diff --git a/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py b/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
index d23a4fc51..e33f7fd03 100644
--- a/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
+++ b/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
@@ -747,9 +747,24 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 converted_parts = []
                 for part in content:
                     if isinstance(part, dict) and part.get('type') == 'image_base64':
-                        part['image_url'] = {'url': part['image_base64']}
-                        part['type'] = 'image_url'
-                        del part['image_base64']
+                        # History trimming (SessionManager) clears image_base64
+                        # on past turns and exclude_none serialization drops
+                        # the key entirely, so the replayed part may carry no
+                        # payload. Prefer the base64 payload; fall back to an
+                        # image_url that survived on the same element; drop
+                        # hollow parts instead of raising KeyError (#2469).
+                        image_b64 = part.get('image_base64')
+                        fallback_url = None
+                        if not image_b64:
+                            raw_image_url = part.get('image_url')
+                            if isinstance(raw_image_url, dict):
+                                fallback_url = raw_image_url.get('url')
+                        if image_b64 or fallback_url:
+                            part['image_url'] = {'url': image_b64 or fallback_url}
+                            part['type'] = 'image_url'
+                            part.pop('image_base64', None)
+                        else:
+                            continue
                     # OpenAI-compatible chat models reject non-image file parts
                     # (audio/document base64 or url). These originate from Voice /
                     # File attachments — including ones replayed from conversation
diff --git a/tests/unit_tests/provider/test_litellm_convert_messages.py b/tests/unit_tests/provider/test_litellm_convert_messages.py
index 87ad2e027..989db9bae 100644
--- a/tests/unit_tests/provider/test_litellm_convert_messages.py
+++ b/tests/unit_tests/provider/test_litellm_convert_messages.py
@@ -91,3 +91,42 @@ def test_convert_messages_plain_string_content_untouched():
     msg = provider_message.Message(role='user', content='just text')
     out = req._convert_messages([msg])
     assert out[0]['content'] == 'just text'
+
+
+def test_convert_messages_replayed_image_without_base64_does_not_crash():
+    """Replayed image parts hollowed out by history trimming must not raise KeyError (#2469).
+
+    SessionManager clears image_base64 on past turns, and URL-less platform
+    images never had a URL, so the replayed part serializes as
+    {'type': 'image_base64'} with no payload keys. The hollow part should be
+    dropped while the sibling text part survives.
+    """
+    req = _make_requester()
+    image = provider_message.ContentElement.from_image_base64('data:image/jpeg;base64,AAAA')
+    # Simulate SessionManager.trim_conversation_messages clearing binary payloads.
+    image.image_base64 = None
+    msg = provider_message.Message(
+        role='user',
+        content=[
+            provider_message.ContentElement.from_text('describe the photo'),
+            image,
+        ],
+    )
+    out = req._convert_messages([msg])
+    assert [p.get('type') for p in out[0]['content']] == ['text']
+
+
+def test_convert_messages_replayed_image_with_url_falls_back_to_url():
+    """When base64 was trimmed but image_url survived, rebuild the OpenAI image_url part from the URL."""
+    req = _make_requester()
+    image = provider_message.ContentElement(
+        type='image_base64',
+        image_base64=None,
+        image_url=provider_message.ImageURLContentObject(url='https://example.com/pic.jpg'),
+    )
+    msg = provider_message.Message(role='user', content=[image])
+    out = req._convert_messages([msg])
+    parts = out[0]['content']
+    assert [p.get('type') for p in parts] == ['image_url']
+    assert parts[0]['image_url'] == {'url': 'https://example.com/pic.jpg'}
+    assert 'image_base64' not in parts[0]