Skip to content

About

Arma 3 SQF scripting reference for AI agents and tools: every command and function from the BI Community Wiki as Markdown, JSON and HTML, updated daily

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

a3ref — Arma 3 scripting reference for AI agents

Every Arma 3 scripting command and function, built first for AI agents, bots and programs.

A daily-updated mirror of every page in Category:Arma 3: Scripting Commands, Category:Arma 3: Functions and the concept pages they rely on (data types, Position, locality, scheduler, event handlers, Description.ext, ...), plus every other Arma 3 page of Category:Arma 3 (mission editing, configs, GUI controls, multiplayer and server, tools, file formats, reference lists, ...) from the Bohemia Interactive Community Wiki. The original wiki sits behind a bot challenge that blocks scripts and agents; this mirror serves the same content as Markdown and JSON at stable, predictable URLs, with no API key, no rate limit and open CORS. A readable HTML version is there for people too.

Site: https://wiki.arma3.io/ · Data: a3ref/data · MCP server: a3ref/mcp · Agent guide: https://wiki.arma3.io/llms.txt · Working on the code: AGENTS.md

The site moved from https://a3ref.github.io/ in November 2026; the old address redirects every path to the same path here.

For AI agents and programs

The easiest way for an agent is the MCP server: search, read pages and lint SQF (catches made-up commands, missing operands, unknown BIS_fnc_* functions, deprecated commands). In Claude Code: /plugin marketplace add a3ref/mcp, then /plugin install a3ref@a3ref.

Without MCP, everything is a plain GET:

You need Fetch
where to start llms.txt: how to use the mirror, topics, groups (a few thousand tokens)
one command https://wiki.arma3.io/commands/<name>.md or .json, e.g. commands/setPos.md
one function https://wiki.arma3.io/functions/<name>.md or .json, e.g. functions/BIS_fnc_spawnGroup.json
a concept https://wiki.arma3.io/topics/<name>.md, e.g. topics/Multiplayer_Scripting.md, topics/Event_Handlers.md
every name names.txt, one per line
one group groups/commands/<group>.md, groups/functions/<group>.md, listed in llms.txt
to search index.json (2 MB): every entry with syntax, summary, groups, locality, status, operator symbol, paths
one event handler events.json: every event of the object, mission, UI and Eden Editor event handler pages with its kind, summary, section heading and #anchor, e.g. AnimStateChanged → topics/Event_Handlers.html#animstatechanged
everything commands.json (7 MB), functions.json (4 MB), topics.json (27 MB): for programs; a model should read single pages instead
a linter's data syntax.json (0.5 MB): forms (nullary/unary/binary), status, replacements and syntax of every name
the record format schema.json (JSON Schema 2020-12)
what changed changes.md, changes.json, Atom feed.xml
version history the a3ref/data repository, one commit per day with changes

Paths use the name as written in scripts (setPos, diag_fps, BIS_fnc_spawnGroup), and lower-case paths work too (commands/setpos.md). Spaces become _; other characters outside A-Za-z0-9.- become ~XX per UTF-8 byte (the operator page a != b is commands/a_~21~3D_b.md; index.json also lists operators by symbol). A missing page returns a 404 that explains how to find the right path.

What the records add on top of the wiki text:

  • status (ok, deprecated, broken) and replaced_by, so agents avoid outdated commands;
  • the type of every parameter and return value normalised for programs (type_expr: Array<String>, Object | Group), whose names are the data type topics;
  • for commands, the syntax form (nullary, unary, binary) with typed left/right operands, and the operator symbol of operator pages;
  • for the ~400 placeholder function pages, the description, parameters, return value and examples parsed from the function's header comment in the game files (docblock, filled_from_docblock).

Markdown pages are the best fit for a model's context: description, every syntax with typed parameters and return value, locality, examples and community notes in one file. Links in Markdown point only to other mirrored pages (agents cannot open the wiki); HTML pages also link to the wiki.

Add it to your agent

Paste this into AGENTS.md, CLAUDE.md, .cursorrules or your system prompt:

## Arma 3 / SQF reference
When writing or reviewing SQF for Arma 3, check commands and BIS_fnc_* functions against
https://wiki.arma3.io/ instead of relying on memory:
- https://wiki.arma3.io/commands/<command>.md (e.g. commands/setPos.md)
- https://wiki.arma3.io/functions/<BIS_fnc_name>.md (e.g. functions/BIS_fnc_spawnGroup.md)
- https://wiki.arma3.io/topics/<topic>.md (e.g. topics/Multiplayer_Scripting.md)
- https://wiki.arma3.io/names.txt lists every name; https://wiki.arma3.io/llms.txt explains the rest
Check syntax, parameter types, return value, locality (local/global argument and effect),
status (deprecated commands name a replacement) and the Arma 3 version that introduced the command.

From code

curl -s https://wiki.arma3.io/commands/setPos.md
curl -s https://wiki.arma3.io/index.json | jq -r '.items[] | select(.name | test("^setPos")) | .syntax[]'
import json, urllib.request

index = json.load(urllib.request.urlopen("https://wiki.arma3.io/index.json"))
entry = next(e for e in index["items"] if e["name"].lower() == "createvehicle")
page = json.load(urllib.request.urlopen("https://wiki.arma3.io/" + entry["json"]))
for s in page["syntaxes"]:
    print(s["syntax"], "->", s["returns"]["type"] if s["returns"] else None)
const index = await (await fetch("https://wiki.arma3.io/index.json")).json();
const fns = index.items.filter((e) => e.type === "function" && e.groups.includes("Multiplayer"));

The same files are also served by raw.githubusercontent.com/a3ref/data/main/… and cdn.jsdelivr.net/gh/a3ref/data@main/….

How it works

The wiki sits behind a Cloudflare managed challenge, so plain HTTP clients get a 403. sync starts the locally installed Google Chrome through patchright, lets it pass the check, then calls the MediaWiki API (/wikidata/api.php) with fetch() from inside the page: 50 pages of wikitext per request, about 110 requests in total. Raw wikitext is cached in raw/*.json of the data clone. Later runs read the wiki's recent changes from the newest edit in the cache (minus a day of overlap) and download only the pages edited, created, moved or deleted since, deciding whether a new page belongs by its categories: a few requests, a few seconds. sync --walk lists every member page instead and downloads those whose revid changed; it also catches pages that join or leave a category through a template, and it is used automatically when the cache is empty or older than the wiki's 90 days of recent changes.

sync also keeps the wiki templates the mirrored pages transclude, recursively, in raw/templates.json: the walk asks the wiki for the templates of every page (50 pages per request, about 220 requests) and downloads those whose revid changed; later runs read the template namespace's recent changes in the same request as the pages', refetch the edited templates and add the new templates of the pages they download.

build then parses each page's {{RV}} template with mwparserfromhell into structured fields, renders the wikitext of each field and concept page with a Python port of the wiki's MediaWiki 1.39 parser (biwiki/mediawiki/) and those templates into the HTML the wiki shows, converts that to Markdown and plain text (biwiki/htmlmd.py), and writes the mirror to out/. Without raw/templates.json it stops and says how to get the file (sync --walk).

Deployment

Three repositories, no secrets:

  • a3ref/data holds the data: raw wikitext of the pages and their templates (raw/*.json) and the parsed commands.json, functions.json, topics.json, index.json, syntax.json. Its workflow runs twice a day, at 04:17 and 16:17 UTC: it checks out this repository for the code, runs sync (real Chrome inside xvfb) and build --dataset, and commits only if something changed. A third run, on Sundays at 10:17 UTC, uses sync --walk.
  • this repository holds only the code. .github/workflows/site.yml builds the site from the latest a3ref/data and deploys it to GitHub Pages:
    • on every push to main (parser or design changes);
    • every hour, but only if the data commit differs from data-version.txt on the live site, so the site follows every data sync, scheduled or run by hand, within the hour;
    • every 6 hours a watchdog run fails (and GitHub emails) if the data sync has not succeeded for 26 hours, since GitHub may drop scheduled runs;
    • manually (Actions → Deploy site → Run workflow).
  • a3ref/mcp is the MCP server and Claude Code plugin; it reads the live site.

Everything in out/ is generated during deployment and never committed.

If Cloudflare ever blocks the GitHub runner, sync locally (python -m biwiki sync && python -m biwiki --dataset data build, then commit and push in the data clone).

Parser problems

Every build writes report.json (also published at https://wiki.arma3.io/report.json) and never stops on a single broken page: the page is left out, reported, and the rest of the site is built.

After each deploy, python -m biwiki report --issue compares the report with known-problems.json, the accepted baseline. Problems are compared by their key (kind, page, detail); where it can be found, each one also says where it is: "md": "line:column" in the page's .md, "wiki": "line:column" in the wikitext, and the text around it.

  • new problems open (or update, with a comment) the issue labelled parser-report, listing each page with links to the site and the wiki; GitHub notifies watchers of the repository;
  • a page the parser crashed on additionally fails the workflow run, which sends the usual failure email;
  • when the new problems are gone, the issue is closed automatically;
  • the run page in Actions shows the same summary.

After fixing the parser or deciding a problem is acceptable, refresh the baseline locally and commit it:

python -m biwiki build
python -m biwiki report            # what is new compared to the baseline
python -m biwiki report --accept   # accept the current state as the baseline

Development

Requires Python 3.10+ and Google Chrome (for sync only). Agents: read AGENTS.md first; it lists the commands, where to change what, the rules and how to verify a change.

pip install -r requirements.txt
git clone https://github.com/a3ref/data data      # the data lives in its own repository; elsewhere: --raw/--dataset/--data

python -m unittest                 # tests on tests/fixtures (under a second, no data clone needed)
python -m biwiki build             # regenerate out/ from the cache only (about 20 s)
python -m biwiki report            # new parser problems compared with known-problems.json
python -m biwiki serve --port 8000 # serve out/ over HTTP
python -m biwiki update            # sync + build (opens a Chrome window for a few seconds)
python -m biwiki sync --walk       # list every member page instead of reading recent changes (about a minute)
python -m biwiki update --hidden   # same, with the Chrome window moved off-screen
python -m biwiki sync --full       # re-download everything, ignoring the cache
python -m biwiki --dataset data build  # also refresh the parsed JSON in the data repository

--only commands / --only functions limits the work to one category. Cloudflare rejects real headless Chrome; if a checkbox appears, run without --hidden and click it.

Output (out/)

File Purpose
index.html the site: home page with search, groups, topics, recent edits
commands/, functions/, topics/ one page per entry as <name>.md (YAML front matter + Markdown), .json and .html, plus lower-case aliases and an index.md/index.html by group; <name>.og.png is the page's 1200×630 preview image for link previews
commands.json, functions.json, topics.json every record of one kind in a single file
index.json lightweight index: name, type, syntaxes, summary, groups, locality, status, operator, aliases, paths
events.json every event handler on the event handler topic pages: name, kind, page, section, anchor, summary (site search and the MCP server)
names.txt every name, one per line, with paths where they do not follow the simple rule
groups/<kind>/<group>.md per-group lists with summaries
syntax.json forms, status, replacements and syntax lines of every name, for linters (used by the MCP server)
llms.txt, llms-full.txt short guide for language models, and the full list of every page
changes.md, changes.json, feed.xml wiki edits, from the git history of a3ref/data (written by python -m biwiki feed)
robots.txt, sitemap.xml, 404.html crawlers and search engines; the 404 page redirects people and explains paths to agents
schema.json JSON Schema of every file above
report.json parser health report: crashes, skipped pages, unknown templates, leftover markup, untyped parameters
assets/ stylesheet, script and search index of the HTML version

JSON record

A command or function record (format 2, "version": 2 in schema.json), shortened:

{
  "name": "createVehicle",            // as written in code (BIS_fnc_... for functions)
  "title": "createVehicle",           // wiki page title
  "type": "command",                  // or "function"
  "url": "https://community.bistudio.com/wiki/createVehicle",
  "revid": 377509, "updated": "2026-01-02T06:50:13Z",
  "groups": ["Object Manipulation"],
  "status": "ok",                     // or "deprecated" / "broken", then also "replaced_by": [...]
  "games": {"ofp": "1.34", "ofpe": "1.00", "arma1": "1.00", ..., "arma3": "0.50"},  // oldest first
  "since": "ofp 1.34",                // first game it appeared in, as the syntax and parameter "since"
  "arma3_since": "0.50",
  "locality": {"effect": "global"},   // argument locality unknown, no server execution needed
  "description": "markdown",
  "syntaxes": [{
    "syntax": "type createVehicle position",
    "params": [
      {"name": "type", "type": "String", "description": "vehicle/object className", "type_expr": "String"},
      {"name": "position", "type": "Position", "description": "desired placement position. ...", "type_expr": "Position"}
    ],
    "returns": {"type": "Object", "type_expr": "Object"},
    "form": "binary", "left": {"text": "type"}, "right": {"text": "position"}
  }, {
    "syntax": "createVehicle [type, position, markers, placement, special]",
    "params": [..., {"name": "markers", "type": "Array of Strings", "optional": true, "default": "`[]`",
                     "description": "if the markers array ...", "type_expr": "Array<String>",
                     "text": "markers: Array of Strings - (Optional, default `[]`) if the markers array ..."}, ...],
    "returns": {"type": "Object", "type_expr": "Object", "description": "created vehicle or objNull if failed"},
    "form": "unary", "right": {"text": "[type, position, markers, placement, special]"}
  }],
  "examples": [{"code": ["raw SQF"]}, {"text": "markdown", "code": ["raw SQF"]}],
  "see_also": ["createVehicleLocal", ...],
  "notes": [{"user": "MrSanchez", "date": "2015-08-22T13:04:00Z", "text": "markdown"}]
}
Field Present Meaning
name, title, type, url, revid, updated, groups, status, description, syntaxes always updated may be null, description empty
operator operator pages the symbol (!=, >>)
replaced_by, aliases, see_also when not empty names
games, since, arma3_since when known game code → version that added it (null if the wiki gives none); since is "<game> <version>"
branch dev/diag/prof-only commands list: ["diag", "prof"]
locality when anything is known argument, effect (local/global), server_exec: true; each key only when known
exec functions spawn / call
syntaxes[] always syntax, params and returns (null when the wiki gives none) always; note, since, effect, server_exec: true (this syntax runs on the server), commands' form, left, right, functions' inferred when set
params[] name always; type, type_expr, types, optional: true, default, description, since, optional_since, text, source when set
examples[], notes[], multiplayer, problems when not empty Markdown; examples also have their SQF code
categories_extra when not empty wiki categories that do not follow from the other fields (WIP, Stubs, AI, Easter Eggs, Arma 3: Remote Execution, ...)
docblock, filled_from_docblock functions with a header comment the header's text, author and meta; which fields were filled from it

Every optional field is left out when it is empty (null, [], {}, false, ""): read a missing field as that empty value. Also left out is what follows from other fields, so it can be rebuilt exactly:

  • a parameter's text when it is name, then : type if there is a type, then - description if there is a description; a return value's text when it is type, then - description the same way;
  • an example's text when it is only its code: each snippet as a ```sqf block, one blank line apart;
  • types (parameters, return values, operands) when it lists the names in type_expr (split at <, >, |); an operand's types are those of the parameter of the same name, or ["Array"] for an array literal;
  • categories: the kind, the games, the groups, locality, server execution and branch categories (rules in schema.json); syntaxes[].operator, which repeated the record's.

Topic records (topics/*.json) have every field, categories included.

Changed in format 2 (the files had no version before): categories became categories_extra; games is a map and since a string; branch is a list (also in syntax.json); a syntax's exec: "server" is server_exec: true (the record's exec stays spawn/call); syntaxes[].operator is gone; docblock keeps text, author and meta; empty, derivable and default values are left out; a return type - is written as Nothing, game codes in since are lower case (Arma3 2.08 → arma3 2.08), an unknown note date is left out instead of "". The command and function bundles are about 40% smaller.

name, type, optional and default of parameters are extracted heuristically from the wiki text (name: Type - (Optional, default x) description); text holds the original, losslessly (left out only when the rule above rebuilds it). A parameter's since is the version that added it; a parameter that became optional later ("(Optional since 2.20, ...)") also has optional_since ("arma3 2.20"), present only then. type is the type as the wiki writes it ("Object or Array of Objects"); type_expr is the same type normalised for programs: alternatives joined with |, element types as Array<...> (Object | Array<Object>, Array<Array<Number>>), qualifiers such as "in range 0..1" dropped. When the wiki lists the alternatives below the parameter (* [[Number]]: ... / * [[Boolean]]: ...), type is missing and type_expr collects them. type_expr is missing when the type is free text. In JSON, links are plain text (fewer tokens); in .md and .html they are real links.

HTML version

A static documentation site with no build tooling and no JS dependencies: serve it with python -m biwiki serve or open out/index.html directly from disk.

  • the other pages of the current group on the left, the table of contents on the right; one page per group (groups/<kind>/<group>.html, next to the Markdown list for agents)
  • search across every command, function and topic (/ to focus, arrow keys, Enter)
  • SQF highlighting where every command and function in code links to its page; Copy button on code blocks
  • Note / Important / Warning boxes, locality badges (filled = global, dashed = local), status, versions
  • light and dark theme following the system; mobile layout

Built to pass the checkers search engines and browsers use, tested on every page: html-validate (all pages), the W3C HTML and CSS validators, axe (WCAG 2.2 AA, light and dark), and Lighthouse (100 for performance, accessibility, best practices, SEO and agentic browsing on the main page types). In practice that means:

  • one self-hosted font (JetBrains Mono, OFL) preloaded with font-display: optional; text uses the system font, so nothing shifts while loading; CSS and JS are minified at build time
  • a strict Content-Security-Policy (no inline scripts; JSON-LD blocks listed by hash)
  • canonical URLs, Open Graph and Twitter cards with a preview image drawn for each page, TechArticle and BreadcrumbList structured data, sitemap.xml, robots.txt, favicon, icons and a web manifest
  • headings never skip a level, every piece of content sits in a landmark, scrollable code is keyboard-focusable

scripts/make_brand_assets.py regenerates the icons and the site-wide preview image; biwiki/ogimage.py draws the per-page ones during the build (build --no-og-images skips them).

Code layout

  • biwiki/fetch.py: browser session, MediaWiki API, incremental cache of pages and templates
  • biwiki/portrender.py, biwiki/htmlmd.py: render wikitext with the MediaWiki parser port, then the wiki's HTML → Markdown and plain text (links, lists, tables, code, notes)
  • biwiki/markdown.py: small Markdown writers (quotes, code spans, tables without spans)
  • biwiki/parse.py: {{RV}} template → record (status, types, syntax forms); concept pages → topic records; biwiki/wikitext.py has its helpers for cutting a page with mwparserfromhell, game names and wiki anchors
  • biwiki/docblock.py: function header comments on placeholder pages → description, parameters, returns, examples
  • biwiki/render.py: writes the mirror files and indexes
  • biwiki/report.py: parser health report, baseline comparison and the GitHub issue
  • biwiki/feed.py: change feed from the data repository's history
  • biwiki/site.py, biwiki/assets/: HTML version (markdown-it-py, SQF highlighting, page templates)
  • biwiki/ogimage.py: the per-page preview images (Pillow, JetBrains Mono), drawn in worker processes
  • biwiki/mediawiki/: a Python port of the MediaWiki 1.39 parser, set up as the BI wiki (see its README); python -m biwiki.mediawiki.parsertests runs MediaWiki's own parser tests against it
  • tests/: unit and end-to-end build tests on a few real pages in tests/fixtures/raw/; tests/mediawiki/: MediaWiki's parser test cases
  • AGENTS.md, CLAUDE.md, .claude/: instructions and Claude Code hooks for AI agents working on the code

License

The code is MIT-licensed (LICENSE), except biwiki/mediawiki/ and tests/mediawiki/, which are derived from MediaWiki and licensed GPL-2.0-or-later (COPYING); the program as a whole is therefore distributed under the GPL. Neither license covers the content: it is the intellectual property of Bohemia Interactive and the BI Community Wiki's contributors, used under the wiki's terms, which allow personal, non-commercial use and require works derived from it to stay non-commercial (it is not Creative Commons). Every generated page links to its source and revision and says so in its footer.

About

Arma 3 SQF scripting reference for AI agents and tools: every command and function from the BI Community Wiki as Markdown, JSON and HTML, updated daily

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages