html-doc
Skill1 starsChanged 3 days agoMITView on GitHub
Uniform HTML documentation of the user's projects (docs/ folder, GitHub Pages) - up to 3 static self-contained pages sharing one style based on the R-Type architecture doc - an optional user guide for non-technical users, the technical documentation (sticky summary, numbered cards, tables, callouts, SVG diagrams) and an interactive 2D/3D project graph of the common languages (C / C++, Python, JS / TS / web, Java / Kotlin, C#, Go, Rust, Zig, Swift, Dart, PHP, Ruby, Lua, Julia, Elixir, Haskell, Objective-C, shell: files, classes, functions, calls, tests, build targets / packages, externals; filters by category/group/relation; update button fetching the GitHub repository) - with a light/dark sun/moon toggle. Use whenever creating, updating or restyling an HTML doc, unless the user explicitly asks for another style.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- secretj/secretj-claude-configtheir own version, first seen Oct 6, 2026
- tonywjs/html-doctheir own version, first seen Sep 15, 2026
- owainlewis/blueprinttheir own version, first seen Aug 15, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- owainlewis/blueprint★ 432same name
- tonywjs/html-doc★ 70same name
- secretj/secretj-claude-config★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository70 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- ask-at-endSkill
- clarifySkill
- continue-workSkill
- exportSkill
- full-permissionSkill
- ideationSkill
- interactive-questionsSkill
- onboardSkill
- pauseSkill
- presentationSkill
- rtkSkill
- setup-contextSkill
- write-like-meSkill
- audit-bugsSkill
- audit-depsSkill
- audit-qualitySkill
- audit-securitySkill
- benchmarkSkill
- cmake-styleSkill
- cpp-classSkill
- cpp-commentsSkill
- cpp-projectSkill
- cpp-styleSkill
- better-explain-docSkill
- better-html-docSkill
- better-html-styleSkill
- better-protocol-docSkill
- docs-updateSkill
- explain-docSkill
- html-styleSkill
- protocol-docSkill
- readme-styleSkill
- reportSkill
- wiki-styleSkill
- ci-cdSkill
- git-commitSkill
- git-conventionsSkill
- pre-commitSkill
- legal-checkSkill
- legal-setupSkill
- licenseSkill
- libutils-exceptionSkill
- libutils-installSkill
- libutils-setupSkill
- libutilsSkill
- python-classSkill
- python-commentsSkill
- python-projectSkill
- python-styleSkill
- auditSkill
- cppSkill
- devSkill
- docSkill
- gitSkill
- legalSkill
- pythonSkill
- setup-projectSkill
- styleSkill
- setup-studioSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 19
Keep the CSS of the templates as is (style of r-type/docs/architecture.html, defined in the html-style skill: tokens, components, rules, …
- Very specific: line 38
No scroll beyond the page (overscroll-behavior: none).
- Very specific: line 71
Before delivering: valid HTML, every tab/summary link points to an existing page/id, readable in both themes and at phone width (860px …
- Very specific: line 163
Every control has a tooltip (short description after 0.45 s of hover) and a help card on double click (longer description + an animated SVG …
- Distinctive: line 214
Check after generation: open it (or screenshot it) and look at the node/link counts printed by the script.
- Distinctive: line 217
Only when the project is a git repository (git rev-parse --is-inside-work-tree): ask (AskUserQuestion, French)
- Very specific: line 242
Add an empty .nojekyll at the root of the branch when missing (files served as they are).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Uniform HTML documentation
Up to 3 pages, each a single static self-contained .html (inline CSS/JS, no CDN, no web font,
no framework, works offline and from file://), in docs/:
| Page | File | When | Template |
| Guide (user documentation) | index.html | only if the tool is meant to be used by non-technical users | templates/user.html |
| Technical documentation | technical.html (index.html when there is no guide) | always | templates/technical.html |
| Project graph | graph.html | always (any supported language) | templates/graph.html + scripts/build_graph.mjs |
Keep the CSS of the templates as is (style of r-type/docs/architecture.html, defined in the html-style skill:
tokens, components, rules, check_style.py), unless the user explicitly asks for another style; a change of the
style is made in html-style/templates/style.css and base.js, then in these templates: technical.html and user.html hold both
word for word (what only they add sits between the Additions of this page comments of the <style>, and in the scripts after base.js: drawings of a small screen, repository
data, tabs of the missing pages), graph.html the same tokens and its own CSS. Only the content changes. Ask (AskUserQuestion) whether the guide page
is needed when the audience of the tool is not obvious.
Common to the 3 pages
- Style: the R-Type doc colors (CSS variables, never changed) merged with the layout of the context-forge site.
- Top bar: mobile menu button (guide / technical), logo + project name + version badge (
{{VERSION}}), page tabs (<nav class="tabs">, hidden on mobile where the menu replaces them), then the grouped actions.top-actions: GitHub icon ({{REPO_URL}}) | divider | language switch, theme switch (same height). - Logo: chosen by the user, always asked for a new site (repository logo when there is one, custom logo made for
the project, or the documentation logo), then applied to every page with
set_logo.py: see "Logo of the brand" inhtml-style. - Repository data fetched:
<html data-repo="{{GITHUB_REPO}}">(owner/repo); at load the page asks the GitHub API (cached 1 h) for the repository name, latest release (or tag), license and URL and replaces the written values in.repo-name(brand, footer, page title),.repo-version,.repo-license,a.repo-link. Offline / private / API limit: the written placeholders stay, so still fill them correctly. - No scroll beyond the page (
overscroll-behavior: none); on wide screens the content column is centered on the page, the contents list sits on its left. - The pages are independent: each declares its kind (
<html data-page="guide|technical|graph">). After creating or removing a page, runpython3 SKILL_DIR/scripts/sync_nav.py docs(shared withexplain-doc, inhtml-style): it rebuilds the tabs of every page from the pages present (Guide, Technical, Graph, then the explanation pages ofexplain-docin the same folder, which get the same tabs, brand, GitHub link and EN / FR switch: both sides are linked whichever is generated last), marks the current one and hides the tabs when there is a single page (build_graph.mjsruns it by itself). At load, each page also checks the other tabs (online:HEADrequest;file://: the page loaded as a hidden script) and removes the tab of a missing page, and the whole bar when a single page remains. File names: guide =index.html; technical =index.htmlwhen there is no guide, elsetechnical.html; graph =graph.html. Generate the technical page and the graph by default (both linked to each other). - Language switch EN | FR (pill at the top right, before the theme switch), English by default, the choice
is saved under
{{STORAGE_KEY}}-langand shared by the pages (also the theme, and the level of theexplain-docpages): the links between the pages carry it (?lang=fr&theme=dark), so it applies to every page even fromfile://where each page has its own storage:- the content is written in both languages: every text node of the page (summary links, titles, paragraphs,
table cells, captions,
{{TITLE}}/{{SUBTITLE}}) as a pair<span data-lang="en">...</span><span data-lang="fr">...</span>(or a block<div data-lang="en">+<div data-lang="fr">for long parts); only the active language is displayed; - UI strings carry
data-i18n(text) /data-i18n-title|aria-label|placeholder(attributes) and are translated bywindow.DOC_FR(English text -> French text) at the top of the page: add the new UI strings there; - scripts translate their generated texts with
window.docI18n.tr()and re-render ondocI18n.onChange()(the graph does it for its legends, status, details, hint and help:HELP/HELP_FR). Another language than EN/FR only on request (add a button and a dictionary).
- the content is written in both languages: every text node of the page (summary links, titles, paragraphs,
table cells, captions,
- Theme switch at the top right: a sliding toggle, blue track with a white sun (light), black track with a yellow
moon and stars (dark),
role="switch"+aria-checked; choice saved under{{STORAGE_KEY}}-theme, default = system preference. - Placeholders:
{{PROJECT}},{{VERSION}}(vX.Y.Z),{{REPO_URL}},{{STORAGE_KEY}}(project slug),{{TITLE}},{{SUBTITLE}},{{LICENSE}}(SPDX id, footer),{{GITHUB_REPO}}(owner/repo). - Page header:
.page-eyebrow(small accent uppercase label: "User guide", "Technical documentation", "Architecture"),h1,p.lead(muted introduction). - Footer
footer.site(guide / technical): "<project>documentation · generated from the repository (version, license)" in both languages; previous / next page cardsnav.pagerabove it (generated bysync_nav.py). - Before delivering: valid HTML, every tab/summary link points to an existing page/
id, readable in both themes and at phone width (860px breakpoint, no horizontal scroll):python3 ../html-style/scripts/check_style.py docs/*.htmlandcheck_render.py docs/*.html(section 7 ofhtml-style): 0 error, then look at it (--shots, the open menu and the panels of the graph too). - Responsive on every platform (phone, tablet, TV, touch laptop; rules in
html-style, already in the three templates - keep them when editing): viewport withviewport-fit=cover, safe areas, off-canvas menu closing on a tap / outside / Escape, no hover-only feature, 44px tap targets, 16px fields. The graph page works by touch: one finger drags / orbits / taps a node, two fingers pinch to zoom and pan (no mouse, Shift or wheel needed), the filter panel is a drawer (closed by its own button, a touch on the graph or Escape: on a phone it covers the whole screen), the details card a bottom sheet over the buttons of the graph. Check each page at 375, 768 and 1280px.
1. Guide (templates/user.html, optional)
For people who just want to use the tool: no jargon, no internal detail, task oriented.
- Sections: About, Installation, First steps, Usage (one sub-part per task, "I want to… / I do…" table), FAQ. Adapt the titles, keep 4-6 sections.
- Components: numbered steps
<ol class="steps">(each<li><strong>Action</strong> detail),<kbd>keys, commands to copy in<pre><code>, FAQ<details class="faq"><summary>, callouts for warnings/results,p.leadfor the intro sentence. - Say what the user sees when it works; one action per step.
2. Technical documentation (templates/technical.html)
- Contents list
<aside class="sidebar">on the left (sticky, off-canvas menu on mobile): title "Contents", one link per section of the page (same numbering, left border), the visible section highlighted. The pages are reached by the top tabs and the previous / next cards. h2/h3with anid(or in asection[id]) get a#anchor link on hover.- One
<section id="...">card per topic, preceded by<!-- N -->,<h2>N. Title</h2>, sub-parts<h3>N.M</h3>. Usual order: overview, split into libraries/modules, each module, network/protocol (the sections ofprotocol-doc: frames, types, codes, handshake sequence diagram, commands, storage), build. 5-9 sections. Components:
Need Component list of elements with attributes <div class="table-wrap"><table>with<thead>and ascopeon eachth(window.docTablesofbase.js: short cells on one line, it scrolls in its box when nothing else fits)remark / info <div class="callout">limit, risk, planned-not-done <div class="callout warn">validated / good practice <div class="callout ok">code, command, path, type <code>inline,<pre><code>block (a Copy button is added on hover)titled code block (file name, command role) <pre data-title="Install"><code>...</code></pre>alternatives (OS, package manager, language) content tabs <div class="ctabs" data-tabs="g"><button data-tab="a">+<div class="ctab-panel" data-tabs="g" data-tab="a">architecture, flow, tree, loop inline <svg>in<figure>+<figcaption>Figure N: ...(on a small screen a drawing keeps 72 % of its size, its labels stay readable, and scrolls in its box)- SVG arrows: computed with
data-from/data-to(seehtml-style;base.jswith its enginewindow.docDiagramis inlined whole in the templates: keep it in sync after a change ofbase.js), never typed coordinates; runpython3 ../html-style/scripts/check_style.py docs/*.html(0 error: arrows on their shapes, text inside). - SVG:
viewBox="0 0 760 H",role="img"+aria-label, only the template classes (box,box-alt,group,lbl,small,title,edge,edge-accent,dash,arrow,arrow-accent), markers once in<defs>. Draw the real mechanism, not decoration. Link to the graph page for the full dependency view. - Short factual paragraphs, present tense, explicit planned vs implemented; every identifier in
<code>; no emoji, no gradient, only the CSS variables.
3. Project graph (templates/graph.html)
Interactive graph drawn on a canvas by an embedded engine (no library); the panel starts with the eyebrow
"Architecture", and the categories / groups / relations are two-column legends (dot swatch for the nodes and the
groups, line swatch for the relations, count on the right, disabled entries faded: a readable grey text, a pale swatch, all / none buttons; a
long name wraps in its column, a very long one takes both columns, nothing is cut; hovering
an entry highlights its nodes / relations; clicking it shows / hides them, a relation toggle doesn't move the
layout): 2D (pan, zoom, drag nodes) and
3D (rotate, Shift+drag to move the target in the screen plane, drag a node to place it, zoom; spheres lit by a
light fixed in the scene, depth fog, floor grid following the camera, X/Y/Z gizmo ball resetting the angle on
click; arrow heads on the relations in both styles; bottom right switch Realistic (default) / Technical = the nodes of context-forge (3d-force-graph): 8 x 8 faceted sphere, Lambert light per pixel with
an ambient 0.8 + a directional 0.6 from the top, opacity 0.95) + cone arrows, no fog / floor,
remembered); key F centers and zooms the camera on the selected node, framed with its close neighbours (2D and 3D) and keeps the focus on it while you only zoom / dezoom (another move of the camera ends it); the 3D orbit pivots around the nearest nodes in front of the camera; 2D: the scroll zooms towards the mouse (or the selected node);
3D: free flight, the scroll moves the camera forward / backward towards the mouse (or the selected node), no limit;
keys (physical positions, labels read from the keyboard layout): arrows = move in the screen plane, ZQSD (WASD on
QWERTY) = move along the floor, A / E (Q / E) = height (zoom in 2D), Shift = faster; click on the axes = reset the angle, double
click on the axes or key R = reset the whole view (position, zoom, angle; R also in 2D);
limits: in 3D the camera stays in a box around the nodes (the floor is its bottom), whose walls appear as an accent
grid when the camera gets close; in 2D the view center stays near the graph and the zoom out is limited; the hover
card follows the view when it moves under a still mouse; hovering a node shows a card (kind in the node colour mixed with the
text colour, readable in both themes, with the file extension, name, path); a name is drawn on the canvas and on the boxes of the diagrams up to 48 characters (the
whole name is in the hover card and the details); slightly transparent nodes in 3D, a near plane (nodes the camera passes are
clipped like behind a WebGL camera, the camera can fly through the graph); no hint text: a "?" button (bottom left, key ? or H) opens a card listing the
controls of the current mode, each with a small drawing of the gesture (mouse / wheel / keys) and its description, search, presets (overview, files & includes, classes & inheritance,
classes & methods, tests, build, all), toggles per category (header, source, test file, class, struct,
interface, abstract, enum, function, method, test, test helper, executable, library, external), per group
(module folder) and per relation (includes, defines, member, inherits, implements, compiled-into, links,
tests, calls: read in the bodies of the functions / methods, resolved by name, by the type of the local variables
and of the class fields, make_unique<T>; the GoogleTest macros are not functions; main() is a node marked entry), colour by category or group, cluster by group, labels, isolate the selection, details panel with
clickable in/out relations, and an "Animate the direction of the relations" option (off by default) moving dots along every
relation from the source to the target.
Diagram mode (third button after 2D / 3D, bar at the top left of the graph): readable diagrams built from the
relations, drawn as boxes with the 2D camera:
- Execution: the calls from an entry point (list:
main()first, then the uncalled functions, the biggest first), one level per call depth, the methods of a class grouped in a dashed box, dashed arrows back to an earlier level; double click a box to start from it; - Inheritance: the class tree of the sources (bases on top, UML hollow triangles), wide levels wrap;
- Flame graph: the static call tree (width = size of the subtree, not a measured time), double click a frame
to zoom into it, the ancestor rows below to go back.
The boxes can be moved (a box, a dashed class group, or several selected boxes: Ctrl / Shift + click, Shift /
Ctrl + drag on the background for an area).
In this mode the filters of the panel (views, categories, groups, relations, graph options, "Color by": the boxes
use the category colours) are greyed (
aria-disabled); the search stays active (matching boxes outlined, Enter centers the first one or starts the diagram from it); on a narrow stage the details card starts under the bar of the views. Every control has a tooltip (short description after 0.45 s of hover) and a help card on double click (longer description + an animated SVG example, the names of its relations placed bydiagram()where nothing is), defined in theHELP/EXtables of the page: when a category, relation or option is added, add its entry there too.
Generation (never fill the snapshot by hand):
node SKILL_DIR/scripts/build_graph.mjs --repo <project> [--out <project>/docs/graph.html] [--ref HEAD] \
[--github owner/repo] [--branch main] [--project name] [--version vX.Y.Z]--repo: the root of a git repository or one of its folders (then only that folder is read). Reads the sources from the git commit (--ref), extracts the graph with theextractGraph()of the page (single source of truth, between the// <extractor>markers), embeds the snapshot (commit hash + date) and the config (GitHub repo fromorigin, branch, version fromproject(... VERSION)).- Re-running it regenerates the page from the template (new features included);
--keep-htmlkeeps the HTML/CSS of an existing page and only replaces the script, the config and the snapshot. - The page shows the embedded snapshot, then once at load (and on the "Update" button) asks the
GitHub API for the last commit of the branch: if it is newer, it downloads the sources
(
raw.githubusercontent.com) and rebuilds the graph in the browser (cached inlocalStorage). Offline, API limit (60 requests/h), private repository or local commits not pushed: the snapshot is kept and the status says why. The extractor (regex parsers, no AST) reads:
Stack Files What is extracted C / C++ .hpp .h .cpp .c...,CMakeLists.txtincludes, namespaces, classes + bases, methods, functions, bodies, CMake targets ( add_subdirectory,foreach,set(SRC ...)), linked librariesPython .py,pyproject.toml,setup.py,requirements*.txtimports (relative / package / src/), classes + bases (ABC / Protocol), methods, functions,if __name__ == "__main__"JS / TS / web .js .ts .jsx .tsx .vue .svelte,.html,.css .scss,package.jsonimport / require / dynamic import ( @/alias), classes / interfaces (extends / implements), functions / arrow functions, methods; pages (script / link / a) and stylesheets (@import, partials)Java / Kotlin .java .kt .scala,pom.xml,build.gradleimports (packages), classes / interfaces / records / objects, methods, mainC# .cs,*.csprojusing(namespaces of the project), classes / interfaces / records, methods,MainGo .go,go.modimports (packages of the module), structs / interfaces, functions, methods (receivers, even in another file) Rust .rs,Cargo.toml(workspaces: one package per crate)mod/use crate::/self::/super::/ sibling modules, structs / enums / traits (+ their methods),impl(trait for type = base, std traits ignored), functions, methods (even in another file),Self::f()calls, raw stringsr#"..."#skipped,#[test]PHP .php,composer.jsonuse(PSR-4 paths), require / include, classes / interfaces / traits, methodsRuby .rb,Gemfile,*.gemspecrequire / require_relative, classes ( <base), modules, methods (def ... end)Zig .zig,build.zig.zon@import,const X = struct / enum / union, functions, methods,test "..."blocksSwift .swift,Package.swiftimports (modules), classes / structs / protocols / actors, extension(methods + conformances), functionsDart / Flutter .dart,pubspec.yamlpackage:/ relative imports, classes (extends / with / implements), mixins, extensions, functionsLua .lua,*.rockspecrequire, function M.f/Class:method(OOP tables,class()/:extend()/:subclass()),... endblocksJulia .jl,Project.tomlinclude / using, modules, structs <:base, abstract types, functions (long and short form)Elixir .ex .exs,mix.exsalias / import / use, defmodule/defprotocol/defimpl,def/defp,Module.fn()callsHaskell .hs,*.cabalimports (modules), data / newtype / class / instance, top-level functions, calls without parentheses Objective-C .m .mm .h,*.podspec#import,@interface/@protocol/@implementation, methods, message sends[obj msg]Shell .sh .bash .zshsource/., functions, commands calling the functionsC / C++ (Meson) meson.buildexecutable()/library(), their sources (lists andfiles()),dependency()Common to all: test files (
tests/,__tests__/,*_test.*,*.test.*,*Test.java...) and test cases (GoogleTest, pytest, Jest / Vitest, Go, Rust, JUnit, xUnit / NUnit, PHPUnit, RSpec / Minitest), the classes / functions of the tests as test helpers, the calls (by name,this/self/ receiver, the type of the local variables and fields, constructions), one package per manifest folder (its files compiled into it, its dependencies as externals, an import of a package of the repository points to it, a package with an entry point is an executable), the externals deduplicated by name, the groups from the folders (generic ones skipped:src,lib,main,java,com...). Ignored folders:node_modules,dist,build,target,venv,vendor,__pycache__,.next,coverage,bin,obj,docs... (extractGraph.wanted(), shared by the script and the update of the page). Another language: add its regexes inextractGraph()of the template and keep the rest.- Check after generation: open it (or screenshot it) and look at the node/link counts printed by the script.
4. Publication (GitHub Pages)
Only when the project is a git repository (git rev-parse --is-inside-work-tree): ask (AskUserQuestion, French)
"Publier la documentation sur la branche gh-pages ?" - Oui, commit + push · Oui, commit local seulement ·
Non (garder docs/ seulement). If yes:
- The Pages URL must open the documentation directly:
https://<owner>.github.io/<repo>/serves theindex.htmlat the root ofgh-pages; pages left only in a sub-folder are never reached by default (404 or the mirror listing). Look at the root of the branch first (git ls-tree --name-only gh-pages), then ask (AskUserQuestion, French) "Où placer la documentation surgh-pages?":Dans docs/ + index à la racine qui redirige (Recommandé): pages indocs/, plus a rootindex.htmlredirecting todocs/(URL of the doc:.../<repo>/docs/, the root link works too);À la racine: pages copied at the root of the branch (URL.../<repo>/); not possible when another rootindex.html/technical.html/graph.htmlexists that isn't the doc: say so and propose the first choice;Toujours dans docs/ (sans redirection): pages indocs/only, the root link does not lead to them (give the.../<repo>/docs/URL and add it to the README). Remember the answer for the project: next publications reuse the same layout without asking again (the layout already on the branch tells it:docs/+ redirect, root pages, ordocs/alone).
- Root redirect (
index.html, only when the root has no otherindex.html; an existing one that isn't a redirect is never overwritten: ask):<!DOCTYPE html> <html lang="en"><head><meta charset="utf-8"><title>{{PROJECT}} documentation</title> <meta http-equiv="refresh" content="0; url=docs/"> <link rel="canonical" href="docs/"> <script>location.replace("docs/" + location.search + location.hash);</script> </head><body><a href="docs/">{{PROJECT}} documentation</a></body></html> - Add an empty
.nojekyllat the root of the branch when missing (files served as they are). - Work in a temporary worktree of
gh-pages(git worktree add <tmp> gh-pages, or--orphan gh-pageswhen it doesn't exist), copy the pages ofdocs/to the chosen place, never delete the other files of the branch (packages mirror,.repo, workflows of libutils/context-forge live there), commit withgit-conventions(docs(pages): update the documentation (<short sha of main>)), push only with the "commit + push" answer, then remove the worktree. - GitHub Pages not enabled yet (
gh api repos/<o>/<r>/pages-> 404): proposegh api -X POST repos/<o>/<r>/pages -f "source[branch]=gh-pages" -f "source[path]=/"(always the root of the branch, the layout above decides where the doc sits). Give the URL that opens the doc (https://<owner>.github.io/<repo>/, or.../<repo>/docs/without redirect) and check it once deployed (gh api repos/<o>/<r>/pages/builds/latest, then aHEADrequest on the URL).
5. Graph explorer (any public repository)
scripts/build_explorer.mjs --out graph.html builds the graph page without embedded project: at the first visit it
asks for a public GitHub or GitLab repository (owner/repo, github.com/...(/tree/<branch>),
gitlab.com/...(/-/tree/<branch>), a self-hosted GitLab, or ?repo= in the address), reads its sources from the
browser (GitHub: 3 API requests per load, raw files from raw.githubusercontent.com; GitLab: API v4) and builds the
graph with the same extractor. The link stays editable in the panel (Open / Update), is remembered by the browser
with the graph of each repository until Ctrl + Shift + R (forgets them and reloads the page, which asks again). It is published on the
gh-pages branch of the ai-utils repository (graph.html + index.html redirecting to it):
https://tsukini22.github.io/ai-utils/ - regenerate it after a change of the template.
- Big graphs (thousands of nodes): Barnes-Hut repulsion above 300 nodes (typed arrays; the exact loop below, same look), the simulation in a Web Worker above 3000 nodes (the page stays fluid while it settles), cached sphere sprites (rebuilt per ~1.5° of camera angle, fog baked in), and above 2000 nodes / 2500 relations: relations batched as hairlines (off-screen and < 3 px skipped, readable arrow heads only), tiny nodes batched per colour, 150 names at most; while the view or the layout moves, a light drawing (square nodes, part of the relations, no heads / names) and the full drawing as soon as it is still. Smaller graphs keep the exact drawing.
- Fetch of the sources: GitHub raw files 32 at a time; GitLab file list pages in parallel and the contents by batches of 90 files through GraphQL (~12 requests for 1000 files; the archive endpoint refuses cross-origin pages), the limits of the APIs waited and retried.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

