diff --git a/agent/chat_completion_helpers.py b/agent/chat_completion_helpers.py
index 364af69085..6a87c7c1e0 100644
--- a/agent/chat_completion_helpers.py
+++ b/agent/chat_completion_helpers.py
@@ -2353,7 +2353,7 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
         for msg in messages:
             api_msg = msg.copy()
             agent._copy_reasoning_content_for_api(msg, api_msg)
-            for internal_field in ("reasoning", "finish_reason", "_thinking_prefill"):
+            for internal_field in ("reasoning", "finish_reason"):
                 api_msg.pop(internal_field, None)
             # Strict OpenAI-compatible gateways (Fireworks-backed OpenCode Go,
             # Mistral, Moonshot/Kimi) reject any message key outside the Chat
@@ -2376,8 +2376,6 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
             # sidecar-carrying message and re-prefilling the whole transcript
             # at exactly the moment the context is largest.
             substitute_api_content(api_msg)
-            for internal_key in [k for k in api_msg if isinstance(k, str) and k.startswith("_")]:
-                api_msg.pop(internal_key, None)
             if _needs_sanitize:
                 # In MoA mode, agent.model is the virtual preset name,
                 # not the actual aggregator model.  Resolve the real
@@ -2411,8 +2409,18 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
 
         # Same safety net as the main loop: drop thinking-only assistant
         # turns so Anthropic-family providers don't 400 the summary call.
+        # _thinking_prefill must survive until here so the drop pass can
+        # recognize stubs after reasoning fields are stripped.
         api_messages = agent._drop_thinking_only_and_merge_users(api_messages)
 
+        # Strip all remaining underscore-prefixed scaffolding keys before the
+        # wire. The summary path calls chat.completions.create() directly,
+        # bypassing the transport's universal underscore-key sweeper.
+        for api_msg in api_messages:
+            if isinstance(api_msg, dict):
+                for internal_key in [k for k in api_msg if isinstance(k, str) and k.startswith("_")]:
+                    api_msg.pop(internal_key, None)
+
         summary_extra_body = {}
         try:
             from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE as _OMIT_TEMP
diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 95648b0698..1a236e6eac 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -1907,13 +1907,8 @@ def run_conversation(
             # Remove finish_reason - not accepted by strict APIs (e.g. Mistral)
             if "finish_reason" in api_msg:
                 api_msg.pop("finish_reason")
-            # ``_thinking_prefill`` deliberately survives here. The reasoning
-            # fields it accompanies were just stripped above, which leaves
-            # ``_is_thinking_only_assistant`` no way to recognize the stub, so
-            # ``_drop_thinking_only_and_merge_users`` below could never fire on
-            # it. The transport strips every leading-underscore key before the
-            # wire (agent/transports/chat_completions.py), so keeping it costs
-            # nothing.
+            # _thinking_prefill survives here intentionally: the drop pass below
+            # needs it. The transport strips all underscore keys before the wire.
             # Strip length-continuation marks; not every transport drops underscore keys.
             api_msg.pop("_length_continuation_fragment", None)
             api_msg.pop("_length_continuation_nudge", None)
diff --git a/run_agent.py b/run_agent.py
index 275cca799c..64d3491a84 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -4626,14 +4626,8 @@ class AIAgent:
             return False
         if msg.get("tool_calls"):
             return False
-        # A prefill stub is thinking-only by construction: the loop only sets
-        # ``_thinking_prefill`` on a turn that had reasoning and no visible
-        # text. It has to be checked ahead of the content inspection below,
-        # because both things this function would otherwise key on are gone by
-        # the time the drop pass runs. The reasoning fields are stripped from
-        # the API copy for providers that don't echo them back, and
-        # ``repair_empty_non_final_messages`` rewrites a non-final stub's empty
-        # content to a placeholder, which would read here as real output.
+        # Prefill stubs are thinking-only by construction; check before content
+        # inspection since repair_empty_non_final_messages may have healed content.
         if msg.get("_thinking_prefill"):
             return True
         # Does it have any actual output?