dsh-bioinf-routed

Dumping all 986 bioinformatics skills flat into every session prompt costs roughly 60 to 80k tokens of context—barely practical in real workflows. dsh-bioinf-routed was built to dissolve that tension: it replaces the monolithic catalog with a search-based routing layer, so you pull exactly the skill you need instead of loading everything up front.

Under the hood it runs a BM25-style weighted scorer (name x6, category x2, whenToUse x1.5, description x1) with built-in CJK bigram tokenization, so both Chinese and English queries resolve cleanly against the full library. A browsable 15-category tree supports domain-level exploration while skill bodies are extracted on demand, keeping a single lookup across all 986 entries well under 10 ms. Three scenario playbooks ship out of the box—literature deep-review with falsifiable hypotheses, bio-data retrieval and analysis, and remote GPU training or prediction—steering the model down the right tool chain rather than letting it guess a skill name that does not exist. An optional remote-exec tool handles SSH-based nvidia-smi checks, nohup training jobs, log polling, and prediction runs on a configured server.

It is aimed at researchers and engineers who already work inside DeepSeek Harness and maintain a large bioinformatics skill inventory, especially teams that need bilingual Chinese-English retrieval or want to kick off remote training and inference runs on a dedicated GPU server.