diff --git a/navi/api/routes/sessions.py b/navi/api/routes/sessions.py
index 84c9f0e..20b5925 100644
--- a/navi/api/routes/sessions.py
+++ b/navi/api/routes/sessions.py
@@ -579,8 +579,19 @@
     if not user_messages:
         return {"name": None}
 
+    # Short first user messages often carry no topic (greetings, one-liners) —
+    # that's why a share of sessions stayed unnamed. The agent's final replies
+    # are where the actual subject lands; fold them into the naming context.
+    assistant_messages = [
+        m.content for m in session.messages
+        if m.role == "assistant" and m.content
+    ][-2:]
+
     backend = backends.get("ollama")
-    name = await generate_session_name(user_messages, backend, settings.ollama_default_model)
+    name = await generate_session_name(
+        user_messages, backend, settings.ollama_default_model,
+        assistant_messages=assistant_messages,
+    )
 
     if name:
         await store.set_name(session_id, name)
diff --git a/navi/core/name_generator.py b/navi/core/name_generator.py
index 584452c..7963a50 100644
--- a/navi/core/name_generator.py
+++ b/navi/core/name_generator.py
@@ -1,27 +1,44 @@
-"""Generate a short session name from user messages via LLM."""
+"""Generate a short session name from the conversation via LLM."""
 
 from navi.llm.base import LLMBackend, Message
 
 _SYSTEM = (
     "You are a session title generator. "
-    "Given the user's messages from a conversation, produce ONE short title (3–6 words, no punctuation at the end). "
+    "Given the user's messages (and, possibly, the assistant's replies) from a conversation, "
+    "produce ONE short title (3–6 words, no punctuation at the end). "
     "The title must reflect the actual topic. "
     "Respond in the same language as the user's messages. "
-    "If the messages contain no clear topic yet (e.g. only greetings, very short or ambiguous text), "
+    "If there's still no clear topic (e.g. only greetings, very short or ambiguous text), "
     "reply with exactly: NO_TITLE"
 )
 
+# Assistant replies can be huge; a tail slice is enough context for a title.
+_ASSISTANT_MSG_CHARS = 600
+_ASSISTANT_MSG_COUNT = 2
+
 
 async def generate_session_name(
     user_messages: list[str],
     backend: LLMBackend,
     model: str,
+    assistant_messages: list[str] | None = None,
 ) -> str | None:
     """Return a short title or None if content isn't substantial enough."""
     if not user_messages:
         return None
 
     combined = "\n".join(f"- {m}" for m in user_messages[:10])
+
+    # Naming runs right after the first exchange, when user messages are often
+    # greetings or one-liners with no topic. The agent's final response is
+    # where the actual subject shows up — include a trimmed tail of it.
+    if assistant_messages:
+        combined += "\n\nAssistant replies:\n" + "\n---\n".join(
+            m[-_ASSISTANT_MSG_CHARS:]
+            for m in assistant_messages[-_ASSISTANT_MSG_COUNT:]
+            if m
+        )
+
     messages = [
         Message(role="system", content=_SYSTEM),
         Message(role="user", content=f"User messages:\n{combined}"),
diff --git a/tests/unit/core/test_name_generator.py b/tests/unit/core/test_name_generator.py
new file mode 100644
index 0000000..2c01f87
--- /dev/null
+++ b/tests/unit/core/test_name_generator.py
@@ -0,0 +1,55 @@
+"""Unit tests for the session name generator."""
+
+import pytest
+
+from navi.core.name_generator import generate_session_name
+from navi.llm.base import LLMBackend, Message
+
+
+class _FakeBackend(LLMBackend):
+    def __init__(self, reply: str):
+        self.reply = reply
+        self.last_prompt = None
+
+    async def complete(self, messages, tools=None, model=None, **kwargs):
+        self.last_prompt = messages[-1].content
+        return type("R", (), {"content": self.reply})()
+
+    async def stream_complete(self, messages, tools=None, model=None, **kwargs):
+        yield {"type": "text", "text": self.reply}
+
+
+@pytest.mark.asyncio
+async def test_folds_assistant_replies_into_context():
+    backend = _FakeBackend("Setup of TLS on the proxy")
+    name = await generate_session_name(
+        ["привет", "ну так продолжай"],
+        backend,
+        "test-model",
+        assistant_messages=["Сделаю шаги: 1) сгенерирую ключи..." + ("x" * 1000)],
+    )
+    assert name == "Setup of TLS on the proxy"
+    # The assistant reply is trimmed to its 600-char tail, not passed whole
+    assert "xxxxx" not in backend.last_prompt.replace("x" * 600, "")
+    tail_len = len("Assistant replies:\n") + 600
+    assert len(backend.last_prompt) - len(backend.last_prompt.split("Assistant replies:\n")[0]) == tail_len
+
+
+@pytest.mark.asyncio
+async def test_no_topic_still_yields_none():
+    backend = _FakeBackend("NO_TITLE")
+    name = await generate_session_name(
+        ["привет", "как дела?"],
+        backend,
+        "test-model",
+        assistant_messages=["Тоже привет!"],
+    )
+    assert name is None
+
+
+@pytest.mark.asyncio
+async def test_without_assistant_replies_unchanged():
+    backend = _FakeBackend("Debugging the deploy")
+    name = await generate_session_name(["fix the deploy"], backend, "test-model")
+    assert name == "Debugging the deploy"
+    assert "Assistant replies:" not in backend.last_prompt
\ No newline at end of file