Routing tried every PNI (rank 1) before any IX port (rank 2), like a
local-pref above peering, so a PNI in London pulled New York's traffic
across the Atlantic even with the same network at the New York IX. PNIs
now route at rank 2 with the IX ports, nearest first. At the same distance
the PNI goes first: an IX port counts as PNI_TIE (0.001) further away,
far below any link weight (at least 1). Distances compare as
(Da-Db)+(dta-dtb) so exits at one distance stay exactly equal, and the
POP-local shortcut skips lists that mix PNIs and IX ports. Backup only and
RouteScript's prefer pni keep their say. The traceroute, the reference
allocation in the tests and the RouteScript help follow.
The image is published to GHCR as the release's version when a release
dispatches docker.yml, and never as latest or per commit. Pushes and pull
requests that touch it build it and try headless (with and without a
token) and browser mode, without publishing.
The image: node:24-slim pinned by digest (the major CI tests on), npm ci
from the lockfile in its own layer, only src, site, scripts and the
harness copied in (an allow-list .dockerignore), root-owned and run as
node, /saves ready for --save-dir. The entrypoint runs the servers with
node directly and, with no SOT_MCP_TOKEN, makes a token for the run and
prints it (MCP_NO_TOKEN=1 to opt out); the docs publish ports to
127.0.0.1 and explain --allow-host. Labels, SBOM and provenance on push.
Every GitHub Action in every workflow is pinned to a commit SHA with its
version, and Dependabot keeps the actions, the image's base and the npm
tools current. tests/docker.test.js guards all of it. The ADR and the
github-pages skill say the image is a tool: GHCR only, never Pages.
From customer level 2, one person runs sales at a time, next to the NOC intern:
- The engineer turned salesdroid ($9k/mo) books you into every network operators' meeting 30 days
ahead at the usual ticket price, brings 40% more offers (half as many again while a meeting's glow
lasts) but never the huge deals, and their customers leave less often at renewal (half as often
for the nerdy ones: HFT, ISPs, universities, L2). Haggles badly.
- The shark ($15k/mo) brings bigger sizes at 25% more, takes a month of each deal as commission,
oversells half of them with a requirement the network does not meet yet, never brings technical
customers, makes selective peering desks want 25% more traffic, wins most haggles, and their
customers leave 50% more often.
- The CEO (free) signs the biggest internet customer the network can take on the golf course every
50-80 days at 30% off, often with a promise nobody checked, signs about a third of the offers you
decline anyway, and has opinions. He cannot be fired: a board meeting brings a VP of Sales a month
later.
Customers remember who signed them (by), which the renewals read. Every roll is a hash of the day
or the deal, so games without a salesperson keep their random stream.
The differential fuzz also leaves out buttons for actions main does not have yet.
The dashboard shows what you have built, not what you haven't. The new tab lists every city you could still open, with its eyeballs, exchange and price, and sorts, filters by region and pages like the other long lists. Open selects the city on the map, so a build still goes through the city panel and its warnings. The tab also works in the NOC window.
Every customer now signs for a term (enterprise and hosting 1-3 years, wholesale and circuits 2-5,
fiber pairs 10-20). When it ends most renew at today's bandwidth price, happy ones often grow to
the next size, a few with competitors around ask for a discount 30 days ahead (match, meet halfway
or let go in Deals; no answer matches), and some leave: unhappy ones because of the service, others
acquired, bust, insourced, moved (they stay if you have a POP there), multihomed, project ended,
a cheaper bid, or now and then one of the odder reasons (never more than one every 60 days).
The bandwidth market is flat for six months, then about 10% cheaper a year, slowed by dearer
hardware, moving in a slow three-year cycle and pushed up during a shortage. New offers and
renewals are priced in it; one renewal moves a price at most 25% down or 10% up.
Nothing happens in a first term. Rolls are hashes of the customer and its term, never Math.random,
so seeded games keep their stream. Old saves get renewals spread over the next year. Happy
customers of a year or more send a referral about every six weeks. The dashboard's Customers tab
shows customers won and lost each month and why, renewals due and the market; customer rows show
the contract end.
The differential fuzz now leaves out state fields only this checkout has, so a new feature's state
does not fail the comparison with main while everything both games share must still match.
- Sea, land, waves, clouds, cables, tubes and packets are drawn by the
graphics card; the land is geometry built once, so panning and zooming
repaint nothing.
- Cities, labels and markers are painted once into a sprite atlas and drawn
in one call (with drawImage in 2D, which speeds up the 2D map too).
- Settings > Map drawing: Auto or Always 2D, per device. Browsers without
WebGL2, a lost context and puzzles keep the 2D map.
- npm run bench:gpu compares 2D and WebGL in a real browser on a real GPU;
?fps shows a frame meter.
A button next to the speed controls opens this page at ?noc=<slot> in a
noopener pop-up, so it runs on its own thread. The window never claims the
slot, saves or runs a day. The game sends it a copy of S and R over a
BroadcastChannel, at most once a second and only while a window is connected,
and runs its clicks through an allowlist of the same ACT.* a click uses.
The window opens on a dark wallboard: a live world map of links and POPs,
traffic and money graphs, alarms, and the hottest ports with the dashboard's
own upgrade buttons (connUp). Clicking a POP filters the ports. Details shows
the full dashboard. Phones do not show the button.
The MCP server gains a second session, --browser: the game in the player's tab, reached through the
agent relay npm run dev hosts at /agent (loopback unless --host; Host and Origin checked, no token).
The player turns it on in Settings → Agent API, a section only the development page has, and picks the
clock: a pace of 30 s to 5 min a day that nobody can pause, or the agent's own. Headless gets the
running clock too (--clock realtime --pace MS) and both get a wait tool.
All of it lives in src/agent/ (development page only). The game offers HOOKS, UI.clockLocked,
setSpeed() and runClock(dt), the frame's day block moved out unchanged, and nothing more: the page
players get grows by under 1 KB gzipped and runs exactly as before.
scripts/mcp.js registers the same tools on @modelcontextprotocol/server (2.4.0, with
@modelcontextprotocol/node 2.1.2 for node:http; dev dependencies, pinned) instead of its own JSON-RPC
loop. stdio works as before; --http [HOST:]PORT serves Streamable HTTP at /mcp, on loopback unless
told otherwise.
Over HTTP every request's Host must be an address of this machine (or --allow-host), an Origin other
than localhost is refused, and a bearer token (SOT_MCP_TOKEN or --token-file) is optional; bound
beyond loopback without one, the server warns. load_save and export_save take JSON only there, or files
under --save-dir and nowhere else. Calls from every client run one at a time, in order.
The command tables and the page-side reads, screens and dispatch (scripts/lib/mcp-commands.js and
mcp-page.js) become src/agent/, the facade window.TUBES.call(tool, args) in the page's own scope.
A new <!-- @include-dev agent --> line in src/index.html pastes it in after the game, but only in the
development page (npm run dev, the MCP server, i18n-check, tests that ask with load({agent:true})).
The build drops the line whole: _site/ is byte for byte what main builds.
Tests load the page players get by default, the build test fails if anything outside src/agent/ names
the facade's api/API_ names, and the pages test checks the built page carries none of it and still
plays. ADR 0004 records the rule.
scripts/route-fuzz.js plays many random games (every mode: presets, sandbox, scenarios,
the weekly, puzzles, architect runs, or a save) on two pages in one process and compares
all of R, R.fl, R.ef, the PNI lists and S after every move and every day, and routing with
the kept tables against routing without them. tests/route-fuzz.test.js plays four.
scripts/mcp.js speaks MCP over stdio and loads the game from src/ in jsdom,
seeded, like the tests. Tools: new_game, load_save, export_save, status,
advance (with per-day timings), network, city, find, traceroute,
peering_check, hot_spots, actions, act, screen, panel, click, open_link.
Actions go through ACT as buttons do; files, slots, settings and the clock
stay out of reach, and nothing evaluates code in the page. Nothing is
hosted and the page is unchanged. See docs/MCP.md.
A subsea cut sends the nearest free ship of its region's maintenance club from its port, along the cable routes at about 500 km a day; the cut lasts while it sails, waits for permits in slow countries and splices (3-5 days shallow, 6-12 deep). Ships stay busy until home (S.fleet). With the club's ships all out the cut waits; a private charter of the quickest free ship costs $1M + $200/km sailed. Own subsea builds show survey/factory, then a lay ship at 150 km/day; a second lay ship can be booked before laying ($3M + $150k/day saved), and crews no longer speed up cables at sea. Ships are drawn on the map and can be tapped.
Review of #205:
- Prefer no longer switches the carrier-direct rule off for the whole network.
A POP with a Prefer-marked transit port keeps what Prefer says; every other
POP keeps the rule. compute(), traceroute and the reference loop in the tests
all use the same per-POP rule, and a new test covers it (New York and Chicago,
Prefer on one POP's Tell-Ya port).
- dayTick() cost. The harness is neutral (the same page on both sides differs by
1%), but the cost was real: the four extra rank passes inside compute()'s main
loop made V8 deoptimise compute() 954 times in 30 game days against 46 on main.
The direct ranks now run in a small function of their own, in the plain way the
reference loop does (their lists are one carrier's ports), so the main loop is
the one main has, and the setup moved out of compute(). 61 deopts instead of 954.
- Italian: "Porta qui" read as "port here" next to "Porta più vicina"; it is
"Instrada qui" now, and "Ogni POP instrada il traffico..." matches.
- Help text: says that real BGP carries the traffic from every POP (the shorter
AS path beats IGP distance), that Carry here does that here, and that Nearest
port is the default. All six languages.
Transit ports at a POP shared load by free capacity whatever the traffic
was for, so Level Tree's own traffic went out of a Tell-Ya port about
half the time. Traffic bound for a carrier, or for a network that buys
transit from it, now goes out of that carrier's port at the same POP
first (the shorter AS path). Each such port is still an ordinary transit
exit, so POPs without one use it when nothing nearer has room.
A new switch on each transit port, Carry here, pulls that traffic from
every POP across your network to this port. The default, Nearest port,
keeps using the closest transit port from POPs that have no port of the
carrier. The switch only shows when another carrier's port could take
the traffic instead.
Prefer-marked ports, Backup only, RouteScript and Auto-TE keep their say,
and any Prefer-marked transit port switches the rule off. Traceroute
follows the same rule and says why it picked an exit. The setting is
repaired in sanitize(); old saves load with it off.
The routing reference test now has several carriers sharing POPs, ports
that carry, backup ports and RouteScript, and its reference loop knows
the new ranks.
Series of Tubes now speaks German. site/i18n/de.js translates all 3,063
labels in informal "du". Networking terms that German engineers say in
English stay English (Peering, Transit, Dark Fiber, Wave, Port, POP, IX,
Route Server, LAG, PNI and so on); plain "fiber" is "Glasfaser" and
"fiber pair" is "Faserpaar". RouteScript syntax stays English, as in the
other languages.
German is registered in the language list and the offline cache, so a
German browser starts in German. Every per-language test now covers it,
and the German style notes are in the add-strings skill and in the
dictionary header.
A backup-only link next to a normal one still carried traffic, because
parallel links merge into one edge. The metric stays on the city pair
(S.pairMet), and any link can go on standby: it adds capacity that
fills only after the active links, so it idles as a backup, takes the
overflow when they are full and carries the traffic when they are down.
Build order no longer matters, and adding a link never idles the
others. Old saves keep each pair's lowest metric with every link
active, which is how they routed before. The panel says what a standby
link is doing, the redundancy list tags it, the map draws it dashed
while idle, and traceroute shows the active link.
The build swaps every label for a short id in the shipped page, the inlined
English and the translation files (scripts/lib/compile-labels.js). Labels
passed to N_() stay whole because saves keep them. Against 1.33: page +4%
gzipped (321 -> 335 KB), translations -40% (it.js 143 -> 85 KB); routing,
day, panels, dashboard, map and boot time unchanged.
- i18n-check rejects labels the build cannot see (outside tr/trn/N_)
- tests/pages.test.js plays the shipped page in English, Italian and French
- docs/adr/0003, add-strings skill, ARCHITECTURE, README, CONTRIBUTING,
CLAUDE.md describe labels
- docs/adr/0001: GitHub Pages is the only host, not open for renegotiation
- docs/adr/0002: source in many files, shipped as one page assembled at build
- tests/pages.test.js: builds stay static files Pages can serve, all local
references resolve, only release.yml deploys and only via the Pages actions
- tests/build.test.js: partial naming and layout rules
- new skills: github-pages, edit-the-game; add-strings, performance, release
and write-tests updated for src/, sorted dictionaries and --base-ref
- CLAUDE.md, ARCHITECTURE.md, CONTRIBUTING.md (incl. porting older branches),
README.md and the PR template describe the new layout
Adds site/i18n/fr.js with all 2541 strings, registers Français in the
language list (picked automatically for fr-* browsers), caches it in the
service worker, and runs the full screen tour and the Signal-link check
in French in the tests.
npm run perf -- --browser times the same workloads in Chromium as a desktop
and as a phone (390x844 at 3x, CPU slowed 4x), one browser per version. Each
run collects garbage first and ends by forcing layout and the map canvas, so
deferred work counts. The jsdom DOM is about 30x slower than a browser's and
has no layout, so the interface rows need this.
Shared machines swing single runs by 30% on identical code, CPU time
included. A row now counts as slower or faster only when it moved 15% and a
Mann-Whitney U test gives p < 0.01 (15 runs a side by default). Identical
code: no flags in three full runs; a routing slowdown planted in a copy is
flagged in jsdom and on the phone.
The panel workload renders both selections every run instead of alternating,
which split its samples in two. CI drives the runner's Chrome through
playwright-core, no browser download.
scripts/perf-bench.js times routing, a game day, an action, panels, the
NOC dashboard and the map on tests/fixtures/big-network.json (104 POPs,
950 PNIs, 832 caches), base and head in alternating turns, and says
whether routing results changed. perf.yml runs it on every PR against
the base; perf-comment.yml posts the table as one updated comment.
New performance skill, CLAUDE.md rule, PR template and docs.
A network with a hundred POPs and a thousand peering ports froze the game for
most of a second every day and after every click, because routing checked all
~1,800 exits for each of ~14,000 flows. Now:
- compute(): exits are asked once per network which networks they reach, each
flow scans only those, and the first ECMP group is found without sorting. The
graph runs on city indices with a typed-array Dijkstra (same heap moves, same
ties), routers and per-link traffic are kept in typed arrays, the diverse-path
check reuses one live-link graph, and R.ef is put together when a link is read.
Results are bit-identical to before.
- Panels and the NOC dashboard are patched in place (morph) instead of being
rebuilt, so the browser no longer restyles 1,500 rows every tick. Buttons whose
action changed are swapped out, keeping #44's tap protection.
- The Networks tab works out each IX port's sessions once instead of once per
network.
- draw() reuses screen-point buffers, a per-frame pair index and cached label
widths, strokes the dotted fiber line on the path it already has, and uses
native roundRect: about a third fewer canvas calls per frame, same output.