dsh-bioinfo

生信模式 (Bioinformatics Mode) — a complete, research-grade bioinformatics agent preset for DeepSeek Harness (DSH), published as a replicable preset kit.

One repository carries everything needed to rebuild, on your own DSH instance, the exact preset this project was developed and acceptance-tested on: preset identity, agent composition + persona, the protein-tools plugin (7 model tools), a 7-skill library with 14 backend scripts, deployment scripts, and a small acceptance-fixture kit that proves your replica behaves like the reference machine.

What you get

Quick start

1. Install DSH — npm install -g @deepseek-ai/dsh (see upstream docs).
2. Deploy the environment — follow docs/INSTALL.md: Python 3.13 + D:\biopython (Biopython 1.87), the D:\bioai toolchain (venv with RDKit/meeko/OpenMM, venv-esm with torch/fair-esm, the Vina binary, WSL2 LocalColabFold, AF2 params).
3. Mount the preset — copy preset.yml, agent.cordis.yml, plugins/, skills/ into <DSH_HOME>/.agent-presets/bioinfo/, then start a session with the 生信模式 preset.
4. Prove the replica — run deploy/run-acceptance.ps1 and scripts/verify-layout.ps1; compare against the reference values in fixtures/README.md.
5. Remember the restart rule — the preset composition is mounted once per DSH host process. After editing any preset file (agent.cordis.yml, plugins/, skills/), restart the DSH host; a new session alone is not enough (symptom: Invalid schema ... got 'type: null' at conversation start while the on-disk file is already fixed).

The reference layout below is part of the contract: a user deploying exactly per docs/INSTALL.md gets an identical preset with zero file edits. Every hard-coded path is also overridable via environment variables for those who deviate.

Configuration (plugin)

Platform notes

• Windows + PowerShell. Every tool command is pwsh; a Linux/bash backend does not exist yet (contributions welcome).
• af2_predict needs WSL2 + an NVIDIA GPU for practical runtimes; CPU works but is slow — esmfold_predict is the zero-setup cloud alternative, but the ESM Atlas has been intermittently down (repeated 504s, 2026-08), so treat it as a fallback; offline fallback is af2_predict with msaMode: "single_sequence".

Known upstream issue (tool schema compilation)

Raw preset file-plugins receive the unsandboxed ctx, whose ctx.tools.register stores definitions verbatim and lets the model layer project definition.parameters straight to the API. A flat per-property map therefore reaches the API without a root type: "object" and gets rejected (Invalid schema ... got 'type: null'). The plugin registers through defineToolDef(): sandboxed loads go through harness.defineTool(...), raw loads hand-compile {type:'object', properties, required} — the same shape dsh-tools emits. New tools must keep using this wrapper.

Tests

npm test                        # node --check + schema validator (cross-platform)
scripts/verify-layout.ps1       # self-consistency of an installed replica
deploy/run-acceptance.ps1       # end-to-end AF2-Multimer acceptance

Release history and user-facing update notices: CHANGELOG.md.

License

Repository code: MIT (LICENSE). Backend toolchain licensing — including ACADEMIC-ONLY components (TM-align/TMalign, PRODIGY, local ESMFold weights) — is itemized in THIRD_PARTY_NOTICES.md.

Contributing & security

Contributions welcome — see CONTRIBUTING.md (tool/backend rules) and CODE_OF_CONDUCT.md. Report vulnerabilities privately via GitHub security advisories (SECURITY.md).

Real-world usage feedback and its disposition (fixed / documented / backlog) is tracked in FEEDBACK.md.