biomed-dispatch

Biomedical Analysis Dispatch

Purpose

Bridge between the OpenClaw conversational interface and Claude Code's scientific execution environment (K-Dense Scientific Skills).

When to use

• Any bioinformatics task: RNA-seq, scRNA-seq, variant calling, sequence analysis
• Drug discovery: molecular docking, virtual screening, ADMET prediction
• Clinical data: survival analysis, variant interpretation, clinical trials search
• Multi-omics: proteomics, metabolomics, pathway enrichment
• Medical imaging: DICOM processing, digital pathology
• Scientific communication: literature review, scientific writing, figure generation
• Any request mentioning specific tools: DESeq2, Seurat, Scanpy, RDKit, BioPython, etc.

Workflow

1. Identify task type from the user's request
2. Locate data files — check if user mentioned a file path; if not, list /workspace/data/ and confirm with user
3. Set up Dashboard — every analysis task must have a live dashboard:

   TASK_DIR=data/<task_name>
   mkdir -p "$TASK_DIR/dashboard" "$TASK_DIR/output"
   cp skills/dashboard/dashboard.html "$TASK_DIR/dashboard/"
   cp skills/dashboard/dashboard_serve.py "$TASK_DIR/dashboard/"
   # Write initial state.json with: progress(0%), 研究概要, 分析计划(list), empty steps
   # Start server
   python "$TASK_DIR/dashboard/dashboard_serve.py" --port <free_port> &
   # Tell user the URL immediately: http://localhost:<port>/dashboard/dashboard.html
   

4. Construct the Claude Code prompt — include dashboard update instructions:
   • Which scientific skill(s) to use
   • Input file path(s)
   • Output directory: always $TASK_DIR/output/
   • Dashboard state.json path and update expectations:
     • Update progress after each step
     • Use step panels with desc, code, code_file, outputs
     • Use {"src": "/output/file.csv"} for table references (NOT inline data)
     • Image paths absolute: /output/fig1.png
   • Expected output format (table, figure, report)
5. Execute via Claude Code CLI:

   claude --dangerously-skip-permissions -p "Use available scientific skills. [TASK]. Input: [PATH]. Outputs: $TASK_DIR/output/. Update dashboard at $TASK_DIR/dashboard/state.json after each step (step panels with code + outputs). Completion: openclaw system event --text 'Done: summary' --mode now"
   

6. Monitor — if the task takes >30s, inform the user it is running in background
7. Report back — summarize results, point user to dashboard URL for details