v2.3.0 (#429)

Co-authored-by: Carlo <carloantonio.patti@mistral.ai>
Co-authored-by: Mathias Gesbert <mathias.gesbert@mistral.ai>
Co-authored-by: Clement Sirieix <clem.sirieix@gmail.com>
Co-authored-by: Michel Thomazo <michel.thomazo@mistral.ai>
Co-authored-by: Clément Drouin <clement.drouin@mistral.ai>
Co-authored-by: Vincent Guilloux <vincent.guilloux@mistral.ai>
Co-authored-by: Thomas Kenbeek <thomas.kenbeek@mistral.ai>
Co-authored-by: Mistral Vibe <vibe@mistral.ai>

tests/test_agent_auto_compact.py

@@ -16,22 +16,13 @@ from vibe.core.types import (
 
 
 @pytest.mark.asyncio
-async def test_auto_compact_triggers_and_batches_observer(
-    telemetry_events: list[dict],
-) -> None:
-    observed: list[tuple[Role, str | None]] = []
-
-    def observer(msg: LLMMessage) -> None:
-        observed.append((msg.role, msg.content))
-
+async def test_auto_compact_emits_correct_events(telemetry_events: list[dict]) -> None:
     backend = FakeBackend([
         [mock_llm_chunk(content="<summary>")],
         [mock_llm_chunk(content="<final>")],
     ])
     cfg = build_test_vibe_config(auto_compact_threshold=1)
-    agent = build_test_agent_loop(
-        config=cfg, message_observer=observer, backend=backend
-    )
+    agent = build_test_agent_loop(config=cfg, backend=backend)
     agent.stats.context_tokens = 2
 
     events = [ev async for ev in agent.act("Hello")]
@@ -50,14 +41,83 @@ async def test_auto_compact_triggers_and_batches_observer(
     assert end.new_context_tokens >= 1
     assert final.content == "<final>"
 
-    roles = [r for r, _ in observed]
-    assert roles == [Role.system, Role.user, Role.assistant]
-    assert observed[1][1] is not None and "<summary>" in observed[1][1]
-    assert observed[2][1] == "<final>"
-
     auto_compact = [
         e
         for e in telemetry_events
-        if e.get("event_name") == "vibe/auto_compact_triggered"
+        if e.get("event_name") == "vibe.auto_compact_triggered"
     ]
     assert len(auto_compact) == 1
+
+
+@pytest.mark.asyncio
+async def test_auto_compact_observer_sees_user_msg_not_summary() -> None:
+    """Observer sees the original user message and final response.
+
+    Compact internals (summary request, LLM summary) are invisible
+    to the observer because they happen inside silent() / reset().
+    """
+    observed: list[tuple[Role, str | None]] = []
+
+    def observer(msg: LLMMessage) -> None:
+        observed.append((msg.role, msg.content))
+
+    backend = FakeBackend([
+        [mock_llm_chunk(content="<summary>")],
+        [mock_llm_chunk(content="<final>")],
+    ])
+    cfg = build_test_vibe_config(auto_compact_threshold=1)
+    agent = build_test_agent_loop(
+        config=cfg, message_observer=observer, backend=backend
+    )
+    agent.stats.context_tokens = 2
+
+    [_ async for _ in agent.act("Hello")]
+
+    roles = [r for r, _ in observed]
+    assert roles == [Role.system, Role.user, Role.assistant]
+    assert observed[1][1] == "Hello"
+    assert observed[2][1] == "<final>"
+
+
+@pytest.mark.asyncio
+async def test_auto_compact_observer_does_not_see_summary_request() -> None:
+    """The compact summary request and LLM response must not leak to observer."""
+    observed: list[tuple[Role, str | None]] = []
+
+    def observer(msg: LLMMessage) -> None:
+        observed.append((msg.role, msg.content))
+
+    backend = FakeBackend([
+        [mock_llm_chunk(content="<summary>")],
+        [mock_llm_chunk(content="<final>")],
+    ])
+    cfg = build_test_vibe_config(auto_compact_threshold=1)
+    agent = build_test_agent_loop(
+        config=cfg, message_observer=observer, backend=backend
+    )
+    agent.stats.context_tokens = 2
+
+    [_ async for _ in agent.act("Hello")]
+
+    contents = [c for _, c in observed]
+    assert "<summary>" not in contents
+    assert all("compact" not in (c or "").lower() for c in contents)
+
+
+@pytest.mark.asyncio
+async def test_compact_replaces_messages_with_summary() -> None:
+    """After compact, messages list contains only system + summary."""
+    backend = FakeBackend([
+        [mock_llm_chunk(content="<summary>")],
+        [mock_llm_chunk(content="<final>")],
+    ])
+    cfg = build_test_vibe_config(auto_compact_threshold=1)
+    agent = build_test_agent_loop(config=cfg, backend=backend)
+    agent.stats.context_tokens = 2
+
+    [_ async for _ in agent.act("Hello")]
+
+    # After compact + final response: system, summary, final
+    assert agent.messages[0].role == Role.system
+    assert agent.messages[-1].role == Role.assistant
+    assert agent.messages[-1].content == "<final>"