From b013ed03e539e94466a0aaafc4d67e40cdeb888d Mon Sep 17 00:00:00 2001
From: Teknium <127238744+teknium1@users.noreply.github.com>
Date: Tue, 14 Jul 2026 05:31:07 -0700
Subject: [PATCH] fix(moa): scope the non-text placeholder to structured
 content only

Follow-up to the cherry-picked empty-user-turn drop: the placeholder
introduced in 8582f35d9 fired for whitespace-only STRING turns too
(content='   ' flattens to non-stripping text but isn't in the
(None, '', []) exclusion set), fabricating an attachment note for a turn
that carried nothing. Gate the placeholder on isinstance(content, list)
so only genuinely structured (e.g. image-only) turns get it; empty and
whitespace-only string turns now fall through to the drop path.

Edge cases verified: trailing empty user turn still ends the view on the
synthetic advisory marker; an all-empty transcript degenerates to [].
---
 agent/moa_loop.py                     |  6 ++++--
 tests/run_agent/test_moa_loop_mode.py | 24 ++++++++++++++++++++++++
 2 files changed, 28 insertions(+), 2 deletions(-)

diff --git a/agent/moa_loop.py b/agent/moa_loop.py
index d8461e921dc2..28707b2a4f14 100644
--- a/agent/moa_loop.py
+++ b/agent/moa_loop.py
@@ -491,7 +491,7 @@ def _reference_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
         if role == "system":
             continue
         if role == "user":
-            if not text.strip() and content not in (None, "", []):
+            if not text.strip() and isinstance(content, list) and content:
                 # Structured content with no extractable text (e.g. an
                 # image-only turn). Emitting an empty user message would be
                 # dropped/rejected by strict providers (Anthropic 400s on
@@ -499,7 +499,9 @@ def _reference_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
                 # mode), and silently skipping the turn would break
                 # user/assistant alternation in the advisory view. Substitute
                 # a placeholder so the reference knows a non-text turn
-                # happened.
+                # happened. Only structured content qualifies — an empty or
+                # whitespace-only STRING turn carries nothing and is dropped
+                # below instead.
                 text = "[user sent non-text content (e.g. an image attachment)]"
             if not text.strip():
                 # Genuinely empty user turn (content="" / None). It carries
diff --git a/tests/run_agent/test_moa_loop_mode.py b/tests/run_agent/test_moa_loop_mode.py
index 659f64da7a94..1c1ab8f71eb7 100644
--- a/tests/run_agent/test_moa_loop_mode.py
+++ b/tests/run_agent/test_moa_loop_mode.py
@@ -1241,3 +1241,27 @@ def test_reference_guidance_appends_text_part_to_decorated_trailing_user():
     assert content[0] == marked_part
     # The guidance rides as a trailing text part outside the cached span.
     assert content[1] == {"type": "text", "text": "\n\nREFERENCE BLOCK"}
+
+
+def test_reference_messages_drops_whitespace_only_string_user_turn():
+    """A whitespace-only STRING user turn is dropped, not placeholdered.
+
+    The non-text placeholder exists for structured content (image-only turns)
+    where a real turn happened that the reference should know about. A bare
+    whitespace string carries nothing — emitting it would 400 strict
+    providers (Kimi/Moonshot 'role user must not be empty'), and
+    placeholdering it would fabricate an attachment that never existed.
+    """
+    from agent.moa_loop import _reference_messages
+
+    messages = [
+        {"role": "user", "content": "   "},
+        {"role": "assistant", "content": "a"},
+        {"role": "user", "content": "real"},
+    ]
+
+    view = _reference_messages(messages)
+
+    assert view[0] == {"role": "assistant", "content": "a"}
+    assert view[-1] == {"role": "user", "content": "real"}
+    assert all(str(m["content"]).strip() for m in view)