The 2D map keeps a canvas per picture, and the list was capped only by its length (1 400). A long zoom on a big
network paints every city again at each step of its scale, so a pinch from 0.6 to 3.2 and back held 181 MB of
pictures (the TTL save: 179 MB) where one frame uses 2 to 23 MB.
At the end of a frame over 48 MB, the pictures it did not use go, those unused the longest first, down to 36 MB, and
each canvas let go is emptied (0×0) so its pixels are freed without waiting for the collector. The frame's own
pictures are never let go, so a screen that needs more keeps what it needs. Finding a picture now marks it as used.
Measured in jsdom on the same pinch (1280×800 at 2×): 43 MB at most (TTL: 40 MB); 4 781 pictures painted against
4 083 (TTL: 5 549 against 6 030); at most 321 painted in one frame, the same as before (the zoom's own steps).
The card's atlas is unchanged: it has its own fixed size.
Given up while its context still lived (memory refused, a setup failing half-way), the card kept the atlas (64 MB at
4096 px), the tubes' points and the rest while the 2D map began painting its own pictures. glKill() now deletes the
biggest and lets the context go (WEBGL_lose_context, where the browser has it), unless the browser already lost it;
glInit() keeps the context from the start so a failed setup is let go too, and the loss event of a context let go on
purpose does not give the card up a second time. Always 2D still keeps the card for later.
The atlas, the land, the tubes and the cables read getError() after they are sent, but a picture sent into the atlas
(texSubImage2D) was marked as sent at once: a card short of memory that only set OUT_OF_MEMORY left it kept and
drawn from an atlas that never got it. A frame that sent new pictures now asks once, at its end and before drawing
them; a refusal gives the map to 2D in the same frame and forgets the card's pictures. A frame that only reuses the
atlas still asks nothing.
The card filled the land by an odd/even count over all the coastlines at once, so two coasts that overlap cancelled
out into sea: Timor, drawn by two coastlines, lost most of its inside and its shadow when zoomed in. The 2D map fills
each coast on its own. The fan now turns every coast the same way and the stencil counts windings (+1 and -1 by the
triangles' facing, filled where not zero), which is the 2D canvas's nonzero rule: an overlap counts 2 and stays land.
Same number of draws.
The 2D map paints its land into an offscreen picture. In Firefox that picture was a software canvas
(willReadFrequently), because Firefox on Android garbled big paths on its graphics card (#43). Firefox on a computer
had it too, and a software canvas is sent to the card again on every frame, at a cost that grows with the picture:
the whole world at mid zoom, four screens of it zoomed in.
Only Firefox on Android keeps it now (softCanvas()). Measured in Firefox 159 on Windows (RTX 4090, full screen
2560x1440), 2D map, frames per second at zoom 0.6 / 1.1 / 1.6 / 3.2:
- big-network fixture, main: 56 / 37 / 34 / 26; with this change: 64 / 49 / 51 / 78;
- the 183-POP save, main: 32 / 25 / 25 / 22; with this change: 36 / 26 / 33 / 55.
(Measured with the change injected into main's page; the land picture is the same code on this branch.)
tests/world-layer-firefox.test.js: which browsers get the software canvas, and that the land is an ordinary one
elsewhere.
Settings › Viewport info is now Diagnostic info, and the frame meter of the WebGL work joins it: one readout in a
corner instead of two, for bug reports. It shows:
- what the screen tells the game, as before (sizes, visual viewport, safe areas, standalone, the tab bar);
- what draws the map: WebGL2 and the graphics card's name, or 2D and why (Always 2D, a puzzle, no WebGL2, a
processor pretending to be a card, a lost context, a card out of memory: glKill() keeps the reason);
- how fast: frames per second, the 95th percentile and slowest frame, draw()'s mean time, the speed and zoom, and on
the card its segments and packets.
?vp and ?fps open it, ?novp and ?nofps close it, five taps on the date toggle it, and a device that had the frame
meter on keeps it. Its text stays technical and in English: it is pasted into reports. On a 360 px phone it wraps
inside the screen and never takes a tap. The setting's name and description are translated.
tests/diagnostic-info.test.js: the setting, the readout's parts, each reason for 2D, the frame figures, the address
flags and the old setting, the taps, and the phone layout.
A card short of memory may refuse a texture or a buffer without throwing: the call only leaves OUT_OF_MEMORY for
getError(), which nothing read. Refused, the 64 MB atlas drew every city, label and marker as a black box while the
map stayed on the card (seen in Chromium by refusing it).
The big uploads now read it (glMem()): the atlas, the land for each zoom level, the tubes once a day, the cables.
The atlas tries 2048 px (16 MB) before giving up; any other refusal is the card's error, so draw() gives WebGL up
for the session and draws the same frame again in 2D. getError() is never read on an ordinary frame: each read waits
for the card.
tests/webgl-out-of-memory.test.js: a 4096 px atlas refused gives a 2048 px one; both refused, or the land or the
tubes refused, leave the map in 2D in that frame with no glitch for the player; panning reads no error.
Cities, markers and labels are now pictures painted once on a canvas of their own. Drawn straight on the map,
main's dashed rings (the selected city, the cities feeding a selected tube, a flood), the cones' tips, a sleeping
POP's "zzz" and a working cable ship's spark took the state the tubes and labels left on the map's canvas: round
line ends and corners, text centred on its middle. A new canvas starts with square ends and text on its baseline, so
the rings came out with square dashes and the "zzz" sat higher. Every picture now starts with that pen (sprPen()),
on the card's shared canvas too.
tests/map-2d.test.js holds the 2D map to main, for players who keep it:
- for the same state, a frame drawn from the kept pictures is, call for call, the frame painted with no cache at
all: over every state a POP can be in, a city without a POP, random views of the big network with days passing,
and seeded random games after every move and day (the route-fuzz driver, two pages in lockstep);
- every tube's width, colour, dash and packets follow main's formulas;
- the 2D map never asks for WebGL, never draws a picture without a canvas, keeps one canvas per picture, draws as
a 2D map after the card is switched off or lost, and paints again at a new pixel density;
- money pop-ups fade as before; every picture starts with the right pen.
Planted bugs (incomplete keys, a stale key memo, the tubes' looks kept without the routing, one shared 2D canvas,
no pen) each fail it. A pixel comparison with main in Chromium (desktop and phone, the fixture, a late-game save and
states from random games) found the pen and nothing else beyond sub-pixel edges.
A frame joined a key for every city, label, quiet dot and stamp it drew, built a closure
to paint each even when the picture was already there, concatenated a prefix onto every
key, made an array per city per copy of the map for its place on screen, and rebuilt
the waves' and the clouds' arrays. Keys are now kept per city (packed into one number for
the yes/no parts), per label text and look, per stamp and scale, and joined again only
when the picture changes; pictures already painted are found before any closure is
made; the card's and the 2D map's pictures are two maps, so keys need no prefix; the
cities' places and the POP lights are computed in place; the waves and the clouds are
written into arrays kept from frame to frame.
Sampled heap profile, TTL save, RTX 4060 Ti: allocations under drawFrame 68 -> 37 MB/s
at x1 and 62 -> 43 MB/s during a pan/zoom sweep; under glBegin 5.4 -> 0.5 MB/s.
On the card the land was most of a frame's work (the stencil fills twice, the sand, the
edge), and it scales with vertices, not pixels. The land is now built per zoom level
(powers of 2) with each coastline simplified (Douglas-Peucker) so that no point moves
more than 0.25 px at the level's zoom; zooming in takes the finer level at once, zooming
out keeps it until the zoom is 10 % under it. COAST, onLand() and the 2D map are
unchanged.
The scale of a city's picture is in steps of 1/24, so a pinch repainted every city every
~4 % of zoom: hundreds of pictures a frame, and the atlas full and reset over and over.
On the card, while the zoom moves, pictures that grow with it (cities, ships, badges) keep
the step of the scale held when it began, stretched by up to 1.25x either way by the
card's filtering, and take a new one beyond that; 250 ms after the zoom stops they are
painted at their own step again, and a still map is drawn once more for it. The 2D map
paints every step as before.
Every sprite kept its own <canvas> once it was in the atlas: a long pan and zoom made
~20,000 of them, then a frame of one to two seconds, then CONTEXT_LOST and a silent
fallback to 2D (4 sweeps out of 5). The card's pictures are now painted on one reused
canvas and sent to the atlas at once (texSubImage2D with a sub-rectangle); a full atlas
restarts and takes the picture at once. The 2D map, and a picture too big for the atlas,
keep a canvas each. The card's and the 2D map's pictures are kept apart, and giving the
card up drops its pictures.
With the perf benchmark now waiting for the surface the map was drawn on,
headless Chromium showed what WebGL costs without a graphics card: its
SwiftShader renderer spends about 8 µs per instance, so the coastlines alone
(29k segments, twice) took about 480 ms and a big map 1.4 s a frame, where the
2D canvas takes 46 ms. Players get the same with a blocked driver, a virtual
machine or Mesa's llvmpipe. glInit() now reads the renderer first and keeps
the 2D map on SwiftShader, llvmpipe, softpipe, lavapipe, WARP or any
"software" renderer; GLX.soft lets the GPU benchmark measure them anyway.
- A sprite or disc drawn as the queue grows was written into the old array and
lost: the record is now taken before the array is read.
- A city's picture without an attack shared its cache key with step 0 of a
DDoS ring (always, with animations off): no attack is now phase -1, and
cityStatic() reads only what the key holds.
- Back at the title (or on a damaged save), drawFrame() returned before hiding
the card's layers, so the last network stayed behind the title: the layers
are now shown or hidden on entry.
- An error from the card (frame start, tube upload, atlas, sprite flush) now
gives WebGL up and draws the same frame in 2D, once; the 2D context, the
sprite queue and GLS.frame are reset whatever fails. Other errors still
reach tickErr() and leave WebGL on.
- Selecting a tube no longer rebuilds and uploads every tube, packet and
point (about 230 KB on the big-network fixture): its halo is its own small
buffer.
- 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.
The stars become three goals; the current one says what is wrong and what to do. A red ring marks a city still short, a pulsing ✂ the link whose cut hurts. Cities far from the puzzle are left off the map, and idle stepping stones are nameless dots until tapped, used or zoomed in on.
ZMAX goes from 16 to 32. A club ship near its home city is drawn at least
46 px (times the icon scale) out from the city's marker, along the line from
the city, and a tap on a city's own marker picks the city before any ship.
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.
Before a game starts the click handler only runs allowlisted actions, so Help's
Keyboard shortcuts button did nothing on the title screen. Allow it, and close
the list with helpClose, so ? returns to the title instead of leaving an empty
map. popStep, zoomStep and keys are views for the MCP server.
A fourth Normal preset, gentler than Change Freeze: calm chaos (so no DDoS
and no chaos monkey), quick builds and no market, plus four new rules that
only it softens: sea cuts mend on their own clock without cable ships, no
conference season, a five-year Ministry of Tubes rescue loan instead of
game over, and customers who sour half as fast. The new rules are written
into S.rules only when they differ from the classic game, so older networks'
rules stay byte for byte what they were.
Quiet notifications is a new setting for every mode, on by default in Best
Effort: toasts the day itself raises pop up only for trouble and milestones,
the rest goes to the NOC log; what a click says still shows. Best Effort
badges count as Freeze-dried. The preset picker is four wide, two by two on
a phone in portrait.
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.
Adapted to main: Globe is main's Globe-Trotter Telecom (now a Philippine
carrier in Cebu and Davao) instead of a second AS4775, the new Philippine
networks buy transit from NTT and Tata like main's other island networks,
cities and networks carry their country, and Philippine transit sits in the
new pricing. Cable names no longer pile up where short cables crowd.
Five sizes from a city's eyeballs (P: under 15, 100, 350, 800 and up). Far view: dots grow with size and megacities glow. Near view: the skyline goes from a hut to five towers with an antenna. cityRoom() lets the selected city, your POPs and cities with an offer, outage or puzzle part claim space first, then the biggest; a city with no room is a small dot without a label until you zoom in, still tappable. IX tags for exchanges you haven't joined are smaller, and on outposts and towns only from zoom 1.5. Phones draw city icons at 82% (taps unchanged).
The eyeballs, content and circuit markers sat on the IX badge of cities without a
POP yet (reported by Thomas Kjær). Architect's mission-critical marker is now the
siren like everywhere else.
Two tubes between the same cities read as one line until clicked (reported by
Thomas Kjær). Every puzzle now draws a shared trench as a bundle with a '2× one
trench' tag and separate paths bowed apart with a '2 paths' tag, at any zoom.
Tubes in one trench are drawn as strands of one bundle with an orange '2× one trench'
pill; separate paths bow well apart with a green '2 paths' pill. Cut tubes no longer
show the backhoe: the monkey does the snipping.
Earned on the monkey's first visit instead of any fiber cut; the badge keeps its id,
so ones already earned carry over. Monkey-proof gets a shield. The monkey's cheer
also cancels a spray still pending from its last hop.
No numbered tags beside cities, no "coming up" plan on a city card and no hint
to build there early: you design for today's brief and find out what comes next.
Now and then a chaos monkey moves in for 120 to 240 days and snips a tube every
couple of weeks, weak spots first, instead of the usual random outages. Outlast
it without a customer noticing for a bonus and the Monkey-proof badge, or throw
it a banana. Drawn on the map in shades; gameplay rule F4 with its test.
A hand-drawn monkey in the map's outline style replaces the emoji: sunglasses,
a spray can, a swaying tail, and poses for running, snipping, dancing and cheering.
Three runs (Euro Run, Coast to Coast, Pacific Rim) on the puzzle machinery: briefs add sites,
raise demand, make sites gold, ask for rings or latency caps, or close a city; earlier builds
stay paid for, and the backhoe review cuts every trench and cable in turn. By tkjaer, from
seriesoftubes-architect.
Every puzzle now carries demand (eyeballs pulling from a content origin, circuits
with latency caps, best-effort flows) routed by compute(). Stars: delivered,
within budget, N-1 under load. Routers, ports, metrics, standby and the
redundancy check work in puzzles; the whole world is open. Weekly schema 2
carries the demand; scripts/lib/puzzle-solver.js sizes and searches builds.
bezGeom() measured every city against every candidate curve, for every link the
first time the map drew it. A curve stays inside the box of its end and control
points, so a city further than the clearance it needs outside that box always
measures as clear; leaving it out gives the same curve. About 4x faster, and
the curves on the big fixture are byte-identical.
The sea waves, hills and the far view's islet list test points against every
coastline polygon while the page loads: about 0.8 s of a throttled phone's load.
A point outside a ring's box is outside the ring, so nearly every test stops
there. Waves, hills and the far view's islands come out byte-identical.