code_flow_graph

Code Flow Graph

Generate interactive node-graph HTML diagrams that visualize code structure and call relationships.

Output

Two files in /docs/code_graph/ (or user-specified / config.json path):

1. code_flow_graph.html — rendering engine (copy verbatim from example/code_flow_graph.html in THIS skill's directory)
2. code_flow_graph_data.js — diagram data (you generate this; use templates/diagram_skeleton.js as starting point)

Workflow

This is a guide, not a script. Adapt steps based on project size, user needs, and context. Skip or merge steps when appropriate.

1. Analyze — Read the project structure; identify architecture type, entry points, key modules. See references/analysis_guide.md for inclusion/exclusion criteria.
2. Generate Overview — Produce a module-level dependency diagram immediately as the first deliverable. One node per module/class — NOT per function.
3. Offer Options — Present deep-dive choices via ask_followup_question with dynamic options based on discovered entry points. Always include "自定义：分析其他函数或模块" as last option.
4. Deep-Dive — Generate selected call-chain / UI event / data type diagrams. See references/deep_dive_types.md.
5. Validate — Run scripts/validate_data.js on the generated file; fix any errors before delivering.
6. Iterate — Append more diagrams on demand without regenerating existing content; update session state.

Config

Check config.json in this skill's directory for user preferences. If missing, use defaults:

• Output: /docs/code_graph/
• Language: Chinese for descriptions, English for code identifiers
• Scope: entire project unless user specifies otherwise

Session Tracking

On first run for a project, create state/code_flow_graph_session.json:

{
  "project": "",
  "created_at": "",
  "updated_at": "",
  "analyzed_entries": ["func1", "func2"],
  "generated_diagrams": ["overview", "func1_callchain"],
  "pending_items": ["func3_callchain", "ui_events"],
  "last_config_hash": ""
}

On subsequent runs, read this file to resume context: skip re-analysis of completed entries, continue from pending_items.

Data Format & Templates

• Full spec: references/data_format.md
• Composable patterns: templates/node_patterns.js (6 node types + connections + groups)
• File skeleton: templates/diagram_skeleton.js (copy as starting point)
• Deep-dive types: references/deep_dive_types.md
• Analysis strategy: references/analysis_guide.md

Locating the HTML Template

The template is at example/code_flow_graph.html relative to THIS skill's SKILL.md file. Use the absolute path based on the skill's installation directory. Do NOT resolve it relative to the user's project root.

Gotchas

1. HTML Template Source Path

Why: The HTML engine lives inside the skill directory, not the user's project. Resolving relative to the project root yields "file not found."

// ❌ Wrong — resolves against user's project (file won't exist there)
fs.copyFileSync('example/code_flow_graph.html', outputDir);

// ✅ Right — resolve against THIS skill's installation directory
fs.copyFileSync(path.join(SKILL_DIR, 'example/code_flow_graph.html'), outputDir);

2. Connection Anchors Must Use offsetTop, Not getBoundingClientRect

Why: The canvas uses CSS transform: scale() for zoom. getBoundingClientRect() returns screen-space coordinates that include the scale factor, causing connections to misalign at any zoom level other than 1x.

// ❌ Wrong — misaligns at zoom ≠ 1x
var relY = (el.getBoundingClientRect().top - nodeEl.getBoundingClientRect().top) + el.offsetHeight / 2;

// ✅ Right — layout-space coordinates, zoom-independent
var relY = el.offsetHeight / 2;
var cur = el;
while (cur && cur !== nodeEl) { relY += cur.offsetTop; cur = cur.offsetParent; }

3. Sync Transform Before Measuring

Why: After switchDiagram() resets scale = 1; panX = 0; panY = 0;, measurements in redrawConnections() read stale CSS transform values unless applyTransform() runs first.

// ❌ Wrong — measurements happen against stale transform
switchDiagram(key);
requestAnimationFrame(() => redrawConnections());

// ✅ Right — apply transform before requesting redraw
switchDiagram(key);
applyTransform();
requestAnimationFrame(() => redrawConnections());

4. callChain[].id Must Match Attr IDs Exactly

Why: The detail panel uses callChain item IDs to highlight the corresponding node/attr on click. A mismatch silently breaks click-to-highlight with no error.

// ❌ Wrong — inconsistent ID format
{ id: 'method_name', name: 'method_name()', ... }  // Missing NodeId prefix

// ✅ Right — format is always NodeId.method_name
{ id: 'ClassName.method_name', name: 'method_name()', ... }
// Must exactly match an existing attr's id in the same diagram

5. Data File Must Be Valid JavaScript

Why: The file is loaded via ``. A single syntax error (missing comma, unescaped quote, unbalanced brace) silently breaks the entire viewer with no user-visible error.

// ❌ Wrong — missing comma between array items, unescaped string
attrs: [
  { id: 'A.foo', name: 'foo()' }   // ← missing comma
  { id: 'A.bar', name: "it's broken" }  // ← unescaped quote
]

// ✅ Right — valid JS, properly escaped
attrs: [
  { id: 'A.foo', name: 'foo()' },
  { id: 'A.bar', name: "it\\'s fixed" },
]

Validation: Always run node scripts/validate_data.js before delivering.

6. Large Codebases: Raise Thresholds

Why: Default thresholds (fan-out >= 2, fan-in >= 3) produce unreadably dense diagrams on large codebases (> 500 lines per diagram page).

// ❌ Wrong — default thresholds on a large project → 80+ nodes
fan_out_threshold: 2, fan_in_threshold: 3

// ✅ Right — raised thresholds for large projects
fan_out_threshold: 3, fan_in_threshold: 4
// Also: collapse deeper calls into `children`, summarize repetitive patterns

7. Overview Must Stay Module-Level

Why: The Overview is for architecture comprehension. Per-function detail belongs in deep-dive diagrams. Mixing levels makes the Overview unreadable.

// ❌ Wrong — Overview with individual functions as nodes
DIAGRAMS.overview.NODES = [
  { id: 'parse_args', type: 'function', ... },
  { id: 'validate_input', type: 'function', ... },
];

// ✅ Right — one node per module/class in Overview
DIAGRAMS.overview.NODES = [
  { id: 'cli', label: 'CLI Module', type: 'module', ... },
  { id: 'core', label: 'Core Engine', type: 'class', ... },
];

8. Don't Modify the HTML Engine

Why: The HTML file contains a tightly-coupled rendering engine with search, tooltips, panels, persistence, and undo. Any modification risks breaking these interconnected features.

// ❌ Wrong — editing the HTML to add a feature
// Adding custom CSS, modifying event handlers, changing DOM structure

// ✅ Right — copy verbatim, all customization goes in the data file
// The engine supports: search, tooltips, callChain panel, fieldDetail panel,
// drag persistence, undo, groups, themes — all driven by data alone.

9. Grid-Aligned Layout Required

Why: Scattered or overlapping node placement makes diagrams unreadable. The engine expects consistent column-based positioning.

// ❌ Wrong — random positions, overlapping nodes
{ id: 'A', x: 47, y: 133, w: 280, ... },
{ id: 'B', x: 52, y: 180, w: 280, ... },  // overlaps A

// ✅ Right — grid-aligned columns (30, 350, 670, 990...) with 40px vertical gaps
{ id: 'A', x: 30, y: 60, w: 280, ... },
{ id: 'B', x: 30, y: 300, w: 280, ... },  // same column, proper gap
{ id: 'C', x: 350, y: 60, w: 280, ... },  // next column

10. Incremental Append — Never Regenerate

Why: Users accumulate diagrams across multiple requests. Regenerating the file loses all previous deep-dive diagrams, breaking the append-only contract.

// ❌ Wrong — overwriting the entire data file
var DIAGRAMS = {};  // Wipes all existing diagrams
DIAGRAMS.new_chain = { ... };

// ✅ Right — read existing file, append new diagram entry, write back
// 1. Read existing code_flow_graph_data.js
// 2. Add new DIAGRAMS.new_chain = { ... }; at the end (before UI_LAYOUT_VIEWS if any)
// 3. Write the complete file back

Engine Features (Built-in, No Data-Side Work Needed)

• Ctrl+K global search (fuzzy, cross-diagram)
• Draggable nodes with position persistence (localStorage)
• Bezier connections with gradient direction indication
• Click-to-highlight with related node dimming
• Collapsible children with connection redirect
• Signature tooltips (hover) and call chain detail panel (click)
• Field detail panel for data type nodes
• Click-to-copy function names
• Ctrl+Z undo (up to 50 steps)
• Catppuccin Mocha dark theme with type-based colors
• Group boxes with auto-sizing
• UI Layout widget-tree viewer (for UI_LAYOUT_VIEWS)