Replies: 2 comments 3 replies
|
Confirmed in source and staged a patch.
const REASONING_EFFORTS = [
{ id: ReasoningEffortId('off'), name: 'Off' },
{ id: ReasoningEffortId('high'), name: 'High' },
{ id: ReasoningEffortId('max'), name: 'Max' },
]No Cherry-pick-ready branch (on current master): https://github.com/zoahdev/deepseek-harness/tree/fix/llm-deepseek-reasoning-low ? adds |
2 replies
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
[Feature Request] deepseek-official 适配器不支持
reasoning_effort: "low"(官方 API 已支持,建议放行并归一 medium/xhigh)摘要 / Summary
DeepSeek 官方文档明确支持
reasoning_effort: low/high/max(medium/xhigh兼容映射为high),且 V4-Pro GA 宣传口径为 "low for simple tasks, high for daily Agent workflows, max for complex tasks"。但 DeepSeek Harness 的官方 providerdeepseek-official(@deepseek-ai/dsh-llm-deepseek)能力白名单只有off / high / max,low在请求发出前就被客户端校验拦截,导致官方模型在 DSH 内无法使用文档推荐的低推理档。English summary: The official DeepSeek API documents and accepts
reasoning_effort: low(plusmedium/xhighcompat-mapped tohigh) fordeepseek-v4-pro/deepseek-v4-flash, but the DSHdeepseek-officialprovider adapter (@deepseek-ai/dsh-llm-deepseek) only whitelistsoff | high | max, so any explicitlowis rejected client-side before the request reaches the API. Suggest addinglowto the whitelist (and normalizingmedium/xhigh→highper the official mapping).复现 / Reproduction
0.1.0-rc.6(macOS arm64,Node v26;Web GUI 与 headless 均复现)deepseek-v4-pro(providerdeepseek-official)low(模型选择器无该档位,需手写配置或脚本传入)对照实验(同一环境同一模型):
lowUNSUPPORTED_REASONING_EFFORT(客户端预检拒绝,无 API 请求)mediumhigh)highmaxrequest/header记录reasoningEffort: "max",会话正常完成)官方文档依据 / The API itself accepts
low直接调用官方 API(
https://api.deepseek.com/chat/completions,模型deepseek-v4-pro),reasoning_effort分别传low/medium/high/max,四次全部 HTTP 200。因此问题不在 API 侧。文档出处:
reasoning_effort取值为 low / high / max,默认high;官方映射表:low→low、medium→high、high→high、xhigh→high、max→max(deepseek-v4-flash 与 deepseek-v4-pro 一致)。reasoning_effort的 Possible values: [low, high, max],原文注明"出于兼容考虑medium、xhigh会映射为high"。根因定位 / Root cause(客户端适配层白名单)
请求在发出前被 DSH 客户端预检拦截,未到达 DeepSeek API:
@deepseek-ai/dsh-llm-deepseek/lib/index.js—reasoningEffort()校验函数只接受"off" | "high" | "max"(L19–22);能力列表REASONING_EFFORTS = [off, high, max](L415–431),resolveModel()只向 harness 暴露这三个档位(L505–526);该适配器注册的 provider 名即"deepseek-official"(L632)。@deepseek-ai/dsh-llm/lib/index.js(L1245 附近)对"请求档位 ∈ 模型注册能力列表"做预检,不通过即抛UNSUPPORTED_REASONING_EFFORT——实测文案与该处逐字一致。thinking: { type: "enabled" }+reasoning_effort: <档位>(snake_case),与文档 OpenAI 格式一致;low无需任何新映射,直接透传即可。packages/llm/llm-deepseek/src/index.tsL95 仍为reasoningEffort: z.union(['off', 'high', 'max']),尚未修复。影响 / Impact
off / high / max,文档推荐用于简单任务的low档不可达(GUI 推理等级选择器也不提供该档位)。low只能绕行:① 用llm-pi-ai手工声明指向api.deepseek.com的路由并逐模型声明reasoningEfforts(含low);② 安装社区插件dsh-thinking-effort(见相关讨论 【插件】dsh-thinking-effort:第三方(pi-ai)模型也能设置思考强度——默认档位自动补齐 + 设置页自定义 #1817)。建议修复 / Suggested fix
"low"加入@deepseek-ai/dsh-llm-deepseek白名单:reasoningEffort()校验(L19–22)、REASONING_EFFORTS(L415–431)、以及 GUI 模型选择器的档位展示;low的线上透传与high/max相同(thinking: enabled+reasoning_effort: "low")。medium/xhigh→ 归一为high(文档明确这是官方兼容行为)。low报错走的是通用层"模型无 reasoning 能力"分支),应一并核对,避免显式档位校验行为不一致。改动很小(一个 union 加一个枚举项 + UI 档位列表),但能对齐官方文档口径。
相关讨论 / Related
#122 第三方模型无法选择推理强度 · #302 自定义 provider 不支持 reasoning effort · #843 自定义 provider 模型无法从 GUI 选档 · #1643 网关选档功能请求 · #1817 dsh-thinking-effort 插件
以上均为第三方/自定义 provider 的档位入口问题;本帖针对的是官方 provider(deepseek-official)自身的档位白名单缺
low,是另一个独立缺口。All reactions