Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: repair orphan tool results in the chat template fallback (#11733)

lorenzozanee69dddaa38ae
diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index a917290a1db..d49bb1b783a 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -2707,6 +2707,50 @@ def _split_parallel_tool_calls(messages: list) -> list:
return out
+def _repair_orphan_tool_results(messages: list) -> list:
+ """Placeholder call before each tool result lacking one (gpt-oss refuses orphans). Fallback only:
+ the model reads it. A text turn takes the calls, since a second assistant turn breaks alternation."""
+ mutated = False
+ out: list = []
+ linked = False
+ repaired_at = None
+
+ for message in messages:
+ role = message.get("role") if isinstance(message, dict) else None
+ if role != "tool":
+ linked = role == "assistant" and bool(message.get("tool_calls"))
+ repaired_at = None
+ out.append(message)
+ continue
+ if linked and repaired_at is None:
+ out.append(message)
+ continue
+
+ call_id = message.get("tool_call_id") or f"replayed_tool_{len(out)}"
+ call = {
+ "id": call_id,
+ "type": "function",
+ "function": {"name": message.get("name") or "tool", "arguments": {}},
+ }
+ if repaired_at is not None:
+ calls = out[repaired_at]["tool_calls"]
+ if all(c.get("id") != call_id for c in calls):
+ out[repaired_at] = {**out[repaired_at], "tool_calls": [*calls, call]}
+ elif out and isinstance(out[-1], dict) and out[-1].get("role") == "assistant":
+ out[-1] = {**out[-1], "tool_calls": [call]}
+ repaired_at = len(out) - 1
+ else:
+ out.append({"role": "assistant", "content": "", "tool_calls": [call]})
+ repaired_at = len(out) - 1
+ if not message.get("tool_call_id"):
+ message = {**message, "tool_call_id": call_id}
+ out.append(message)
+ linked = True
+ mutated = True
+
+ return out if mutated else messages
+
+
_MARKUP_BY_TOKENIZER: "weakref.WeakKeyDictionary" = weakref.WeakKeyDictionary()
@@ -3375,6 +3419,14 @@ def apply_chat_template_for_generation(
return _render_with_fallback(candidate)
except Exception:
continue
+ # Last and lazy: a history an earlier candidate renders never pays for the scan.
+ repaired = _repair_orphan_tool_results(split)
+ if repaired is not split:
+ try:
+ # Split again: gpt-oss renders only tool_calls[0] and names every later result after it.
+ return _render_with_fallback(_split_parallel_tool_calls(repaired))
+ except Exception:
+ pass
raise
diff --git a/studio/backend/tests/test_chat_template_tool_arguments.py b/studio/backend/tests/test_chat_template_tool_arguments.py
index 9b2169e06c9..998f6376728 100644
--- a/studio/backend/tests/test_chat_template_tool_arguments.py
+++ b/studio/backend/tests/test_chat_template_tool_arguments.py
@@ -67,6 +67,27 @@ class _StrictTemplateTokenizer:
return "RENDERED"
+class _StrictToolOrderTokenizer:
+ def __init__(self):
+ self.seen_messages = []
+
+ def apply_chat_template(
+ self,
+ messages,
+ *,
+ tokenize = False,
+ add_generation_prompt = True,
+ **kw,
+ ):
+ self.seen_messages.append(messages)
+ for index, message in enumerate(messages):
+ if message.get("role") == "tool":
+ previous = messages[index - 1] if index else {}
+ if previous.get("role") not in ("assistant", "tool"):
+ raise ValueError("A tool message must follow an assistant or tool message.")
+ return "RENDERED"
+
+
def test_string_arguments_are_parsed_to_dict():
out = _normalize_tool_call_arguments(_conv('{"query": "sweden"}'))
args = out[1]["tool_calls"][0]["function"]["arguments"]
@@ -89,6 +110,142 @@ def test_render_succeeds_on_strict_template_with_string_arguments():
assert result == "RENDERED"
+def test_render_repairs_orphan_tool_result_after_native_template_failure():
+ tok = _StrictToolOrderTokenizer()
+ messages = [
+ {"role": "user", "content": "weather?"},
+ {"role": "tool", "name": "web_search", "content": "21C sunny"},
+ ]
+
+ assert apply_chat_template_for_generation(tok, messages) == "RENDERED"
+ repaired = tok.seen_messages[-1]
+ assert [message["role"] for message in repaired] == ["user", "assistant", "tool"]
+ assert repaired[1]["tool_calls"][0]["id"] == "replayed_tool_1"
+ assert repaired[2]["tool_call_id"] == "replayed_tool_1"
+
+
+class _CallLinkedToolTokenizer:
+ """gpt-oss rule: each result answers a call of the preceding assistant; no two assistant turns in a row."""
+
+ def __init__(self):
+ self.seen_messages = []
+
+ def apply_chat_template(
+ self,
+ messages,
+ *,
+ tokenize = False,
+ add_generation_prompt = True,
+ **kw,
+ ):
+ self.seen_messages.append(messages)
+ call_ids = None
+ previous_role = None
+ for message in messages:
+ role = message.get("role")
+ if role == "assistant" and previous_role == "assistant":
+ raise ValueError("Conversation roles must alternate")
+ if role == "assistant":
+ call_ids = {call.get("id") for call in message.get("tool_calls") or ()}
+ elif role == "tool":
+ if not call_ids or message.get("tool_call_id") not in call_ids:
+ raise ValueError(
+ "Message has tool role, but there was no previous assistant message with a tool call!"
+ )
+ else:
+ call_ids = None
+ previous_role = role
+ return "RENDERED"
+
+
+def test_render_gives_an_assistant_text_turn_the_orphan_call():
+ tok = _CallLinkedToolTokenizer()
+ messages = [
+ {"role": "user", "content": "weather?"},
+ {"role": "assistant", "content": "checking"},
+ {"role": "tool", "tool_call_id": "c1", "name": "web_search", "content": "21C sunny"},
+ ]
+
+ assert apply_chat_template_for_generation(tok, messages) == "RENDERED"
+ repaired = tok.seen_messages[-1]
+ assert [message["role"] for message in repaired] == ["user", "assistant", "tool"]
+ assert repaired[1]["content"] == "checking"
+ assert [call["id"] for call in repaired[1]["tool_calls"]] == ["c1"]
+ assert "tool_calls" not in messages[1]
+
+
+class _FirstCallNamesResultsTokenizer(_CallLinkedToolTokenizer):
+ def apply_chat_template(self, messages, **kw):
+ super().apply_chat_template(messages, **kw)
+ lines, last = [], None
+ for message in messages:
+ if message.get("tool_calls"):
+ last = message["tool_calls"][0]["function"]["name"]
+ elif message.get("role") == "tool":
+ lines.append(f"functions.{last}: {message['content']}")
+ return "\n".join(lines)
+
+
+def test_render_gives_each_orphan_result_its_own_call():
+ tok = _FirstCallNamesResultsTokenizer()
+ messages = [
+ {"role": "user", "content": "weather?"},
+ {"role": "tool", "tool_call_id": "c1", "name": "web_search", "content": "21C sunny"},
+ {"role": "tool", "name": "web_fetch", "content": "rain later"},
+ ]
+
+ rendered = apply_chat_template_for_generation(tok, messages)
+ assert rendered == "functions.web_search: 21C sunny\nfunctions.web_fetch: rain later"
+ repaired = tok.seen_messages[-1]
+ assert [message["role"] for message in repaired] == [
+ "user",
+ "assistant",
+ "tool",
+ "assistant",
+ "tool",
+ ]
+ assert [m["tool_calls"][0]["id"] for m in repaired if m.get("tool_calls")] == [
+ "c1",
+ "replayed_tool_3",
+ ]
+ assert repaired[4]["tool_call_id"] == "replayed_tool_3"
+
+
+def test_orphan_repair_is_idempotent_and_leaves_linked_results_alone():
+ from core.inference.chat_template_helpers import _repair_orphan_tool_results
+
+ messages = [
+ {"role": "user", "content": "weather?"},
+ {"role": "tool", "name": "web_search", "content": "21C sunny"},
+ {"role": "assistant", "content": "", "tool_calls": [{"id": "c2"}]},
+ {"role": "tool", "tool_call_id": "c2", "content": "rain"},
+ ]
+ snapshot = json.loads(json.dumps(messages))
+
+ repaired = _repair_orphan_tool_results(messages)
+ assert [message["role"] for message in repaired] == [
+ "user",
+ "assistant",
+ "tool",
+ "assistant",
+ "tool",
+ ]
+ assert repaired[3:] == messages[2:]
+ assert _repair_orphan_tool_results(repaired) is repaired
+ assert messages == snapshot
+
+
+def test_render_keeps_valid_tool_result_history_unchanged():
+ tok = _StrictToolOrderTokenizer()
+ messages = [
+ {"role": "assistant", "content": "", "tool_calls": [{"id": "c1"}]},
+ {"role": "tool", "tool_call_id": "c1", "content": "21C sunny"},
+ ]
+
+ assert apply_chat_template_for_generation(tok, messages) == "RENDERED"
+ assert tok.seen_messages == [messages]
+
+
class _RecordingTokenizer:
"""Lenient template: renders whatever arguments it is given (string or dict)."""
@@ -359,6 +516,20 @@ def test_render_succeeds_on_single_call_template_with_parallel_calls():
assert result == "RENDERED"
+def test_orphan_repair_is_not_built_when_an_earlier_fallback_renders(monkeypatch):
+ # Llama 3.x: the split renders parallel calls, so the orphan scan must not run.
+ import core.inference.chat_template_helpers as helpers
+
+ def _must_not_run(messages):
+ raise AssertionError("orphan repair built although the split rendered")
+
+ monkeypatch.setattr(helpers, "_repair_orphan_tool_results", _must_not_run, raising = False)
+ assert (
+ apply_chat_template_for_generation(_SingleToolCallTokenizer(), _parallel_conv())
+ == "RENDERED"
+ )
+
+
def test_string_arguments_and_parallel_calls_are_repaired_together():
conv = _parallel_conv()
for call in conv[1]["tool_calls"]: