fix(models): omit Codex tool results with blank call IDs (#6234)

* fix(models): omit Codex tool results with blank call IDs

* fix(models): validate Codex call and result IDs symmetrically

Use one strip-based validity check for assistant calls and tool results while preserving non-blank IDs verbatim. Add paired regressions for ordinary and invalid tool calls, plus non-blank correlation controls.

Validation: 126 focused tests passed; 210 blocking-I/O tests passed; backend lint and format checks passed. Paired regression fails on the previous implementation for whitespace-only IDs.

* fix(models): log dropped Codex tool results

---------

Co-authored-by: destire-mio <248462155+destire-mio@users.noreply.github.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

backend/packages/harness/deerflow/models/AGENTS.md

@@ -58,6 +58,17 @@ the same policy first so run metadata reports the effective values. Design note:
 `tests/test_reasoning_contract.py`, the contract section of
 `tests/test_model_factory.py`, `tests/test_models_router_reasoning.py`.
 
+### Codex tool-call/result serialization (`packages/harness/deerflow/models/openai_codex_provider.py`)
+
+`_convert_messages` uses `_is_valid_call_id` for both assistant tool calls
+(including `invalid_tool_calls`) and tool results, omitting empty or
+whitespace-only call IDs even when direct provider use bypasses middleware repair.
+Non-blank IDs remain byte-for-byte unchanged so calls and results keep their
+correlation; serialization does not trim IDs or rewrite the input messages.
+Each omitted tool result emits a warning with its normalized content length;
+the warning never includes the result content.
+Coverage: `tests/test_codex_provider.py`.
+
 ### Codex SSE termination (`packages/harness/deerflow/models/openai_codex_provider.py`)
 
 `response.completed` ends stream consumption immediately, before transport EOF;

backend/packages/harness/deerflow/models/openai_codex_provider.py

@@ -29,6 +29,11 @@ logger = logging.getLogger(__name__)
 CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
 
 
+def _is_valid_call_id(call_id: Any) -> bool:
+    """Check call/result correlation IDs without rewriting non-blank values."""
+    return isinstance(call_id, str) and bool(call_id.strip())
+
+
 def _build_usage_metadata(oai_usage: dict) -> dict:
     """Convert Codex/Responses API usage dict to LangChain usage_metadata format.
 
@@ -170,7 +175,7 @@ class CodexChatModel(BaseChatModel):
                     call_id = tc.get("id")
                     if not (isinstance(name, str) and name):
                         continue
-                    if not (isinstance(call_id, str) and call_id):
+                    if not _is_valid_call_id(call_id):
                         continue
                     args = tc.get("args")
                     input_items.append(
@@ -182,11 +187,16 @@ class CodexChatModel(BaseChatModel):
                         }
                     )
             elif isinstance(msg, ToolMessage):
+                content = self._normalize_content(msg.content)
+                # A blank ID cannot identify the call this result answers.
+                if not _is_valid_call_id(msg.tool_call_id):
+                    logger.warning("Dropping tool result with blank call_id (content %d chars)", len(content))
+                    continue
                 input_items.append(
                     {
                         "type": "function_call_output",
                         "call_id": msg.tool_call_id,
-                        "output": self._normalize_content(msg.content),
+                        "output": content,
                     }
                 )
 

backend/tests/test_codex_provider.py

@@ -11,8 +11,10 @@ Covers:
 from __future__ import annotations
 
 import json
+import logging
 from unittest.mock import patch
 
+import pytest
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 
 from deerflow.models.credential_loader import CodexCliCredential
@@ -202,13 +204,85 @@ def test_convert_messages_ai_with_tool_calls():
     assert any(item.get("type") == "function_call" and item["name"] == "search" for item in items)
 
 
-def test_convert_messages_tool_message():
+@pytest.mark.parametrize("call_id", ["tc1", "0", " tc1 "])
+def test_convert_messages_tool_message(call_id, caplog):
     model = _make_model()
-    tool_msg = ToolMessage(content="result data", tool_call_id="tc1")
-    _, items = model._convert_messages([tool_msg])
+    tool_msg = ToolMessage(content="result data", tool_call_id=call_id)
+    with caplog.at_level(logging.WARNING, logger="deerflow.models.openai_codex_provider"):
+        _, items = model._convert_messages([tool_msg])
     assert items[0]["type"] == "function_call_output"
-    assert items[0]["call_id"] == "tc1"
+    assert items[0]["call_id"] == call_id
     assert items[0]["output"] == "result data"
+    assert caplog.record_tuples == []
+
+
+@pytest.mark.parametrize("call_id", ["tc1", "0", " tc1 "])
+def test_convert_messages_preserves_paired_call_ids(call_id):
+    model = _make_model()
+    ai_msg = AIMessage(content="", tool_calls=[{"name": "search", "args": {"q": "foo"}, "id": call_id}])
+    tool_msg = ToolMessage(content="result data", tool_call_id=call_id)
+
+    _, items = model._convert_messages([ai_msg, tool_msg])
+
+    assert items == [
+        {"type": "function_call", "name": "search", "arguments": '{"q": "foo"}', "call_id": call_id},
+        {"type": "function_call_output", "call_id": call_id, "output": "result data"},
+    ]
+    assert ai_msg.tool_calls[0]["id"] == call_id
+    assert tool_msg.tool_call_id == call_id
+
+
+@pytest.mark.parametrize("blank_id", ["", "   ", "\t\r\n"])
+@pytest.mark.parametrize("call_field", ["tool_calls", "invalid_tool_calls"])
+def test_convert_messages_omits_paired_calls_and_results_with_blank_ids(blank_id, call_field):
+    model = _make_model()
+    args = {"q": "foo"} if call_field == "tool_calls" else '{"q":'
+    ai_msg = AIMessage(content="Searching.", **{call_field: [{"name": "search", "args": args, "id": blank_id}]})
+    tool_msg = ToolMessage(content="result data", tool_call_id=blank_id)
+
+    _, items = model._convert_messages([ai_msg, tool_msg])
+
+    assert items == [{"role": "assistant", "content": "Searching."}]
+    assert getattr(ai_msg, call_field)[0]["id"] == blank_id
+    assert tool_msg.tool_call_id == blank_id
+
+
+@pytest.mark.parametrize("blank_id", ["", "   ", "\t\r\n"])
+@pytest.mark.parametrize(
+    ("content", "content_length"),
+    [
+        ("private result", 14),
+        ([{"type": "text", "text": "private"}, {"type": "text", "text": "result"}], 14),
+        ("", 0),
+    ],
+)
+def test_convert_messages_omits_tool_results_with_blank_call_ids(blank_id, content, content_length, caplog):
+    model = _make_model()
+    orphaned_result = ToolMessage(content=content, tool_call_id=blank_id)
+    original_result = orphaned_result.model_dump()
+    messages = [
+        SystemMessage(content="Follow the instructions."),
+        HumanMessage(content="Search for foo."),
+        AIMessage(content="", tool_calls=[{"name": "search", "args": {"q": "foo"}, "id": "tc1"}]),
+        orphaned_result,
+        ToolMessage(content=[{"type": "text", "text": "result data"}], tool_call_id="tc1"),
+        AIMessage(content="Done."),
+    ]
+
+    with caplog.at_level(logging.WARNING, logger="deerflow.models.openai_codex_provider"):
+        instructions, items = model._convert_messages(messages)
+
+    assert instructions == "Follow the instructions."
+    assert items == [
+        {"role": "user", "content": "Search for foo."},
+        {"type": "function_call", "name": "search", "arguments": '{"q": "foo"}', "call_id": "tc1"},
+        {"type": "function_call_output", "call_id": "tc1", "output": "result data"},
+        {"role": "assistant", "content": "Done."},
+    ]
+    assert orphaned_result.model_dump() == original_result
+    assert caplog.record_tuples == [
+        ("deerflow.models.openai_codex_provider", logging.WARNING, f"Dropping tool result with blank call_id (content {content_length} chars)"),
+    ]
 
 
 def test_convert_messages_keeps_placeholder_result_paired_with_invalid_tool_call():