mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-22 16:25:58 +00:00
60811ced37
Kimi's Anthropic-compatible endpoints (api.moonshot.cn/anthropic, api.kimi.com/coding) implement the adaptive thinking contract — they accept thinking.type=adaptive + output_config.effort (all of low, medium, high, xhigh, max verified live) and return thinking blocks, and the replay-validation 400s that originally motivated dropping the parameter (#13848) no longer occur. _supports_adaptive_thinking() now returns True for Kimi-family models, so they get thinking={type: adaptive, display: summarized} + output_config.effort via ADAPTIVE_EFFORT_MAP instead of nothing, and the blanket drop of the thinking parameter for Kimi-family endpoints is removed. MiniMax and other non-adaptive third parties keep the manual budget_tokens path; Claude behavior is unchanged.
209 lines
8.3 KiB
Python
209 lines
8.3 KiB
Python
"""Kimi / Moonshot thinking behavior on the Anthropic-Messages wire.
|
|
|
|
Contract (changed from the original #13848 mitigation):
|
|
|
|
- Kimi-family endpoints receive ``thinking`` in **adaptive** form
|
|
(``thinking.type="adaptive"`` + ``output_config.effort``) — never manual
|
|
``budget_tokens``. Their Anthropic-compatible endpoints
|
|
(``api.moonshot.cn/anthropic``, ``api.kimi.com/coding``) accept the
|
|
field set, and the replay-validation 400s that originally motivated
|
|
dropping the parameter (#13848) no longer occur.
|
|
|
|
- ``convert_messages_to_anthropic`` still preserves unsigned
|
|
reasoning_content-derived thinking blocks on replay for this family, so
|
|
multi-turn tool-call history round-trips.
|
|
|
|
Kimi on the chat_completions route handles ``thinking`` via ``extra_body``
|
|
in ``ChatCompletionsTransport`` (#13503).
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
|
|
class TestKimiCodingAnthropicThinking:
|
|
"""Kimi-family thinking on the Anthropic wire (incl. /coding)."""
|
|
|
|
def test_kimi_coding_with_explicit_disabled_omits_thinking(self) -> None:
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="kimi-k2.5",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": False},
|
|
base_url="https://api.kimi.com/coding",
|
|
)
|
|
assert "thinking" not in kwargs
|
|
|
|
def test_non_kimi_third_party_still_gets_thinking(self) -> None:
|
|
"""MiniMax and other third-party Anthropic endpoints must retain thinking."""
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="MiniMax-M2.7",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": True, "effort": "medium"},
|
|
base_url="https://api.minimax.io/anthropic",
|
|
)
|
|
assert "thinking" in kwargs
|
|
assert kwargs["thinking"]["type"] == "enabled"
|
|
|
|
def test_native_anthropic_still_gets_thinking(self) -> None:
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="claude-sonnet-4-20250514",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": True, "effort": "medium"},
|
|
base_url=None,
|
|
)
|
|
assert "thinking" in kwargs
|
|
|
|
|
|
class TestKimiFamilyGetsAdaptiveThinking:
|
|
"""Kimi-family endpoints get adaptive thinking + output_config.effort."""
|
|
|
|
@pytest.mark.parametrize(
|
|
"base_url,model",
|
|
[
|
|
# Official Kimi / Moonshot hosts (all URL shapes)
|
|
("https://api.kimi.com/coding", "kimi-k2.5"),
|
|
("https://api.kimi.com/coding/v1", "kimi-k2.5"),
|
|
("https://api.kimi.com/coding/anthropic", "kimi-k2.5"),
|
|
("https://api.kimi.com/v1", "kimi-k2.5"),
|
|
("https://api.moonshot.ai/anthropic", "moonshot-v1-32k"),
|
|
("https://api.moonshot.cn/anthropic", "moonshot-v1-32k"),
|
|
("https://api.moonshot.cn/anthropic/v1", "kimi-0714-preview"),
|
|
# Custom / proxied hosts with a Kimi-family model (#17057)
|
|
("http://my-kimi-proxy.internal", "kimi-2.6"),
|
|
("https://llm.example.com/anthropic", "moonshotai/kimi-k2.5"),
|
|
],
|
|
)
|
|
def test_kimi_family_endpoint_gets_adaptive_thinking(
|
|
self, base_url: str, model: str
|
|
) -> None:
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model=model,
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": True, "effort": "high"},
|
|
base_url=base_url,
|
|
)
|
|
assert kwargs.get("thinking", {}).get("type") == "adaptive", (
|
|
f"Kimi-family endpoint ({base_url}, {model}) must receive "
|
|
f"adaptive thinking, got {kwargs.get('thinking')!r}"
|
|
)
|
|
assert "budget_tokens" not in kwargs["thinking"]
|
|
assert kwargs["output_config"] == {"effort": "high"}
|
|
# Adaptive mode must not force temperature or inflate max_tokens
|
|
# (those are manual-budget-mode side effects).
|
|
assert "temperature" not in kwargs
|
|
assert kwargs["max_tokens"] == 4096
|
|
|
|
@pytest.mark.parametrize(
|
|
"hermes_effort,wire_effort",
|
|
[
|
|
("minimal", "low"),
|
|
("low", "low"),
|
|
("medium", "medium"),
|
|
("high", "high"),
|
|
("xhigh", "xhigh"),
|
|
("max", "max"),
|
|
("ultra", "max"),
|
|
],
|
|
)
|
|
def test_kimi_effort_mapping(self, hermes_effort: str, wire_effort: str) -> None:
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="kimi-0714-preview",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": True, "effort": hermes_effort},
|
|
base_url="https://api.moonshot.cn/anthropic/v1",
|
|
)
|
|
assert kwargs["thinking"] == {"type": "adaptive", "display": "summarized"}
|
|
assert kwargs["output_config"] == {"effort": wire_effort}
|
|
|
|
def test_kimi_thinking_disabled_omits_parameter(self) -> None:
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="kimi-0714-preview",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": False},
|
|
base_url="https://api.moonshot.cn/anthropic/v1",
|
|
)
|
|
assert "thinking" not in kwargs
|
|
assert "output_config" not in kwargs
|
|
|
|
def test_custom_endpoint_non_kimi_model_keeps_thinking(self) -> None:
|
|
"""Custom endpoint with a non-Kimi model must keep thinking intact.
|
|
|
|
Guards against over-broad model-family matching — only model names
|
|
starting with a Kimi/Moonshot prefix should route to adaptive.
|
|
"""
|
|
from agent.anthropic_adapter import build_anthropic_kwargs
|
|
|
|
kwargs = build_anthropic_kwargs(
|
|
model="MiniMax-M2.7",
|
|
messages=[{"role": "user", "content": "hello"}],
|
|
tools=None,
|
|
max_tokens=4096,
|
|
reasoning_config={"enabled": True, "effort": "medium"},
|
|
base_url="https://my-llm-proxy.example.com/anthropic",
|
|
)
|
|
assert "thinking" in kwargs
|
|
assert kwargs["thinking"]["type"] == "enabled"
|
|
|
|
def test_kimi_family_replay_preserves_unsigned_thinking(self) -> None:
|
|
"""On a custom Kimi endpoint, unsigned reasoning_content thinking
|
|
blocks must survive the third-party signature-stripping pass so
|
|
the upstream's message-history validation passes.
|
|
"""
|
|
from agent.anthropic_adapter import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "hi"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "planning the tool call",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"type": "function",
|
|
"function": {"name": "skill_view", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "ok"},
|
|
]
|
|
_, converted = convert_messages_to_anthropic(
|
|
messages,
|
|
base_url="http://my-kimi-proxy.internal",
|
|
model="kimi-2.6",
|
|
)
|
|
# The assistant message still carries the unsigned thinking block
|
|
# synthesised from reasoning_content (required by Kimi's history
|
|
# validation). A plain third-party endpoint would have stripped it.
|
|
assistant_msg = next(m for m in converted if m["role"] == "assistant")
|
|
assistant_blocks = assistant_msg["content"]
|
|
thinking_blocks = [
|
|
b for b in assistant_blocks
|
|
if isinstance(b, dict) and b.get("type") == "thinking"
|
|
]
|
|
assert len(thinking_blocks) == 1
|
|
assert thinking_blocks[0]["thinking"] == "planning the tool call"
|