Integration-kit/book/03-sidecar.md

3. The sidecar

Your module may not open a connection to a game server. Not a game socket, not an RCON channel, not a query port, not an engine's admin API. It talks to a sidecar, and the sidecar talks to the game.

That is a rule in the contract (MODULE_API.md §2.7, as of MODULE_API_VERSION 1.4.0) rather than advice this kit is offering. It is also the rule most likely to feel like ceremony when your game already exposes a perfectly good remote-control protocol and your module is fifty lines from working. This chapter is why it is not.

It is the one rule in that list with no CI behind it. An outbound socket is not statically detectable the way an internal require is. So it is enforced by review, and by you having read this.

---

What a sidecar is

A small, separate service that owns the connection to your game, keeps a durable copy of what the game said, and exposes an HTTP + WebSocket API that the website's backend reads.

   your game server  ──dials out──▶  your sidecar  ──HTTP + WS──▶  website core
                                          │                          (your module)
                                          ▼
                                    its own store

Three properties, and each is doing real work.

1. The game dials out; the sidecar listens

The sidecar binds the listener. The game connects to it, and the game opens no listening port at all.

This is the inversion people find surprising and it is the load-bearing one. The website is the internet-facing process; your game is not, and must not become reachable because a web app knows how to reach it. A module holding the connection makes the public web app the thing the game trusts, and puts the game's address inside the same process as every request from the internet.

In uo-link, that listener is sidecar/src/shard.rs — serve binds a loopback address and accepts shard connections forever, handling one at a time and looping back to accept the next. The game plugin does the dialling, with its own backoff. Loopback, in that deployment, because the sidecar runs on the game host: the only socket the game speaks over never leaves the machine.

Only the website's backend talks to the sidecar, and it authenticates. uo-link's web.rs requires a token on every request — accepted as a bearer header, an API-key header, or a query parameter, that last one only because browser WebSocket clients cannot set handshake headers — and compares it in constant time. Auth is always on; there is no unauthenticated mode to accidentally deploy.

2. Persist before you forward

This is the property that makes a sidecar worth having even when your game is already remote-controllable, and the one a message-passing diagram never conveys.

The sidecar owns the durable copy. It writes what the game said into its own store, and answers reads from that store — not by round-tripping the game.

uo-link does this in sidecar/src/store.rs: SQLite, holding event history, the latest snapshot of every board the site renders, the economy series and the published ruleset. insert_event is called for every live event as it is broadcast; the upsert_* functions keep one current row per board; the REST read paths query that store.

What it buys, concretely:

• A website that is down, restarting or mid-deploy loses nothing. Events that arrive while nothing is listening are still recorded. Without a store they are simply gone, and your first deploy of the week is a hole in your data.
• A page renders the last thing the game said rather than going blank. A rules page that empties itself because the game restarted is worse than a stale one.
• The live feed is allowed to be lossy. uo-link's WebSocket fan-out drops frames for a consumer that has fallen behind and logs that it did — deliberately, because durability is the store's job and not the socket's. A feed that instead buffered without limit for a slow client would eventually take the sidecar down.

That last point is the reasoning to carry into your own design. Once the store is authoritative, every other component is allowed to be best-effort, and each of them gets simpler. Skip the store and you find yourself trying to make a socket reliable, which is the hard version of this problem.

A module cannot do any of this from inside the website process. There is nowhere to put what arrives while the website is not running, because the website not running is exactly the case.

2a. The other direction, if you ever want events

Everything above is about data leaving the game. Skip this section until you want chapter 5 — but read it before you build the sidecar rather than after, because retrofitting it is more work than allowing for it.

An event on the website is core telling your module "do this to the world now", and your module telling your sidecar, and your sidecar telling the game. That is a command — a request with a reply, going the way nothing above goes. It needs three things the read path does not.

Request/reply correlation. A command is not a broadcast: the caller waits for an answer and has to know which answer is theirs. uo-link does this in sidecar/src/rpc.rs — an id on the way out, a map of pending calls, the reply matched back and the waiter woken. You need it for reads that ask the game a live question too, so it is often already there; commands are what make it load-bearing.

An idempotency key, executed at most once, stored where the game is. Core hands your module a key that is a function of the step's identity and never of the attempt, so every retry carries the same one. The far end must execute a given key once and answer a repeat with the reply the first attempt produced — not by running the command again.

That store belongs as close to the game as the state it protects. A store in the sidecar is right for a command whose effect is the sidecar's own; a command that changes the world needs the store where the world is, because the case it exists for is the game restarting mid-run. And a repeat arriving while the original is still in flight is its own answer — "busy", transient by construction, because the work is happening.

Without this, a command that arrived, ran, and whose acknowledgement was lost is indistinguishable from one that never arrived. The only safe policy is then never to retry, which means a game restarting mid-event writes the step off.

A deadline the game enforces on its own. A borrowed value — a doubled gather rate, a raised spawn cap — carries an expiry down the wire, and the game side must restore the baseline when it passes without being asked again. The website's copy of that deadline is for the console. The game's copy is the fail-safe: if the website is never heard from again, the value still comes back.

Two details that are easy to get wrong and expensive to change later. Send the deadline as a duration, not an absolute time — two machines' clocks are two clocks. And if the borrowed value lives in the game's own save file, the hold must be persisted and the timer re-armed at load; a restart preserves the change and destroys only the thing that would have undone it.

2b. The game host already has the files your site wants

There is a third kind of traffic, and it is worth knowing about before you decide your sidecar only ever forwards live state. Most games keep content on the host that a website wants to show: sprites, icons, portraits, localisation tables, map or spawn definitions. It is static, it is large, and it changes only when an operator patches the game.

The tempting answer is to make the operator's problem: export it on a desktop with some third-party tool, upload the result, repeat after every patch. It works once and rots immediately, because nothing reminds anyone to redo it.

The better answer costs less than it sounds like: the game host already has those files, and you already have a channel to the game host. Route them over it.

Four design notes, all learned the expensive way in uo-link's protocol 8 (the "asset bridge", v8.md):

• This is request/reply, never events. A sidecar that persists and broadcasts every event would write megabytes of sprite into its own store and fan it out to every connected client. Content must ride the same correlated round-trip a query uses — see chapter 5 for the shape.
• Serve one at a time, and say so in the protocol. Decoding assets costs the game host real memory. One in-flight request with an explicit "busy" answer is simpler and safer than a queue, and a caller that treats busy as flow control rather than failure gets a working import out of it.
• Two stages: what exists, then what changed. A cheap call that returns a list with a hash per item and no content, then a second that fetches only the hashes that moved. The common case — a restart that changed nothing — must cost one small round trip, not a re-download of everything.
• Version your derivation, separately from the protocol. If you improve how you read a file, the bytes you produce change while the source file's hash does not. uo-link carries an EXTRACTOR_VERSION for exactly that, and a consumer treats a change in it like a changed hash.

And one operational note, because it is the part that surprises people: do not import on boot. A patch is an event the operator knows about and your website does not. Re-reading hundreds of megabytes on every restart to discover that nothing changed pays for the rare case forever; a button an operator presses after they patch costs nothing and is honest about who knows what.

3. The wire is a versioned contract, not a build dependency

Your sidecar and your module ship separately, on different schedules, to hosts you do not control. So the wire between them is a compatibility contract with a version on it.

uo-link declares PROTOCOL_VERSION in sidecar/src/main.rs, stamps X-UOLink-Version onto every response from web.rs, and refuses a request whose declared version does not match rather than parsing it optimistically. A refusal is a clear failure an operator can act on; a mis-parse is a wrong number on a page with nobody to tell.

Two habits come with that:

• Bump the version in the same change that changes a message shape, on every side that declares it. In this project a protocol bump has three declaration sites — the sidecar, the game-side overlay's manifest, and the documented spec — and the tooling refuses to pair components that disagree.
• Version the shape, not the content. Adding a new event kind that old consumers ignore is not a break. Changing what a field means is, even when the JSON still parses.

The worked example

uo-link is a complete implementation of everything above, and it is small enough to read:

File	What it owns
sidecar/src/shard.rs	The listener the game dials into; one connection at a time, then accept the next.
sidecar/src/store.rs	SQLite: event history, per-board snapshots, the series and the ruleset.
sidecar/src/web.rs	HTTP + WebSocket for the website, the auth middleware, the version header and the lossy live fan-out.
sidecar/src/rpc.rs	Request/reply correlation, so a website read can ask the game a question and match the answer.
sidecar/src/config.rs	The config file, including a token generated on first run rather than defaulted.

The protocol it speaks is specified in link/PLAN.md and link/INTEGRATION.md. Those are normative for that sidecar; your game is not Ultima Online and your messages will not be its messages. What transfers is the structure — a listener the game dials into, a store written before anything is forwarded, a lossy live feed, an authenticated read API with a version on it.

"But my game already speaks a remote-control protocol"

Then your sidecar is thin, not absent — and be sure the surface you are thinking of actually carries what your module needs, because that is where this question usually goes wrong.

A remote-control channel is built for an operator typing commands: it tells you what you asked about, when you ask. What a website needs is what happened — every kill, every join, every departure, delivered whether or not anyone was listening at that moment. Those are different products, and a channel that answers the first can only approximate the second by polling it, which turns "someone left the clan at 14:02" into "the count was different at 14:03".

So the honest test is not does my game expose a protocol but does it deliver events. If it does, your sidecar keeps that connection, its credentials and its reconnect loop out of an Express process, keeps a store so the site is not blank whenever the game restarts, and presents your module one versioned HTTP + WS shape. That is what thin means: less code, the same architecture.

Rust is the worked example, and it is not that case. The dry run in rust-dryrun.md designs a module for it precisely because it shares so little with Ultima Online — and its answer is an Oxide plugin: C# loaded by the mod framework a modded Rust server already runs, hooking the game's events and dialling out to a sidecar, exactly as the ServUO overlay does. The Rust server is a binary, where ServUO is source a shard owner compiles, so the way in is a published hook API rather than a file you edit. The three-part shape survives that unchanged, which is the more useful finding: the plugin-dials-out arrangement is not a property of having source access.

It also pairs one sidecar to one game server, on that server's own host, rather than one sidecar fronting a community's several — because those servers sit on separate machines, and a shared sidecar would be reached across a network by plugins that are supposed to talk to it over loopback. Worth knowing before you design yours: if your game runs as a fleet, the question "how many sidecars" is answered by where the loopback boundary is, not by how many processes you would rather run. Your module holding several clients is the cheaper end of that trade, and core never learns there is more than one.

That document reached the game over RCON until 2026-08-19, and before that concluded there should be no sidecar at all. It carries both corrections, dated, rather than having been quietly rewritten — the value of a dry run is the record of what it found, including where it was overruled.

Building yours

There is no template for a sidecar in this kit; it is your program, in your language, and the surface it must expose is the surface your module reads. What to settle before writing code:

1. Which direction does the connection go? The game dials out. If your game cannot — if it only accepts connections — then your sidecar is the client to the game and the listener for the website, and the rule that stands is the one that matters: the address of the game is known to the sidecar and to nothing else.
2. What is durable? Everything a page must still render when the game is down. Write it before you forward it.
3. What is a snapshot and what is an event? They are different storage problems: an event is appended and read back as history, a board is one current row per subject that you overwrite. uo-link's store holds both, and keeping them separate is why a restart does not replay a year of events at a page.
4. What is the version, and where is it declared? One place, on every response, refused on mismatch.
5. How does the website authenticate? A token, generated rather than defaulted, always required.

Then chapter 4, if your game needs code inside it — which is the part where getting it wrong takes the game down rather than the website.