DSH-ReasoningEffort-Fix

针对DSH：添加第三方API时无法调节思考强度问题（DeepSeek Harness小完善计划其二）

DSH（DeepSeek Harness）第三方模型思考等级配置技能（Agent Skills 标准格式）。

An Agent Skill (standard Agent Skills format) for configuring reasoning effort on third-party models in DSH (DeepSeek Harness).

中文文档 · English

中文文档

技能标识：02-reasoning-efforts ｜ 系列：DeepSeek Harness 小完善计划（其一：上下文窗口 · 其三：视觉能力解禁）

这是什么

一个符合 Agent Skills 标准（SKILL.md + YAML frontmatter + references/ 渐进披露）的独立技能仓库。它告诉 AI：如何让手工接入 DSH 的第三方模型支持思考/推理——关键是 reasoningEfforts（声明档位）与 reasoning（路由默认档）两个字段缺一不可。

项目目的

DSH 0.1.0-rc.7 手工接入第三方模型（官方直连或中转站）时有两个缺陷：模型会话没有思考等级选择器，且思考被默认关闭（请求按非推理模型发送，不产生思考内容）。根因是手工模型条目缺少 reasoningEfforts 声明，在配置解析时被判为非推理模型。

本仓库是一个 Agent Skill——给 AI 看的"维修手册"，不是可执行软件。AI 加载 SKILL.md 后，即可按规程自行定位 settings.yaml、核实配置、补全 reasoningEfforts + reasoning 两个字段，并指导你验证修复效果。

症状速查

遇到以下任一现象，即命中本技能解决的问题：

• 给第三方模型发消息，回复里没有思考内容；
• UI 里没有思考等级选择器；
• 请求报错 UNSUPPORTED_REASONING_EFFORT；
• 想让第三方模型默认 high 档（或按需开放 low / medium / max 等档位）；
• 选 max 档发现和 high 没区别——那是预算式路径的收敛（xhigh/max→high）；真 max 需换目录自适应模型（如 kimi-coding 的 k3），手工路由开不了自适应。

仓库结构

02-reasoning-efforts/
├── SKILL.md                # 技能入口（AI 读这个）
├── README.md               # 仓库说明（人读这个）
├── AGENTS.md               # 维护守则（给维护本仓库的 AI 看）
└── references/
    ├── deep-dive.md        # 闸门源码、预算机制、协议差异、排障表
    └── examples.yaml       # 完整配置示例（含 thinkingBudgets、多协议）

使用方法

1. 安装：把本目录整体放入所用 AI 的 skills 目录（各 AI 的技能目录位置不同，以其官方文档为准）。
2. 触发：对 AI 描述症状或贴报错即可，例如：
   • "模型回复没有思考内容" / "界面上没有思考等级选择器"；
   • "选思考档后请求报错 UNSUPPORTED_REASONING_EFFORT"；
   • "想让这个模型默认用 high 档思考"。
3. AI 会做什么：定位 settings.yaml → 查证现有配置 → 补写 reasoningEfforts（档位声明）与 reasoning（路由默认档）→ 提示你如何验证（选择器出现、思考内容产生）。
4. 需要你配合：
   • settings.yaml 在工作区外，AI 写入时会请求提升权限，请审批；
   • 修改期间尽量关闭 DSH 界面——界面打开时配置可能被并发重写；
   • 改完自行备份：通过界面修改供应商设置会重写整个配置段，手工字段（含这两个字段）可能被静默抹掉，丢了让 AI 按本技能补回。

版本基线

内容基于 DSH 0.1.0-rc.7 + @earendil-works/pi-ai 0.82.1（2026-08 实战验证）。DSH 升级后字段名可能变化，请按 references/deep-dive.md 的源码定位入口重新核实并更新 frontmatter 的 metadata.baseline。

许可

文档类内容。发布到 GitHub 前建议添加 LICENSE（如 CC BY 4.0）。

English

Skill ID: 02-reasoning-efforts | Series: DeepSeek Harness Improvement Plan (Part 1: Context Window · Part 3: Vision Input Unlock)

What This Is

A standalone skill repository in the standard Agent Skills layout (SKILL.md + YAML frontmatter + references/ for progressive disclosure). It tells the AI how to enable thinking/reasoning for third-party models hand-declared in DSH — the key is that reasoningEfforts (declares the levels) and reasoning (route-level default) must both be present.

Project Purpose

When a third-party model (direct API or relay gateway) is hand-declared in DSH 0.1.0-rc.7, two defects appear: the conversation has no reasoning effort selector, and thinking is off by default (requests are sent as a non-reasoning model, so no thinking content is produced). Root cause: a hand-declared model entry without reasoningEfforts is judged non-reasoning during config resolution.

This repository is an Agent Skill — a "repair manual" for AI, not executable software. After loading SKILL.md, the AI can follow the procedure to locate settings.yaml, inspect the configuration, fill in the two fields reasoningEfforts + reasoning, and guide you through verifying the fix.

Symptoms

This skill applies when you hit any of the following:

• You message a third-party model and the reply contains no thinking content;
• The UI shows no reasoning effort selector;
• Requests fail with UNSUPPORTED_REASONING_EFFORT;
• You want the model to think at high by default (or to open more levels such as low / medium / max);
• max behaves exactly like high — that is the budget-based path clamping xhigh/max to high; a real max requires a catalog adaptive model (e.g. kimi-coding's k3), and hand-declared routes cannot opt into adaptive thinking.

Repository Structure

02-reasoning-efforts/
├── SKILL.md                # Skill entry point (the AI reads this)
├── README.md               # Repo overview (for humans)
├── AGENTS.md               # Maintenance rules (for the AI maintaining this repo)
└── references/
    ├── deep-dive.md        # Gate source code, budget mechanism, protocol differences, troubleshooting
    └── examples.yaml       # Full config examples (thinkingBudgets, multiple protocols)

Usage

1. Install: drop this whole directory into your AI's skills directory (the location varies by AI; follow its official docs).
2. Trigger: just describe the symptom or paste the error, e.g.:
   • "The model's replies have no thinking content" / "There's no reasoning effort selector in the UI";
   • "Requests fail with UNSUPPORTED_REASONING_EFFORT after I pick a thinking level";
   • "I want this model to think at high by default".
3. What the AI does: locate settings.yaml → inspect the current configuration → add reasoningEfforts (level declarations) and reasoning (route-level default) → tell you how to verify (selector appears, thinking content shows up).
4. What you need to do:
   • settings.yaml lives outside the workspace; approve the elevated-permission prompt when the AI writes to it;
   • Keep the DSH UI closed while editing — an open UI may rewrite the config concurrently;
   • Back up after the fix: changing provider settings through the UI rewrites the whole config section and may silently drop hand-added fields (these two included); if lost, ask the AI to re-apply this skill.

Version Baseline

Content verified against DSH 0.1.0-rc.7 + @earendil-works/pi-ai 0.82.1 (field-validated 2026-08). Field names may change as DSH evolves; re-verify via the source-code entry points in references/deep-dive.md and update metadata.baseline in the frontmatter.

License

Documentation content. Adding a LICENSE (e.g. CC BY 4.0) is recommended before publishing to GitHub.