diff --git a/tests/test_triggers.py b/tests/test_triggers.py
new file mode 100644
index 0000000..544f605
--- /dev/null
+++ b/tests/test_triggers.py
@@ -0,0 +1,86 @@
+"""Regression tests for the whole-word chat triggers (Redmine #49439).
+
+Commit 86c924e converted the last twelve bare-substring triggers in
+``TwitchBot._handle_triggers`` to the ``_word_in`` whole-word helper. These
+tests exercise the real trigger path with the socket-facing methods stubbed,
+and assert three things per converted trigger:
+
+1. NEGATIVE CONTROL: the old bare-substring check *does* match the
+   false-positive message, so the bug being fixed was real and the case is
+   not vacuously green.
+2. The fixed code does not respond to that false-positive message at all.
+3. The fixed code still responds to a legitimate whole-word message.
+"""
+
+import pytest
+
+import twitch
+
+
+@pytest.fixture
+def fire(monkeypatch):
+    """Return ``fire(message) -> (messages_sent, sang_major_general)``.
+
+    Everything patched here goes through ``monkeypatch`` so it is undone at
+    the end of the test rather than leaking into the rest of the session.
+    The inter-message delays are zeroed instead of stubbing ``time.sleep``,
+    which would otherwise be a process-wide change to the stdlib module.
+    """
+    monkeypatch.setattr(twitch, "MESSAGE_DELAY", 0)
+    monkeypatch.setattr(twitch, "LONG_MESSAGE_DELAY", 0)
+
+    bot = twitch.TwitchBot()
+    sent: list[str] = []
+    sang: list[bool] = []
+    monkeypatch.setattr(bot, "send_message", lambda _channel, message: sent.append(message))
+    monkeypatch.setattr(bot, "_sing_major_general", lambda: sang.append(True))
+
+    def _fire(message: str) -> tuple[list[str], bool]:
+        sent.clear()
+        sang.clear()
+        bot._handle_triggers(message)
+        return list(sent), bool(sang)
+
+    return _fire
+
+
+# (trigger, false-positive message, legitimate message, expected reply fragment,
+#  whether the trigger answers with the Major-General song instead of a message)
+CASES = [
+    ("lemons", "nice lemonsqueezer", "when life gives you lemons", "When life gives you lemons", False),
+    ("rainbow", "double rainbows everywhere", "sing rainbow connection", "the rainbow connection", False),
+    ("racist", "reading antiracist books", "that's racist", "Everyone's a little bit racist", False),
+    ("hitler", "studying hitlerism", "hitler jokes again", "Springtime for Hitler", False),
+    ("nazi", "the rise of nazism", "grammar nazi", "join the Nazi party", False),
+    ("major-general", "the major-generals arrived", "do the major-general song", None, True),
+    ("thanks", "happy thanksgiving", "thanks bot", "So what can I say except you're welcome?", False),
+    ("thank you", "go thank yourself", "thank you so much", "my way of being me", False),
+    ("boobs", "boobsledding season", "boobs", "BOOBS!", False),
+    ("boobies", "a boobiesque painting", "boobies", "BOOBIES!", False),
+    ("yay", "yayyy", "yay!", "Yay! ^_^", False),
+    ("crazy", "welcome to crazytown", "that's crazy", "Crazy? I was crazy once", False),
+]
+
+
+@pytest.mark.parametrize(
+    ("word", "fp_message", "ok_message", "reply_fragment", "sings_major_general"),
+    CASES,
+    ids=[case[0] for case in CASES],
+)
+def test_trigger_matches_whole_words_only(fire, word, fp_message, ok_message, reply_fragment, sings_major_general):
+    # 1. Negative control: the pre-fix bare-substring check fired on this message.
+    assert word in fp_message.lower(), f"{word!r} is not a substring of {fp_message!r}"
+
+    # 2. The false positive must no longer draw any response.
+    sent, sang = fire(fp_message)
+    assert sent == [], f"{word!r} still fires on {fp_message!r}"
+    assert not sang, f"{word!r} still sings the Major-General on {fp_message!r}"
+
+    # 3. A legitimate whole-word use must still fire.
+    sent, sang = fire(ok_message)
+    if sings_major_general:
+        assert sang, f"{word!r} no longer sings the Major-General on {ok_message!r}"
+    else:
+        assert any(reply_fragment in message for message in sent), (
+            f"{word!r} no longer fires on {ok_message!r}: sent={sent}"
+        )