dsh-biodsh

中文 | English

BioDSH for DeepSeek Harness: a bundle that adds a BioDSH agent preset for bioinformatics work. Your other presets are not changed.

The BioDSH preset is the standard preset of your dsh release plus:

Part	What it does
BioDSH persona	Answers for doctors and wet-lab scientists, in the user's language; keeps raw data untouched; reports what was actually computed. When the working folder contains a machine-readable output specification (TABLE_CONTRACT.json, CLAIMS.json, …), a one-line persona and a specification-first instruction replace it.
BioDSH skills	17 validated BioDSH workflows in the session skill catalog: single-cell QC, clustering, annotation and treatment response; bulk RNA-seq limma-voom; survival (KM, Cox), regression with robust SEs and ANOVA designs, each checked against R; workflow building, and more.
Skill library	2000+ community bioinformatics, clinical and data skills. Kept out of the catalog so it stays small and cacheable; the model searches it with the skill_library tool, and you can still run any of them with /name.
Skill selection	On each new message, a model reads the whole catalog (BioDSH skills and library) and injects the one or two skills that fit the request, in the same form as /name.
Convergence guard	In a long turn, asks the agent to save its code and finish the needed checks (900 s, or 60 steps), then to report honestly what is unfinished (1100 s, or 80 steps); ends a turn past 90 steps. It never asks the agent to guess.
Delivery gate	For tasks that come with a public output specification (table contract, claims), checks the saved outputs against it, checks declared columns against the original inputs, and replays the analysis script in a clean folder before the turn may end; on failure it returns the concrete errors for repair.
BioDSH-MAGI	Experimental, off by default. Multi-Analyst Gated Inspection. When an analysis with an output specification shows a risk signal (the delivery gate needed repairs, a source check failed, or the run was long), a second analyst redoes it from the original inputs in a separate folder without seeing the first result, and an arbiter compares both and delivers only after its own executable, cell-by-cell checks pass the delivery gate. Modes: off (default), adaptive, always. In our evaluation it rescued only a few tasks while roughly doubling time and cost. The first analyst's result is kept under .biodsh-magi/.

Install

Requires DeepSeek Harness 0.2.1 or later and pnpm on PATH (dsh's plugin manager runs pnpm).

• Command line

  dsh plugin --profile web add dsh-biodsh
  

• Web UI: open Plugins, choose Install, enter dsh-biodsh.

• Ask the agent (Creator mode): "install the dsh-biodsh bundle".

Restart dsh after installing, then pick BioDSH in the agent preset menu, or make it the default:

# ~/.dsh/cordis.patch.yml
- id: agent-preset-registry
  config:
    default: biodsh

BioDSH (PTC) (biodsh-ptc) is the same preset with programmatic tool calling, like dsh's own ptc preset: the model calls tools from one program per step, so writing a script, running it and checking its output take one step instead of three. On 15 development tasks it cut the median from 23 to 16 steps and from 129 s to 97 s at the same accuracy.

Configure

The bundle adds a settings namespace biodsh (a Host row, so you never edit the preset itself). Change it in ~/.dsh/settings.yaml, on the Settings page, or with settings.mutate; most fields apply to the next step without a restart.

# ~/.dsh/settings.yaml
biodsh:
  selectProvider: deepseek-official   # skill selection model; empty = same route as the agent
  selectModel: deepseek-flash
  selectMaxInject: 2
  guardSoftSeconds: 900               # save code, finish needed checks (time is the main signal; steps are a backstop)
  guardSoftSteps: 60
  guardHardSeconds: 1100              # report what is unfinished; no guessing
  guardHardSteps: 80
  guardMaxSteps: 90                   # end the turn (0 = no cap)
  magiMode: off                       # BioDSH-MAGI (experimental): off | adaptive | always
  magiSteps: 60                       # adaptive: review runs with at least this many steps
  magiSeconds: 900                    # ...or this many seconds
  library: true                       # 2000+ skill library and the skill_library tool (restart to apply)

Versions

A dsh agent preset restates its full plugin list, so each dsh-biodsh release is built against one dsh release (dshTarget in package.json) by scripts/build.mjs, which copies the standard preset of that release and changes only the persona and the three BioDSH rows. scripts/install-test.mjs installs a build through dsh's own plugin manager in a fresh home and runs a real turn on the preset.

Licenses

BioDSH code and skills: MIT. Skills under skills/library come from community repositories and keep their own licenses; see each skill's folder.

---

中文说明

为 DeepSeek Harness 增加一个 BioDSH 智能体预设，用于生信分析。你原有的预设不受影响。

BioDSH 预设 = 你当前 dsh 版本的标准预设，再加上：

• BioDSH 人设：面向医生和湿实验科学家，跟随用户语言作答，不改原始数据，如实报告实际算了什么。工作目录里有机器可读的输出规范（TABLE_CONTRACT.json、CLAIMS.json 等）时，换成一句话人设和“按规范工作”的说明。
• BioDSH 官方技能：17 个经过验证的流程（单细胞质控/聚类/注释/治疗响应、bulk RNA-seq limma-voom、生存分析 KM/Cox、稳健标准误回归、方差分析设计等，统计类均与 R 逐项对照），出现在会话技能目录里。
• 技能库：2000+ 社区技能，不放进目录（目录保持小而稳定，利于缓存）。模型用 skill_library 工具检索，你也可以用 /技能名 直接调用。
• 技能精选：每条新消息，用模型对照全部技能选出最合适的一两个，以 /name 同样的形式注入。
• 收敛护栏：长任务到 900 秒（或 60 步）提醒先保存代码、做完必要核验；1100 秒（或 80 步）要求如实报告未完成的部分；超过 90 步停止本轮。不会要求它猜答案。
• 交付检查：带输出规范（表格合同、结论声明）的任务，在结束前核对输出结构、把声明的列对照原始数据，并在干净目录重放分析脚本；不通过就把具体错误交回去修。
• BioDSH-MAGI 复核（实验性功能，默认关闭；Multi-Analyst Gated Inspection，多分析员门控核查）：带输出规范的分析如果出现风险信号（交付检查要返修、与原始数据核对失败、运行过长），会另开一位分析员在独立目录里从原始数据重做一遍，看不到第一份结果；再由裁决员对照两份结果，写出可执行的逐格核查并通过交付检查后才交付。三档：off 关闭（默认）、adaptive 自适应、always 每次都复核。评测中它只救回少数题，时间和费用约翻倍。第一份结果保存在 .biodsh-magi/ 里。

设置：在 ~/.dsh/settings.yaml 写 biodsh: 一段（字段见上方英文说明），或在设置页修改，大多数改动下一步就生效。

安装：需要 DeepSeek Harness 0.2.1 及以上，并且 PATH 里有 pnpm。

• 命令行：dsh plugin --profile web add dsh-biodsh
• 网页：「插件」→「安装」，输入 dsh-biodsh

装好后重启 dsh，在智能体预设菜单里选 BioDSH，或按上面的 YAML 设为默认。

BioDSH (PTC)（biodsh-ptc）是同一预设加上程序化工具调用（同 dsh 自带的 ptc 预设）：模型每步写一段程序来调用工具，写脚本、运行、检查输出一步完成。在 15 道开发题上，步数中位数从 23 降到 16，耗时从 129 秒降到 97 秒，正确率不变。