tools-only Execution Flows and Tool Chains
This document maps how tools chain together in different execution scenarios.
High-Level Dependency Graph
User Request
│
├─► Intent Classification (System Layer)
│
├─► Base Chat Path ───────────────────────────────────────────────┐
│ ├─► web_search ──┐ │
│ ├─► web_open_url │ │
│ ├─► ipython ─────┼─► External Data/Computation │
│ ├─► memory_space─┤ │
│ └─► datasources ─┘ │
│ │
└─► OK Computer Path ─────────────────────────────────────────────┤
│ │
├─► Skill Detection ──┐ │
│ ├─► docx ─────────┼─► read_file(SKILL.md) │
│ ├─► xlsx ─────────┤ │
│ ├─► pdf ──────────┤ │
│ └─► webapp ───────┘ │
│ │
├─► Tool Selection │
│ ├─► read_file ◄─── Required first for skills │
│ ├─► todo_read ──── Session state check │
│ └─► todo_write ─── Task planning │
│ │
├─► Data Acquisition │
│ ├─► web_search │
│ ├─► get_data_source │
│ ├─► search_image_by_text │
│ └─► search_image_by_image │
│ │
├─► Browser Automation (if needed) │
│ ├─► browser_visit ◄─── Entry point │
│ ├─► browser_click │
│ ├─► browser_find │
│ ├─► browser_input │
│ ├─► browser_scroll_* │
│ ├─► browser_screenshot │
│ └─► browser_state │
│ │
├─► Content Generation │
│ ├─► ipython ◄─────── Primary computation │
│ ├─► write_file │
│ ├─► edit_file │
│ └─► shell ◄─────── Build orchestration │
│ │
├─► Media Generation │
│ ├─► generate_image │
│ ├─► generate_speech │
│ └─► generate_sound_effects │
│ │
├─► Asset Extraction (Web Replication) │
│ ├─► screenshot_web_full_page │
│ ├─► find_asset_bbox │
│ └─► crop_and_replicate_assets_in_image │
│ │
└─► Delivery │
├─► deploy_website │
├─► slides_generator │
└─► KIMI_REF tags (automatic) │
Skill-Specific Tool Chains
DOCX Skill Tool Chain
read_file(SKILL.md) ─► read_file(Example.cs)
│
▼
shell: ./docx init ──► Setup environment
│
▼
ipython: Generate Program.cs ──► Meta-programming
│
▼
shell: ./docx build ──► Compilation pipeline:
├─► dotnet build
├─► dotnet run
├─► python fix_element_order.py
├─► ./validator/Validator
├─► python validate_docx.py
└─► pandoc verification
│
▼
read_file(output.docx) ──► Verification
│
▼
KIMI_REF output.docx
XLSX Skill Tool Chain
read_file(SKILL.md)
│
▼
Per-Sheet Loop:
├─► ipython: openpyxl creation
├─► ipython: wb.save()
├─► shell: KimiXlsx recheck ──┐
├─► shell: reference-check ───┤ Validation gates
└─► Error? ──► Fix & retry ◄──┘
│
▼
[If PivotTable needed]:
├─► shell: KimiXlsx pivot (MUST be last)
│
▼
shell: KimiXlsx validate ──► Exit code 0 required
│
▼
KIMI_REF output.xlsx
PDF Skill Tool Chain (HTML Route)
read_file(SKILL.md) ─► read_file(routes/html.md)
│
▼
ipython: Generate HTML + CSS
├─► matplotlib charts (if needed)
└─► KaTeX math (if needed)
│
▼
write_file: /tmp/input.html
│
▼
shell: pdf.sh html input.html ──► Playwright + Paged.js
├─► Mermaid rendering (if present)
├─► Pagination stability detection
└─► PDF export (scale: 1.5)
│
▼
KIMI_REF output.pdf
PDF Skill Tool Chain (LaTeX Route)
read_file(SKILL.md) ─► read_file(routes/latex.md)
│
▼
ipython: Generate .tex source
├─► Document structure
├─► Math formulas
└─► Bibliography (.bib)
│
▼
write_file: /tmp/main.tex
│
▼
shell: compile_latex.py main.tex --runs 2
├─► tectonic run 1 (generate aux)
└─► tectonic run 2 (resolve references)
│
▼
KIMI_REF output.pdf
WebApp Skill Tool Chain
read_file(SKILL.md)
│
▼
shell: init-webapp.sh "Title"
├─► Copy 73-file template
└─► npm install (26,082 files)
│
▼
ipython: Component architecture planning
│
▼
write_file: src/components/*.tsx
├─► React components
├─► TypeScript types
└─► shadcn/ui usage
│
▼
shell: npm run build ──► Vite bundler:
├─► Tree-shaking
├─► Code splitting
├─► Asset optimization
└─► dist/ generation
│
▼
deploy_website(dist)
│
▼
Public URL returned
Tool Synergy Patterns
Pattern 1: Data Pipeline
get_data_source ──► ipython (pandas analysis) ──► write_file (CSV)
│ │
└─► Source citation └─► KIMI_REF
Pattern 2: Web Research
web_search ──► browser_visit ──► browser_scroll ──► browser_screenshot
│ │
└─► Initial discovery └─► Visual verification
Pattern 3: Asset Extraction
screenshot_web_full_page ──► find_asset_bbox ──► crop_and_replicate
│ │ │
└─► Full page capture └─► Detection └─► Extraction
Pattern 4: Media Production
generate_image ──► ipython (Pillow processing) ──► write_file
│ │
└─► Base image └─► Final asset
Pattern 5: Document Assembly
read_file (multiple) ──► ipython (analysis) ──► Skill-specific generation
│ │
└─► Source material └─► shell/ipython pipeline
│
▼
KIMI_REF
Key Insights
Skill Loading is Always First: Every specialized workflow starts with
read_file(SKILL.md)Validation Gates: XLSX skill enforces per-sheet validation; DOCX validates after build
Build Orchestration: Complex skills (DOCX, WebApp) use shell as build system, not just command executor
Route Selection: PDF skill branches early (HTML vs LaTeX) with different tool chains
Meta-Programming: DOCX skill uses IPython to generate C# that generates documents
External Toolchains: WebApp skill manages 26K+ node_modules files through npm/vite