seriesoftubes/docs/ARCHITECTURE.md
Frederic Jung 74e8611f8e
The sales desk answers renewal haggles, each salesperson their way
With the new switch at the sales desk, whoever runs sales answers a
renewal haggle three days after it opens, unless the player did first.
The engineer gives business customers their discount and keeps nerdy
ones at the market price (nobody leaves); the shark holds the old price,
often with a fresh promise, bills a month of the saving, and loses some
(nerds most); the CEO gives 10% more off for a double term, says yes, or
picks a fight. Every answer is a hash of the contract. The switch is
stored only when on, so existing saves are unchanged.
2026-11-06 20:56:50 +01:00

25 KiB
Raw Blame History

Architecture

Series of Tubes is a static web game hosted on GitHub Pages, and only there (ADR 0001). Its source is split into many small files under src/; the build joins them into one self-contained index.html (one inline <style>, one inline <script>), exactly the page players have always loaded (ADR 0002).

src/                         the game page, in pieces
  index.html                 shell: <head>, static markup, the language loader, the @include lines
  i18n/en.js                 the English for every label → inlined at the top of the page's <script>
  styles/NN-*.css            → pasted into the page's <style>, in path order
  game/NN-area/NN-*.js       → pasted into the page's <script>, in path order
  agent/NN-*.js              → the agent facade the MCP server drives: development page only (@include-dev)
site/                        static files copied to the build as they are (dictionaries, icons, sw.js…)
scripts/lib/page.js          the assembler: <!-- @include DIR --> → every file under src/DIR, sorted;
                             <!-- @include-dev DIR --> the same, in the development page only
scripts/build.js             _site/ = site/ + the assembled page, version-stamped, release notes baked in

Because the partials are concatenated into one classic script, they share one top-level scope, just as when they were one file: S, CITY or compute() are visible everywhere, function declarations are hoisted across files, and top-level statements run in path order. Find code by name (grep -rn "function compute(" src/game); the folders follow the game's layers:

Folder under src/game/ What lives there
00-core/ tr()/trn()/N_() and the dictionary; version, release notes and the analytics counter
10-catalog/ Prices and gear: POP and port costs, routers (RG), supply chain, port slots, dark fiber systems and optics (OPT), reach and subsea upgrades
20-world/ Map projection, landmasses, coastline, city coordinates (CITY_LL), CITIES, CABLES
30-networks/ Other ASes (NETS), rival exchanges (XIX), the long tail, traffic shares (NSH)
40-progress/ Customer types (BIZ_T), LEVELS and their requirements, GOALS; the per-save ruleset (S.rules, rule(), PRESETS, chaosMul(), leadDays())
50-state.js The globals S, R, UI; emptyR()
60-features/ Banks and loans, scenarios and the weekly challenge, puzzles and the published weekly puzzle, conferences, the tutorial, traceroute, the share card, newGame(), the offer trial (autoOffer()), standing orders (S.ops, opsDay(), run from dayTick()), the sales desk (S.sales, salesDay(), salesShape() on new offers, orbit ones through pieSlShape(); signing for you: S.slAuto, slAutoDay(); renewal haggles: S.slAuto.ren, slRenDay())
70-rules/ Formatting helpers and market rules: exchanges, PNIs, the peering desk (requests answered by policy), peering policy and snubs, DE-KICKS, sold-out routes, hot ports and the Tier 1/2/3 rank
80-simulation/ Dijkstra, RouteScript, where a DDoS flood goes (ddosLoad()/ddosAfter(), called by compute()), compute(), monthly prices, sanitize()/migrate(), customer requirements, events, the customer lifecycle (contracts, renewals, the bandwidth market, why customers leave), geography pricing and latency, second sites and service
85-incidents/ addInc(), DDoS attacks (targets, sizes, scoring, the NOC card) and scrubbing, fireFailure(), clearInc(), fireEvent(), the chaos monkey, cable ships (50-cable-ships.js: a subsea cut sends a club ship from its port along the cable routes; voyages in S.fleet, drawn by 96-map/88-cable-ships.js)
88-days/ dayTick(), new offers, badges, checkGoals()
90-actions/ The ACT map, one file per area (const ACT in 10-interface.js, Object.assign(ACT,…) in the rest), and their helpers
91-save.js, 92-audio.js Save, export and import; sound effects
93-ui/ DOM helpers ($, btn, morph), the phone tab bar, install pitch, HUD, left panel, city and link panels, decommissioning
94-dashboard/ Scenario and level progress, where to meet a network, customers and connections overviews, cities with no POP yet, lease swaps, the NOC dashboard and its own window for a second display (90-noc-window.js: this page at ?noc=<slot>, a read-only copy of S and R over a BroadcastChannel, clicks sent back to the game; 91-noc-wall.js: its wallboard), the bank, hints
95-modals/ Toasts, notifications, modals, title screen, what's new, help, the link and own-cable modals
96-map/ The canvas: resize, land and clouds, settings, level of detail, link geometry, how each tube looks, the WebGL map (the default, 2D as fallback: context, background, tubes, and the sprite atlas cities, labels and markers are drawn from, in 2D too), draw(), the world layer, cities, input and click delegation
99-loop.js frame() and boot()

src/styles/ holds the CSS the same way: base, HUD, drawers, toasts, modal, tablet, then phone.

src/agent/ is the agent facade (window.TUBES): the typed commands, reads and screens the MCP server plays through. It is assembled into the development page only (npm run dev, the MCP server, tests that ask for it); the page players get never has it (ADR 0004). Its top-level names all start with api/API_, and nothing outside src/agent/ may use them. Where it must reach into the game (agent mode in the player's tab locks the clock and draws its badge and Settings section), the game offers HOOKS and the flag UI.clockLocked (50-state.js), empty and false in the page players get, and runClock(dt) (99-loop.js).

Rules for the source (checked by tests/build.test.js and the assembler): files are named NN-lower-case.js/.css, end with a newline and never close their <script>/<style>; top-level code may only use top-level values from earlier files; tables people add to have one entry per line, the id-keyed ones sorted. New features get their own file in the right folder (edit-the-game skill). Development runs on npm run dev (assembles the development page on every request, reloads on save); npm run page -- -o FILE writes the page players get, npm run page -- --source --dev -o FILE the development page.

Other files

Path What it is
src/i18n/en.js English: label → text, one entry per line, sorted; inlined into the page by <!-- @include i18n -->
site/i18n/it.js, pt.js, fr.js, es.js, de.js Translations: label → text, the same layout; loaded only for that language
site/sw.js Service worker: network-first for pages, stale-while-revalidate for assets. Cache name is stamped with the version
site/site.webmanifest, icons, 404.html, CNAME, robots.txt, sitemap.xml, llms.txt PWA and GitHub Pages files, copied as they are
scripts/lib/page.js Assembles the page from src/ (also from a git ref with --ref, for performance baselines)
scripts/lib/compile-labels.js Shrinks labels to short ids in the shipped page and dictionaries
scripts/build.js Builds _site/ for GitHub Pages: copies site/, writes the page, stamps __VERSION__/__BUILD_DATE__, bakes release notes, writes version.json
scripts/dev.js npm run dev: development server (loopback unless --host), serving the development page and hosting the agent relay
scripts/test.js npm test: runs every tests/*.test.js
scripts/release-notes.js Extracts a version's CHANGELOG bullets
scripts/perf-bench.js, perf-browser.js, perf-stats.js, perf-fixture.js Big-network benchmark, base vs head, headless and in Chromium as a desktop and a throttled phone, with a significance test per row (posted on every PR by perf.yml + perf-comment.yml), and the generator of its save, tests/fixtures/big-network.json
scripts/gpu-bench.js, scripts/lib/gpu-bench-client.js npm run bench:gpu -- --save FILE: the map in your own browser on your own graphics card, 2D against WebGL, scenario by scenario; the table shows in the page and in the terminal
scripts/mcp.js, scripts/lib/mcp-game.js, scripts/lib/mcp-headless.js, src/agent/ The game as a local MCP server, on the official SDK over stdio or HTTP: typed tools to play, inspect and debug it from an AI client, headless or in the player's own tab (MCP.md)
scripts/agent/relay.js, scripts/lib/mcp-browser.js, scripts/lib/relay-client.js The agent relay npm run dev hosts at /agent, and the MCP server's side of it: how a tool reaches the game in the player's tab
scripts/i18n-check.js, scripts/i18n-add.js, scripts/lib/dict.js Translation coverage and dictionary order; adding, renaming and removing entries
scripts/legacy/ The splitter that made src/ from the old site/index.html (and gave its text labels), and port-branch.js, which carries older branches over
tests/harness.js Loads the assembled page in jsdom with a seeded Math.random and a fake canvas: the players' page, or the development page with {agent: true}
.github/workflows/ci.yml Tests and build on every push and PR
.github/workflows/auto-release.yml On a package.json version bump on main: test, tag, GitHub release from the CHANGELOG, start the deploy
.github/workflows/release.yml Builds and deploys to GitHub Pages, the only workflow that can
.github/workflows/weekly-puzzle.yml Sunday 03:37 UTC: makes the week's puzzle, commits it to main (data only, no release) and starts release.yml for the live release tag with weekly=true
scripts/weekly-puzzle.js, site/puzzles/weekly/ The weekly puzzle generator and its output, one <key>.json per week plus index.json (see below)
scripts/weekly-scenario.js, scripts/lib/scenario-bot.js, site/scenarios/weekly/ The weekly scenario generator, its reference player and its output (see below)
docs/adr/ Decisions: GitHub Pages only (0001), one page assembled at build (0002), translations by label (0003), development-only code (0004)

Globals

  • S: the whole saved game (one JSON object). Cash, day, POPs, links, transit, IX ports, PNIs, caches, customers (biz), offers, incidents (inc), live events (ev), stats, settings. Saved to localStorage key sot-save-v1:<slot> (one per network, see below).
  • R: results of the last compute(): per-link load, per-exit use, served and demanded traffic, reach, costs and revenue. Never saved; always rebuilt.
  • UI: view state (selection, open tabs, speed, confirmations). Not game state.
  • cam: map camera (x, y, zoom).

Lookup tables built at load: CITY[id], NET[id], CABLE, FAB, BIZ_T, REQS.

The life of a day

frame() (requestAnimationFrame) accumulates time; every DAY_MS × speed it calls dayTick(), at most four times per frame, each inside a try/catch so one bad tick never freezes the game.

dayTick():

  1. Advances S.day, runs scheduled world events (conferences, IX price wars, licences).
  2. Runs compute() to route today's traffic.
  3. Records history for 95th-percentile billing, stats and streaks.
  4. Rolls failures (fireFailure()) and events (fireEvent()) when their timers run out, and expires incidents. Then the NOC intern's standing orders (opsDay()), if one is hired: each order clicks the same ACT actions a player would, muted, within its monthly budget, and the shift recomputes once at the end, only if a job changed the network. An order that throws is paused; the day goes on. The NOC gets one notice a week. Then the sales desk (slAutoDay()), if someone runs sales and the player lets them sign: deals are Sales', never the intern's. Each salesperson signs in their own style (a routing trial a day, or the shark's sign-everything).
  5. Updates customer mood, churn and new offers; IX members join; loans tick.
  6. Every 30 days: bills the month (revenue minus costs) and checks for bankruptcy.
  7. Checks goals, level, badges and scenario end conditions.

compute(): the traffic model

  1. Builds a graph from your live links: parallel links between the same two cities merge into one edge with summed capacity. The routing metric belongs to the city pair (S.pairMet) and sets the edge's weight; Dijkstra gives shortest paths from each POP. Links on standby (l.standby) add their capacity to the edge, but when loads are handed back to links the active ones fill first and standby links take the rest.
  2. Builds the exit list E: embedded caches (rank 0), PNIs and IX ports (2, both peering), transit (3). Each exit has a capacity, the networks it can reach (ok(net)) and an egress preference (high, normal, low = backup only). A transit port can also be set to carry (t.carry): see step 4.
  3. Builds flows: for every POP, its demand (eyeballs and customers) split across destination networks by their traffic share (NSH, with regional variants for China, Russia and Korea). Incidents can add junk flows or scale demand.
  4. Routes flows rank by rank: the nearest usable exit first, ECMP across equal exits (a PNI goes before an IX port at the same distance: PNI_TIE, #298), consuming link, router (nodeCap) and exit capacity. A carrier's own transit port also goes first, before plain transit, for traffic bound for that carrier or for a network that buys transit from it (25-direct-egress.js: the shorter AS path): always at its own POP, and from every POP when the port is set to carry. Such a port is still an ordinary transit exit for the other POPs. Backup, RouteScript and Auto-TE rules keep their say, and so does a Prefer-marked port at its own POP (the other POPs keep the rule). With Auto-TE, overflow spills onto alternate paths; RouteScript rules (teRules()) can prefer, avoid, cap, drain or blackhole.
  5. Layer-2 customers (L2VPN, SD-WAN, waves) are carved out of link capacity first.
  6. Totals costs (R.cost) and revenue (R.rev) per month.

It runs every day and after every action, so it is built for big networks: the graph works on city indices (CIDX), each exit is asked once per network what it can reach, and R.ef (what crosses each link) is only put together for a link when something reads it.

What follows from the network's shape alone is kept from one call to the next (29-route-cache.js): the shortest paths from every POP, and who reaches whom with each network's exit lists. Each has a key made of everything it is built from (the live graph with its weights; every exit with what it reaches and how it is preferred, RouteScript and the events that move routes), so most days, when only demand changes, they are reused, and any other change rebuilds them. Only indices are kept: exits, lists and paths are made fresh every call, because routing writes on them. tests/route-cache.test.js checks the keys, and the equivalence tests route with and without the kept tables. scripts/route-fuzz.js plays hundreds of random seeded games on main's page and this checkout's at once, comparing all of R and S after every move and day, with the kept tables and without (tests/route-fuzz.test.js plays a few).

Each loop of compute() runs in a function of its own (the stages are closures, the busiest is rankPass; the money is computeMoney()). V8 optimises a small function quickly and keeps it; compute() as one 33 KB function never stayed optimised: it ran in baseline code, boxing every number, and was compiled again and again in the background only to be thrown away, which cost more CPU than the routing itself. Keep it that way: a new loop goes in a function.

Served share per city (sfOf) drives eyeball revenue and customer mood.

Routing in a worker

During play, each day's compute() runs off the main thread (src/game/80-simulation/28-route-worker.js), so a big network does not freeze the page once a day.

  • Built from the page itself. The worker is the page's own inline game script (document.currentScript.textContent) run again from a Blob URL, after a short prelude of inert stand-ins for document, storage and the other DOM globals, and the player's dictionary. Nothing is fetched: GitHub Pages still serves one page, and a release cannot mismatch. In the worker RW_IN is true and boot() never runs; the catalogs are rebuilt by the same top-level code.
  • The day split. dayTick() is dayPre() (the world moves on), compute(), dayPost() (bills, moods, offers, goals). In play, frame() calls rwDay() instead: dayPre(), then S and what routing reads of yesterday's R (R.load, R.l2route, R.efg) are posted, and the next day waits until R comes back and dayPost() has run. So a day sees exactly the routing it would have had in one go: same results, same random numbers. Half a day is never saved.
  • What crosses. R.ef and R.fl are lazy getters; their pieces (R.efp, R.flc) travel instead, typed arrays transferred, and efBuild()/flDefine() (in 30-compute.js) rebuild the same lazy R on the page.
  • Actions mid-day. Every compute() bumps RW.gen. If an action routed while a day was in flight, the worker's answer is stale: it is dropped and the day ends with the action's R.
  • Fallbacks. Tests (__TUBES_TEST__, and jsdom has no Worker), window.__TUBES_SYNC__, a browser without workers, a post that fails or a worker that errors take the synchronous path (rwKill() routes the day in flight on the page).
  • Still on the main thread, because the caller reads R right away: actions (after()), loading a game, failures and events that fire during a day, standing orders, the chaos monkey, RouteScript what-ifs, the link panel's cut check and automatic offers.

Actions and rendering

  • Player actions live in ACT. Buttons render as btn(label, act, args, cls, disabled) and a delegated click handler calls ACT[act](JSON args).
  • An action validates its inputs (they come from the DOM), mutates S, shows a toast() and calls after().
  • after() = purgeInc() + compute() + checkGoals() + renderAll() + save().
  • renderLeft() draws the left tabs (deals, NOC, customers…), renderPanel() the selection panel (city, link, network). On phones these become bottom sheets controlled by renderTabbar().
  • These panels and the dashboard are rebuilt as HTML every tick and written with morph(), which patches only what changed. Buttons whose action changed are replaced, so taps stay safe (#44).
  • The map is a canvas drawn every frame by draw(). Zoomed out (lodFar()) it uses a simplified style; parallel waves are grouped by linkGroups().

The weekly puzzle

A puzzle week runs from Sunday 06:00 Berlin time to the next Sunday 06:00, and is named after the ISO week of its Monday (the week starting Sunday 1 November 2026 is 2026-W45); pzWeekKey() and pzWeekStart() in src/game/60-features/93-puzzles-weekly.js handle summer time and the year boundary.

Every Sunday weekly-puzzle.yml runs scripts/weekly-puzzle.js, which loads the real game headless, picks a theme the last three weeks did not have (region tour, ring, detour around a hub, ocean crossing, island hopping, latency sprint) and cities the last four weeks did not use, gives them traffic (content in the first city, eyeballs in the others, a circuit with a latency cap for a sprint), searches for the cheapest topology that connects them and the cheapest one that survives any single cut (a whole cable, or a trench), then sizes both with the game's own routing (scripts/lib/puzzle-solver.js: links and routers big enough for every flow, after any cut too, then as small as they go). The budget sits just above that N-1 build. Both references are judged by the game's own star code and must earn one and three stars, or the puzzle is thrown away. The file holds only ids, numbers and translation labels with their fill-ins ({"v":2,"key","theme","cities","src","dem","via","avoid","budget","twist","name":{"l","p"},"joke":{"l","p"},"ref","rev"?}). --check re-validates every file the way the game does.

The game fetches puzzles/weekly/<key>.json (relative, same origin) only when the Puzzles list opens, and checks it field by field (pzWkValid()): known fields, playable cities, labels from the theme's own list, sane numbers. When it is missing, offline or wrong, the seeded weekly (pzWeekly()) stands in, so the weekly puzzle always works. The service worker fetches these files network-first and keeps a copy for offline play.

The weekly ships without a release: the workflow commits only site/puzzles/weekly/ to main (no version bump, so auto-release.yml stays quiet) and starts release.yml from the live release's tag with weekly=true, which rebuilds that exact release with every weekly file from main and deploys it. Every release build takes the weekly files from main too, so no deploy ever drops a published week.

The weekly scenario is its twin, in the same weeks and the same workflow run: scripts/weekly-scenario.js writes site/scenarios/weekly/<key>.json. A template (plain data in the script) picks a theme, a home city, rules from the scenario vocabulary (scnRule keys, WEEK_MODS) and a goal type with parameters; each goal type (WS_GOALS in src/game/60-features/94-scenarios-weekly.js) is implemented once in the game, with its progress card. A reference player (scripts/lib/scenario-bot.js) plays it through ACT, seeded with the scenario's seed; the goal's number is calibrated from that run, the scenario must not be won on day 0, the reference must win with a tenth of the days to spare, and its finishing day is the par. The file keeps the reference plan (ref), which --check replays. The game fetches the file when the Scenarios list opens, checks it with wsValid() (again in sanitize() on load), and falls back to the in-browser weekly().

Saves, imports and migration

  • save() writes S to localStorage every few seconds and after actions, into this window's slot (UI.slot, src/game/91-slots.js): sot-save-v1:<id>, plain JSON or gz:-compressed above a million characters, with a stamp sv:{seq,at} appended. Then the index sot-slots (name, ASN, preset, day, tier, cash, last played, size, seq) is updated; the title screen lists networks from the index alone.
  • Each tab keeps its slot in sessionStorage['sot-slot']. One window per slot: a Web Lock (sot-slot:<id>) where supported, plus a heartbeat sot-open:<id>; BroadcastChannel sot-slots asks another window to hand over, and the storage event stops a window whose slot was saved newer elsewhere (higher seq) or deleted. A window never writes a slot whose index seq is newer than its own.
  • The single save of older versions (sot-save-v1) is copied to slot main and read back before the index lists it; the old key is removed only by that slot's next successful save.
  • boot() loads the save through migrate() → sanitize(). If loading throws, the slot is kept and marked damaged in the index (it can still be exported or deleted).
  • Puzzles never write a slot; while one runs, save() writes the network set aside (UI.pzKeep).
  • sanitize(s) is the gate for all untrusted state (saves and imported files): it drops unknown ids, clamps numbers, repairs references and fills defaults. Every new field in S needs a line there.

Translations

Text is keyed by label (ADR 0003):

  • tr('area.label', vars) looks up DICT['area.label'] and fills {placeholders}; trn('area.label', n) picks one or other of a plural; N_('area.label') marks a label kept in a table or a save, translated where it is shown.
  • DICT is the English (src/i18n/en.js, inlined into the page) with the player's language (site/i18n/<lang>.js) laid over it. English text from saves of 1.33 and older still finds its translation.
  • Static markup uses data-i18n="label", data-i18n-aria-label, data-i18n-title; i18nDom() swaps it at boot.
  • The build swaps labels for short ids in the shipped page and dictionaries (scripts/lib/compile-labels.js); N_() labels stay whole because saves keep them.
  • npm run i18n:check checks the English against the code and every translation against the English.

Tests

tests/*.test.js use node:test and tests/harness.js, which assembles src/ in memory, so tests never run a stale build. npm test runs every file in tests/; a new suite needs no registration:

const { load } = require('./harness');
const { ev, errors } = load();            // seeded, deterministic
ev("ACT.start({});S.cash=1e8;ACT.launchBB({city:'fra'})");
ev("compute()");
assert.equal(ev("R.reach>0"), true);
assert.deepEqual(errors, []);

ev() runs code inside the page. To force a specific random failure, patch the picker, e.g. evaluate fireFailure with .replace('const e=pick(opts);', "const e='x';"). Use window.__reseed(n) to reset the random sequence mid-test.