fix(middleware): drop orphan ToolMessages so strict providers don't 400 (#4080)

* fix(middleware): drop orphan ToolMessages with no matching AIMessage tool_call

The rebuild loop only skipped ToolMessages whose tool_call_id matched a
known AIMessage tool_call (to be re-emitted after it). An orphan ToolMessage
whose tool_call_id has no matching AIMessage tool_calls fell through and was
kept, leaving a dangling tool result that strict providers reject. Drop
orphan ToolMessages as well, logging at debug.

* fix(dangling): demote orphan-drop logs, add tool_call_id=None test

- Update module/class docstrings to mention orphan ToolMessage handling
- Accumulate orphan drop_count and emit a single logger.warning
  instead of per-message logger.debug calls
- Simplify early-return logic: return None only when no patching
  AND no orphans were dropped
- Add test_tool_call_id_none_orphan_is_dropped — a ToolMessage
  with tool_call_id=None is always an orphan and must be dropped

Closes #4080

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(test): use model_construct for None tool_call_id test to bypass pydantic validation

ToolMessage content='ghost' tool_call_id=None fails pydantic validation at
construction. Use model_construct to simulate a corrupt/edge-case payload
without tripping the string-only guard.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

backend/packages/harness/deerflow/agents/middlewares/dangling_tool_call_middleware.py

@@ -1,12 +1,16 @@
-"""Middleware to fix dangling tool calls in message history.
+"""Middleware to fix dangling tool calls and orphan tool results in message history.
 
 A dangling tool call occurs when an AIMessage contains tool_calls but there are
 no corresponding ToolMessages in the history (e.g., due to user interruption or
-request cancellation). This causes LLM errors due to incomplete message format.
+request cancellation). An orphan ToolMessage occurs when a tool result exists
+without a matching AIMessage tool_call (e.g., after summarization/branching
+dropped the upstream AIMessage). Both cause strict-provider rejections.
 
-This middleware intercepts the model call to detect and patch such gaps by
-inserting synthetic ToolMessages with an error indicator immediately after the
-AIMessage that made the tool calls, ensuring correct message ordering.
+This middleware intercepts the model call to:
+- Insert synthetic ToolMessages with an error indicator for each dangling AIMessage
+  tool_call, ensuring correct message ordering
+- Drop orphan ToolMessages whose originating tool_call is no longer present in the
+  request, preventing strict OpenAI-compatible backends from returning HTTP 400
 
 Note: Uses wrap_model_call instead of before_model to ensure patches are inserted
 at the correct positions (immediately after each dangling AIMessage), not appended
@@ -47,11 +51,14 @@ def _has_invalid_tool_name(name: object) -> bool:
 
 
 class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
-    """Inserts placeholder ToolMessages for dangling tool calls before model invocation.
+    """Inserts placeholder ToolMessages for dangling tool calls and drops orphan
+    ToolMessages (tool results whose originating AIMessage tool_call is gone).
 
-    Scans the message history for AIMessages whose tool_calls lack corresponding
-    ToolMessages, and injects synthetic error responses immediately after the
-    offending AIMessage so the LLM receives a well-formed conversation.
+    Scans the message history for:
+    - AIMessages whose tool_calls lack corresponding ToolMessages, and injects
+      synthetic error responses immediately after the offending AIMessage
+    - ToolMessages with no matching AIMessage tool_call (orphans), and drops
+      them so strict OpenAI-compatible backends do not reject the request
     """
 
     @staticmethod
@@ -238,8 +245,17 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
 
         patched: list = []
         patch_count = 0
+        drop_count = 0
         for msg in messages:
-            if isinstance(msg, ToolMessage) and msg.tool_call_id in tool_call_ids:
+            if isinstance(msg, ToolMessage):
+                if msg.tool_call_id in tool_call_ids:
+                    continue  # Will be re-emitted after its AIMessage
+                # Orphan: ToolMessage whose originating AIMessage tool_call is
+                # no longer in the request (e.g. removed by summarization).
+                # Drop it silently from the model request so strict providers
+                # do not reject it with HTTP 400. Persisted state is untouched;
+                # this only affects the single model call.
+                drop_count += 1
                 continue
 
             sanitized_msg = self._sanitize_ai_message_tool_names(msg)
@@ -271,11 +287,14 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
                     )
                     patch_count += 1
 
-        if patched == messages:
+        if patched == messages and not drop_count:
             return None
-
-        if patch_count:
-            logger.warning(f"Injecting {patch_count} placeholder ToolMessage(s) for dangling tool calls")
+        if drop_count or patch_count:
+            logger.warning(
+                "DanglingToolCallMiddleware: %d orphan(s) dropped, %d placeholder(s) injected",
+                drop_count,
+                patch_count,
+            )
         return patched
 
     @override

backend/tests/test_dangling_tool_call_middleware.py

@@ -464,7 +464,14 @@ class TestBuildPatchedMessagesPatching:
         assert isinstance(patched[4], ToolMessage)
         assert patched[4].tool_call_id == "call_2"
 
-    def test_orphan_tool_message_is_preserved_during_grouping(self):
+    def test_orphan_tool_message_is_dropped_during_grouping(self):
+        """An orphan ToolMessage — one whose tool_call_id has no matching AIMessage
+        tool_call — is dropped from the patched output.
+
+        Behavior intentionally changed: strict OpenAI-compatible providers reject a
+        ToolMessage that does not follow an assistant tool_call, so an orphan left
+        over from interruption/compaction must not be forwarded.
+        """
         mw = DanglingToolCallMiddleware()
         orphan = _tool_msg("orphan_call", "orphan")
         msgs = [
@@ -477,12 +484,57 @@ class TestBuildPatchedMessagesPatching:
         patched = mw._build_patched_messages(msgs)
 
         assert patched is not None
+        # The orphan is dropped; call_1's result is regrouped right after its AIMessage.
+        assert isinstance(patched[0], AIMessage)
+        assert isinstance(patched[1], ToolMessage)
+        assert patched[1].tool_call_id == "call_1"
+        assert isinstance(patched[2], HumanMessage)
+        assert orphan not in patched
+        assert patched.count(orphan) == 0
+        assert len(patched) == 3
+
+    def test_leading_orphan_tool_message_is_dropped(self):
+        """A ToolMessage that leads the transcript with no preceding tool_call is an
+        orphan and must be dropped (leaving a valid grouped transcript)."""
+        mw = DanglingToolCallMiddleware()
+        leading_orphan = _tool_msg("stale_call", "stale")
+        msgs = [
+            leading_orphan,
+            _ai_with_tool_calls([_tc("bash", "call_1")]),
+            _tool_msg("call_1", "bash"),
+        ]
+
+        patched = mw._build_patched_messages(msgs)
+
+        assert patched is not None
+        assert leading_orphan not in patched
+        assert isinstance(patched[0], AIMessage)
+        assert isinstance(patched[1], ToolMessage)
+        assert patched[1].tool_call_id == "call_1"
+        assert len(patched) == 2
+
+    def test_tool_call_id_none_orphan_is_dropped(self):
+        """A ToolMessage whose tool_call_id is None is always an orphan —
+        no valid tool call uses ``None`` as its id — and must be dropped."""
+        mw = DanglingToolCallMiddleware()
+        # Use model_construct to bypass pydantic validation (ToolMessage requires
+        # a string tool_call_id at construction, but a corrupt serialized payload
+        # or edge-case provider could still produce None at runtime).
+        none_id_orphan = ToolMessage.model_construct(content="ghost", tool_call_id=None)
+        msgs = [
+            _ai_with_tool_calls([_tc("bash", "call_1")]),
+            none_id_orphan,
+            _tool_msg("call_1", "bash"),
+        ]
+
+        patched = mw._build_patched_messages(msgs)
+
+        assert patched is not None
+        assert none_id_orphan not in patched
+        assert len(patched) == 2
         assert isinstance(patched[0], AIMessage)
         assert isinstance(patched[1], ToolMessage)
         assert patched[1].tool_call_id == "call_1"
-        assert patched[2] is orphan
-        assert isinstance(patched[3], HumanMessage)
-        assert patched.count(orphan) == 1
 
     def test_invalid_tool_call_is_patched(self):
         mw = DanglingToolCallMiddleware()