mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-05-18 04:41:56 +00:00
cd9470f416
The cherry-picked PR #15251 from @tw2818 correctly identified the DeepSeek 400 root cause but placed the fix in the legacy fallback path of `build_kwargs`, which DeepSeek never reaches — DeepSeek has a registered ProviderProfile and goes through `_build_kwargs_from_profile` instead. The legacy-path block was therefore dead code. This commit pivots the fix to where it actually fires: - New `DeepSeekProfile` in `plugins/model-providers/deepseek/__init__.py` overrides `build_api_kwargs_extras` to emit DeepSeek's expected wire format (mirrors `KimiProfile`): {"reasoning_effort": "<low|medium|high|max>", "extra_body": {"thinking": {"type": "enabled" | "disabled"}}} - Model gating: only `deepseek-v4-*` and `deepseek-reasoner` emit thinking control. `deepseek-chat` (V3) is untouched — current behavior. - Effort mapping: low/medium/high passthrough, xhigh/max → max, unset → omitted (DeepSeek server applies its own default). - Revert the legacy-path additions from PR #15251 — they were dead code, and the `_copy_reasoning_content_for_api` strip block specifically would have nullified the existing reasoning_content padding machinery (`_needs_deepseek_tool_reasoning` → space-pad on replay) that the active provider already relies on for replay correctness. - Unit tests pin the wire-shape contract and the model gating rules (26 tests, all passing). Existing transport + provider profile suites (321 tests) continue to pass. - AUTHOR_MAP: map twebefy@gmail.com → tw2818 for release notes credit. Closes #15700, #17212, #17825. Co-authored-by: tw2818 <twebefy@gmail.com>
99 lines
3.6 KiB
Python
99 lines
3.6 KiB
Python
"""DeepSeek provider profile.
|
|
|
|
DeepSeek's V4 family (and the legacy ``deepseek-reasoner``) defaults to
|
|
thinking-mode ON when ``extra_body.thinking`` is unset. The API then returns
|
|
``reasoning_content`` and starts enforcing the contract that subsequent turns
|
|
echo it back; combined with how Hermes replays history this lands on the
|
|
notorious HTTP 400 ``reasoning_content must be passed back`` error after the
|
|
first tool call (#15700, #17212, #17825).
|
|
|
|
This profile overrides :meth:`build_api_kwargs_extras` to mirror the Kimi /
|
|
Moonshot wire shape that DeepSeek's OpenAI-compat endpoint expects:
|
|
|
|
{"reasoning_effort": "<low|medium|high|max>",
|
|
"extra_body": {"thinking": {"type": "enabled" | "disabled"}}}
|
|
|
|
Non-thinking models (only ``deepseek-chat`` today, which is V3) are left as
|
|
no-ops so we don't perturb the V3 wire format.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
from providers import register_provider
|
|
from providers.base import ProviderProfile
|
|
|
|
|
|
def _model_supports_thinking(model: str | None) -> bool:
|
|
"""DeepSeek thinking-capable model families.
|
|
|
|
Currently covers the V4 family (``deepseek-v4-pro``, ``deepseek-v4-flash``,
|
|
and any future ``deepseek-v4-*`` variants) and the legacy
|
|
``deepseek-reasoner`` (R1). ``deepseek-chat`` is V3 with no thinking mode.
|
|
"""
|
|
m = (model or "").strip().lower()
|
|
if not m:
|
|
return False
|
|
if m.startswith("deepseek-v") and not m.startswith("deepseek-v3"):
|
|
# deepseek-v4-*, deepseek-v5-*, etc. — every V4+ generation has
|
|
# thinking. v3 explicitly excluded.
|
|
return True
|
|
if m == "deepseek-reasoner":
|
|
return True
|
|
return False
|
|
|
|
|
|
class DeepSeekProfile(ProviderProfile):
|
|
"""DeepSeek — extra_body.thinking + top-level reasoning_effort."""
|
|
|
|
def build_api_kwargs_extras(
|
|
self, *, reasoning_config: dict | None = None, model: str | None = None, **context
|
|
) -> tuple[dict[str, Any], dict[str, Any]]:
|
|
extra_body: dict[str, Any] = {}
|
|
top_level: dict[str, Any] = {}
|
|
|
|
if not _model_supports_thinking(model):
|
|
# V3 / unknown — leave wire format untouched, current behavior.
|
|
return extra_body, top_level
|
|
|
|
# Determine enabled/disabled. Default is enabled to match DeepSeek's
|
|
# API default; the API requires this to be set explicitly to avoid the
|
|
# reasoning_content echo trap on subsequent turns.
|
|
enabled = True
|
|
if isinstance(reasoning_config, dict) and reasoning_config.get("enabled") is False:
|
|
enabled = False
|
|
|
|
extra_body["thinking"] = {"type": "enabled" if enabled else "disabled"}
|
|
|
|
if not enabled:
|
|
return extra_body, top_level
|
|
|
|
# Effort mapping. Pass low/medium/high through; xhigh/max → max.
|
|
# When no effort is set we omit reasoning_effort so DeepSeek applies
|
|
# its server default (currently high).
|
|
if isinstance(reasoning_config, dict):
|
|
effort = (reasoning_config.get("effort") or "").strip().lower()
|
|
if effort in ("xhigh", "max"):
|
|
top_level["reasoning_effort"] = "max"
|
|
elif effort in ("low", "medium", "high"):
|
|
top_level["reasoning_effort"] = effort
|
|
|
|
return extra_body, top_level
|
|
|
|
|
|
deepseek = DeepSeekProfile(
|
|
name="deepseek",
|
|
aliases=("deepseek-chat",),
|
|
env_vars=("DEEPSEEK_API_KEY",),
|
|
display_name="DeepSeek",
|
|
description="DeepSeek — native DeepSeek API",
|
|
signup_url="https://platform.deepseek.com/",
|
|
fallback_models=(
|
|
"deepseek-chat",
|
|
"deepseek-reasoner",
|
|
),
|
|
base_url="https://api.deepseek.com/v1",
|
|
)
|
|
|
|
register_provider(deepseek)
|