  ...promptingLiveMessage,
+  content: [
+    {
+      type: "steering",
+      id: "note-1",
+      text: "also check the tests",
+      createdAt: TS,
+    },
+  ],
+}
+
+function seed(
+  turns: MessageTurn[],
+  overrides: {
+    liveMessage?: LiveMessage | null
+    liveOwnsActiveTurn?: boolean
+    localTurns?: MessageTurn[]
+    inFlightUserTurnId?: string | null
+  }
+) {
+  const detail: DbConversationDetail = {
+    summary: {
+      id: CID,
+      folder_id: 1,
+      title: "t",
+      title_locked: false,
+      agent_type: "claude_code",
+      status: "in_progress",
+      kind: "regular",
+      model: null,
+      git_branch: null,
+      external_id: null,
+      message_count: turns.length,
+      child_count: 0,
+      created_at: TS,
+      updated_at: TS,
+      pinned_at: null,
+    },
+    turns,
+    in_flight_user_turn_id: overrides.inFlightUserTurnId ?? null,
+  }
+  const next = new Map(useConversationRuntimeStore.getState().byConversationId)
+  next.set(CID, {
+    conversationId: CID,
+    externalId: null,
+    dbConversationId: null,
+    detail,
+    detailLoading: false,
+    detailError: null,
+    acpLoadError: null,
+    localTurns: overrides.localTurns ?? [],
+    backgroundTurns: [],
+    pendingBackgroundSettlements: [],
+    optimisticTurns: [],
+    liveMessage: overrides.liveMessage ?? null,
+    syncState: "idle" as const,
+    activeTurnToken: null,
+    lastTurnOwned: false,
+    liveOwnsActiveTurn: overrides.liveOwnsActiveTurn ?? false,
+    delegationKickoffText: null,
+    sessionStats: null,
+    historyAssistantBaseline: null,
+    batchBoundaryIndex: null,
+    batchBoundaryPrefixHash: null,
+    loadingOlderTurns: false,
+    olderTurnsPrependEpoch: 0,
+    pendingCleanup: false,
+  })
+  useConversationRuntimeStore.setState({ byConversationId: next })
+}
+
+const timelineIds = () => getTimelineTurns(CID).map((t) => t.turn.id)
+
+afterEach(() => {
+  resetConversationRuntimeStore()
+})
+
+describe("persisted-tail strip vs. a live message that renders nothing", () => {
+  it("keeps the child's reply while the new turn has produced nothing yet", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveOwnsActiveTurn: true,
+      liveMessage: promptingLiveMessage,
+    })
+    expect(timelineIds()).toEqual(["u1", "a1"])
+  })
+
+  it("keeps the child's reply when the live message holds only an empty block", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveOwnsActiveTurn: true,
+      liveMessage: emptyTextLiveMessage,
+    })
+    expect(timelineIds()).toEqual(["u1", "a1"])
+  })
+
+  it("still strips the persisted copy once the live message shows the reply", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveOwnsActiveTurn: true,
+      liveMessage: replyLiveMessage,
+    })
+    expect(timelineIds()).toEqual(["u1", `live-${CID}-m1`])
+  })
+
+  it("still strips for a promoted reply, which renders on its own", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveOwnsActiveTurn: true,
+      liveMessage: promptingLiveMessage,
+      localTurns: [turn("promoted", "assistant")],
+    })
+    expect(timelineIds()).toEqual(["u1", "promoted"])
+  })
+})
+
+describe("in-flight partial suppression vs. a live message that renders nothing", () => {
+  it("keeps the persisted partial while the turn has produced nothing yet", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveMessage: promptingLiveMessage,
+      inFlightUserTurnId: "u1",
+    })
+    expect(timelineIds()).toEqual(["u1", "a1"])
+  })
+
+  it("keeps every persisted reply of the round, not just the newest", () => {
+    seed(
+      [
+        turn("u1", "user"),
+        turn("a1", "assistant"),
+        turn("a2", "assistant"),
+        turn("a3", "assistant"),
+      ],
+      { liveMessage: emptyTextLiveMessage, inFlightUserTurnId: "u1" }
+    )
+    expect(timelineIds()).toEqual(["u1", "a1", "a2", "a3"])
+  })
+
+  it("still hides the persisted partial once the live message shows the reply", () => {
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveMessage: replyLiveMessage,
+      inFlightUserTurnId: "u1",
+    })
+    expect(timelineIds()).toEqual(["u1", `live-${CID}-m1`])
+  })
+
+  it("keeps the persisted reply when the live message carries only a steer", () => {
+    // A mid-turn message is the user's, not a rendering of the reply, so it
+    // cannot stand in for the persisted copy it would otherwise hide.
+    seed([turn("u1", "user"), turn("a1", "assistant")], {
+      liveMessage: steeringOnlyLiveMessage,
+      inFlightUserTurnId: "u1",
+    })
+    expect(timelineIds()).toEqual(["u1", "a1", `live-${CID}-m1`])
+  })
+})