[Bugfix] Drop incomplete tool-call markup in non-streaming to match streaming (#47562)

Signed-off-by: Suvrakamal Das <jyb2084.usa@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Ben Browning <56071+bbrowning@users.noreply.github.com>
This commit is contained in:
Suvrakamal Das
2026-09-01 18:11:36 +00:00
committed by GitHub
co-authored by Claude Ben Browning
parent 2f01039666
commit e7cf4730d6
6 changed files with 399 additions and 1 deletions
+19
View File
@@ -131,6 +131,17 @@ class Parser:
self._reasoning_parser is None
or self._reasoning_parser.engine_based_streaming
) and (self._tool_parser is None or self._tool_parser.engine_based_streaming)
if (
self._reasoning_parser is None
and self._tool_parser is not None
and hasattr(self._tool_parser, "skip_reasoning_parsing")
):
# With no reasoning parser configured, reasoning markup is
# plain content: an engine-based tool parser should pass it
# through verbatim where its grammar allows, not consume or
# reclassify it. The engine ignores the flag for markers
# shared with non-reasoning structure.
self._tool_parser.skip_reasoning_parsing = True
self._stream_state = StreamState(
tool_call_id_type=(
get_tool_call_id_type(model_config)
@@ -508,6 +519,14 @@ class DelegatingParser(Parser):
or (isinstance(content, str) and not content.strip())
):
return [], None
# No complete tool calls: for engine-based parsers, return
# the tool parser's content, which drops incomplete
# tool-call markup (e.g. a <tool_call> opener truncated by
# max_tokens or a stop string), so the non-streaming path
# matches streaming. Legacy parsers keep their existing
# behavior of returning the raw content.
if self._engine_based and tool_call_info is not None:
return None, tool_call_info.content or None
return None, content
return tool_calls, content