mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-06-17 09:41:58 +00:00
fix: route minimax m3 reasoning controls through profile
Follow up PR #46609's api.minimax.io reasoning report by moving the behavior out of the broad run_agent host gate and into the MiniMax provider profile. Only MiniMax-M3 on the documented OpenAI-compatible /v1 route gets reasoning_split/thinking/reasoning_effort; Anthropic-format MiniMax and non-M3 models keep their existing wire shapes. Co-authored-by: goku94123 <gooku94123@gmail.com>
This commit is contained in:
Teknium
parent
ba3883cd18
commit
49e743985a
5 changed files with 171 additions and 14 deletions
|
|
@ -531,6 +531,7 @@ class ChatCompletionsTransport(ProviderTransport):
|
|||
supports_reasoning=params.get("supports_reasoning", False),
|
||||
qwen_session_metadata=params.get("qwen_session_metadata"),
|
||||
model=model,
|
||||
base_url=params.get("base_url"),
|
||||
ollama_num_ctx=params.get("ollama_num_ctx"),
|
||||
session_id=params.get("session_id"),
|
||||
)
|
||||
|
|
|
|||
|
|
@ -1,13 +1,65 @@
|
|||
"""MiniMax provider profiles (international + China).
|
||||
|
||||
Both use anthropic_messages api_mode — their inference_base_url
|
||||
ends with /anthropic which triggers auto-detection to anthropic_messages.
|
||||
The default API-key routes use anthropic_messages because their base URLs end
|
||||
with /anthropic. Users can opt MiniMax-M3 into the OpenAI-compatible endpoint
|
||||
with base_url=https://api.minimax.io/v1; that route needs MiniMax-specific
|
||||
reasoning controls in extra_body.
|
||||
"""
|
||||
|
||||
from typing import Any
|
||||
from urllib.parse import urlparse
|
||||
|
||||
from providers import register_provider
|
||||
from providers.base import ProviderProfile
|
||||
|
||||
minimax = ProviderProfile(
|
||||
|
||||
def _is_minimax_global_openai_base_url(base_url: str | None) -> bool:
|
||||
parsed = urlparse(str(base_url or "").strip())
|
||||
if (parsed.hostname or "").lower() != "api.minimax.io":
|
||||
return False
|
||||
path = parsed.path.rstrip("/").lower()
|
||||
return path == "/v1"
|
||||
|
||||
|
||||
def _is_minimax_m3(model: str | None) -> bool:
|
||||
normalized = str(model or "").strip().lower()
|
||||
return normalized in {"minimax-m3", "minimax/minimax-m3"}
|
||||
|
||||
|
||||
class MiniMaxProfile(ProviderProfile):
|
||||
"""MiniMax — M3 OpenAI-compatible reasoning controls."""
|
||||
|
||||
def build_api_kwargs_extras(
|
||||
self,
|
||||
*,
|
||||
reasoning_config: dict | None = None,
|
||||
model: str | None = None,
|
||||
base_url: str | None = None,
|
||||
**context: Any,
|
||||
) -> tuple[dict[str, Any], dict[str, Any]]:
|
||||
"""Emit M3 reasoning controls for api.minimax.io/v1.
|
||||
|
||||
MiniMax-M3's OpenAI-compatible endpoint keeps thinking inline unless
|
||||
``reasoning_split`` is sent, so always request the split format on that
|
||||
route. ``thinking`` controls the M3 mode; Hermes' effort levels are not
|
||||
a MiniMax depth knob here, so they only select adaptive vs disabled.
|
||||
"""
|
||||
if not _is_minimax_global_openai_base_url(base_url) or not _is_minimax_m3(model):
|
||||
return {}, {}
|
||||
|
||||
extra_body: dict[str, Any] = {"reasoning_split": True}
|
||||
|
||||
if isinstance(reasoning_config, dict) and reasoning_config.get("enabled") is False:
|
||||
extra_body["thinking"] = {"type": "disabled"}
|
||||
return extra_body, {}
|
||||
|
||||
if reasoning_config is not None:
|
||||
extra_body["thinking"] = {"type": "adaptive"}
|
||||
|
||||
return extra_body, {}
|
||||
|
||||
|
||||
minimax = MiniMaxProfile(
|
||||
name="minimax",
|
||||
aliases=("mini-max",),
|
||||
api_mode="anthropic_messages",
|
||||
|
|
@ -17,7 +69,7 @@ minimax = ProviderProfile(
|
|||
default_aux_model="MiniMax-M3",
|
||||
)
|
||||
|
||||
minimax_cn = ProviderProfile(
|
||||
minimax_cn = MiniMaxProfile(
|
||||
name="minimax-cn",
|
||||
aliases=("minimax-china", "minimax_cn"),
|
||||
api_mode="anthropic_messages",
|
||||
|
|
@ -27,7 +79,7 @@ minimax_cn = ProviderProfile(
|
|||
default_aux_model="MiniMax-M3",
|
||||
)
|
||||
|
||||
minimax_oauth = ProviderProfile(
|
||||
minimax_oauth = MiniMaxProfile(
|
||||
name="minimax-oauth",
|
||||
aliases=("minimax_oauth", "minimax-oauth-io"),
|
||||
api_mode="anthropic_messages",
|
||||
|
|
|
|||
|
|
@ -4792,15 +4792,6 @@ class AIAgent:
|
|||
return bool(github_model_reasoning_efforts(self.model))
|
||||
except Exception:
|
||||
return False
|
||||
if base_url_host_matches(self._base_url_lower, "api.minimax.io"):
|
||||
# MiniMax (api.minimax.io): enable reasoning extra_body
|
||||
# (reasoning_split, thinking, reasoning_effort) for both the
|
||||
# Anthropic-format and OpenAI-format endpoints. Without this the
|
||||
# safety gate strips those fields before they reach the API, so M3
|
||||
# leaks thinking into response content and burns output tokens. M3
|
||||
# specifically benefits from the OpenAI-compatible endpoint
|
||||
# (/v1/chat/completions) for prompt caching.
|
||||
return True
|
||||
if (self.provider or "").strip().lower() == "lmstudio":
|
||||
opts = self._lmstudio_reasoning_options_cached()
|
||||
# "off-only" (or absent) means no real reasoning capability.
|
||||
|
|
|
|||
|
|
@ -1318,6 +1318,7 @@ AUTHOR_MAP = {
|
|||
"agentsmithlaor@gmail.com": "oferlaor", # PR #22356 salvage (cron origin sender identity)
|
||||
"jhin.lee@unity3d.com": "leehack", # PR #22053 salvage (telegram DM topic reply fallback)
|
||||
"caojiguang@gmail.com": "caojiguang", # PR #35117 carries #31853 (weixin _api_post/_api_get wait_for)
|
||||
"gooku94123@gmail.com": "goku94123", # PR #46609 salvage (MiniMax reasoning extra_body)
|
||||
# pander: empty email, salvaged via PR #19665 from #16126 by @ms-alan
|
||||
"ayman.a.kamal@hotmail.com": "A-kamal", # PR #18678 (xAI image resolution fix)
|
||||
# Kanban bug-fix batch salvage (May 2026)
|
||||
|
|
|
|||
|
|
@ -118,3 +118,115 @@ class TestMinimaxAuxModelNotHighspeed:
|
|||
"is a -highspeed variant — that costs 2x for the same model and "
|
||||
"broke #4082 the first time. Revert to plain M2.7 or M3."
|
||||
)
|
||||
|
||||
|
||||
class TestMinimaxM3OpenAIReasoningWireShape:
|
||||
"""MiniMax-M3 on api.minimax.io/v1 gets MiniMax's OpenAI-compatible knobs."""
|
||||
|
||||
def test_m3_openai_route_requests_reasoning_split_by_default(self):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
extra_body, top_level = profile.build_api_kwargs_extras(
|
||||
reasoning_config=None,
|
||||
model="MiniMax-M3",
|
||||
base_url="https://api.minimax.io/v1",
|
||||
)
|
||||
assert extra_body == {"reasoning_split": True}
|
||||
assert top_level == {}
|
||||
|
||||
def test_m3_openai_route_maps_explicit_effort_to_adaptive_only(self):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
extra_body, top_level = profile.build_api_kwargs_extras(
|
||||
reasoning_config={"enabled": True, "effort": "high"},
|
||||
model="MiniMax-M3",
|
||||
base_url="https://api.minimax.io/v1",
|
||||
)
|
||||
assert extra_body == {
|
||||
"reasoning_split": True,
|
||||
"thinking": {"type": "adaptive"},
|
||||
}
|
||||
assert top_level == {}
|
||||
|
||||
def test_m3_openai_route_does_not_send_reasoning_effort(self):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
extra_body, _top_level = profile.build_api_kwargs_extras(
|
||||
reasoning_config={"enabled": True, "effort": "xhigh"},
|
||||
model="MiniMax-M3",
|
||||
base_url="https://api.minimax.io/v1/",
|
||||
)
|
||||
assert extra_body == {
|
||||
"reasoning_split": True,
|
||||
"thinking": {"type": "adaptive"},
|
||||
}
|
||||
|
||||
def test_m3_openai_route_can_disable_thinking(self):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
extra_body, top_level = profile.build_api_kwargs_extras(
|
||||
reasoning_config={"enabled": False, "effort": "high"},
|
||||
model="MiniMax-M3",
|
||||
base_url="https://api.minimax.io/v1",
|
||||
)
|
||||
assert extra_body == {
|
||||
"reasoning_split": True,
|
||||
"thinking": {"type": "disabled"},
|
||||
}
|
||||
assert top_level == {}
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"model,base_url",
|
||||
[
|
||||
("MiniMax-M2.7", "https://api.minimax.io/v1"),
|
||||
("MiniMax-M3", "https://api.minimax.io/anthropic"),
|
||||
("MiniMax-M3", "https://api.minimaxi.com/v1"),
|
||||
],
|
||||
)
|
||||
def test_non_m3_or_non_global_openai_routes_emit_no_openai_reasoning_knobs(
|
||||
self, model, base_url
|
||||
):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
extra_body, top_level = profile.build_api_kwargs_extras(
|
||||
reasoning_config={"enabled": True, "effort": "high"},
|
||||
model=model,
|
||||
base_url=base_url,
|
||||
)
|
||||
assert extra_body == {}
|
||||
assert top_level == {}
|
||||
|
||||
def test_transport_threads_base_url_to_profile(self):
|
||||
import model_tools # noqa: F401
|
||||
import providers
|
||||
from agent.transports.chat_completions import ChatCompletionsTransport
|
||||
|
||||
profile = providers.get_provider_profile("minimax")
|
||||
assert profile is not None
|
||||
kwargs = ChatCompletionsTransport().build_kwargs(
|
||||
model="MiniMax-M3",
|
||||
messages=[{"role": "user", "content": "ping"}],
|
||||
tools=None,
|
||||
provider_profile=profile,
|
||||
reasoning_config={"enabled": True, "effort": "medium"},
|
||||
base_url="https://api.minimax.io/v1",
|
||||
)
|
||||
assert kwargs["extra_body"] == {
|
||||
"reasoning_split": True,
|
||||
"thinking": {"type": "adaptive"},
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue