diff --git a/docs/cascade_runbook.md b/docs/cascade_runbook.md
index 338d2a7cf..9d0ee527d 100644
--- a/docs/cascade_runbook.md
+++ b/docs/cascade_runbook.md
@@ -109,9 +109,10 @@ naturally.
 
 ## One-shot replay: `everos cascade sync [PATH]`
 
-Use this when the watcher missed an event (WSL mount, network share,
-external editor with no inotify) or when you want a deterministic
-flush before, say, a smoke test:
+Use this when no server is running and you want the index caught up
+with the markdown — after batch edits, before a smoke test, or on a
+mount where the watcher misses events (WSL mount, network share,
+external editor with no inotify) while the daemon is down:
 
 ```bash
 everos cascade sync                           # drain everything pending
@@ -120,10 +121,16 @@ everos cascade sync users/u1/episodes/X.md    # re-enqueue + drain
 
 The CLI builds the same `CascadeOrchestrator` as the daemon but only
 calls `sync_once` / `drain_once` — no watcher / scanner background task.
-Its drain still runs the same compaction + version-cleanup (`prune`) as
-the daemon, but `prune` uses `delete_unverified=False`, so it never
-deletes a file another process may be mid-commit on. Safe to run in
-parallel with a live `everos server`.
+It holds the OME lock for the whole run and **refuses to start (exit code
+3) while a server holds it**: two processes writing the same LanceDB
+tables cannot see each other's snapshot and both insert the row (4–5 %
+duplicate rows after a 10-hour soak with two concurrent `sync` processes
+next to a server). The same rule applies to `cascade fix --apply` and
+`cascade rebuild`; `cascade status` and `cascade fix` (listing) are
+read-only and work alongside a server. A running server projects every
+markdown change itself, so nothing is lost by waiting for it — unless it
+was started with `EVEROS_DISABLE_CASCADE=1` or has been quiesced, in
+which case stop it before syncing.
 
 ## Rebuild the index: `everos cascade rebuild`
 
@@ -135,12 +142,12 @@ everos cascade rebuild          # prompts for confirmation
 everos cascade rebuild --yes    # non-interactive
 ```
 
-> **Stop the `everos server` first.** Unlike `cascade sync`, rebuild
-> **drops and recreates** the active backend's tables or collections. A running
-> daemon holds
-> cached table handles that would keep pointing at (and writing to) the
-> dropped dataset, corrupting the rebuild. This is the one cascade
-> command that is **not** safe to run alongside a live server.
+> **Stop the `everos server` first.** Like every index-writing cascade
+> command, rebuild refuses to run while a server holds the memory root
+> (exit code 3) — and it has the strongest reason: it **drops and
+> recreates** the active backend's tables or collections. A running daemon
+> holds cached table handles that would keep pointing at (and writing to)
+> the dropped dataset, corrupting the rebuild.
 
 What it does, in order:
 
@@ -231,7 +238,10 @@ Workarounds:
 - Rely on the scanner — at default 30 s interval, throughput is
   bounded but eventually-consistent.
 - Drop the scan interval to ~5 s if the memory root is small.
-- Run `everos cascade sync` explicitly after batch edits.
+- With no server running, run `everos cascade sync` explicitly after batch
+  edits. A running server picks them up itself, and `sync` refuses to run
+  next to it (exit code 3): two processes writing the same index insert
+  rows twice.
 
 ### Daemon process crash mid-batch
 
@@ -373,6 +383,7 @@ is a deployment-side change with no schema work.
   in the entry inline. Tracked separately.
 - **Reference-file change detection (agent_skill)**: edits to
   `references/*.md` siblings won't trigger a re-index — only changes
-  to `SKILL.md` itself fire the watcher. Workaround: run
-  `everos cascade sync agents/<a>/skills/skill_<n>/SKILL.md` after
-  editing references.
+  to `SKILL.md` itself fire the watcher. Workaround: touch or re-save
+  `SKILL.md` so the watcher fires; with the server stopped,
+  `everos cascade sync agents/<a>/skills/skill_<n>/SKILL.md` re-enqueues
+  it directly.
diff --git a/docs/how-memory-works.md b/docs/how-memory-works.md
index be6fda712..7bb95acf5 100644
--- a/docs/how-memory-works.md
+++ b/docs/how-memory-works.md
@@ -263,8 +263,8 @@ Two paths, two guarantees:
 
 So a `/search` immediately after the `/flush` that produced a record may
 miss it. The markdown is durable regardless; index lag never loses data. If
-you need read-your-write, retry with backoff, or force the queue with
-`everos cascade sync`.
+you need read-your-write, retry with backoff (the running server is the
+only index writer; `everos cascade sync` is for when no server is running).
 
 Integrity is anchored by a few invariants (details in
 [storage_layout.md](storage_layout.md)): the frontmatter `id` /
diff --git a/src/everos/entrypoints/cli/commands/cascade.py b/src/everos/entrypoints/cli/commands/cascade.py
index d8095f56e..105935e7e 100644
--- a/src/everos/entrypoints/cli/commands/cascade.py
+++ b/src/everos/entrypoints/cli/commands/cascade.py
@@ -5,7 +5,7 @@
 
 - ``cascade sync [PATH]`` — flush the work queue. With ``PATH`` the
   command first force-enqueues that single file (used after a manual
-  md edit when waiting for the watcher is impractical), then drains.
+  md edit with no server running), then drains.
 - ``cascade status`` — print the queue + LSN summary that the daemon
   sees right now.
 - ``cascade fix`` — list every ``failed`` row. With ``--apply``, also
@@ -30,6 +30,7 @@
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import enum
 import os
 from collections.abc import AsyncIterator
@@ -47,6 +48,7 @@
 from everos.core.persistence import MemoryRoot
 from everos.entrypoints.cli._log_setup import configure_cli_logging
 from everos.entrypoints.cli.commands._backfill_cmd import run_backfill
+from everos.infra.ome.exceptions import EngineLockHeldError
 from everos.infra.persistence.index import (
     connect,
     drop_business_tables,
@@ -61,6 +63,7 @@
 )
 from everos.memory.cascade import (
     CascadeOrchestrator,
+    hold_ome_lock,
     match_kind,
     ome_lock_is_free,
 )
@@ -136,15 +139,21 @@ def _apply_verbose_logging(verbose: bool | None) -> None:
 
 
 @asynccontextmanager
-async def _runtime(*, verify: bool = True, ensure: bool = True) -> AsyncIterator[None]:
+async def _runtime(
+    *, verify: bool = True, ensure: bool = True, exclusive: bool = False
+) -> AsyncIterator[None]:
     """Stand up sqlite + lancedb the same way the API lifespan would.
 
     The CLI uses the same lazy, process-wide singletons the API lifespan
     does. They are **per-process**: a running daemon has its own
-    connection and table-handle cache, so read/write traffic interleaves
-    safely, but a change to the table *set* made here (drop / recreate)
-    is invisible to the daemon's cached handles — which is why
-    ``rebuild`` refuses to run while a server holds the OME lock.
+    connection and table-handle cache, and reads its own LanceDB
+    snapshot. Reads interleave safely; writes do not — a second process
+    upserting the same table cannot see what the daemon just committed
+    (nor the other way round) and both insert the row, so every command
+    that writes the index (``sync``, ``fix --apply``, ``rebuild``) passes
+    ``exclusive=True`` and holds the OME lock for its whole run. Refused
+    with exit code 3 while a server (or another exclusive CLI phase) holds
+    it; a server starting meanwhile fails at its own lock instead.
 
     ``verify=False`` skips :func:`verify_business_schemas` — required by
     ``cascade rebuild``, whose whole purpose is to recover from a table
@@ -159,19 +168,39 @@ async def _runtime(*, verify: bool = True, ensure: bool = True) -> AsyncIterator
     Rebuild recreates the tables and their indexes itself after dropping,
     so skipping the pre-drop pass loses nothing.
     """
-    engine = get_engine()
-    async with engine.begin() as conn:
-        await conn.run_sync(SQLModel.metadata.create_all)
-    await connect()
-    if verify:
-        await verify_business_schemas()
-    if ensure:
-        await ensure_business_indexes()
+    lock = hold_ome_lock() if exclusive else contextlib.nullcontext()
     try:
-        yield
+        lock.__enter__()
+    except EngineLockHeldError:
+        typer.echo(
+            "error: another process holds this memory root's OME lock — a "
+            "running `everos server`\n"
+            "  (or another exclusive CLI phase). Two processes writing the "
+            "same index insert rows\n"
+            "  twice, so this command needs the root to itself. A server "
+            "projects markdown changes\n"
+            "  on its own unless it was started with EVEROS_DISABLE_CASCADE=1 "
+            "or quiesced; stop it\n"
+            "  first, then re-run.",
+            err=True,
+        )
+        raise typer.Exit(code=3) from None
+    try:
+        engine = get_engine()
+        async with engine.begin() as conn:
+            await conn.run_sync(SQLModel.metadata.create_all)
+        await connect()
+        if verify:
+            await verify_business_schemas()
+        if ensure:
+            await ensure_business_indexes()
+        try:
+            yield
+        finally:
+            await shutdown()
+            await dispose_engine()
     finally:
-        await shutdown()
-        await dispose_engine()
+        lock.__exit__(None, None, None)
 
 
 def _build_orchestrator() -> CascadeOrchestrator:
@@ -215,12 +244,20 @@ def sync(
         typer.Option("--verbose", "-v", help=_VERBOSE_OPTION_HELP),
     ] = None,
 ) -> None:
-    """Drain the cascade queue (and optionally re-enqueue a path first)."""
+    """Drain the cascade queue (and optionally re-enqueue a path first).
+
+    Holds the OME lock for the run (see :func:`_runtime`): a second process
+    writing the same LanceDB tables inserts rows twice — the server reads
+    its own table snapshot and cannot see what the CLI process just
+    committed, so both decide the row is new (4-5 % duplicate rows after a
+    10-hour soak with two concurrent ``cascade sync`` processes). Refused
+    with exit code 3 while a server holds the root.
+    """
     _apply_root_env(root)
     _apply_verbose_logging(verbose)
 
     async def _run() -> None:
-        async with _runtime():
+        async with _runtime(exclusive=True):
             orchestrator = _build_orchestrator()
             if path is not None:
                 rel = _resolve_relative(path)
@@ -315,7 +352,7 @@ def fix(
     _apply_verbose_logging(verbose)
 
     async def _run() -> None:
-        async with _runtime():
+        async with _runtime(exclusive=apply):
             rows = await md_change_state_repo.list_failed()
             if not rows:
                 typer.echo("no failed rows")
@@ -495,7 +532,7 @@ async def _run() -> None:
         # to fix; the startup guard would abort before we could rebuild.
         # ensure=False: the pre-drop migration pass would raise on exactly
         # the damage we are here to repair (see _runtime).
-        async with _runtime(verify=False, ensure=False):
+        async with _runtime(verify=False, ensure=False, exclusive=True):
             # Reset the queue FIRST so every crash window converges on
             # "queue pending → next scan re-indexes". Doing it after the
             # drop leaves a window where a crash yields empty tables with
diff --git a/src/everos/memory/cascade/__init__.py b/src/everos/memory/cascade/__init__.py
index 7ff6b1bbf..4c659c6dc 100644
--- a/src/everos/memory/cascade/__init__.py
+++ b/src/everos/memory/cascade/__init__.py
@@ -21,6 +21,7 @@
 from ._backfill import BackfillPhase as BackfillPhase
 from ._backfill import BackfillPresenter as BackfillPresenter
 from ._backfill import NullBackfillPresenter as NullBackfillPresenter
+from ._backfill import hold_ome_lock as hold_ome_lock
 from ._backfill import ome_lock_is_free as ome_lock_is_free
 from .orchestrator import CascadeConfig as CascadeConfig
 from .orchestrator import CascadeHealth as CascadeHealth
@@ -38,6 +39,7 @@
     "CascadeOrchestrator",
     "KindSpec",
     "NullBackfillPresenter",
+    "hold_ome_lock",
     "match_kind",
     "ome_lock_is_free",
 ]
diff --git a/src/everos/memory/cascade/_backfill.py b/src/everos/memory/cascade/_backfill.py
index 11f2ac850..ac0516332 100644
--- a/src/everos/memory/cascade/_backfill.py
+++ b/src/everos/memory/cascade/_backfill.py
@@ -25,9 +25,10 @@
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import dataclasses
 import datetime as dt
-from collections.abc import Callable
+from collections.abc import Callable, Iterator
 from pathlib import Path
 from typing import Any, Protocol
 from uuid import uuid4
@@ -1095,6 +1096,32 @@ def _probe_ome_lock_available() -> bool:
         handle.close()
 
 
+@contextlib.contextmanager
+def hold_ome_lock() -> Iterator[None]:
+    """Hold the OME jobstore lock for the duration of a CLI write phase.
+
+    Same file and flags as :meth:`OfflineEngine._acquire_lock`, so a server
+    that starts meanwhile fails at startup with :class:`EngineLockHeldError`
+    instead of becoming a second index writer. Raises
+    :class:`EngineLockHeldError` when another process already holds it.
+    """
+    root = MemoryRoot.resolve()
+    lock_path = Path(str(root.ome_db) + ".lock")
+    lock_path.parent.mkdir(parents=True, exist_ok=True)
+    handle = open(lock_path, "a+")  # noqa: SIM115
+    try:
+        try:
+            portalocker.lock(handle, portalocker.LOCK_EX | portalocker.LOCK_NB)
+        except portalocker.LockException as exc:
+            raise EngineLockHeldError(f"another process holds {lock_path}") from exc
+        try:
+            yield
+        finally:
+            portalocker.unlock(handle)
+    finally:
+        handle.close()
+
+
 def _build_cluster_engine() -> OfflineEngine:
     """Construct (but do not start) the throw-away OME engine Phase 2 drives.
 
diff --git a/tests/integration/test_cascade_cli_integration.py b/tests/integration/test_cascade_cli_integration.py
index c89ad7f84..86fc729b6 100644
--- a/tests/integration/test_cascade_cli_integration.py
+++ b/tests/integration/test_cascade_cli_integration.py
@@ -15,11 +15,13 @@
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import datetime as _dt
 import re
 from collections.abc import Iterator
 from pathlib import Path
 
+import portalocker
 import pytest
 from typer.testing import CliRunner
 
@@ -313,6 +315,75 @@ def test_rebuild_refuses_to_run_while_a_server_holds_the_lock(
     assert "rebuild complete" not in combined
 
 
+@contextlib.contextmanager
+def _server_holds_the_lock(root: Path) -> Iterator[None]:
+    """Hold the OME jobstore lock the way a running ``everos server`` does."""
+    lock_path = root / ".index" / "sqlite" / "ome.db.lock"
+    lock_path.parent.mkdir(parents=True, exist_ok=True)
+    with open(lock_path, "a+") as handle:
+        portalocker.lock(handle, portalocker.LOCK_EX | portalocker.LOCK_NB)
+        try:
+            yield
+        finally:
+            portalocker.unlock(handle)
+
+
+def test_sync_refuses_to_run_while_a_server_holds_the_lock(
+    cli_runtime: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """``sync`` is an index writer; next to a running server it exits 3
+    before opening anything (a second writer inserts rows twice).
+    """
+
+    def _boom() -> None:
+        raise AssertionError("the runtime must not open the DB when refusing")
+
+    monkeypatch.setattr(cascade_mod, "get_engine", _boom)
+    with _server_holds_the_lock(cli_runtime):
+        result = CliRunner().invoke(cascade_mod.app, ["sync"])
+
+    assert result.exit_code == 3, result.output
+    assert "holds this memory root" in result.stderr.lower()
+    assert "stop it" in result.stderr.lower()
+    assert "sync complete" not in result.output
+
+
+def test_fix_apply_refuses_but_fix_and_status_still_run_next_to_a_server(
+    cli_runtime: Path,
+) -> None:
+    """Only the writing commands need the root to themselves."""
+    with _server_holds_the_lock(cli_runtime):
+        applied = CliRunner().invoke(cascade_mod.app, ["fix", "--apply"])
+        listed = CliRunner().invoke(cascade_mod.app, ["fix"])
+        status = CliRunner().invoke(cascade_mod.app, ["status"])
+
+    assert applied.exit_code == 3, applied.output
+    assert "holds this memory root" in applied.stderr.lower()
+    assert listed.exit_code == 0, listed.output + listed.stderr
+    assert status.exit_code == 0, status.output + status.stderr
+
+
+def test_sync_holds_the_lock_while_it_drains(
+    cli_runtime: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """The lock is held for the whole drain, not probed and released — a
+    server starting meanwhile must fail at its own lock, not join in.
+    """
+    seen: list[bool] = []
+
+    class _Orchestrator:
+        async def sync_once(self) -> int:
+            seen.append(cascade_mod.ome_lock_is_free())
+            return 0
+
+    monkeypatch.setattr(cascade_mod, "_build_orchestrator", lambda: _Orchestrator())
+    result = CliRunner().invoke(cascade_mod.app, ["sync"])
+
+    assert result.exit_code == 0, result.output + result.stderr
+    assert seen == [False]  # the lock was ours during the drain
+    assert cascade_mod.ome_lock_is_free()  # and released afterwards
+
+
 # Reduce false negatives on date drift.
 def test_resolve_relative_via_command_arg(cli_runtime: Path) -> None:
     """An absolute path under the root works through ``cascade sync <path>``."""