Gérald LONLASandClaude Opus 4.8
f6a5cc1b02
The non-greedy `<think>.*?</think>` regex removes only up to the first closing tag. When a model nests `<think>` inside its own reasoning (e.g. `<think><think>inner</think></think>`), the outer `</think>` was never consumed, causing it to appear verbatim in the visible output returned to the caller. Added a third pass (`_THINK_CLOSE_RE`) that removes any `</think>` tags still present after the block and dangling-open passes; such tags can only be orphaned closing tags at that point. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
82 lines
2.9 KiB
Python
82 lines
2.9 KiB
Python
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
import client
|
|
|
|
|
|
def test_normalize_server_url_trims_and_strips_v1() -> None:
|
|
assert client.normalize_server_url(" http://10.168.168.7:1234/v1/ ") == "http://10.168.168.7:1234"
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"value,expected",
|
|
[
|
|
(None, None),
|
|
(42, 42),
|
|
(42.9, 42),
|
|
("17", 17),
|
|
(b"99", 99),
|
|
([123], 123),
|
|
((456,), 456),
|
|
],
|
|
)
|
|
def test_coerce_seed_supported_values(value, expected) -> None:
|
|
assert client.coerce_seed(value) == expected
|
|
|
|
|
|
@pytest.mark.parametrize("value", [[], [1, 2], "abc", float("nan")])
|
|
def test_coerce_seed_invalid_values(value) -> None:
|
|
with pytest.raises(ValueError):
|
|
client.coerce_seed(value)
|
|
|
|
|
|
def test_coerce_seed_empty_string_returns_none() -> None:
|
|
# Empty string is treated as "no seed provided" → None → random seed downstream.
|
|
assert client.coerce_seed("") is None
|
|
|
|
|
|
def test_resolve_request_seed_negative_one_generates_positive_seed() -> None:
|
|
resolved = client.resolve_request_seed(-1)
|
|
assert isinstance(resolved, int)
|
|
assert resolved > 0
|
|
|
|
|
|
def test_build_responses_input_text_shape() -> None:
|
|
payload = client.build_responses_input_text("hello")
|
|
assert payload[0]["role"] == "user"
|
|
assert payload[0]["content"][0]["type"] == "input_text"
|
|
assert payload[0]["content"][0]["text"] == "hello"
|
|
|
|
|
|
def test_build_chat_messages_handles_none_prompts() -> None:
|
|
messages = client.build_chat_messages(None, None)
|
|
assert messages == [{"role": "user", "content": ""}]
|
|
|
|
|
|
def test_strip_think_content_removes_reasoning_blocks() -> None:
|
|
# Use real newlines — as actual LLM output would contain.
|
|
text = "Final answer\n<think>internal chain of thought</think>\nVisible output"
|
|
assert client.strip_think_content(text) == "Final answer\n\nVisible output"
|
|
|
|
|
|
def test_strip_think_content_removes_dangling_start_tag() -> None:
|
|
# strip() removes the trailing real newline left after the dangling block is excised.
|
|
text = "Visible output\n<think>unfinished"
|
|
assert client.strip_think_content(text) == "Visible output"
|
|
|
|
|
|
def test_strip_think_content_nested_think_tags_no_orphan_close() -> None:
|
|
# When a think block contains nested <think> tags (e.g. a model quoting the tag
|
|
# in its own reasoning), the non-greedy regex matches up to the *first* </think>,
|
|
# which would leave the outer closing tag as an orphan in the visible output.
|
|
# The fix strips any remaining </think> tags after the block/dangling passes.
|
|
assert client.strip_think_content("<think><think>nested</think></think>") == ""
|
|
# Regression: visible text surrounding nested blocks must be preserved intact.
|
|
result = client.strip_think_content(
|
|
"Before <think>reasoning <think>deep</think></think> after"
|
|
)
|
|
assert "</think>" not in result
|
|
assert "Before" in result
|
|
assert "after" in result
|