[Bugfix][Frontend] Apply output_config and thinking in Anthropic count_tokens (#59180)

Signed-off-by: Kang Nahoon <ejrtks102020@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
This commit is contained in:
Kang Nahoon
2026-10-08 00:24:20 +00:00
committed by GitHub
co-authored by Flora Feng
parent 1283362e76
commit 1995c0fcd0
3 changed files with 32 additions and 4 deletions
@@ -1987,6 +1987,32 @@ class TestThinkingConfig:
assert result.include_reasoning is True
assert result.thinking_token_budget is None
@pytest.mark.parametrize(
"config",
[
pytest.param({"output_config": {"effort": "low"}}, id="effort"),
pytest.param({"thinking": {"type": "disabled"}}, id="disabled"),
pytest.param(
{"thinking": {"type": "enabled", "budget_tokens": 2048}},
id="enabled",
),
],
)
def test_count_tokens_matches_messages(self, config):
"""Effort and thinking change the rendered prompt, so count_tokens must
apply them the same way /v1/messages does."""
messages = [{"role": "user", "content": "Hello"}]
expected = _convert(
AnthropicMessagesRequest(
model="test-model", max_tokens=4096, messages=messages, **config
)
)
result = _convert(
AnthropicCountTokensRequest(model="test-model", messages=messages, **config)
)
assert result.reasoning_effort == expected.reasoning_effort
assert result.thinking_token_budget == expected.thinking_token_budget
class TestProbeDisabledThinkingEffort:
"""``auto`` falls back to ``low`` when ``none`` cannot turn thinking off."""