Cantorite: Harrowfall — Devlog

An npm-workspaces monorepo with a Vite + Three.js client and an empty simulation package; an orthographic RTS camera you can pan, zoom and rotate.

The founding requirement: a browser RTS whose simulation never touches rendering, so networking can be added later without a rewrite.

Click to select a unit, click to move it; it finds its way around obstacles with A* on a grid.

Roadmap: prove the command pipeline (input produces commands, the sim applies them) on the smallest possible slice.

Drag-box multi-select, simple formation moves, and basic combat: attack-move, health bars and death.

Roadmap: the core loop needs fighting before it needs anything else.

Resource nodes and a harvester that gathers, returns and earns credits on its own.

Roadmap: the economy half of the core loop.

Place buildings, watch them construct, and queue units at a factory.

Roadmap: base building and production.

An AI opponent that runs its own economy, builds a factory and sends attack waves.

Roadmap: single-player versus a simple AI was the proof-of-concept opponent.

Placeholder boxes become stock low-poly models loaded through an asset registry, tinted per team, with a distinct model set per faction.

Roadmap: swap primitives for free CC0 art without gameplay code ever naming a file.

Every interaction gets a touch equivalent: multi-touch camera, tap to select, cancel placement.

Requirement from the start: mouse first, but touch must be additive, not a rewrite.

A HUD, a win/lose condition, sound stubs and environment polish: the proof of concept is complete.

Roadmap: close out the first playable milestone.

Height tiers that change pathfinding and give the higher unit a damage bonus; later, ramps became the only way up.

Confirmed with the user up front: elevation is gameplay, not decoration.

More unit and building types from the full Space Kit, plus world decoration props.

Roadmap: the first "next level" expansion after the prototype.

A right-hand build panel covering the whole roster, with icons rendered from each model.

Reordered ahead of resource fields: a roster nobody can reach through the UI is not done.

The single crystal becomes a field of many crystals that deplete as they are harvested.

Roadmap: make harvesting look and play like a classic RTS.

Seeded, symmetric procedural maps and a pre-game setup screen (size, resources, credits, seed), built on the planet-select menu.

Roadmap: replayability beyond one hand-made map.

Rocks, craters, containers and barriers scattered over both the static and generated maps.

Roadmap: the board looked empty.

The debug grid goes, the ground gets a painted texture and real rock cliffs around every height change.

“Are you able to make the ground for the map more visually appealing…”

A separate, non-playable showcase map: two bases, rock formations, a monorail and a mixed city built from stock tiles.

A direct ask from the user: a purely visual map using every available asset.

Generated maps can include mirrored cities whose buildings block movement and whose roads are open ground.

“Can the game engine now use this to generate semi-random cities for the maps?”

Units acquire targets by themselves, attack-move and stances, muzzle flashes and explosions, real audio and under-attack alerts.

The first full playtest found two armies could stand inside each other forever: nothing ever set an attack target.

A default opening that actually earns money, right-click orders, and on-screen control hints.

Playtest finding: the default start had zero income and no way to learn the controls.

Team-colour decals, crystals that no longer look like scenery, shadows and legible elevation.

Playtest finding: you could not tell whose base was whose, or what was a resource.

Edge scrolling, zoom toward the cursor, pan bounds, selection modifiers and rally points.

Playtest finding: the camera and ordering felt unfinished.

The roster is re-costed, resources scale with map size, buildings have a build radius, and the AI defends and diversifies.

Playtest finding: economy and AI were flat.

Pause, a pause menu and end screen, elimination victory, a responsive layout, a minimap, settings and key remapping.

Playtest finding: the match had no frame around it.

Image-based lighting, a low sun with visible shadows, ambient occlusion, bloom and anti-aliasing, and units that face where they move.

“the graphics of the game are very basic… how do we go to the next level.”

Trees, spires and ground cover from the Low Poly Space Alien Worlds pack, and a mountain horizon outside the map.

“can we use these assets for world building — the bases and buildings can stay as they are”

The first level becomes the night Glacier world: indigo ground, glowing flora and a real planet behind the menu.

“make the game's first level look more like this style… really try your best”

Every plateau edge grows rocky outcrops, so raised ground reads as a ridge instead of a stamped step.

“Please also make the edges of elevated parts of the map look like mountains using the assets”

All nine biomes become selectable planets, and the menu restyles itself to the world you pick.

“this is the art style I want to work towards, different biomes etc.”

Each biome's ground slabs become floor patches and ledge caps, scree gathers under drops, and alien structures stand as landmarks.

“can we look at potential floor tiling in the biome's asset packs too? I think that might take it to the next level.”

You only see what your units see; unexplored ground is black, explored ground dims, and enemy buildings are remembered.

The last item from the original playtest that changed the game itself: the enemy base was visible from frame one.

Real MSAA, a subtle rim light, per-biome colour grading and moving cloud shadows.

“are there easy ways to make it a bit more nice looking? Eg post processing, edge smoothing…”

A subtle outline around your units and buildings only, so they separate from the scenery.

“Outlines for player interactive stuff might be good. Make it subtle.”

A miniature-style tilt-shift band and lit windows on buildings; a softer-shadow experiment was tried and rejected.

“As for the rest of the style improvements: let's do it.”

One biome-driven skin across the HUD, and a sidebar that shows cost, what a thing does, and why you cannot build it yet.

“Would be good if it follows the style of the rest of the game.”

Generated cities are built from the alien pack's structures, tinted to each world, instead of stock city blocks.

“Lets use the buildings from this asset pack for the city generation.”

A tone-mapping bug that made every biome's exposure setting inert is fixed, and the glacier ground gets contrast.

“The colors of the ground match everything else too closely… Maybe make it slightly lighter?”

The showcase becomes a set of biome-aware maps with a labelled parade of every model in the game and a citadel map around a user-supplied hero asset.

Neither presentation map showed the roster the game actually uses.

The citadel becomes both factions' Vehicle Factory, and model importing becomes a documented tool with a cameo generator.

The user asked for the citadel as the Vehicle Factory, and for the next linked model to take an hour rather than a day.

Large maps 2.9x bigger, a Relaxed AI that is actually relaxed, units that defend nearby allies, and crystal fields that regrow.

“the maps are way too small. A large map should actually be large. Like really large.”

No more units spawning stranded, a home-based turret for Relaxed, longer turret range, faint range rings, and shorter decorative rocks.

Direct feedback after playing the phase 39 build.

Maps grow again, models are rescaled against each other, and the camera opens tilted over your own base.

“maps should be way bigger. A small map should be the current large.”

Vehicles fire on the move, outposts work anywhere, fields spawn near every base, and phantom shadows from off-screen objects are fixed.

A large round of direct feedback with three annotated screenshots.

About 6,000 draw calls down to 1,114, with no shader recompiles during play and an identical frame; plus a render-scale setting.

Measured: two-thirds of all JavaScript time was the renderer recompiling shaders.

A roster plan for Directorate, Swarm and Umbra; Meshy-generated models first, then a switch to units and buildings built entirely in code.

“Let's make a style for each roster/faction. So each faction gets their own buildings and units.”

A Meshy-generated, animated Directorate trooper replaces the stock soldier, after two skeleton-cloning bugs were fixed.

The first model from the new art direction to reach a match.

Directorate vehicles, factory and harvester plus a Heavy Infantry mech; the Barracks and Vehicle Factory split production.

Continuing the phase 44 roster; found a rotation bug that had silently done nothing for units since phase 38.

5.49M triangles down to 745K: the new trooper went through the import pipeline, and unseen skeletons stopped animating.

Measured: 86% of the frame's triangles were 43 soldiers.

About 8.5 ms a frame recovered by culling the static world ourselves instead of walking 12,000 objects every frame.

Measured: 12,072 of the scene's 12,243 objects never change.

The model pipeline emits simpler versions and the game swaps them by on-screen size: 62% fewer model triangles at default zoom.

Measured: a 40-pixel soldier was drawn with 7,109 triangles.

Designed, measured and deferred: moving the sim to a Web Worker would cost more per frame than the sim itself.

Measured before building: a 1.13 ms round trip against a sim that costs under 1 ms.

A reported dark mass beside buildings turned out to be ground plants under newly placed buildings; they are now cleared.

The user reported it with an annotated screenshot.

An opt-in WebGPU renderer behind a flag, with a parity mode for fair comparisons.

three.js already shipped WebGPU sharing the same scene classes, so it could be tried without forking the scene graph.

The five custom shader patches move to TSL so WebGPU can become the only renderer.

The user reported that WebGPU was "rendering just fine, fps seems stable".

Real idle and run animation, a weapon and sound per role, muster points, randomised starts, a bigger economy, the Directorate HQ and an announcer.

“It looks okay but I have no feeling like I do using [a classic RTS].”

One sculpted heightfield with natural plateaus and cliffs, a painted classic ground, and crystal fields that grow organically.

“See how this world feels? The elevations feel natural, structures feel natural…”

Cities grow organically over flat ground with irregular blocks, parks and car parks, drawn from a new procedural civilian building set.

“I don't like rectangle/square blocks of cities only. It should be more dynamic.”

Six new units (artillery, APC, railgun skimmer, Colossus, gunship, bomber) and eight buildings, including aircraft, anti-air, repair, walls and an orbital superweapon.

“Each of the units that's not yet in the roster should be buildable/trainable.”

Eight parallel streams: fewer living cities, procedural world props, per-biome haze, bigger models, the menu as front door, aiming turrets, mountain edges and RTS camera controls.

Direct, with eight reference screenshots.

Layered combat effects, a living battlefield, a classic command interface, a perspective camera, veterancy, garrisons, new infantry, a smarter AI and several performance rounds.

“I'm giving you free reign… make it the best RTS you can imagine.”

Up to four players (2v2, free-for-all) with an alliance table, shared vision, per-player colours and AIs that defend their allies.

“Would it be possible (also for performance) to have more than 2 players? And teaming?”

Each faction gets its own passive, upgrades, buildings, units and powers; first the Field Commander hero, the Recovery Crawler, walls and the Transit Relay.

“What kind of upgrades/buildings/units/powers would you implement?… Write to a plan.”

The Swarm's look is set with two procedural creatures (Larvum skirmisher and Bio-launcher), shown in the showcase only for now.

Phase 61's plan puts the Swarm, its core roster and its regrowth specials next.

Every push is type-checked, sim-checked, built and held to an asset budget (74 MB, 683 files; the raw Meshy exports stay out). Pushes to p59 deploy a preview at a preview address; master will be production.

“Say i want this hosted somewhere else.. what would you propose?”

The Larvum is no longer cute: a lean, spiky four-legged hunter with a tusked jaw and green eyes, 7.7k triangles, four animations, team-coloured plates. The Bio-launcher is on its last bake.

“they need to be more fear-inducing. Especially the slug looks kinda cute.”

After a first all-Blender armour body (below), the user found a free hand-modelled, rigged cyborg soldier; it is being cut from 337k to about 12k triangles, repainted in team colours and wired in. Only the reduced game model goes into the repo.

“It still looks a bit messy for me. Would it help if you could download a base model and work from there?”

Eight fixes on three branches: garrisoned buildings glow in the occupant's colour, harvesters prefer fields over 100 credits, huge artillery and Rail Spire range, building repair for credits, a hover fix, railways that follow slopes, and homeworld resupply when every harvester is gone.

user's playtest list; homeworld resupply is the one exception to "no free income for whoever is behind".

Eleven research documents became eighteen phases: foundations, world events, a living map, economy, tech structures, map types, Boot Camp, the Harrow invaders, modes and replays, spectacle, water and air, a thirteen-mission campaign, minor factions, Conquest, and last, abilities and bases that move. The work is split into a world-and-rules track for the second developer and a factions-and-presentation track.

“It should become a full game in the end. You are the creative mind behind this.”

Every slot picks a faction; Command Points earned in combat pay for powers; Powers and Upgrades tabs in the sidebar; share codes carry the setup.

The plan for faction specials needed one frame all three factions plug into.

Seventeen creatures sculpted in Blender, from the Larvum to the Hive Titan, plus fifteen buildings grown in code: hive, pools, spines and resin walls.

“The more elite units should be more fear inducing like giants. Or elite soldiers. Not all 4/6 legged same body creatures.”

Thirty-one sleek black-and-magenta models, and the Veil: Umbra units that stand still cloak, and only detectors see them.

“Please do not make it too trashy and rusted. It is still a high tech game.”

The Aegis Pylon dome, the burrowing Sinkmaw and the Shade Obelisk; then 96 painted classic sidebar icons, one per slot per faction, from model screenshots.

“Something else I'd like in the roster is a unique defensive building per faction." And: "The icons should pop…”

The Directorate mech rebuilt in Blender on the infantry's own body: bulkier armour, an arm cannon and a shoulder cannon, a stomping walk.

“Probably base it on the infantry, just make it a bit more bulky. The current model is not it.”

The non-playable invaders: faceted ivory and quartz constructs with a gold glow, from the Gleaner to the Tiller titan, now real units in the sim with seams and Anchors you can tame. The wave director followed that evening (see the next page).

“In game it could do with some kind of golden glow or accents so it stands out.”

Four new worlds: Meridian (the Directorate's Earth), Hivemere (Swarm), Penumbra (Umbra) and Fallow (Harrow), each with its own civilian buildings, props and dormant structures. Fallow is guarded by dormant Harrow wards and a Bastion that sends waves.

“I'd like to also have a few faction specific worlds… Also civilian/noncombat buildings.”

A director that rolls events on a budget, with Calm, Normal and Chaotic presets. The Cant Storm grounds aircraft and kills the minimap; the Meteor Shower leaves new crystal; a derelict crashes midway between the bases for anyone to claim.

“Would you also be able to start on a first world event? E.g. [a storm] or something.”

Upgrades, passives and four powers per faction (the Directorate's Rapid Deploy drops a squad from a dropship), the Sensor Array, the Swarm's Clutch Well and larvae, the Umbra Infiltrator and Hologram Projector, and a classic skirmish AI with a new Easy level.

The phase 61 plan, built mostly by the second agent while this one made the art.

Defences build 2.5× further out, a match-end rule: lose your bases, lose the match, enemy pings, superweapon timers, diagonal walls, units that hold at weapon range, bomber sorties, faster crystal regrowth and a Drill per faction.

The user's playtest feedback, split between the two agents.

A replayable late-game bench, then no shader churn (83.1) and a shadow pass on a budget (83.2): cached static shadows and one simple shadow shape per unit, about 5–7 ms off a busy frame.

A late-game four-player match ran at 47–58 ms a frame; the target is 25.

The project gets its own name, after the crystal everyone fights over and the invaders who come to reap it. Names borrowed from older games were replaced with our own (the Swarm's Clutch Well and Sinkmaw, the Cant Storm), and a Credits screen lists what the game is built with and under which licence.

A game that might one day be sold needs a name and an identity of its own.

Watchtowers to hold, salvage crates, guardian lairs of native creatures, rocks and spires you can bring down, and crystal fields that grow when nobody mines them. Tech structures to capture: a Cant Siphon for income, a Relay Mast for vision, a Field Clinic and a Motor Pool to heal and repair. And the Harrow's wave director: mining too much draws them, seams tear open in the crystal, and their waves come mirrored so no side is favoured.

The Vesper Reach roadmap, phases 66 to 71.

Crater Arena, Crystal Rush, King's Hill and Twin Rivers, each checked for fairness with a badge on the preview. Six quick modes (King of the Crystal, Regicide, Superweapon Race, Treaty, Frenzy, Sudden Death), a long list of mutators, and replays you can scrub and watch from any side.

Variety between matches, without a map editor yet.

Five short lessons for new players, then Hold the Light: thirteen Directorate missions across the nine worlds on a star map, with a fork in the road, a choice that changes who fights beside you, veterans who carry over, and upgrades between missions. After the first playtest, every mission marks the units and places its orders are about, on the ground and with a ping on the map.

The first fifteen minutes decide whether someone keeps playing.

Each Directorate role gets an ability on one key: frag grenades, smoke, jump jets, star shells, a field hub on the APC. Units take more damage from behind and can retreat. Around it: a slow-motion final blow, music that follows the fight, named veterans, and a profile with commander levels, 28 medals and cosmetics you earn.

Depth per unit, and a reason to come back.

Shallows, deep water, tidal flats, ice and lava, crossed by bridges and by hover units. Day turns to night, the Saltmarsh tide rises, Ember's vents erupt, and the Apex sleeps in its den until something wakes it. Six minor factions of the Vesper Reach live in settlements with their own moods; win them over or make them enemies.

The world, round two.

Every base now runs on lumen: generators power it, and too little power slows production and drops the defences, so generators become worth raiding. Buildings can be sold for part of their price, and the HQ can fold up, drive somewhere else and unfold.

Bases that are more than a wall of turrets.

A new interface called Bezel: a quiet graphite frame, your faction's accent colour, and one typeface throughout. A new main menu around a turning cantorite crystal, a loading screen, a skirmish briefing, a pause sheet and a debrief. The build menu moved into a tray at the bottom, and phones and tablets get their own layouts, with a compact action bar and a More button.

Designed as mockups first, then approved by the user before a line of it was built.

The rest of the Harrow roster, and the Tiller: a walking titan with three phases, Tines that lock onto a target, a sweep you can see coming, and a boss bar. Fallen Chaff rise again, and elite Harrow carry marks that show what makes them dangerous. A sandbox page lets anyone try the fights.

The invaders needed a proper boss.

The Colossus, the Hive Titan and the Effigy: one giant each, announced to every player when it is being built. After the playtest they are about a third bigger and a quarter stronger, while a fully promoted one is no longer close to unbeatable.

“The strongest unit - the giants - … all the way upgraded to the highest rank … It's almost unbeatable.”

At night only your own base lights the ground around it, and each faction can build a light tower that shows the dark around it. After the playtest its light reaches twice as far.

Night should hide things, and you should be able to fight that.

Small maps are now 128 cells across and Medium 256 (they were 88 and 176), so Medium sits between Small and Large. Bridges were rebuilt so units cross them properly, camp militias defend themselves, Harrow pylons under a dome can be brought down, units stop firing at what they cannot hurt, the Refinery's build area grows with the map, and the Drill pays a little more.

“A medium map should be a tad bigger. The difference between large and medium is huge.”

A base short of power now shows which buildings stopped and what to build to fix it. Each faction can also protect its generators with an upgrade: the Directorate's Generator Aegis puts a shield on each one, the Swarm's Deep Roots let its glands burrow out of sight, and the Umbra's Phantom Reliquaries raise a hologram twin beside each generator.

“can we have a way to protect power a bit more?”

A bomber run now hurts a base instead of deleting its HQ, and each faction gets a heavy anti-air defence after research: the Skyguard Battery, the Hailspore Colony and the Nightfall Choir, about five times as deadly to bombers as the old turret. Detectors also notice a cloaked unit only closer in than they can shoot it.

“1 or 2 bombers can destroy an HQ.”

Buildings can repair themselves five seconds after a fight, for 10 % more than a manual repair; switch it on for the whole base or per building. And units no longer chase a fast raider across the map: past a generous leash they give up and walk back to where they stood.

“Units will infinitely follow attacking units from their starting point.”

AI matches that ended in a draw two times out of three now do so a quarter to two fifths of the time. The Umbra AI, which never won, now wins about a third of the matches that are decided; the Aggressive AI beats the Standard one; and every AI answers an enemy experimental.

A skirmish should end.

Three new models, one per faction, for an HQ on the move: it folds up, drives and unfolds on the spot. Painted icons for the Drills, the light towers, the Black Market, the Morphling and the Devourer.

The HQ that moves needed a body of its own.

The cause: every wreck stayed in the scene as its full, unmerged model (about 500 parts for a harvester) until it sank, so a few fights doubled the work of every frame. Wrecks off camera are now skipped like live units, and they merge back into one piece once they settle: twelve wrecks cost 3.9 ms a frame instead of 15.1 off-screen, and 9.2 instead of 20.5 in view. A slow machine now plays a little slower instead of freezing. Earlier that day, cloak detection got faster too, taking a late-game simulation step from 20.9 to 12.6 ms.

“It becomes unbearably slow after a while.”

Cantorite now also builds as a Windows app. On a laptop with two graphics chips it picks the fast one by itself, which made frames 3.5 to 4 times faster than in the browser on the built-in chip, and it saves to files. A preview for now; not released yet.

The first step towards a store release.

The match from the report was replayed at real speed for sixteen game minutes, in the browser and in the desktop app, with every part of a frame timed. The simulation turned out to be cheap: under 2.5 ms a step all match long, with nothing in it growing. Nothing leaked either. The problem was the work of preparing each frame for drawing: the main thread ran at 100 % while the graphics card sat nearly idle at 1 %. That is why lowering the quality from High to Medium hardly helped: pixels were never the bottleneck. From minute 6 frames took 33 ms (30 fps), and in the big fight at minute 15 they took 50 ms (20 fps).

“It becomes unbearably slow after a while.”

What grew was the number of things in the scene, and it jumped after every fight. A destroyed vehicle left its whole model behind, broken back into its separate parts so it could fall apart (about 500 for a harvester, 491 for a heavy tank), for around 50 seconds. And unlike living units, wrecks were never skipped when they were off screen. Twelve wrecks took a frame from 11 to 20 ms. Now a wreck merges back into one piece three seconds after it falls, and wrecks off camera are skipped: twelve of them cost 3.9 ms instead of 15.1 off screen, and 9.2 instead of 20.5 in view. In the desktop app over the full sixteen minutes, minutes 6 to 14 went from about 30 fps to a steady 60, and the big fight at minute 15 went from 20 to 30 fps.

Found by counting what was in the scene every thirty seconds.

Every separate thing the game hands to the graphics card costs time on the main thread, so the game "bakes" models: parts that never move relative to each other are merged, and identical models are drawn in one batch. A new switch turns all of it off, to show what it is worth. A late-game base costs 3 932 draws and 78 ms a frame without it, and 454 draws and 31 ms with it. On a phone-sized screen it is 5 839 draws against 350. After the wrecks, the largest remaining cost was the scenery: rocks and plants were drawn once per part in every patch of the map. They are now merged per part across everything in view, which took a late base from 638 to 433 draws with an identical picture. Health bars now take two draws in total instead of one each, and the fog-of-war check went from about 2 ms a frame to under 0.1.

Every change was checked against a pixel-by-pixel comparison, so nothing looks different.

On a machine that cannot keep up, the game used to try to catch up by running more and more simulation steps each frame, which made the next frame even later. Now it runs at most four steps a frame and lets the game slow down a little instead, with a small "SIM BEHIND" tag in the corner while that happens. Phones and other low-memory devices draw at 30 fps, which cut the main thread's load from 77 % to 46 % without slowing the game. On Medium, unit shadows update every second frame (4.7 ms down to 2.1 to 3.0), and on Medium and Low the resolution steps down a little only when the graphics chip is what holds the frame back.

A slow device should play slower, not stop.

On an iPhone, a web page that uses too much memory is simply closed by the system, and Safari reloads it. The loading screen built and uploaded every model before the match, which peaked at 1 023 MB. A new low-memory mode switches on by itself on iPhones, iPads, phones and devices with little memory: Medium quality, smaller textures on units, no copies kept of images once the graphics chip has them, fewer helper threads, and only the models each faction in the match actually needs. The peak is now 445 MB. Measuring it also turned up 1.2 GB of throwaway data created while working out the size of animated models; that is now computed without it, with identical results, on every device. Desktops look exactly as before.

“On iOS phones the game seems to crash before loading completes often.”

A common suggestion for speed is to move the simulation to its own thread. Measured on the late game: the simulation and the AI together take 0.7 ms a step, and copying the whole game state across to another thread costs about 1 ms. On a desktop that would gain almost nothing. On a slow phone, where a step takes around 30 ms, it will matter, so it is planned for phones, sending a compact 3 KB summary each step instead of the whole state.

Measure before moving work around.

The research question: would wrapping the game in an app, or rewriting it natively, make it faster? Mostly no. App wrappers are packaging: they bring files, store listings and a proper window, but they run the same browser engine, and the desktop app measured exactly the same as the browser. The one real gain is on laptops with two graphics chips, where the desktop app always picks the fast one (3.5 to 4 times faster frames than the browser on the built-in chip). On iPhones every app that shows web content runs on the same engine as Safari, with the same memory limits, so no wrapper could have fixed the crash; only using less memory could. A native rewrite would take months for problems the fixes above already solve. The plan that came out of it: keep the desktop app for the PC release, bring Android first through a light wrapper that runs the live site in the phone's browser engine (the project for it is ready), and follow with iOS once the memory fix has passed on a real iPhone.

“Research our options to wrap this game in an app wrapper and to use more resources like a native app would.”

A side is out once it has no base left: no HQ, no production or economy buildings, no Mobile HQ on the road. Turrets, walls and stray units no longer keep a beaten player alive; they self-destruct, so the map is clean when the match ends. A side down to its last base building has it revealed on everyone's map, so there is never a hunt for one hidden refinery. AI test matches that used to run to a draw now all finish, in about 16 minutes on small maps.

“I destroyed everything. Nothing was left as far as I could see and still didn't win.”

The HQ has three times the health (3 000), and a free guard turret that hits both ground and air and rebuilds itself 45 seconds after it falls. Losing the HQ no longer ends everything: production and research run at half speed until you rebuild it, for the price of a Mobile HQ.

“HQ should be way tougher. Like 3x … maybe a free turret/anti air next to it that rebuilds automatically.”

Units now see half again as far: a soldier 19.5 instead of 13, a scout 25.5, a gunship 27. Enemies no longer vanish the moment they step out of range. Buildings, night and cloak detection are unchanged.

From the playtest: units saw less than the camera shows.

Three new World options for infantry fights. Light cover: soldiers beside rocks, walls, rubble or a wreck take 30 % less from direct fire. Suppression: machine-gun fire pins infantry down, halving their speed and slowing their fire until it stops. Thickets: patches of brush on the forest and ice worlds hide soldiers from anyone not close by, until they fire; tanks flatten a path through, and bombs burn them away.

Infantry needed more to do than stand in the open.

Camps start with three defenders and grow to six over the match, refilling their losses. The armed camps now raid players they are hostile to with war bands sized by the match time and the map, one band at a time, and a warning on the ticker. Peaceful camps never raid. Still an idea for later: one or two walled strongholds per map to win over.

From the playtest: two or three soldiers per camp is nothing in this game.

Capturable tech structures and minor-faction camps used to pile up in the middle of the map. They are now spread in every direction, near and far from each base, and still mirrored so every player gets the same chances.

From the playtest: everything clustered in the centre.

The music is now composed and rendered in advance instead of synthesised live: one crystal motif in every faction's voice, with calm, tension and combat loops, a menu theme and victory and defeat cues. When the invaders arrive, their own theme of choir, crystal and drums takes over, and a quiet hum rises as your mining draws their attention. Fallen Chaff now visibly climb back to their feet.

The Harrow should sound as alien as they look.

Online play now does two against two: the host picks 1v1 or 2v2, each player picks a team, and the match starts when everyone is ready. Tested with four browsers playing a full minute in perfect step. Built, but not live yet: it switches on once its server is.

The simulation was kept deterministic from day one for this.

The devlog moved out of the game into a site of its own, so its screenshots no longer weigh on the game's download. Old links still lead here.

Keeping the game lean.

On the alien worlds the cities are now the ruins of an ancient civilisation, built in code from the world's own stone and growth: causeways that crumble toward the edge, walled precincts, fallen blocks. The earth-like worlds keep the European city, now marked by the war (burnt-out cars, craters, checkpoints, rubble), and every city's edge wears into the land around it. Fallow and Penumbra got calmer, more readable looks, and street lamps pool light at night.

The owner's ask: cities should fit the map they stand on.

Cities are laid out to a plan that suits their world. The ancient ruins centre on a paved court with an obelisk and monuments facing it, and an avenue running toward the middle of the map. European towns have a tight old core round a market square with its church. Courtyards fill the insides of blocks, some lots are bombed-out ruins, and the roads out of town lead to a resource field or up a ramp.

The owner took four of five proposed generator changes and asked for them as the default, with no new menu option.

Short lines like "Yes, sir" and "Mustering" were recorded inside a longer carrier sentence and cut out of it, and most kept the first sound of the next word. All 209 of them are re-cut in the pause before that word, from their original takes, and a new check listens to the last moment of every shipped line.

The owner heard "Yes, sir. Thhh".

When iOS takes the graphics memory back (often while Safari is in the background), the game now restores the match in place in a few seconds instead of asking for a reload: textures are redrawn, lighting, shadows and the minimap rebuilt, and the match is paused and saved while the graphics are gone. Phones also use less memory while loading.

The owner, playing on an iPhone: "That needs to go asap."

From an iPhone playtest: no accidental page zoom, a building panel you can close, a compact view of a large selection, one button for all powers, autosave with a Continue button, and a filter that selects only fighting units.

In progress.

Android in Google Play through the light wrapper, which is ready; the iOS app follows once the low-memory mode has passed on a real iPhone. A desktop release is prepared alongside.

User: release on the App Store and Google Play.

For phones: the simulation and the AI move to a second thread and send a small summary each step, so drawing gets the main thread to itself. Drawing on its own thread may follow.

From the research above.

Every spoken line as recorded audio, then an announcer of its own for each faction.

The voices have been chosen.

Cursors in the game's style, a marker on the ground while you aim a power or an ability, and a gold button for superweapons.

From the playtest.

A grid in the World options to switch each world event on or off. On the table as an idea: one or two walled minor-faction strongholds per map to win over.

The rest of the world, round two.

The standard AI keeps its power up instead of running its base dim, and the aggressive AI plays as the harder opponent for every faction.

From test matches between the AIs.

A guard model of its own for each faction, a gun and a missile pod, instead of the basic turret it borrows now.

The guard hits air and ground and should look it.

Survival, Last Stand and Crisis modes around the invaders, then Conquest: a campaign across some two dozen territories against rival commanders.

The rest of the long roadmap.

A Claude model plays against you: it sets strategy every 10-20 s from a fog-of-war report, and the scripted AI carries it out.

“would integrating an AI agent in the game itself be an option to let it play against me?”