diff --git a/emrg/server/scheduler.py b/emrg/server/scheduler.py
index 574e11b5..d565cc47 100644
--- a/emrg/server/scheduler.py
+++ b/emrg/server/scheduler.py
@@ -3713,7 +3713,29 @@ def _ensure_emrg_agent_entry(self) -> None:
         # The write goes through `AgentRegistry`, so the register keeps one writer
         # and one temp prefix; only the row's *meaning* (the evolution workspace) is
         # the scheduler's to decide.
-        reg = AgentRegistry(agents_file(config_dir()))
+        #
+        # Read and write must address the *same* file, and it must be the one the
+        # register is actually at. `agents_file` names where the register belongs,
+        # and a register still at its pre-rename name is what `register_for_reading`
+        # answers — the derivation every reader in this package now uses (#2546).
+        # Asking `agents_file` here instead put the two halves at two different
+        # files: `AgentRegistry._read` maps a missing file to `[]`, so this saw no
+        # `emrg` row, upserted one, and created a *second* register holding only
+        # that row — which, from the next call on, is the file `register_for_reading`
+        # prefers, so every other entry vanished from every reader. A write that
+        # cannot address the register it just read is worse than no write, so when
+        # the two names disagree this declines and leaves the move to
+        # `migrate_legacy_registry`, which the daemon runs at startup, before this
+        # function exists.
+        register = register_for_reading(config_dir())
+        if register != agents_file(config_dir()):
+            logger.info(
+                "TaskScheduler: agents.yml is still at %s — leaving the move to the "
+                "startup migration rather than writing a second register",
+                register.name,
+            )
+            return
+        reg = AgentRegistry(register)
         target = str(Path.home() / ".emrg" / "evolution" / "emrg")
         try:
             row = next((e for e in reg.entries() if e.get("name") == "emrg"), None)
diff --git a/tests/test_scheduler.py b/tests/test_scheduler.py
index 06bcdb70..a1ffe703 100644
--- a/tests/test_scheduler.py
+++ b/tests/test_scheduler.py
@@ -1892,6 +1892,55 @@ def test_ensure_emrg_agent_entry_other_entries_preserved(tmp_path):
     assert "emrg" in names
     assert len(names) == 2
 
+
+def test_ensure_emrg_agent_entry_declines_when_the_register_is_at_the_legacy_name(tmp_path):
+    """The read half and the write half must address the same file.
+
+    A register still at the pre-rename name is a state the move can be stuck in
+    (a read-only home, a permission boundary, a concurrent caller) and the state
+    `register_for_reading` exists for. Asking `agents_file` alone here did not
+    read that register — `AgentRegistry._read` maps a missing file to `[]` — so
+    the emrg row looked absent, and the upsert created a *second* register under
+    the new name holding only that row. Worse than the write itself: the new file
+    is the one `register_for_reading` prefers from then on, so the entries left
+    at the old name disappeared from every reader at once.
+
+    The emrg row here is deliberately a dead path, which is the shape that gets
+    past the "an existing entry with a live directory is left alone" early return
+    and really reaches the write.
+    """
+    from emrg.config import register_for_reading
+    from emrg.server import scheduler as mod
+    from emrg.server.scheduler import TaskScheduler
+
+    legacy = tmp_path / "projects.yml"
+    before = yaml.safe_dump([
+        {"name": "emrg", "path": str(tmp_path / "gone" / "emrg")},  # dead on purpose
+        {"name": "other", "path": str(tmp_path / "other")},
+    ])
+    legacy.write_text(before, encoding="utf-8")
+
+    sched = TaskScheduler(InstanceIdentity())
+
+    orig_config = mod.config_dir
+    try:
+        mod.config_dir = lambda: tmp_path
+        sched._ensure_emrg_agent_entry()
+    finally:
+        mod.config_dir = orig_config
+
+    assert not (tmp_path / "agents.yml").exists(), (
+        "a second register was created: the write half did not address the file "
+        "the read half had just read"
+    )
+    assert legacy.read_text(encoding="utf-8") == before, "the legacy register is untouched"
+    register = register_for_reading(tmp_path)
+    assert register == legacy
+    assert [e["name"] for e in yaml.safe_load(register.read_text(encoding="utf-8"))] == [
+        "emrg", "other",
+    ], "every entry the register held is still there"
+
+
 # ── TaskHandler cycle truncation detection ──────────────────
 # mem-repo lesson (tool-call truncation must be flagged, not silently
 # treated as a successful/empty cycle — #523 applied it to the chat UI;