| name | sync-docs |
| description | Update the docs under docs/ (and the site nav) so they match the current app — edit the right pages, take missing screenshots, lint the set for faults that live between pages, sweep every Markdown file in the repo and the plan tree next door off the disk, then regenerate the SEO/AIO/LLM discovery files. Run before a release or whenever app behavior changes. Use when the user says "sync the docs", "update the docs", "lint the docs", "update the sitemap/llms.txt", or "make the docs match the code". |
| argument-hint | [topic | since-ref] |
| user-invocable | true |
Sync Docs
Keep the user-facing documentation in docs/ truthful to the app. This is a docs-only task — Markdown, the pictures the pages ask for, and the generated discovery files; nothing in src/. Never run git: releasing is /git-release's.
The docs are served at leaftext.com/docs by the static SPA in docs/ (index.html + docs.js + docs.css). Each page is a plain .md file; docs.js draws it with the app's own renderer, fetched as a module (site/leaftext-core.js) and routes by #/<path> (a route is the file path under docs/ without .md).
When to run, and on what
Before cutting a release, so the published docs describe what shipped; and any time app behavior changes — a new feature, a changed shortcut, a renamed setting, a new theme, a new dependency, a changed install or release flow.
The argument is optional and is either of two things: a topic (sync-docs themes), which jumps straight to the pages step 2 maps it to, or a git ref (v0.1.200) to diff against. With neither, read the working tree plus recent commits.
Process
1. Find what changed
If a topic was given, skip to step 2 for that area. Otherwise determine the changed app surface (read-only):
git diff --name-only HEAD~5..HEAD
git status --porcelain
Look at the actual source for anything relevant — the docs must describe real behavior, not guesses.
Run this pass even if you edited docs while building the feature. Editing the one section you were thinking about is not the same as auditing every page the change maps to. "I already touched the docs, so there's no drift" is exactly how stale rows survive — do the diff and the per-page audit anyway.
2. Map changes to doc pages
Pages carry an ordering prefix (docs/01-features/03-library.md); the prefix is stripped from the route, so that page is #/01-features/03-library in the SPA and 03-library.md from a sibling.
| App area / source | Doc page(s) to update |
|---|
Render pipeline, GFM/CommonMark coverage, syntax languages, emoji, alerts, math, footnotes, local images, sanitizer allowlist (src/markdown/, src/xml.rs, src/tei.rs, src/data.rs, src/eml.rs) | docs/01-features/01-rendering.md |
A new readable format — one arm in src/format.rs, plus its extensions | docs/01-features/01-rendering.md, docs/01-features/03-library.md#file-types, docs/02-installation.md#file-associations, and the format lists in README.md |
Tabs, document/scroll history, ScrollAnchor, live reload, recent files, shortcuts (src/app/workspace.rs, src/app/history.rs, src/app/watch.rs, src/assets/shell/theme.js) | docs/01-features/02-navigation.md |
Library pane, vaults, the in-memory corpus, search, the graph, GitHub sync (src/folder_tree.rs, src/vault_corpus.rs, src/doc_graph.rs, src/store/vaults.rs, src/git.rs, src/app/vaults.rs, src/app/vault_git.rs) | docs/01-features/03-library.md |
Minimap model and behavior (src/minimap.rs, src/assets/shell/minimap.js) | docs/01-features/04-minimap.md |
Themes, theme sources, token contract (LEAF_SEMANTIC_TOKEN_CONTRACT, reading_mode_css, themes/*.md) | docs/01-features/06-themes.md and docs/02-development/04-theming.md |
Anything under design/ — a color, a token, an icon, a component | nothing to edit by hand. is generated from those files by , and fails if it has drifted. Edit , run the bundler |
Update every page a change touches. A renamed setting, for example, may appear in both 05-settings.md and 03-library.md.
The README is a doc page for this purpose. It carries its own copy of the feature tour, the network/privacy claims, the install steps and the format lists, so a change to any of those lands there as well as in docs/. Claims about what reaches the network, what is written to disk, and when an update installs itself are the ones worth re-deriving from the source rather than trusting: they are the promises a reader is most entitled to, and the easiest to leave behind.
3. Edit the page(s)
Sweep each touched page for stale enumerations — don't just append. A page usually carries one or more enumerations that mirror the code: a Summary/overview table, a keyboard-shortcut list, an IPC-command or settings table, a feature matrix. Adding a new section below does not fix a row that is now wrong or missing in a table above — that drift is silent and is the most common miss. For every touched page, find each such table/list and re-derive it from the source, not from memory. Known enumerations to re-check whenever the relevant area changes:
docs/01-features/03-library.md — the Summary table (one row per library capability), the File actions table (one row per right-click menu item in src/assets/shell/context-menu.js), and the Facts table.docs/02-development/01-architecture.md — the IPC command table (one row per IpcCommand variant in src/app/events.rs, grouped where the code groups them) and the source-file roles list.docs/01-features/02-navigation.md and docs/03-quickstart.md — the keyboard-shortcut lists (the handlers are in src/assets/shell/theme.js, code-view.js, selection-toolbar.js and navigation.js).docs/01-features/05-settings.md — the settings table (one row per field in Settings in src/lib.rs).docs/01-features/07-editing.md — the Summary table and the block-kind list the insert row offers (MARKDOWN_INSERTS in block-controls.js).docs/01-features/06-themes.md and README.md — the family list and count (themes/*.md, one file per family).docs/02-development/07-workflow.md — the skill table (one row per folder under .agents/skills/) and the where a test goes table (the same rows /sync-tests holds).docs/02-development/02-building.md — the Individual tasks table and the verify recipe (one row per check-* in the Justfile).
A useful check: enumerate the source (e.g. the IpcCommand variants, the menu items, the settings fields) and confirm the doc table has exactly those rows — no extras, none missing.
Link every concept that has a home, and deep-link to the section rather than the page. When prose names a feature, view, setting, theme or concept that is documented, make it a link, and aim it at page#section-slug where a matching section exists. That covers links being written now and page-top links already there: Data paths → See [Settings](features/settings.md) should be features/settings.md#paths, and the [library](library.md) pane stays current (about live updates) should be library.md#live-updates. A slug is the heading text lowercased, spaces to hyphens, punctuation dropped ("Mermaid diagrams" → mermaid-diagrams). From the README that is a relative docs/<route>.md#<anchor>; between doc pages a relative <page>.md#<anchor>, which docs.js intercepts. Don't over-link: the first, most relevant mention in a passage, never every repetition, and never a word linked to the page it sits on. Page-level links stay page-level where they genuinely mean the whole page — "Next" lists, overview tables, and the deliberate relative-link demo.
- A paragraph is one line, in every file the sweep in step 7 reaches and not only the page being edited — a wrapped file is fixed where it is found. A break doing real work keeps two trailing spaces, or the file carries
<!-- keep-wrapping --> on a line of its own.
- A single
# Title H1, followed by a one-line > tagline blockquote, then the intro paragraph.
- Plain, factual prose. No marketing fluff, no changelog entries ("now supports…"). State current behavior.
- Keep version numbers and counts (e.g. "last 8 files", "4 parse workers", "2 MB limit", current
Cargo.toml version) matching the code.
One page is generated and must not be hand-edited: docs/02-development/05-design-system.md. Its every count is read out of design/, so an edit here is lost on the next run and just check-design-docs fails first. To change what it says, change design/ (see /design-tokens) and run just bundle-design-docs.
Renderer constraints — the docs are drawn by the app's own renderer, which supports a GFM subset. Use only:
- Headings, paragraphs, bold/italic, lists (nested), tables, blockquotes,
inline code, fenced code blocks, links, images, task lists, footnotes, emoji shortcodes, math ($…$, $$…$$), Mermaid fences.
- GitHub alerts via blockquote markers:
> [!NOTE], > [!TIP], > [!IMPORTANT], > [!WARNING], > [!CAUTION].
Do NOT use Mintlify/MDX components (<Tabs>, <Tab>, <Steps>, <Step>, <Card>, <CardGroup>, <Accordion>, <Note>, <Tip>, <Warning>) or theme={null} on code fences. Convert those concepts to plain Markdown: tabs/steps/accordions → ### subheadings or **1. …** numbered bold lines; cards → a bullet list of links; callouts → the > [!TYPE] alerts above.
Cross-page links must be relative .md paths, written with the prefixed file name, so they work both on GitHub and in the SPA: from docs/02-installation.md link 03-quickstart.md; from docs/01-features/02-navigation.md link a sibling as 06-themes.md and a top-level page as ../02-installation.md. docs.js intercepts these and turns them into #/route navigations, stripping the prefixes.
4. If you add or remove a page
The nav builds itself. docs.js derives the sidebar, the mobile dropdown and the pager from the folder listing (site/docs-nav.js), so there is no table to edit — the NN- prefix on the file name is what sets its place in the order, and a folder's README.md is that folder's index. Two places to keep in sync:
- The
.md file under docs/ (create or delete), named with the prefix that puts it where you want it. Renumber its siblings if it has to land between two of them.
- The Documentation list in the root
README.md — keep the docs/<path>.md relative links current.
5. Take the pictures the pages ask for
A page that names a screenshot nobody took renders a broken frame at leaftext.com. just check-doc-images is in just verify, so one written and left fails the build — this step is where it is taken rather than where it is caught, and a reference added without its picture stops the release rather than shipping.
node scripts/doc-images.mjs
node scripts/doc-images.mjs --missing
Take each missing one. Two commands, per picture — the app photographs itself, then the same encoder the diagram export uses writes the PNG:
pwsh scripts/capture-screenshot.ps1 -Doc <the document to open> -Out shot.bmp
just squeeze-png shot.bmp imgs/<name>.png --palette
- The file name is already chosen — it is whatever the page asks for. Write it there; never rename the reference to match a file you happened to make.
-Doc is what should be on screen, and a screenshot of a feature has to show that feature. Omit it for the home screen. -LibraryOpen opens the library pane; -Vault <folder> registers one, which is the only way the search box and the vault switcher exist at all; -Recents <files> fills the home screen's list; -Unlocked lifts the padlocks, which typing in the page or the source needs. -ThemeFamily and -ThemeMode pin the theme (leave them at Fern light unless the picture is about a theme, so the set looks like one app). -Width/-Height change the window, and the page lays out at the size the window was created with.-Do drives the window before the shot, so a picture can show a menu, a sheet, a hover or a selection rather than only a document. Steps are click:X,Y, rclick:X,Y, move:X,Y, drag:X1,Y1,X2,Y2, hold:… (a drag caught mid-gesture), scroll:X,Y,NOTCHES, type:text, key:{ESC}, wait:MS. A drag or a hold takes two more numbers — drag:X1,Y1,X2,Y2,MOVES,GAP — for how many mouse moves to walk in and how many milliseconds apart, which is how a picture catches a fast gesture; a hold written that way is photographed where the walk stops rather than after the usual settle. Coordinates are pixels in the captured image, so measure them off the last shot you took at the same size — take one plain shot first, look at it, then aim.-Crop "X,Y,W,H", same pixels. Detail shots ship cropped: a whole window around a 200 px control is a picture of the window.--palette cuts the image to 256 colors: it halves the file and is the only step that moves a pixel. Use it for every screenshot.- The shot runs against a throwaway profile under
-Work, never the owner's — their settings, recent files and vault registry are not read or written. Nothing here needs settings.json hand-edited.
- Anything the window cannot be made to show is not faked, and the reference cannot be left standing either. fails the build on one, so a picture nobody can take is a decision rather than a note in the hand-back: take it another way — a state staged through the page's own setter, a shot of another program taken by hand, both of which now carries rows for — or cut the reference and write the sentence the picture was going to say. Either way the call goes in the ticket.
A new reference is part of the same edit that adds it — writing  and moving on is what built the backlog this step exists to drain.
6. Lint the whole set, not just the pages you touched
Steps 2 and 3 keep a page truthful. This step keeps the set coherent — the faults that live between pages and that no single edit can see. Run it before a release, or any time a change spanned more than one page.
- Contradictions. Two pages describing the same behavior differently. The pairs worth checking are the ones that share a subject:
README.md against 01-introduction.md (the pitch and the feature list), 03-library.md against 05-settings.md (a setting named in both), 02-installation.md against 02-development/03-releasing.md (paths and versions), 02-navigation.md against 03-quickstart.md (the shortcut lists). Where they differ, the source settles it — not whichever page reads better.
- Stale claims. A sentence that was true at some version and quietly is not: a count, a limit, a "this is not supported yet", a named default. These survive because nothing links to them and nothing tests them. Re-derive the number.
- Orphans. A page nothing links to. The nav reaches every page by folder listing, so an orphan still renders — it just never gets found from the page a reader is already on. Give it an inbound link from the page whose subject leads to it.
- A concept with no home. A word the docs use across several pages as though it were defined somewhere, and it is not — no section, no glossary row. Either it gets a
docs/GLOSSARY.md row or the page that owns the subject gets a section, and the other mentions link there. This is what stops a term meaning something slightly different on each page. The same fault has a second home: a word the tickets, rankings and skills lean on with no row in ../docs/GLOSSARY.md drifts the same way, and step 7 is where that is swept.
- What was learned and never written down. If this session derived something real about how the app behaves — a limit, an order of operations, a reason a thing works the way it does — and no page says it, that belongs on a page now. Otherwise the next person re-derives it. This is the most common way the docs fall behind while every individual edit was correct.
Say what the lint found in the hand-back, including anything left unfixed and why.
7. Every Markdown file, not just the published pages
The page map in step 2 covers what the app publishes. It is not the whole set of writing this session can have made untrue, and a map written by hand is exactly what misses a folder somebody added last week. Read the set off the disk:
just check-docs
node scripts/check-docs.mjs --list
The list is generated by walking the folders, so a page added anywhere shows up without anyone editing anything. Work down it and ask one question of each file: did this session make a word of it untrue? Only files where the answer is yes get edited — this is a sweep, not a rewrite of 150 documents.
Every file in both trees is a guide, not a log — AGENTS.md says so for the whole repo, and this sweep is where it is enforced. A page, a design/ table, a skill: no changelog entry, no "what this pass found", no count of what was audited, no paragraph explaining why the file now says what it says. Cut one where the sweep finds it. What a build turned up belongs in its ticket. The one exception is a rule that was paid for in a version number, which is evidence the rule is real.
What each role means in practice:
| Role | What makes it untrue, and what to do |
|---|
published (docs/) | Behavior changed. Step 2's map says which page; step 3 says how to edit it |
source of a token / a color (design/, themes/) | Never edited here. A value changes in design/ and the bundlers regenerate — see /design-tokens |
a repeatable job (.agents/skills/) | The steps a skill names moved, or a rule it enforces changed. A skill describing a script or a path that no longer exists is worse than no skill. A change here is also published: Workflow is the account of these skills a contributor reads, so it is swept in the same pass |
the guide (AGENTS.md, and README.md) | A standing rule changed, or the repo grew a file the guide's tables do not reach. Never a log of what shipped |
a license notice (src/assets/*-MIT.md) | Only when what is vendored changes |
installer text (wix/) | Only when the install flow changes |
| any of them, always | A hard-wrapped paragraph, and the same in a comment in the code. just check-wrapping --fix joins them across both trees, whatever else the sweep found |
plan (../docs/features/, ../docs/refactor/, ../docs/fixes/, each grouped into subject folders) | The one this step exists for. A plan for something that now ships is the most misleading writing in either tree: move the file into the matching subject folder under ../docs/done/, and move its row in ../docs/README.md under Shipped saying what shipped. A plan half-built gets the built part struck through, not deleted |
A plan with every box ticked and still filed as live work fails just check-docs and names itself. The move above is remembered by two skills and was still missed once — v0.1.462 shipped scroll-position and left it in ../docs/fixes/, so the running order went on calling it next up. The check runs inside /check, so a release cannot go out past it.
A file with no role fails just check-docs and names itself. That is the backstop: add a new kind of document and the suite stops until this table says who keeps it true. Fixing it means adding the folder to ROLES in scripts/check-docs.mjs and a row above — never just the script.
Never link from docs/ into ../docs/: those pages are published and that folder is not. Links only go the other way. Two files are named GLOSSARY.md and they are not interchangeable — docs/GLOSSARY.md is published and defines app words, ../docs/GLOSSARY.md is not published and defines planning words. A row filed in the wrong one is published writing about tickets, or a ticket word nobody reading the app can find.
8. Regenerate the SEO / AIO / LLM discovery files
The files AI crawlers and search engines read are generated from the docs, not hand-maintained. After editing docs — and always after adding or removing a page — regenerate them:
node scripts/seo-gen.mjs
It rewrites five files at the repo root (the deployed site root) from README.md + the docs/ tree:
robots.txt — allows the major search + AI crawlers (Googlebot, Bingbot, GPTBot, OAI-SearchBot, ChatGPT-User, CCBot, PerplexityBot, ClaudeBot, Google-Extended) and points at the sitemap.sitemap.xml — every address a fetcher can actually ask for, each with a git-derived <lastmod>. A doc page's advertised address is its raw .md: a #/route never reaches the server, so advertising one is 18 addresses that all answer with the docs shell. The router still serves those routes for people who share them; they are simply not what a crawler is pointed at.sitemap-md.txt — one .md source URL per line.llms.txt — a concise index: page title → .md link.llms-full.txt — a fuller enumeration: title, page URL, Markdown URL, and a one-line description per page.
Page list, titles, summaries and dates are all derived from the current files, and the output is byte-identical for the same tree, so a no-op run leaves git untouched. (<lastmod> reads the last commit date per file; it refreshes on the next run after you commit.)
Forgetting this step is caught rather than shipped. check-site.mjs runs the generator in memory on every just verify and names any committed file that disagrees, along with the address it should gain or lose. Dates are not compared: a file's <lastmod> is its own last commit date, which the commit that changes it cannot know in advance.
9. Verify
-
just check-wrapping — no paragraph or comment broken across lines in either tree, the stylesheets included.
-
Grep the changed files for leftovers: no <Tabs, <Card, <Step, <Note, <Tip, <Warning, <Accordion, or theme={null}.
-
node scripts/doc-images.mjs — every picture a touched page asks for is there. just verify refuses one that is not, so this is where it is found rather than at the release.
-
Re-run node scripts/seo-gen.mjs and confirm it leaves the discovery files unchanged (a dirty tree here means step 8 was skipped or a doc changed after it ran).
-
Every enumeration on a touched page matches the source one-for-one — no stale, missing or extra rows.
-
Every internal link resolves to a real .md / route, each #anchor matches a real heading slug on the target page, and step 3's linking pass has actually been made over the touched pages and the README intro.
-
Optional but preferred: serve and click through.
python -m http.server 8000
Or smoke-test rendering without a browser:
just build-web
node docs/render-docs-check.mjs
10. Hand back — do NOT release
Leave the changes uncommitted. The whole reply is the owner's message, word for word; what the lint found and left open is a ticket, not a sentence. Publishing is a separate, explicit /git-release the owner asks for (site-only: no version bump).
Anything this pass finds that it is not here to do is a ticket, written before the hand-back — /ticket, its row in ../docs/README.md, /pm.
Anything the work in front of this pass hints could be faster is a performance finding: file it as a ticket in the same turn, without stopping for a benchmark or fixing it in passing, and never name it in the reply. A page or documentation pass that repeats work, reads more than its result needs or visibly waits has supplied enough of a hint; the new ticket measures and fixes it.
Reference
docs/docs.js — the shell, the routing and the link interception. Its nav comes from site/docs-nav.js, which reads the folder listing, so there is no page list in it.docs/index.html, docs/docs.css — the docs shell and chrome.site/leaftext-core.js — loads the app's own renderer, which defines what Markdown the docs may use.README.md — the Documentation section with relative docs/<route>.md links.- Building — the published account of how a picture gets taken.
/git-release — the separate skill that commits and pushes (site-only changes don't bump the version).
