P

peterwangze/dsh-reasoning-level

“DeepSeek Harness 统一推理等级插件：一个 llm-reasoning 设置项动态配置 DSH 所有模型的默认推理等级（思考等级）——自动为手写声明的模型补齐推理能力、按路由设置默认等级（仅当全部模型支持）、同步 DeepS…”

OverviewREADME

TL;DR

DeepSeek Harness 统一推理等级插件：一个 llm-reasoning 设置项动态配置 DSH 所有模型的默认推理等级（思考等级）——自动为手写声明的模型补齐推理能力、按路由设置默认等级（仅当全部模型支持）、同步 DeepSeek 官方路由，改设置即生效。

30 days of GitHub stars

+1 stars

+100% observed change

Oct 15+0 stars1 totalOct 16+0 stars (estimated)1 totalOct 17+0 stars (estimated)1 totalOct 18+0 stars (estimated)1 totalOct 19+0 stars (estimated)1 totalOct 20+1 stars2 totalOct 21+0 stars (estimated)2 totalOct 22+0 stars (estimated)2 totalOct 23+0 stars (estimated)2 totalOct 24+0 stars (estimated)2 totalOct 25+0 stars (estimated)2 totalOct 26+0 stars (estimated)2 totalOct 27+0 stars (estimated)2 totalOct 28+0 stars (estimated)2 totalOct 29+0 stars (estimated)2 totalOct 30+0 stars (estimated)2 totalOct 31+0 stars (estimated)2 totalNov 1+0 stars (estimated)2 totalNov 2+0 stars (estimated)2 totalNov 3+0 stars (estimated)2 totalNov 4+0 stars (estimated)2 totalNov 5+0 stars2 totalNov 6+0 stars (estimated)2 totalNov 7+0 stars (estimated)2 totalNov 8+0 stars (estimated)2 totalNov 9+0 stars (estimated)2 totalNov 10+0 stars2 totalNov 11+0 stars2 totalNov 12+0 stars2 total

Oct 14Oct 28Nov 12

Also in this category

4 items

E

Models

Ebony-Vinyl/dsh-our-free-model

Adds free model lanes to DeepSeek Harness — DeepSeek V4.1 Flash, Kimi K3 and others — with no login, sign-up or API key, plus a local OpenAI-compatible gateway, fallback routing and a usage dashboard.

★ 70K

E

Models

Ebony-Vinyl/dsh-our-free-model

Our Free Model — 在 dsh 里装上这个插件即可，无需登录、注册或填 API Key，就能使用包括 Muse Spark 1.3、MiMo V2.6 在内的前沿模型，完全免费不限量；模型清单随上游更新、地区可用性自动探测、思考强度真实生效，另附 Token 热力图与吞吐看板和 OpenAI 格式本地转发端口。

★ 70K

A

Models

alchaincyf/nuwa-skill

女娲：把任何人蒸馏成可运行的思维 Skill。

★ 38.1K

Z

Models

zhu1090093659/dsh-web

LiangShen agent preset plus its composer lever for the dsh web GUI. The host half declares the bundled preset to the harness's agent-preset registry at activation (declare = enabled), so the preset is selectable for new sessions without writing anything into the harness home; the preset keeps a minimal persona with standing working discipline while the harness's own user-role agent-instructions injection delivers the workspace instructions (the optional `instructionSource: system-prompt` mode lifts them into the prompt instead), and declares one tool presentation for the session's whole lifetime: `ptc` by default (the wire collapses to `run_code` and every other tool is reached through the generated SDK), with `native` as the explicit assembled-roster option. High-fan-out tool families (default `['mcp__*', 'browser_*', 'task_board_*', 'team_*', 'squad*', 'ui_*', 'ssh_*']`, while every host builtin stays resident; customizable via `pagedToolPatterns`) stay paged out of the wire and generated SDK until the preset's own `tool_activate` tool loads their namespace (LRU cap three, state replayed from the durable session event stream), and the tool catalog — exactly the reachable tools plus paged-out namespace summaries — and a one-line working-context projection travel as durable user messages republished only on change. The browser half adds a slot-machine lever left of the model selector on the new-session screen to toggle the mode. Hot-pluggable via cordis.patch.yml + profile symlink, no dsh source changes.

★ 10.4K