From 47a47bf71d0104226add2cbb67b4539a02b7298e Mon Sep 17 00:00:00 2001
From: nubs <nubs@nubs.site>
Date: Fri, 5 Jun 2026 18:57:36 +0000
Subject: [PATCH] fix(llm): guard against null arguments in streaming tool-call
 accumulator (#2923)

---
 src/llm_core.py                  |  7 ++++++-
 tests/test_llm_core_streaming.py | 20 ++++++++++++++++++++
 2 files changed, 26 insertions(+), 1 deletion(-)

diff --git a/src/llm_core.py b/src/llm_core.py
index f8664eb09..6a4c81e52 100644
--- a/src/llm_core.py
+++ b/src/llm_core.py
@@ -1690,7 +1690,12 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                                             if func.get("name"):
                                                 _tc_acc[idx]["name"] = func["name"]
                                             if "arguments" in func:
-                                                _tc_acc[idx]["arguments"] += func["arguments"]
+                                                # Guard against a null arguments delta: `func` can be
+                                                # {"arguments": None} (JSON null), and a raw `+= None`
+                                                # raises TypeError that the broad except swallows,
+                                                # silently dropping the rest of the chunk. Matches the
+                                                # Anthropic accumulator (`partial = ... or ""`) above.
+                                                _tc_acc[idx]["arguments"] += func["arguments"] or ""
                                                 # Stream tool arg deltas for doc tools
                                                 if func["arguments"] and _tc_acc[idx].get("name") in ("create_document", "update_document", "edit_document"):
                                                     yield f'data: {json.dumps({"type": "tool_call_delta", "index": idx, "name": _tc_acc[idx]["name"], "arg_delta": func["arguments"]})}\n\n'
diff --git a/tests/test_llm_core_streaming.py b/tests/test_llm_core_streaming.py
index 447628695..637b94b9d 100644
--- a/tests/test_llm_core_streaming.py
+++ b/tests/test_llm_core_streaming.py
@@ -149,3 +149,23 @@ def test_sparse_integer_indices_then_null_do_not_collide(monkeypatch):
     events = _drive(monkeypatch, lines)
     calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
     assert sorted(c["name"] for c in calls) == ["f0", "f2", "fn"], f"collision: {calls}"
+
+
+def test_null_arguments_delta_does_not_drop_sibling_calls(monkeypatch):
+    # A gateway can emit a tool_call delta whose `arguments` is JSON null. The
+    # accumulator did `"" += None`, raising TypeError caught by the broad except
+    # that wraps the whole chunk — so it abandoned the rest of the tool_calls
+    # loop, silently dropping every LATER call in the same delta. Here the first
+    # call has arguments: null; the second (same delta) must still survive.
+    lines = [
+        _sse({"tool_calls": [
+            {"index": 0, "id": "a", "type": "function",
+             "function": {"name": "first", "arguments": None}},
+            {"index": 1, "id": "b", "type": "function",
+             "function": {"name": "second", "arguments": "{}"}},
+        ]}),
+        "data: [DONE]",
+    ]
+    events = _drive(monkeypatch, lines, model="gpt-4o-test")
+    calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
+    assert sorted(c["name"] for c in calls) == ["first", "second"], calls