Streaming tool calls tolerate argument fragments under a fresh index; empty or invalid tool arguments repaired in history (fixes HTTP 400 cascade)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

tai.py

@@ -100,7 +100,7 @@ def _fromisoformat(text):
         return parsed.replace(tzinfo=tzinfo)
     raise ValueError("invalid datetime, use ISO like 2026-10-08T09:00")
 
-VERSION = "1.36.0"
+VERSION = "1.36.1"
 WORKER_STEPS = 12
 CREATE_SKILL_STEPS = 40
 FORK_MAX_DEPTH = 2
@@ -1033,6 +1033,9 @@ def clean_history(messages):
                 message = dict(message, tool_calls=good or None)
                 if not good and not message.get("content"):
                     continue
+            fixed = [dict(call, function=dict(call["function"], arguments=tool_args_valid(call["function"].get("arguments")))) if isinstance(call.get("function"), dict) and tool_args_valid(call["function"].get("arguments")) != call["function"].get("arguments") else call for call in (message.get("tool_calls") or [])]
+            if fixed:
+                message = dict(message, tool_calls=fixed)
         kept.append(message)
     return [message for message in kept if not (message.get("role") == "tool" and message.get("tool_call_id") in dropped)]
 
@@ -1044,6 +1047,15 @@ def transient_failure(exc):
     return "CLI not found" not in str(exc)
 
 
+def tool_args_valid(arguments):
+    """Arguments text strict backends accept: empty or non-JSON becomes an empty object."""
+    try:
+        json.loads(arguments)
+    except (TypeError, ValueError):
+        return "{}"
+    return arguments
+
+
 def tool_args_text(raw_arguments):
     if not raw_arguments:
         return "{}"
@@ -5580,6 +5592,7 @@ class ChatClient:
         content_parts = []
         reasoning_parts = []
         calls = {}
+        last_pos = 0
         try:
             for raw in response:
                 line = raw.decode("utf-8", "replace").strip()
@@ -5604,10 +5617,16 @@ class ChatClient:
                 if think:
                     reasoning_parts.append(think)
                 for call in delta.get("tool_calls") or []:
-                    slot = calls.setdefault(call.get("index", 0), {"id": "", "name": "", "arguments": ""})
+                    func = call.get("function") or {}
+                    pos = call.get("index", 0)
+                    if pos not in calls and calls and not call.get("id") and not func.get("name"):
+                        # Off-spec streams send an argument fragment under a fresh index
+                        # with no id or name: it continues the open call.
+                        pos = last_pos
+                    last_pos = pos
+                    slot = calls.setdefault(pos, {"id": "", "name": "", "arguments": ""})
                     if call.get("id"):
                         slot["id"] = call["id"]
-                    func = call.get("function") or {}
                     if func.get("name"):
                         slot["name"] += func["name"]
                     if func.get("arguments"):
@@ -5620,6 +5639,7 @@ class ChatClient:
         ordered = [calls[pos] for pos in sorted(calls)]
         for slot in ordered:
             slot["name"] = tool_name_clean(slot["name"])
+            slot["arguments"] = tool_args_valid(slot["arguments"])
             if not slot["id"]:
                 slot["id"] = uuid.uuid4().hex
         return {"role": "assistant", "content": "".join(content_parts), "reasoning": "".join(reasoning_parts), "tool_calls": ordered, "backend": label}

test_tai.py

@@ -4809,6 +4809,38 @@ class RetrySummaryTests(unittest.TestCase):
         self.assertEqual(cleaned[2]["tool_call_id"], "b")
         self.assertEqual(len(history[1]["tool_calls"]), 2)
 
+    def test_clean_history_repairs_empty_arguments(self):
+        history = [
+            {"role": "user", "content": "hi"},
+            {"role": "assistant", "content": None, "tool_calls": [
+                {"id": "a", "type": "function", "function": {"name": "shell", "arguments": ""}}]},
+            {"role": "tool", "tool_call_id": "a", "content": "x"},
+        ]
+        cleaned = tai.clean_history(history)
+        self.assertEqual(cleaned[1]["tool_calls"][0]["function"]["arguments"], "{}")
+        self.assertEqual(history[1]["tool_calls"][0]["function"]["arguments"], "")
+
+    def test_streaming_joins_argument_fragment_under_fresh_index(self):
+        chunks = [
+            {"choices": [{"delta": {"tool_calls": [{"index": 0, "id": "c1", "function": {"name": "shell", "arguments": ""}}]}}]},
+            {"choices": [{"delta": {"tool_calls": [{"index": 1, "function": {"arguments": "{\"command\":\"ls\"}"}}]}}]},
+        ]
+        body = "".join("data: %s\n\n" % json.dumps(chunk) for chunk in chunks) + "data: [DONE]\n\n"
+
+        class Response:
+            def __iter__(self):
+                return iter(body.encode("utf-8").splitlines(True))
+
+            def close(self):
+                pass
+
+        chat = tai.ChatClient.__new__(tai.ChatClient)
+        with mock.patch("urllib.request.urlopen", return_value=Response()):
+            reply = chat.streaming("http://x/chat/completions", "", {"model": "m"}, "t", lambda piece: None)
+        self.assertEqual(len(reply["tool_calls"]), 1)
+        self.assertEqual(reply["tool_calls"][0]["name"], "shell")
+        self.assertEqual(reply["tool_calls"][0]["arguments"], "{\"command\":\"ls\"}")
+
 
 class TuiLayoutTests(unittest.TestCase):
     def make_app(self, rows=10, cols=20):