diff --git a/backend/AGENTS.md b/backend/AGENTS.md
index 03faf4389..bcbc6995e 100644
--- a/backend/AGENTS.md
+++ b/backend/AGENTS.md
@@ -533,6 +533,20 @@ full and delta checkpoint modes. Only an explicitly absent legacy parent link ma
 use chronological compatibility lookup; cycles, dangling links, and depth-limit
 exhaustion fail closed. Existing single-checkpoint branches are never repaired by
 copying a raw checkpoint because delta state is not self-contained in one tuple.
+Both lookups additionally require the replay base to be a **settled** checkpoint
+(`has_pending_tasks` — no scheduled `next` tasks). A checkpoint with pending tasks
+is a mid-run snapshot: resuming from it replays the writes of the node that was
+about to run. Message ids alone cannot exclude those, because middleware may
+rewrite a message's id inside the run that produced it — `DynamicContextMiddleware`
+moves the first user turn to `{id}__user` and gives `{id}` to the injected
+reminder, so every checkpoint written before it holds the same prompt under an
+unmatched id. Selecting one of those re-added the original prompt *after* the
+edited one, and the model answered the question the edit was replacing (#4531).
+`next` is not derivable on the degraded raw-checkpoint read path, which reports no
+tasks; absence of evidence stays permissive there rather than failing closed.
+Edit replay resolves its base through the same lineage-first path as regenerate;
+it must pass `head_checkpoint` or it silently degrades to the chronological scan
+that cannot tell sibling branches apart.
 
 ### Sandbox System (`packages/harness/deerflow/sandbox/`)
 
diff --git a/backend/app/gateway/checkpoint_lineage.py b/backend/app/gateway/checkpoint_lineage.py
index 6b121c318..275784c7d 100644
--- a/backend/app/gateway/checkpoint_lineage.py
+++ b/backend/app/gateway/checkpoint_lineage.py
@@ -45,6 +45,23 @@ def is_duration_only_checkpoint(checkpoint_tuple: Any) -> bool:
     return isinstance(writes, dict) and "runtime_run_duration" in writes
 
 
+def has_pending_tasks(checkpoint_tuple: Any) -> bool:
+    """Return whether *checkpoint_tuple* still has graph work scheduled.
+
+    A replay base must be a state the thread was at rest in. A mid-run
+    checkpoint owns the writes of the node that was about to run, and resuming
+    from it replays them — re-adding the very turn the replay is meant to
+    replace. Message ids alone cannot detect this, because middleware may
+    rewrite a message's id in the same run that produced it.
+
+    ``next`` is not derivable on the degraded raw-checkpoint read path, which
+    reports no tasks at all. Absence of evidence therefore stays permissive:
+    those reads keep selecting the same base they always did.
+    """
+
+    return bool(getattr(checkpoint_tuple, "next", None))
+
+
 def _message_id(message: Any) -> str | None:
     value = getattr(message, "id", None)
     if value is None and isinstance(message, dict):
@@ -96,7 +113,8 @@ async def find_checkpoint_before_message(
     Following ``parent_config`` is important after a regenerate: a thread can contain
     sibling checkpoint branches, and a global time-ordered scan can otherwise select
     a checkpoint from the wrong branch. Duration-only metadata checkpoints do not
-    represent an addressable conversation state and are skipped.
+    represent an addressable conversation state and are skipped, and so are
+    checkpoints that still have pending tasks (see :func:`has_pending_tasks`).
     """
 
     if message_id not in {_message_id(message) for message in checkpoint_messages(head_checkpoint)}:
@@ -132,7 +150,7 @@ async def find_checkpoint_before_message(
             continue
 
         parent_message_ids = {_message_id(message) for message in checkpoint_messages(parent)}
-        if message_id not in parent_message_ids:
+        if message_id not in parent_message_ids and not has_pending_tasks(parent):
             return parent
         current = parent
 
@@ -148,8 +166,9 @@ def find_checkpoint_before_message_chronologically(
     This is a compatibility fallback for imported or legacy checkpoints that do
     not carry ``parent_config`` links. Callers must prefer the lineage walk when
     links are available because a chronological scan cannot distinguish sibling
-    checkpoint branches. Duration-only checkpoints are ignored, and only
-    checkpoints with an addressable id can become the replay base.
+    checkpoint branches. Duration-only checkpoints are ignored, and only settled
+    checkpoints (see :func:`has_pending_tasks`) with an addressable id can become
+    the replay base.
     """
 
     previous_checkpoint = None
@@ -159,6 +178,6 @@ def find_checkpoint_before_message_chronologically(
         message_ids = {_message_id(message) for message in checkpoint_messages(checkpoint_tuple)}
         if message_id in message_ids:
             return previous_checkpoint, True
-        if checkpoint_configurable(checkpoint_tuple).get("checkpoint_id"):
+        if checkpoint_configurable(checkpoint_tuple).get("checkpoint_id") and not has_pending_tasks(checkpoint_tuple):
             previous_checkpoint = checkpoint_tuple
     return None, False
diff --git a/backend/app/gateway/routers/thread_runs.py b/backend/app/gateway/routers/thread_runs.py
index 5a0c226ee..c12fa4f1f 100644
--- a/backend/app/gateway/routers/thread_runs.py
+++ b/backend/app/gateway/routers/thread_runs.py
@@ -38,6 +38,7 @@ from app.gateway.pagination import trim_run_message_page
 from app.gateway.run_models import RunCreateRequest
 from app.gateway.services import build_checkpoint_state_accessor, build_thread_checkpoint_state_accessor, sse_consumer, start_run, wait_for_run_completion
 from app.gateway.utils import sanitize_log_param
+from deerflow.agents.middlewares.dynamic_context_middleware import strip_injected_user_message_id_suffix
 from deerflow.runtime import CancelOutcome, RunRecord, RunStatus, serialize_channel_values_for_api
 from deerflow.runtime.secret_context import redact_config_secrets, redact_metadata_secrets
 from deerflow.utils.messages import ORIGINAL_USER_CONTENT_KEY, get_original_user_content_text, message_to_text
@@ -344,7 +345,12 @@ def _clean_human_message_for_regenerate(message: Any) -> dict[str, Any]:
         "content": [{"type": "text", "text": content}],
         "additional_kwargs": additional_kwargs,
     }
-    message_id = _message_id(message)
+    # Replay the id the client originally sent. The dynamic-context reminder
+    # re-keys the first user message of a thread to `{id}__user`, and replaying
+    # that persisted id into a state that has no reminder yet makes the
+    # middleware treat the turn as already injected, silently dropping the date
+    # and memory block the original turn had.
+    message_id = strip_injected_user_message_id_suffix(_message_id(message))
     if message_id:
         clean_message["id"] = message_id
     name = _message_name(message)
@@ -653,7 +659,12 @@ async def _prepare_edit_regenerate_payload(
     if not source_ai_id:
         raise HTTPException(status_code=409, detail="The source assistant message is missing an id")
 
-    base_checkpoint_tuple = await _find_base_checkpoint_before_human(thread_id, source_human_id, request)
+    base_checkpoint_tuple = await _find_base_checkpoint_before_human(
+        thread_id,
+        source_human_id,
+        request,
+        head_checkpoint=latest_checkpoint,
+    )
     target_run_id = await _find_target_run_id(thread_id, source_ai_id, source_ai, source_human, request)
     source_record = await _require_successful_source_run(thread_id, target_run_id, request)
     checkpoint = _checkpoint_response(base_checkpoint_tuple)
diff --git a/backend/packages/harness/deerflow/agents/middlewares/dynamic_context_middleware.py b/backend/packages/harness/deerflow/agents/middlewares/dynamic_context_middleware.py
index e4e35d23a..599a35ec5 100644
--- a/backend/packages/harness/deerflow/agents/middlewares/dynamic_context_middleware.py
+++ b/backend/packages/harness/deerflow/agents/middlewares/dynamic_context_middleware.py
@@ -60,6 +60,23 @@ _DYNAMIC_CONTEXT_REMINDER_KEY = "dynamic_context_reminder"
 # so it is never exposed to user-influenceable memory content.
 _REMINDER_DATE_KEY = "reminder_date"
 _SUMMARY_MESSAGE_NAME = "summary"
+# Suffix the ID-swap gives the real user message; the reminder SystemMessage
+# takes the original id so ``add_messages`` can replace it in place.
+INJECTED_USER_MESSAGE_ID_SUFFIX = "__user"
+
+
+def strip_injected_user_message_id_suffix(message_id: str | None) -> str | None:
+    """Return the id *message_id* had before the reminder ID-swap.
+
+    Replaying a persisted user turn must feed the graph the id the client
+    originally sent: a ``{id}__user`` message is skipped as an injection target,
+    so replaying one into a state that has no reminder yet silently drops the
+    date and memory block for that turn.
+    """
+
+    if isinstance(message_id, str) and message_id.endswith(INJECTED_USER_MESSAGE_ID_SUFFIX):
+        return message_id[: -len(INJECTED_USER_MESSAGE_ID_SUFFIX)] or message_id
+    return message_id
 
 
 def _extract_date(content: str) -> str | None:
@@ -120,7 +137,7 @@ def _is_user_injection_target(message: object) -> bool:
     # (id__user__user__user...) and ghost-message re-execution.
     # Using endswith (not substring "in") avoids false positives on IDs that
     # happen to contain "__user" in the middle.
-    if message.id and str(message.id).endswith("__user"):
+    if message.id and str(message.id).endswith(INJECTED_USER_MESSAGE_ID_SUFFIX):
         return False
     return True
 
@@ -232,7 +249,7 @@ class DynamicContextMiddleware(AgentMiddleware):
         messages.append(
             HumanMessage(
                 content=original.content,
-                id=f"{stable_id}__user",
+                id=f"{stable_id}{INJECTED_USER_MESSAGE_ID_SUFFIX}",
                 name=original.name,
                 additional_kwargs=original.additional_kwargs,
             )
diff --git a/backend/tests/test_checkpoint_lineage.py b/backend/tests/test_checkpoint_lineage.py
new file mode 100644
index 000000000..10d560924
--- /dev/null
+++ b/backend/tests/test_checkpoint_lineage.py
@@ -0,0 +1,149 @@
+"""Replay-base resolution rules shared by regenerate, edit replay and branching.
+
+The load-bearing rule here is that a replay base must be a checkpoint the
+thread was actually at rest in. A mid-run checkpoint still owns the interrupted
+node's pending writes, so resuming from it replays them.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from types import SimpleNamespace
+
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
+
+from app.gateway.checkpoint_lineage import (
+    CheckpointParentMissingError,
+    find_checkpoint_before_message,
+    find_checkpoint_before_message_chronologically,
+)
+
+THREAD_ID = "thread-1"
+
+
+def _snapshot(checkpoint_id: str, messages: list[object], *, next_tasks: tuple[str, ...] = (), parent_id: str | None = None):
+    parent_config = None
+    if parent_id is not None:
+        parent_config = {"configurable": {"thread_id": THREAD_ID, "checkpoint_ns": "", "checkpoint_id": parent_id}}
+    return SimpleNamespace(
+        values={"messages": messages},
+        config={
+            "configurable": {
+                "thread_id": THREAD_ID,
+                "checkpoint_ns": "",
+                "checkpoint_id": checkpoint_id,
+                "checkpoint_map": None,
+            }
+        },
+        metadata={},
+        parent_config=parent_config,
+        next=next_tasks,
+    )
+
+
+class _Accessor:
+    """Minimal accessor over a fixed snapshot list, keyed by checkpoint id."""
+
+    def __init__(self, snapshots: list[object]) -> None:
+        self.snapshots = snapshots
+
+    async def aget(self, config):
+        checkpoint_id = config.get("configurable", {}).get("checkpoint_id")
+        return next(
+            (item for item in self.snapshots if item.config["configurable"]["checkpoint_id"] == checkpoint_id),
+            SimpleNamespace(values={}, config={}, metadata=None, parent_config=None, next=()),
+        )
+
+
+def _first_turn_history() -> list[object]:
+    """Newest-first history of a thread whose only turn is still its first.
+
+    ``DynamicContextMiddleware`` swaps the first user message's id mid-run:
+    the injected reminder takes ``{id}`` and the real user message becomes
+    ``{id}__user``. So the pre-injection checkpoints hold the very same prompt
+    under an id the replay-base lookup does not recognise (#4531).
+    """
+    system = SystemMessage(id="h1", content="<system-reminder>date</system-reminder>")
+    swapped_human = HumanMessage(id="h1__user", content="question")
+    raw_human = HumanMessage(id="h1", content="question")
+    ai = AIMessage(id="ai-1", content="answer")
+    return [
+        _snapshot("ckpt-head", [system, swapped_human, ai], parent_id="ckpt-after-inject"),
+        _snapshot("ckpt-after-inject", [system, swapped_human], next_tasks=("LoopDetectionMiddleware.before_agent",), parent_id="ckpt-mid"),
+        _snapshot("ckpt-mid", [raw_human], next_tasks=("DynamicContextMiddleware.before_agent",), parent_id="ckpt-input"),
+        _snapshot("ckpt-input", [], next_tasks=("__start__",), parent_id="ckpt-empty"),
+        _snapshot("ckpt-empty", []),
+    ]
+
+
+def test_chronological_scan_skips_checkpoints_with_pending_tasks():
+    history = _first_turn_history()
+
+    base, found = find_checkpoint_before_message_chronologically(history, "h1__user")
+
+    assert found is True
+    assert base is not None
+    assert base.config["configurable"]["checkpoint_id"] == "ckpt-empty"
+
+
+def test_lineage_walk_skips_checkpoints_with_pending_tasks():
+    history = _first_turn_history()
+    accessor = _Accessor(history)
+
+    base = asyncio.run(find_checkpoint_before_message(accessor, history[0], "h1__user", max_depth=10))
+
+    assert base.config["configurable"]["checkpoint_id"] == "ckpt-empty"
+
+
+def test_chronological_scan_prefers_the_previous_turn_boundary():
+    """A later turn resolves to the previous run's settled tail, not its input checkpoint."""
+    human1 = HumanMessage(id="h1__user", content="first")
+    ai1 = AIMessage(id="ai-1", content="first answer")
+    human2 = HumanMessage(id="h2", content="second")
+    history = [
+        _snapshot("ckpt-turn2-head", [human1, ai1, human2, AIMessage(id="ai-2", content="second answer")]),
+        _snapshot("ckpt-turn2-mid", [human1, ai1, human2], next_tasks=("model",)),
+        _snapshot("ckpt-turn2-input", [human1, ai1], next_tasks=("__start__",)),
+        _snapshot("ckpt-turn1-tail", [human1, ai1]),
+    ]
+
+    base, found = find_checkpoint_before_message_chronologically(history, "h2")
+
+    assert found is True
+    assert base.config["configurable"]["checkpoint_id"] == "ckpt-turn1-tail"
+
+
+def test_lineage_walk_reports_missing_parent_when_no_settled_ancestor_exists():
+    """Fail closed rather than fork a mid-run checkpoint."""
+    human = HumanMessage(id="h1__user", content="question")
+    history = [
+        _snapshot("ckpt-head", [human, AIMessage(id="ai-1", content="answer")], parent_id="ckpt-mid"),
+        _snapshot("ckpt-mid", [HumanMessage(id="h1", content="question")], next_tasks=("DynamicContextMiddleware.before_agent",)),
+    ]
+    accessor = _Accessor(history)
+
+    with pytest.raises(CheckpointParentMissingError):
+        asyncio.run(find_checkpoint_before_message(accessor, history[0], "h1__user", max_depth=10))
+
+
+def test_unknown_pending_tasks_do_not_block_selection():
+    """Raw full-mode reads cannot derive tasks; absence of evidence stays permissive."""
+    human = HumanMessage(id="h1", content="question")
+    history = [
+        SimpleNamespace(
+            values={"messages": [human]},
+            config={"configurable": {"thread_id": THREAD_ID, "checkpoint_ns": "", "checkpoint_id": "ckpt-head"}},
+            metadata={},
+        ),
+        SimpleNamespace(
+            values={"messages": []},
+            config={"configurable": {"thread_id": THREAD_ID, "checkpoint_ns": "", "checkpoint_id": "ckpt-base"}},
+            metadata={},
+        ),
+    ]
+
+    base, found = find_checkpoint_before_message_chronologically(history, "h1")
+
+    assert found is True
+    assert base.config["configurable"]["checkpoint_id"] == "ckpt-base"
diff --git a/backend/tests/test_thread_regenerate_prepare.py b/backend/tests/test_thread_regenerate_prepare.py
index 24ff667a9..56ed3e16e 100644
--- a/backend/tests/test_thread_regenerate_prepare.py
+++ b/backend/tests/test_thread_regenerate_prepare.py
@@ -7,7 +7,7 @@ from unittest.mock import AsyncMock, patch
 
 import pytest
 from fastapi import HTTPException
-from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 from langgraph.checkpoint.base import empty_checkpoint, uuid6
 from langgraph.checkpoint.memory import InMemorySaver
 
@@ -21,6 +21,7 @@ def _checkpoint(
     *,
     metadata: dict | None = None,
     goal: dict | None = None,
+    next_tasks: tuple[str, ...] = (),
 ):
     channel_values = {"messages": messages}
     if goal is not None:
@@ -36,6 +37,7 @@ def _checkpoint(
         },
         checkpoint={"channel_values": channel_values},
         metadata=metadata or {},
+        next=next_tasks,
     )
 
 
@@ -117,6 +119,7 @@ class FakeAccessor:
             config=checkpoint.config,
             metadata=checkpoint.metadata,
             parent_config=getattr(checkpoint, "parent_config", None),
+            next=getattr(checkpoint, "next", ()),
         )
 
     async def aget(self, config):
@@ -629,7 +632,30 @@ def test_prepare_edit_regenerate_payload_returns_new_human_and_edit_metadata():
     }
 
 
-def _edit_source_run_fixtures() -> tuple[FakeEventStore, FakeRunManager]:
+def _first_turn_checkpointer() -> FakeCheckpointer:
+    """First-turn history where the user message's id is swapped mid-run.
+
+    ``DynamicContextMiddleware`` moves the first user message to ``{id}__user``
+    and gives ``{id}`` to the injected reminder, so every checkpoint written
+    before that node ran holds the same prompt under an id the replay-base
+    lookup cannot match (#4531).
+    """
+    system = SystemMessage(id="human-1", content="<system-reminder>date</system-reminder>")
+    swapped_human = HumanMessage(id="human-1__user", content="original question")
+    raw_human = HumanMessage(id="human-1", content="original question")
+    ai = AIMessage(id="ai-1", content="answer v1")
+    return FakeCheckpointer(
+        [
+            _checkpoint("ckpt-head", [system, swapped_human, ai]),
+            _checkpoint("ckpt-after-inject", [system, swapped_human], next_tasks=("LoopDetectionMiddleware.before_agent",)),
+            _checkpoint("ckpt-mid", [raw_human], next_tasks=("DynamicContextMiddleware.before_agent",)),
+            _checkpoint("ckpt-input", [], next_tasks=("__start__",)),
+            _checkpoint("ckpt-empty", []),
+        ]
+    )
+
+
+def _answer_run_fixtures() -> tuple[FakeEventStore, FakeRunManager]:
     event_store = FakeEventStore(
         [
             {
@@ -655,7 +681,7 @@ def test_prepare_edit_regenerate_payload_preserves_a_rename_the_replay_base_pred
     base.checkpoint["channel_values"]["title"] = "auto generated title"
     latest = _checkpoint("ckpt-ai", [human, ai])
     latest.checkpoint["channel_values"]["title"] = "User renamed title"
-    event_store, run_manager = _edit_source_run_fixtures()
+    event_store, run_manager = _answer_run_fixtures()
 
     response = asyncio.run(
         thread_runs._prepare_edit_regenerate_payload(
@@ -683,7 +709,7 @@ def test_prepare_edit_regenerate_payload_lets_an_untitled_base_name_the_edited_t
     ai = AIMessage(id="ai-1", content="answer v1")
     latest = _checkpoint("ckpt-ai", [human, ai])
     latest.checkpoint["channel_values"]["title"] = "title of the replaced question"
-    event_store, run_manager = _edit_source_run_fixtures()
+    event_store, run_manager = _answer_run_fixtures()
 
     response = asyncio.run(
         thread_runs._prepare_edit_regenerate_payload(
@@ -698,6 +724,83 @@ def test_prepare_edit_regenerate_payload_lets_an_untitled_base_name_the_edited_t
     assert "title" not in response.input
 
 
+def test_prepare_regenerate_payload_replays_the_pre_swap_user_message_id():
+    """Replaying `{id}__user` would make the reminder middleware skip the turn.
+
+    The replay base predates the injection, so the turn must re-enter the graph
+    under the id the client originally sent or it loses its date/memory block.
+    """
+    from app.gateway.routers import thread_runs
+
+    event_store, run_manager = _answer_run_fixtures()
+
+    response = asyncio.run(
+        thread_runs._prepare_regenerate_payload(
+            "thread-1",
+            "ai-1",
+            _request(_first_turn_checkpointer(), event_store, run_manager=run_manager),
+        )
+    )
+
+    assert response.checkpoint["checkpoint_id"] == "ckpt-empty"
+    assert response.input["messages"][0]["id"] == "human-1"
+
+
+def test_prepare_edit_regenerate_payload_skips_mid_run_replay_base_on_first_turn():
+    """The replay base must predate the turn, not sit inside the run that produced it.
+
+    ``ckpt-mid`` still holds the original prompt (under its pre-swap id) and owns
+    the injection node's pending writes, so replaying from it re-adds the prompt
+    the edit is meant to replace.
+    """
+    from app.gateway.routers import thread_runs
+
+    event_store, run_manager = _answer_run_fixtures()
+
+    response = asyncio.run(
+        thread_runs._prepare_edit_regenerate_payload(
+            "thread-1",
+            "human-1__user",
+            "edited question",
+            _request(_first_turn_checkpointer(), event_store, run_manager=run_manager),
+        )
+    )
+
+    assert response.checkpoint["checkpoint_id"] == "ckpt-empty"
+    assert response.metadata["regenerate_checkpoint_id"] == "ckpt-empty"
+
+
+def test_prepare_edit_regenerate_payload_prefers_checkpoint_lineage():
+    """Edit replay must resolve its base the same lineage-first way regenerate does.
+
+    A chronological scan cannot tell sibling branches apart (#4358), so the head
+    checkpoint has to reach the lineage walk.
+    """
+    from app.gateway.routers import thread_runs
+
+    event_store, run_manager = _answer_run_fixtures()
+    checkpointer = _first_turn_checkpointer()
+    from app.gateway.checkpoint_lineage import CheckpointParentMissingError
+
+    walk = AsyncMock(side_effect=CheckpointParentMissingError("no parent link"))
+
+    with patch.object(thread_runs, "find_checkpoint_before_message", walk):
+        response = asyncio.run(
+            thread_runs._prepare_edit_regenerate_payload(
+                "thread-1",
+                "human-1__user",
+                "edited question",
+                _request(checkpointer, event_store, run_manager=run_manager),
+            )
+        )
+
+    assert walk.await_count == 1
+    head_checkpoint = walk.await_args.args[1]
+    assert head_checkpoint.config["configurable"]["checkpoint_id"] == "ckpt-head"
+    # Legacy checkpoints without parent links still degrade to the bounded scan.
+    assert response.checkpoint["checkpoint_id"] == "ckpt-empty"
+
+
 @pytest.mark.parametrize(
     ("replacement_text", "detail"),
     [
diff --git a/frontend/src/core/threads/hooks.ts b/frontend/src/core/threads/hooks.ts
index 9399b357c..f10455c5e 100644
--- a/frontend/src/core/threads/hooks.ts
+++ b/frontend/src/core/threads/hooks.ts
@@ -918,6 +918,27 @@ function getMessagesAfterBaseline(
   });
 }
 
+/**
+ * Human-message baseline for a prepared replay (regenerate / edit).
+ *
+ * A replay masks the turn it supersedes, so those messages leave the live
+ * message list the moment the mask is applied. Baselining on the pre-mask count
+ * means the replacement only ever restores the count instead of exceeding it,
+ * and the optimistic copy is never recognised as confirmed. That matters for
+ * the first turn of a thread, where the runtime re-keys the replacement message
+ * so identity comparison cannot stand in for the count either.
+ */
+export function countHumanMessagesExcludingSuperseded(
+  messages: Message[],
+  supersededMessageIds: readonly string[],
+): number {
+  const superseded = new Set(supersededMessageIds);
+  return messages.filter(
+    (message) =>
+      message.type === "human" && (!message.id || !superseded.has(message.id)),
+  ).length;
+}
+
 export function getVisibleOptimisticMessages(
   optimisticMessages: Message[],
   previousHumanMessageCount: number,
@@ -2068,6 +2089,10 @@ export function useThreadStream({
         const prepared = await prepare();
         preparedSupersededRunId = prepared.target_run_id;
         preparedSupersededMessageIds = getSupersededMessageIds(prepared);
+        prevHumanMsgCountRef.current = countHumanMessagesExcludingSuperseded(
+          persistedMessages,
+          preparedSupersededMessageIds,
+        );
         const replacementHumanMessageId =
           "replacement_human_message_id" in prepared &&
           typeof prepared.replacement_human_message_id === "string"
diff --git a/frontend/tests/unit/core/threads/message-merge.test.ts b/frontend/tests/unit/core/threads/message-merge.test.ts
index 643267efa..444a99874 100644
--- a/frontend/tests/unit/core/threads/message-merge.test.ts
+++ b/frontend/tests/unit/core/threads/message-merge.test.ts
@@ -7,6 +7,7 @@ import {
   buildVisibleHistoryMessages,
   areOptimisticMessagesConfirmed,
   computeSummarizationTransientMessages,
+  countHumanMessagesExcludingSuperseded,
   flattenThreadHistoryPages,
   getSummarizationMiddlewareMessages,
   getThreadHistoryNextPageParam,
@@ -387,6 +388,58 @@ test("mergeMessages shows server human instead of optimistic duplicate after fir
   ]);
 });
 
+test("edit replay of the only turn hides the optimistic copy once the server human arrives", () => {
+  // The runtime re-keys the first user message of a thread, so the persisted
+  // replacement never matches the optimistic id and only the count can confirm
+  // it. Masking the superseded turn drops the live count to zero first.
+  const supersededHuman = {
+    id: "human-1__user",
+    type: "human",
+    content: "introduce Li Bai",
+  } as Message;
+  const optimisticHuman = {
+    id: "replacement-1",
+    type: "human",
+    content: "introduce Du Fu",
+  } as Message;
+  const serverHuman = {
+    id: "replacement-1__user",
+    type: "human",
+    content: "introduce Du Fu",
+  } as Message;
+
+  const baseline = countHumanMessagesExcludingSuperseded(
+    [supersededHuman],
+    ["human-1__user", "ai-1"],
+  );
+  expect(baseline).toBe(0);
+
+  expect(getVisibleOptimisticMessages([optimisticHuman], baseline, 0)).toEqual([
+    optimisticHuman,
+  ]);
+  expect(getVisibleOptimisticMessages([optimisticHuman], baseline, 1)).toEqual(
+    [],
+  );
+  expect(mergeMessages([], [serverHuman], [])).toEqual([serverHuman]);
+});
+
+test("countHumanMessagesExcludingSuperseded keeps turns the replay does not supersede", () => {
+  const keptHuman = { id: "human-1", type: "human", content: "one" } as Message;
+  const supersededHuman = {
+    id: "human-2",
+    type: "human",
+    content: "two",
+  } as Message;
+  const ai = { id: "ai-1", type: "ai", content: "answer" } as Message;
+
+  expect(
+    countHumanMessagesExcludingSuperseded(
+      [keptHuman, ai, supersededHuman],
+      ["human-2", "ai-2"],
+    ),
+  ).toBe(1);
+});
+
 test("getVisibleOptimisticMessages keeps optimistic user input until server human arrives", () => {
   const optimisticHuman = {
     id: "opt-human-1",