diff --git a/backend/apps/agents/manager/run/RunOptions.py b/backend/apps/agents/manager/run/RunOptions.py
index 38d281bb..1b84664a 100644
--- a/backend/apps/agents/manager/run/RunOptions.py
+++ b/backend/apps/agents/manager/run/RunOptions.py
@@ -320,6 +320,31 @@ class RunOptions(AgentManagerProtocol):
             if distilled:
                 fenced = wrap_platform_note(f"Summary of earlier conversation (older turns compacted):\n{distilled}")
                 history = f"{fenced}\n\n{history}" if history else fenced
+            elif session.compacted_through_msg_id and p_mode != "none":
+                # A rebuild that drops old turns AND carries no gist of them is the agent's memory
+                # loss, and until now it happened in total silence: the distiller returns "" for five
+                # different reasons (feature off, cutoff gone from the branch, empty body, the aux
+                # call raising, the aux call returning nothing) and only one of them logs, at DEBUG.
+                # The recap never carries the model's own replies by design, so this summary IS the
+                # memory; without it the agent knows what was ASKED and which tools ran, but not what
+                # it concluded, which is exactly the "confidently describes work that isn't there"
+                # report. Say it out loud and stamp it, so the next report is diagnosable instead of
+                # unexplainable (the ENG-397 move).
+                logger.warning(
+                    f"Agent {session_id}: rebuilding past a compaction cutoff with NO distilled "
+                    f"summary; this turn loses the gist of its own earlier work (prefix={p_mode})"
+                )
+                try:
+                    from backend.apps.service.client import submit_diagnostic
+                    submit_diagnostic({
+                        "kind": "recap_summary_missing",
+                        "session_id": session_id,
+                        "model": session.model,
+                        "prefix_mode": p_mode,
+                        "recap_chars": len(history or ""),
+                    })
+                except Exception:
+                    logger.debug("submit_diagnostic recap_summary_missing failed", exc_info=True)
             if history:
                 # SYSTEM channel, not the user message (ENG-358 structural fix): a transcript recap
                 # inside user content is byte-for-byte what anti-distillation filters hunt, and no
diff --git a/backend/tests/test_recap_summary_missing_is_loud.py b/backend/tests/test_recap_summary_missing_is_loud.py
new file mode 100644
index 00000000..fbbe2785
--- /dev/null
+++ b/backend/tests/test_recap_summary_missing_is_loud.py
@@ -0,0 +1,60 @@
+"""The agent's memory loss had no signal at all, which is why it kept being unexplainable.
+
+A fresh-session rebuild sends the recap (the user's asks + the tool trail, never the model's own
+replies, by design) plus a model-written distilled summary of the dropped span. That summary IS the
+agent's memory of its own earlier conclusions. `distilled_history_summary` returns "" for FIVE
+different reasons, and only one of them logged, at DEBUG:
+
+    feature off | cutoff no longer on the branch | empty body | aux call raised | aux returned ""
+
+The aux call is a real LLM call, so it fails whenever aux is unavailable (observed live:
+"No AI provider connected for auxiliary LLM call"). When it does, the rebuilt turn knows what was
+ASKED and which tools ran, but not what it CONCLUDED, which is exactly the field report: the agent
+confidently described work that did not exist, then retracted it after re-reading the workspace.
+
+Rule: a guard may never disable itself in silence.
+"""
+import inspect
+
+from backend.apps.agents.manager.run import RunOptions as RO
+from backend.apps.agents.manager.session import distill_history
+
+
+def p_src():
+    return inspect.getsource(RO)
+
+
+def test_a_rebuild_with_no_summary_warns_and_reports():
+    src = p_src()
+    i = src.index('elif session.compacted_through_msg_id and p_mode != "none":')
+    block = src[i:i + 1600]
+    assert "logger.warning" in block, "silence is the bug; it must say so"
+    assert '"kind": "recap_summary_missing"' in block, "and it must be queryable on the fleet"
+    assert "session_id" in block, "ENG-397's lesson: an envelope with no session is untestable"
+
+
+def test_it_only_fires_when_memory_was_actually_expected():
+    """No cutoff means nothing was dropped, so nothing was lost; `none` means a policy block already
+    stripped the recap deliberately. Neither is the memory-loss case."""
+    src = p_src()
+    i = src.index('elif session.compacted_through_msg_id and p_mode != "none":')
+    cond = src[i:i + 90]
+    assert "compacted_through_msg_id" in cond
+    assert 'p_mode != "none"' in cond
+
+
+def test_it_sits_on_the_else_of_the_summary_being_present():
+    """It must be the ELSE of `if distilled:`, or it would fire on healthy rebuilds too."""
+    src = p_src()
+    assert src.index("if distilled:") < src.index('elif session.compacted_through_msg_id')
+
+
+def test_the_distiller_really_has_five_silent_exits():
+    """If this count ever changes, the comment above the warning is stale and should be re-read."""
+    src = inspect.getsource(distill_history.distilled_history_summary)
+    assert src.count('return ""') == 5, f"expected 5 silent-empty exits, found {src.count('return ')}"
+
+
+def test_the_feature_is_on_by_default():
+    """A default-off distiller would make every long chat forget, which is worth knowing loudly."""
+    assert distill_history.DISTILL_ENABLED is True