xjtulyc/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. 拆分长任务 — 如果任务包含多个阶段（如：文献搜索 + 写大纲 + 写正文 + 做图 + 编译），必须拆成多个 Claude Code session，每个 session 只做一件事：
   • Phase 1: 文献搜索 + 大纲
   • Phase 2: 写正文（或分章节）
   • Phase 3: 生成图表
   • Phase 4: 编译 PDF
   • 原因： 单次 CC session 超过 10 分钟大概率卡住（上下文窗口满、API 超时、生成超长文本）
5. Construct the Claude Code prompt — 短而聚焦，包含 dashboard 更新指令：
   • Which scientific skill(s) to use（明确指定 skill 路径，如 先读 ~/next-medgeai/MedgeClaw/scientific-skills/scientific-skills/scientific-writing/SKILL.md）
   • 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)
6. Execute via Claude Code CLI（推荐 stream-json + hooks）:

   # 推荐：stream-json 模式（可观测）
   cd "$TASK_DIR" && claude -p "短任务描述。先读 skill 文件。完成后: openclaw system event --text 'Done: 摘要' --mode now" \
     --output-format stream-json \
     --verbose \
     --dangerously-skip-permissions \
     2>/dev/null | tail -1
   
   # 旧方式（不推荐，无可观测性）
   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"
   

7. Monitor — 用 hooks 的 progress.json 判断进度：
   • 如果 last_update 超过 5 分钟没变 → 大概率卡了，kill 掉重来
   • 如果任务 >30s，告知用户后台运行中
8. Report back — 总结结果，指向 dashboard URL

科学写作任务的特殊处理

文献综述 / 论文写作必须拆分：

# Phase 1: 文献搜索 + 大纲（5-10 分钟）
cd writing_outputs/<task_name> && claude -p "读 ~/next-medgeai/MedgeClaw/scientific-skills/scientific-skills/literature-review/SKILL.md 和 scientific-writing/SKILL.md。按 literature-review 流程搜索文献，创建 outline.md（Stage 1）。" \
  --output-format stream-json --verbose --dangerously-skip-permissions 2>/dev/null | tail -1

# Phase 2: 写正文（分章节，每章 5-10 分钟）
cd writing_outputs/<task_name> && claude -p "读 outline.md 的第 1-3 节。用 Edit 工具在 manuscript.tex 中补充这些章节的正文。写完整的学术散文。" \
  --output-format stream-json --verbose --dangerously-skip-permissions 2>/dev/null | tail -1

# Phase 3: 创建 BibTeX + 添加引用（5 分钟）
cd writing_outputs/<task_name> && claude -p "读 manuscript.tex 和 outline.md。创建 references/references.bib（至少 30 篇），在 tex 中添加 \cite{}（每节至少 5 处），在 \end{document} 前加 \bibliographystyle{unsrt} 和 \bibliography{references}。" \
  --output-format stream-json --verbose --dangerously-skip-permissions 2>/dev/null | tail -1

# Phase 4: 生成图表（5-10 分钟）
cd writing_outputs/<task_name> && claude -p "在 figures/ 下创建 5 个 Python 脚本生成图表 PDF。中文标签用 Noto Sans CJK SC 字体。" \
  --output-format stream-json --verbose --dangerously-skip-permissions 2>/dev/null | tail -1

# Phase 5: 编译 PDF（手动或简单 CC）
cd writing_outputs/<task_name>/drafts && xelatex -output-directory=../final manuscript.tex
cd ../final && bibtex manuscript && cd ../drafts && xelatex -output-directory=../final manuscript.tex && xelatex -output-directory=../final manuscript.tex

为什么必须拆分：

• 单次让 CC 做完所有步骤（搜文献 + 写大纲 + 写正文 + 做图 + 编译）会导致：
  • 上下文窗口满（74 篇摘要 + task.md + skill 文件 + LaTeX 模板 = 超大上下文）
  • Opus 生成超长 LaTeX 文本（数千 token）需要 10+ 分钟，容易超时
  • 中途卡住后无法恢复，只能重来
• 拆分后每个 phase 独立，失败了只需重跑该 phase

Output handling

• Tables → summarize top rows, mention full file path
• Figures → send the image file to the user directly
• Reports → send the PDF/HTML file to the user directly
• Errors → show the error message and suggest a fix

Example dispatches

Clinical data analysis (complete flow with dashboard):

# 1. Setup
TASK_DIR=data/charls_ace
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/"
# 2. Write initial state.json
# 3. Start dashboard server
python "$TASK_DIR/dashboard/dashboard_serve.py" --port 7790 &
# 4. Dispatch to Claude Code
claude --dangerously-skip-permissions -p "分析 CHARLS 队列中 ACE 与 CVD 的关联。Input: data/charls_ace/charls.dta. Output: data/charls_ace/output/. 每步更新 dashboard state.json（step panels with code + outputs）。完成后: openclaw system event --text 'Done: ACE-CVD分析完成' --mode now"

RNA-seq differential expression:

claude --dangerously-skip-permissions -p "Use DESeq2 scientific skill. Run differential expression. Counts: /workspace/data/counts.csv, metadata: /workspace/data/meta.csv, contrast: treatment vs control. Save to /workspace/data/rnaseq/output/. Update dashboard at /workspace/data/rnaseq/dashboard/state.json."

Single-cell RNA-seq:

claude --dangerously-skip-permissions -p "Use Scanpy scientific skill. Analyze 10X data at /workspace/data/10x/. QC, clustering, markers. Save to /workspace/data/10x/output/. Update dashboard state.json with step panels."

输出路径约束（重要）

所有任务输出必须写入指定目录：

| 任务类型 | 输出路径 | 说明 | |---------|---------|------| | 数据分析 | data/<task_name>/output/ | CSV、图表、报告 | | Dashboard | data/<task_name>/dashboard/ | state.json, dashboard.html, serve.py | | 科学写作 | writing_outputs/<date>_<topic>/ | LaTeX、PDF、BibTeX、figures/ | | 临时文件 | data/<task_name>/temp/ | 中间产物 |

禁止写入：

• ❌ 项目根目录（~/next-medgeai/MedgeClaw/）
• ❌ /workspace/outputs/（已废弃）
• ❌ /workspace/data/（只读，用户输入数据）

Important rules

• Never modify raw data files in /workspace/data/
• If the user's request is ambiguous, ask one clarifying question before dispatching
• If Claude Code returns an error about a missing package, retry with uv pip install [package] prepended to the command
• 涉及中文可视化时，在 prompt 中加入：绘图前先导入 skills/cjk-viz/scripts/setup_cjk_font.py 执行字体检测，不要硬编码字体名