Repository navigation
qwen3 reasoning parser re-enters reasoning on a later <think> and deletes the markers #87
Description
After </think>, streaming treats the first <think> it sees as a new reasoning block, even inside a tool-call argument. It strips the markers and moves the text between them into reasoning_content. Non-streaming instead deletes every <think> from the text and splits at the first </think>. Asking the model to edit code that quotes the markers (such as this parser) therefore corrupts the tool call, and differently in each mode.
from freetoken.server.reasoning_parser import ReasoningParser
toks = ["plan", "</think>", "\n\n", "Edit:", "<tool_call>", "\n<function=edit>\n<parameter=new>\n",
'START = "', "<think>", '"\nEND = "', "</think>", '"\n', "</parameter>\n</function>\n", "</tool_call>"]
p = ReasoningParser("qwen3", force_reasoning=True); R = C = ""
for t in toks: r, c = p.parse_stream_chunk(t); R += r; C += c
print(repr(R)) # 'plan"\nEND = "' <- argument text leaked into reasoning
print(ReasoningParser("qwen3", force_reasoning=True).parse_non_stream("".join(toks))) # <think> deletedFixed on 5080deploy by 362719c (merged in 1542d52); stream/non-stream parity tests are in tests/server/test_streaming_model_matrix.py.
Activity
gdevenyi commentedon Oct 16, 2026
Upstream: FlashML-org#711