diff --git a/tests/agents/test_tool_use_agent.py b/tests/agents/test_tool_use_agent.py
new file mode 100644
index 000000000..6b6cd5a0e
--- /dev/null
+++ b/tests/agents/test_tool_use_agent.py
@@ -0,0 +1,88 @@
+from agentlab.agents.tool_use_agent.tool_use_agent import StructuredDiscussion
+from agentlab.llm.response_api import AnthropicAPIMessageBuilder
+
+
+def make_msg(text: str) -> AnthropicAPIMessageBuilder:
+    msg = AnthropicAPIMessageBuilder.user().add_text(text)
+    msg._cache_breakpoint = False
+    return msg
+
+
+def make_discussion(
+    group_specs: list[tuple[int, bool]], keep_last_n_obs: int | None = None
+) -> StructuredDiscussion:
+    """Build a discussion with one group per spec of (n_messages, has_summary)."""
+    discussion = StructuredDiscussion(keep_last_n_obs=keep_last_n_obs)
+    for group_idx, (n_messages, has_summary) in enumerate(group_specs):
+        discussion.new_group(f"group_{group_idx}")
+        for msg_idx in range(n_messages):
+            discussion.append(make_msg(f"g{group_idx}_m{msg_idx}"))
+        if has_summary:
+            discussion.set_last_summary(make_msg(f"g{group_idx}_summary"))
+    return discussion
+
+
+def texts(messages) -> list[str]:
+    return [msg.content[0]["text"] for msg in messages]
+
+
+def test_flatten_empty_discussion():
+    assert StructuredDiscussion().flatten() == []
+
+
+def test_flatten_without_keep_last_n_obs_keeps_everything_verbatim():
+    discussion = make_discussion([(2, True), (1, True), (1, False)])
+
+    assert texts(discussion.flatten()) == ["g0_m0", "g0_m1", "g1_m0", "g2_m0"]
+
+
+def test_flatten_summarizes_all_but_the_tail():
+    discussion = make_discussion([(2, True), (2, True), (2, True)], keep_last_n_obs=1)
+
+    assert texts(discussion.flatten()) == ["g0_summary", "g1_summary", "g2_m0", "g2_m1"]
+
+
+def test_flatten_keeps_the_last_n_groups_verbatim():
+    discussion = make_discussion([(1, True), (1, True), (1, True), (1, True)], keep_last_n_obs=2)
+
+    assert texts(discussion.flatten()) == ["g0_summary", "g1_summary", "g2_m0", "g3_m0"]
+
+
+def test_flatten_falls_back_to_messages_when_group_has_no_summary():
+    discussion = make_discussion([(2, False), (1, True), (1, True)], keep_last_n_obs=1)
+
+    assert texts(discussion.flatten()) == ["g0_m0", "g0_m1", "g1_summary", "g2_m0"]
+
+
+def test_flatten_keeps_everything_when_keep_last_n_obs_exceeds_group_count():
+    discussion = make_discussion([(1, True), (1, True)], keep_last_n_obs=5)
+    messages = discussion.flatten()
+
+    assert texts(messages) == ["g0_m0", "g1_m0"]
+    assert [msg._cache_breakpoint for msg in messages] == [False, False]
+
+
+def test_flatten_sets_a_single_cache_breakpoint_at_the_tail_boundary():
+    discussion = make_discussion([(1, True), (1, True), (2, True)], keep_last_n_obs=1)
+    # pre-set a stale breakpoint on an earlier message to check it gets cleared
+    discussion.groups[0].summary._cache_breakpoint = True
+
+    messages = discussion.flatten()
+
+    assert texts(messages) == ["g0_summary", "g1_summary", "g2_m0", "g2_m1"]
+    # the breakpoint sits on the first message of the tail group, earlier ones are cleared
+    assert [msg._cache_breakpoint for msg in messages] == [False, False, True, False]
+
+
+def test_flatten_marks_the_first_message_when_the_whole_discussion_is_the_tail():
+    discussion = make_discussion([(1, True), (1, True)], keep_last_n_obs=2)
+    messages = discussion.flatten()
+
+    assert [msg._cache_breakpoint for msg in messages] == [True, False]
+
+
+def test_flatten_is_stable_across_calls():
+    discussion = make_discussion([(1, True), (1, True), (1, False)], keep_last_n_obs=1)
+
+    assert texts(discussion.flatten()) == texts(discussion.flatten())
+    assert [msg._cache_breakpoint for msg in discussion.flatten()] == [False, False, True]