dsh-tier
Skill79 starsChanged 14 days agoMITView on GitHub
给本机 DeepSeek-V4.1-flash 做智力分档(low/high/max)并按任务难度自动切换。开工判档、收工回落到 high。
In this repository26 other files it gives its agents.
- doc-tidySkill
- dsh-backupSkill
- dsh-claude-styleSkill
- dsh-code-craftSkill
- dsh-crewSkill
- dsh-data-craftSkill
- dsh-deep-researchSkill
- dsh-diagramSkill
- dsh-doc-publisherSkill
- dsh-doc-searchSkill
- dsh-doc-tidySkill
- dsh-excel-copilotSkill
- dsh-flashgouSkill
- dsh-lifeSkill
- dsh-mail-craftSkill
- dsh-meeting-copilotSkill
- dsh-memory-syncSkill
- dsh-mobileSkill
- dsh-reviewSkill
- dsh-scrum-masterSkill
- dsh-slidesSkill
- dsh-storyboardSkill
- dsh-videoSkill
- dsh-voiceSkill
- dsh-web-crawlerSkill
- dsh-write-craftSkill
The file
Files
智力分档(low / high / max)
机制(已实测,不是推测)
引擎 RPC session/selectModel 可以当场改会话的推理档,不用重启引擎、不用重装:
node /sdcard/DeepSeekHarness/tools/tier.mjs low|high|max
node /sdcard/DeepSeekHarness/tools/tier.mjs --auto "<用户这句话>" # 启发式判档
node /sdcard/DeepSeekHarness/tools/tier.mjs --show # 当前档- 底层:
node tools/rpc-call.mjs session/selectModel '{"sessionId":…,"provider":"deepseek-official","model":"deepseek-flash","reasoningEffort":"low"}' - 生效范围:下一个 LLM 请求开始(同一轮里紧接着的那一步就用新档);
选择会持久化进 profile 的
agent-default-model条目 → 之后一直是这个档,直到下次改。 - 实测证据:设
low后,会话日志下一条request/header里reasoningEffort=low(sessions/…/session.v4.jsonl.zstd,多帧 zstd 要逐帧解)。 - 深档只有三档 + off:
deepseek-flash的reasoningEfforts= low / high / max。
何时用哪一档
| 档 | 判据(满足其一) | 典型 |
| low | 只是取数/看一眼/搬运;不需要权衡 | 查用量、看通知、读文件、列目录、翻译、设提醒、简单发消息 |
| high | 默认。多步操作、要读代码、要判断但路径清楚 | 常规改代码、写文档、排简单报错、跑脚本 |
| max | 不可逆或代价高;要在两个方案里取舍;要说服自己"没编" | 装机/发版/删改数据、内核与协议改动、架构设计、根因分析、会审、写进记忆的"事实" |
判错只差一档,宁可判高(low 判错会误事,max 判错只是多花几分钱)。
纪律
- 一轮只切一次,在动手前切;
tier.mjs自带"同档不重复写"的判断。 - 任务变重/变轻可以中途切(下一步就生效),但别来回抖。
- 一轮结束回到 high(休息档),别把
max留在默认位上烧钱。 - 切换会写 profile 文件(
profiles/web/cordis.patch.yml)—— 这是预期,也是它持久的原因; 不要手改那个文件里的 effort,走tier.mjs。 - 子代理(subagent)默认继承父档;想让子代理单跑更低/更高的档,
要在
dsh-tool-subagent上开modelSelectionSettings: true(需重启引擎,见中枢总计划 H2)。
省钱账(为什么值得做)
账单 ≈ 上下文 × 轮次,其中思考 token 按输出价计。日常大量轮次是"取数型",
放在 max 上纯属浪费;反过来把架构/发版放在 low 上,判错一次的返工比省下的贵得多。
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

