Teknium
f7d90c9410
The #89503/#70058/#74295/#87279 bug class kept regenerating because every transport and provider profile hand-rolled its own effort translation map (9 sites, 4 distinct policies). New agent/reasoning_effort.py is the single source of truth: - EFFORT_LADDER: canonical low->high ordering (superset check against VALID_REASONING_EFFORTS pinned by test) - clamp_effort(): one policy — supported passes verbatim, otherwise nearest WEAKER supported level (never escalate, never invert the ladder), floor when nothing weaker, 'none' never a degradation target, declared vendor-documented overrides win, bespoke names pass through - declared wire vocabularies as data: OpenAI-compat, Codex Responses, xAI (4.6/legacy), Actual relays, Kimi K3/K2, TokenHub, GLM-5.2, DeepSeek V4, Ollama Cloud, Meta, Solar Converted sites (all behavior-preserving except noted): - chat_completions chokepoint, Kimi + TokenHub paths - codex transport (backend branches now pick a declared set) - auxiliary_client Responses path - hermes_cli.models clamp_reasoning_effort_to_supported -> thin wrapper - plugins: kimi-coding, zai, opencode-zen, deepseek, ollama-cloud, meta-ai, upstage, custom (copilot already routes via the wrapper) Behavior fixes the shared policy surfaces: - ollama-cloud/opencode-go 'minimal' now degrades to 'low' instead of being dropped (drop left the server default = MORE thinking than asked) New tests: ladder contract (every configurable level is clamped by every declared wire set; monotonicity across the full ladder for every set).
112 lines
4.1 KiB
Python
112 lines
4.1 KiB
Python
"""DeepSeek provider profile.
|
|
|
|
DeepSeek's V4 family 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 (``deepseek-v3-*`` variants) are left as no-ops so we
|
|
don't perturb the V3 wire format.
|
|
|
|
The legacy aliases ``deepseek-chat`` / ``deepseek-reasoner`` were retired on
|
|
2026-07-24. Use ``deepseek-v4-flash`` or ``deepseek-v4-pro``; Hermes remaps
|
|
the retired IDs in ``hermes_cli.model_normalize``.
|
|
"""
|
|
|
|
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). Retired aliases are remapped
|
|
before requests leave Hermes, so they are not listed here.
|
|
"""
|
|
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
|
|
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 via the shared vocabulary in agent.reasoning_effort
|
|
# (DeepSeek V4: low/medium/high/max, xhigh rounds up to max). When no
|
|
# effort is set we omit reasoning_effort so DeepSeek applies its
|
|
# server default (currently high).
|
|
if isinstance(reasoning_config, dict):
|
|
from agent.reasoning_effort import (
|
|
DEEPSEEK_V4_EFFORTS,
|
|
DEEPSEEK_V4_OVERRIDES,
|
|
clamp_effort,
|
|
)
|
|
|
|
effort = (reasoning_config.get("effort") or "").strip().lower()
|
|
if effort and effort != "none":
|
|
clamped = clamp_effort(
|
|
effort, DEEPSEEK_V4_EFFORTS, DEEPSEEK_V4_OVERRIDES
|
|
)
|
|
if clamped in DEEPSEEK_V4_EFFORTS:
|
|
top_level["reasoning_effort"] = clamped
|
|
|
|
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-v4-pro",
|
|
"deepseek-v4-flash",
|
|
),
|
|
base_url="https://api.deepseek.com/v1",
|
|
default_aux_model="deepseek-v4-flash",
|
|
)
|
|
|
|
register_provider(deepseek)
|