dsh-academic-research-skills
Manifest validCommunity-maintained ARS port for DeepSeek Harness with research and writing skills, native Team review and pipeline workflows, and workspace-scoped status and integrity tools.
dsh-academic-research-skills
DeepSeek Harness(DSH)插件:把 Claude Code 的
Academic Research Skills (ARS)
(v3.24.2,CC-BY-NC-4.0)移植为 DSH 技能,并随包携带完整 ARS 核心工具层
(scripts/、shared/、根 agents/ 与 MODE_REGISTRY.md)。
安装后自动获得:
- 5 个模型可调用技能:
deep-research(13-agent 调研团队,8 种模式)、academic-paper(12-agent 论文写作,11 种模式)、academic-paper-reviewer(5 席位同行评审面板)、academic-pipeline(10 阶段全流水线编排)、sr-screener(协议驱动的系统综述记录筛选,8 种模式) - 2 个原生 DSH 团队技能(0.4.4 原生层):
ars-team用真实subagent跑五席位评审;ars-pipeline-team把 10 阶段流水线映射成 Team board 的 stage task graph - 5 个原生工具(0.4.4):
ars_search(OpenAlex 检索)、ars_lookup(DOI 元数据,OpenAlex/Crossref)、ars_integrity(调用 vendoredcheck_pipeline_integrity.py)、ars_pipeline_status(流水线阶段产物盘点)与ars_manifest(pinned upstream/内容计数);仅在tools服务挂载时注册,否则自动跳过 - 可选 ARS Team agent preset(0.4.4):在
agentPresets服务可用时注册, 新会话可选择ars-team预设,组合ars-team技能与 Team board - 映射、验证与市场草稿:
native/team-map.md、docs/native-validation.md、docs/market/listing.md - 完整 ARS 核心工具层:根
scripts/(Python 检查器、ledger、citation gate、 adapters)、shared/(schema/合同)、根agents/与MODE_REGISTRY.md, 与上游对应文件逐字节一致 - 16 个用户可调用命令:
/ars-full、/ars-plan、/ars-reviewer、/ars-lit-review、/ars-abstract、/ars-revision、/ars-outline、/ars-citation-check、/ars-disclosure、/ars-format-convert、/ars-rebuttal-audit、/ars-revision-coach、/ars-cache-invalidate、/ars-mark-read、/ars-unmark-read、/ars-3w
署名 / Attribution:本项目为 ARS 的 DSH 移植,技能内容与命令衍生自 Cheng-I Wu(Imbad0202)的 academic-research-skills, 全部内容(含移植层)遵循 CC-BY-NC-4.0(非商业用途),详见 NOTICE.md 与 LICENSE。
本仓库是由 TY-Cheng 维护的社区 DSH 移植,源自 nullptr-DZF 的移植版本, 不是 ARS 官方发行版;上游仍是 Imbad0202 的 ARS。
安装
当前版本为 v0.4.6;详见发布说明。
要求:DeepSeek Harness(dsh)0.2.x(本次 e2e 在 0.2.2-rc.5 上验证),
pnpm 可用。完整确定性工具链另需 Python 3.9+;建议在插件目录建 .venv:
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txt
# 可选:PDF 内容分类器
.venv/bin/python -m pip install -r requirements-pdf-content-classifier.txt
# 方式一:GitHub 仓库(当前 v0.4.6)
dsh plugin --profile web add github:TY-Cheng/dsh-academic-research-skills#v0.4.6
# 提交上游审查的 e2e 使用 v0.3.0(不含 0.4.4 原生层)
dsh plugin --profile web add github:TY-Cheng/dsh-academic-research-skills#v0.3.0
# 方式二:本地路径(开发调试)
dsh plugin --profile web add <本仓库绝对路径>
# 把 web 换成实际 profile(本次 e2e 用 desktop);重启对应 profile 后生效
安装后 dsh.profile.bundles 会追加 dsh-academic-research-skills,启动时
lib/startup.js 把 23 个入口(5 个 vendored 技能 + 2 个原生技能 + 16 个命令)
注册进 ctx.skills;manifest.json 记录上游 pin,SYNC_POLICY.md 说明
vendored 边界。
使用
| 用法 | 操作 |
|---|---|
| 模型自动路由 | 直接说“做学术研究 / 文献综述 / 帮我审稿 / 写论文”等(技能描述带触发词) |
| 技能注入 | 输入 /deep-research、/academic-pipeline 等技能名 |
| 斜杠命令 | 输入 /ars-full(全流程)、/ars-reviewer(审稿)、/ars-plan(规划)等 |
与上游的差异
- hooks 不移植:上游 Claude Code 的按角色写入 guard 与 SessionStart
播报在此移植版中不可用。macOS 上,DSH
0.2.2-rc.5的文件写入与 Bash 后端在实测的workspace-write策略下允许 workspace 内写入、 拒绝向测试用的相邻目录写入。这是宿主写入边界,不等同于 ARS 的按角色 guard。见实测与适用范围。 - 完整核心已 vendor:
skills/、scripts/、shared/、根agents/、MODE_REGISTRY.md与上游对应文件逐字节一致;commands/是生成的 DSH 适配层。Python 依赖见上。 - 路径解析:注册时会在每个技能正文前加入 ARS package root 提示,
DSH_ARS_ROOT指向插件包根目录;sr-screener自己的scripts/仍相对 技能目录解析。这样python3 scripts/...不会误落到用户论文项目目录。 - 原生层(0.4.4):
native/skills/ars-team把五席位评审 映射为真实subagent;native/skills/ars-pipeline-team把 10 阶段流水线 映射为 Team board 的 stage task graph;lib/startup.js在tools服务 可用时注册ars_search/ars_lookup/ars_integrity/ars_pipeline_status/ars_manifest,在agentPresets可用时注册可选ars-teampreset。 vendored 内容不变;原生层验证与提交上游审查的 v0.3.0 e2e 分开。 - 子代理:5 个 vendored 技能内共 42 个 agent 提示文件(另有 4 个根
agents/文件)默认内联运行;ars-team用subagent分工跑五席位。 角色分离是流程控制,不是独立误差过程的证明。 - 命令包装为 DSH 适配版:由上游
commands/ars-*.md再生成并去掉 Claude Code 专属的Skill工具调用与${CLAUDE_PLUGIN_ROOT}路径,产物可直接在 DSH 中执行(见下节)。
重新生成命令包装
commands/ 由上游 commands/ars-*.md 转换而来(加入
disable-model-invocation: true + user-invocable: true):
node tooling/regenerate-commands.mjs <上游仓库路径>
再生成时会自动去掉 Claude Code 专属的 Skill 工具调用和
${CLAUDE_PLUGIN_ROOT} 路径。
日常同步与校验:
node tooling/sync-upstream.mjs <上游仓库路径> # 复制核心 + 生成命令 + 写 manifest
npm run verify # 离线哈希校验 + 23 个注册断言
English
A DeepSeek Harness plugin
port of Claude Code's
Academic Research Skills (ARS)
(v3.24.2, CC-BY-NC-4.0). The package also carries the full ARS core:
scripts/, shared/, root agents/, and MODE_REGISTRY.md.
Provides, out of the box:
- 5 model-invocable skills:
deep-research,academic-paper,academic-paper-reviewer,academic-pipeline,sr-screener - 2 native team skills (0.4.4 native layer):
ars-team, which runs the five-seat panel as realsubagents, andars-pipeline-team, which maps the 10 pipeline stages onto a Team-board stage task graph - 5 native tools (0.4.4):
ars_search(OpenAlex search),ars_lookup(DOI metadata via OpenAlex/Crossref),ars_integrity(runs the vendoredcheck_pipeline_integrity.py),ars_pipeline_status(stage-artifact inventory), andars_manifest(pinned upstream/content counts); registered only when thetoolsservice is mounted, skipped otherwise - Optional ARS Team agent preset (0.4.4): registered when
agentPresetsis mounted; a new session can select thears-teampreset, which composes thears-teamskill with the Team board - Full ARS core: root
scripts/(Python checkers, ledger, citation gate, adapters),shared/(schemas/contracts), rootagents/, andMODE_REGISTRY.md, byte-identical to upstream - 16 user-invocable commands:
/ars-full,/ars-plan,/ars-reviewer,/ars-lit-review,/ars-abstract,/ars-revision,/ars-outline,/ars-citation-check,/ars-disclosure,/ars-format-convert,/ars-rebuttal-audit,/ars-revision-coach,/ars-cache-invalidate,/ars-mark-read,/ars-unmark-read,/ars-3w
Attribution: this is a DSH port of ARS; skill content and commands are derived from Cheng-I Wu's academic-research-skills. Everything in this repo (including the porting layer) is licensed CC-BY-NC-4.0 — non-commercial use only. See NOTICE.md and LICENSE.
This community DSH port is maintained by TY-Cheng and builds on nullptr-DZF's original port. It is not an official ARS release; upstream remains Imbad0202's ARS.
Install
The current version is v0.4.6. See the release notes.
Requires DeepSeek Harness 0.2.x (tested on 0.2.2-rc.5) and pnpm. The
deterministic tooling also needs Python 3.9+; a plugin-local .venv is
recommended:
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txt
# optional PDF content classifier
.venv/bin/python -m pip install -r requirements-pdf-content-classifier.txt
# current release (v0.4.6)
dsh plugin --profile web add github:TY-Cheng/dsh-academic-research-skills#v0.4.6
# v0.3.0 used for the e2e submitted for upstream review (no 0.4.4 native layer)
dsh plugin --profile web add github:TY-Cheng/dsh-academic-research-skills#v0.3.0
# or from a local checkout:
dsh plugin --profile web add /path/to/this/repo
# replace web with your profile (this e2e used desktop); restart it
The bundle patch mounts lib/startup.js, which registers 23 entries
(5 vendored skills, 2 native skills, and 16 commands) into ctx.skills at boot.
Differences from upstream
- Hooks are not ported: the upstream Claude Code per-role write-scope
guard and SessionStart announcement are unavailable in this port. On macOS,
DSH
0.2.2-rc.5's file-write and Bash backends under the testedworkspace-writepolicy allowed writes inside the workspace and denied the tested writes to a sibling directory. This is host-level confinement, not the ARS per-role guard. See the test and its limits. - Full core is vendored:
skills/,scripts/,shared/, rootagents/, andMODE_REGISTRY.mdare byte-identical to upstream;commands/is the generated DSH adapter layer. Python dependencies are listed above. - Path resolution: every registered skill body is prefixed with an ARS
package-root note;
DSH_ARS_ROOTpoints at the plugin package root.sr-screener's ownscripts/still resolve relative to its skill directory, sopython3 scripts/...does not fall through to the paper project. - Native layer (0.4.4):
native/skills/ars-teammaps the five-seat review to realsubagents;native/skills/ars-pipeline-teammaps the 10 stages onto a Team-board stage task graph;lib/startup.jsregistersars_search/ars_lookup/ars_integrity/ars_pipeline_status/ars_manifestwhen thetoolsservice is mounted, and an optionalars-teampreset whenagentPresetsis mounted. Vendored content is unchanged; the native validation is separate from the v0.3.0 e2e submitted for upstream review. - Subagents: the 42 skill-agent prompt files (plus the 4 root
agents/files) run inline by default;ars-teamusessubagents for the five seats. Role separation is a process control, not proof of independent error processes. - Command wrappers are DSH-adapted: the wrappers are regenerated from the
upstream
commands/ars-*.mdand stripped of Claude Code'sSkilltool calls and${CLAUDE_PLUGIN_ROOT}paths, so they run as-is under DSH.
Regenerating command wrappers
commands/ is generated from upstream commands/ars-*.md:
node tooling/regenerate-commands.mjs <upstream-checkout-path>
Regeneration also strips Claude Code's Skill tool call and
${CLAUDE_PLUGIN_ROOT} paths.
For a routine sync and offline check:
node tooling/sync-upstream.mjs <upstream-checkout> # copy core + regenerate commands + manifest
npm run verify # hashes + 23 registration assertions
Comments
Loading…
Similar plugins
by nullptr-DZF
Ported from the ARS Claude Code plugin (44k+ stars on GitHub): academic research skills for DeepSeek Harness — a deep-research agent team, a paper-writing pipeline, a multi-perspective peer-review pan
★ 34
JavaScript
Aug 27, 2026
dsh plugin --profile web add dsh-academic-research-skillsby lobster-liu
Multi-agent A-share research pipeline skill for DeepSeek Harness (6 roles, dual-blind risk review, source-cited reports, PII gate)
★ 0
MIT
Python
Nov 4, 2026
dsh plugin --profile web add dsh-invest-researchby coffee-man666
Repository and agent-runtime analysis skills as an installable DeepSeek Harness (dsh) plugin
★ 0
MIT
TypeScript
Aug 24, 2026
dsh plugin --profile web add @deepseek-ai/dsh-lensby qkycir-123
Automatically turn successful DeepSeek Harness sessions into reusable, reviewable Agent Skills.
★ 109
↓ 442/wk
MIT
TypeScript
Nov 15, 2026
dsh plugin --profile web add dsh-run2skillby lau4tin1
Skill routing for DeepSeek Harness: embeds skills and each user task with a local model, keeps the clearly-relevant skills with a gap-based selection rule, and auto-injects their full bodies into the
★ 0
NOASSERTION
TypeScript
Oct 1, 2026
dsh plugin --profile web add dsh-skill-routerCoding agent preset plus an embedded coding-discipline skill for DeepSeek Harness: verify-first edits, minimal diffs, real test runs and evidence-based reports.
★ 0
dsh plugin --profile web add dsh-coding-discipline