Alex CoulombePresents
← Harborfolk

Development diary · browser RTS

People make
the harbor.

A small strategy game for my kid, built around believable scale and people you can see doing the work. Notes from the playable builds, with the experiments and rough edges kept in view.

v0.3 Towing and dockside repair · shipped

Permalink ↗

Homeward bound

Alex wanted crippled ships to make it home, with people patching the hull and pumping out the water. Homeward Bound starts with a damaged 128-metre Leviathan whose machinery is disabled. Select your ships, press T and click the ironclad. Five fully crewed cutters approach separate positions on the open-water side, send line handlers to stern winches, rig their ropes and begin the slow tow. The hull keeps its dimensions, damage and crew throughout the journey.

The pulling rules give a cutter one towing unit, a Sovereign three and a Leviathan eight. A cutter needs one unit, a Sovereign three and a Leviathan five. Selection uses enough eligible ships to do the job. Helm and propulsion still need their living crew, and handling a line takes someone away from the guns. Underpowered or crewless groups cannot move the target. The tow progresses at 0.8 metres per second for an ironclad in calm water. Flooding, rough seas and incoming fire continue to matter, so reaching safety takes preparation.

Each side now has a 160-by-four-metre repair pier, one berth, six named dockworkers and eight thousand repair supplies. Four shipwrights walk to lifting platforms beside the damaged hull; two pump operators reach their pumps and run hoses aboard. People retain the same 1.717-metre body dimensions as the sailors. Open Shipyard and click a worker’s name to watch closely. The repair camera looks down the gap between hull and pier so the hammer and work platform remain visible.

Shipwrights spend one supply for each hull point restored, while their patches reduce the leak. Pumps remove actual water, letting the unchanged hull rise in the sea. Empty stores stop repairs while staffed pumps continue. Machinery recovers once the hull is at least 50% of maximum hull health and its leaks are closed. Lost sailors stay lost. The berth holds one ship at a time; tugs release their lines on delivery, and the repaired vessel can cast off and sail under its own power. An intact ship can also return to the dock without a tow.

Towlines, rigging jobs, machinery, worker positions, hull work, water and supplies save together. Both co-op players can issue towing and cast-off orders. The multiplayer checks save during the tow, disconnect, load and invite again; then repeat that cycle during dock repair. Each stage resumes the same ships and people. Towing, pumping and dock work can finish after a battle result while cannon combat remains stopped.

Recordings includes Homeward bound: six minutes and thirty seconds in 5,542 bytes. Four legal tow, resume, cast-off and movement commands take the authored damaged ship through recovery, repair and departure. The starting damage belongs to scenario setup. Subsequent changes come from the real simulation, including finite stock consumption and the final sailing order. New recordings use .7. Frozen .1 through .6 recordings keep their earlier simulation paths, including resets and fleet reconfiguration; earlier fire, storm and rescue examples remain unchanged.

All 346 automated tests and the production build pass. Six public browser gates cover real click/keyboard orders, visible rigging and repairs, exact saves, WebGL/WebGPU, narrow controls, departure and complete Node-to-Chromium replay. Four real two-process co-op gates cover shared workers and stock, guest orders, named saves in both stages and cast-off. Existing replay and flooding browser checks pass, and public runtime files match the tested build. Fresh review corrected the repair camera and stale recording-version wording. A final geometry check caught a tug crossing the pier; the corrected formation keeps to open water, and collision tests cover the pier as well as other hulls. The rates and pulling-power rules are authored for play; there is no rope elasticity, force, torque or buoyancy solver. Dockworkers are a separate named roster and are not combat targets. Harbor economy docks keep their previous system. Physical older devices, sustained large-fleet performance and repeated family balance tests remain open; screenshot FPS is not a benchmark.

Five small crewed cutters towing the full-size Azure Leviathan toward its repair pier
Five eight-metre cutters tow the 128-metre ironclad. Rigging, crew and hull dimensions remain in metres.
A named shipwright standing on a work platform and hammering a patch beside the ironclad hull
A shipwright reaches the hull-side platform before repairs consume supplies and close leaks.
A dock pump operator working a pump with a hose running to the damaged ship
Dock pumps remove actual floodwater. Hull repairs and pumping are separate jobs.
Shipyard controls showing named workers, hull repairs, pumped water and remaining repair supplies
Inspect the named workers, remaining stock and work in progress, then cast off the same repaired ship.

v0.3 Flooding, lifeboats and survivor rescue · shipped

Permalink ↗

All hands. Abandon ship.

Alex wanted damaged ships to take on water and sink gradually, with crew operating pumps, launching lifeboats and rescuing survivors. Naval battles now save water volume, leaks and safety priorities for each hull. Cannon damage opens leaks, rough seas increase inflow, and existing finite-stock repairs can patch them. Pumps remove water rather than granting hull health. As water rises, the full-size hull sits lower in the sea. These are authored flood reserves and pump rates, not a physical buoyancy or stability solver.

Select a ship and open Safety. Automatic safety sends a small pump party; Pumps first assigns more hands, while Guns staffed leaves pumps uncrewed. Sailors walk to the machinery before it removes water. A person pumping cannot simultaneously run an engine, serve a gun or repair the hull. The close-up shows named engineers working beside human-sized pumps. This creates another choice alongside firefighting: which duties can the ship afford to leave unstaffed?

At zero hull, a vessel begins foundering rather than disappearing. Automatic evacuation follows after three seconds, or twenty with Pumps first. You can order Abandon selected ships sooner. Sailors walk to the rail, board their allocated boats, and lower an occupied boat over eight seconds. Abandonment is irreversible once underway. When the flood reserve fills, the hull descends over fifty-five seconds until its tallest parts disappear. Anyone who misses the boats enters the water with a finite, weather-sensitive exposure timer.

Every person stays 1.717 metres tall. Capital-ship lifeboats are six metres long and 2.1 metres wide, carrying eight people. Cutters carry a three-metre boat with three seats. A Leviathan has eight boats for its sixty-four playable sailors; a Sovereign has one for its eight. They carry the original named crew, not extra decorative people. Occupied boats row toward intact friendly ships, route around their original hull and wait or steer around one another. Visual review caught converging boats merging their crews; a spacing check now follows them through a two-minute rescue.

Select an intact crewed ship and click Rescue or press R. It approaches friendly boats or people overboard, then a sailor reaches the rail and works a rescue line. One survivor is hauled aboard every two seconds. Rescued people remain passengers, preserving their names and IDs without becoming free gunners or replacement crew. A cutter can carry three rescued passengers; a capital ship can carry its normal crew count again. Passengers go back into the water if their rescue ship sinks. Rescue and evacuation continue after victory, while cannon combat stops.

Mayday · All Hands starts paused with twelve ships, 158 named people and an authored emergency. Both fleets hold fire so you can practice. Pump the damaged Azure ironclad, try an evacuation, and send the cutters to collect its boats. Recordings includes All hands rescue: two minutes in 4,829 bytes, using four legal pump, rescue, resume and abandon commands. Initial damage belongs to scenario setup; subsequent events come from the simulation. New recordings use .5. Frozen .1 through .4 files retain their previous rules, and replay Safety controls are read-only. Older saved battles can explicitly enable the new system.

All 325 automated tests and the production build pass. Six public browser gates cover pumping, finite evacuation, visible sinking, actual passengers, exact reload, WebGL/WebGPU, narrow Safety controls and complete Node-to-Chromium rescue playback. Four real two-process co-op gates verify guest orders and identical boats, jobs, flood volumes and people through a named save, disconnection and fresh invitation. The previous weather and replay browser checks also pass. All thirty-five hosted runtime files match the tested build. Fresh review corrected button contrast, evacuation wording, boat spacing and replay documentation. Harbor trade boats keep their earlier system. Physical older devices, sustained large-fleet profiling and repeated family balance tests remain open.

Named engineers operating bilge pumps aboard the full-size Azure Leviathan
Real engineers leave their usual duties and walk to pumps. Water volume falls while the same crew remain identifiable.
Original Azure sailors seated inside an open pointed six-metre lifeboat beside the ironclad
A six-metre boat carries eight original sailors. The people and boat retain their metre dimensions.
The Crimson ironclad partway through sinking after its flood reserve fills
The hull descends gradually instead of disappearing. Its dimensions stay fixed as the wreck slips below the water.
An intact Azure cutter with rescued passengers and nearby occupied lifeboats
Survivors join a cutter as passengers, preserving their identities without filling combat posts. Lifeboats keep their spacing.

v0.3 Wind, rain and stormy seas · shipped

Permalink ↗

Reading the weather

Alex wanted weather to change sailing and the way fires spread, with rain helping crews and stormy nights giving the battles more atmosphere. Weather is now part of naval game state. Open Weather to choose Calm, Trade wind, Rain or Storm, then the compass bearing the wind blows toward. The host controls the conditions; both co-op players share them. The existing Lighting controls remain local viewing preferences.

Sailing Sovereigns make more progress with the wind than straight into it. Rowed cutters are less sensitive, while steam ironclads retain more headway. Rough seas reduce speed. Gusts follow simulation time, so pausing freezes them and saving preserves the same point in the weather. Calm keeps the earlier baseline movement. These are authored response curves; you can queue tacking courses yourself, but captains do not yet find optimal sailing routes.

Wind favors heat transfer toward downwind deck sections and carries severe-fire embers farther between nearby hulls. Rain cools actual heat, weakens new ignition and slows spread. Heavy rain can quench an unattended fire. It never restores hull health or refills repair stock. The comparison capture shows a deck with its flames out and its hull still damaged. There is now a reason to wait for a squall or keep hands at the guns while rain assists the fire parties.

Stormwatch Armada starts paused with twenty-four ships, 82 named crew, two 69-metre sailing flagships and authored starting fires. Choose Lighting, then Night. Warm lanterns and fire stand out against storm shade, rain, wind-blown smoke and moving water. People remain 1.717 metres tall. Rain trails are fourteen millimetres wide and ninety centimetres long to represent falling motion, and wave amplitude is bounded at 0.65 metres. Hulls and crew heave together; boarders interpolate between the moving decks.

The effects stay bounded: High draws at most 8,192 rain trails and 128 foam patches, with 1,024 trails and forty patches on lower profiles. Water tessellation caps at 128 segments per axis on High and 96 on lower profiles; distant people, ships and rain do not grow to fill pixels. Procedural wind, rain and distant thunder use the effects controls and become quiet when paused or in menus. Weather is uniform across the arena. This adds neither regional fronts nor physical flooding, roll or buoyancy simulation.

Recordings offers Riding out the storm: fifty seconds in 4,332 bytes. Seven legal orders show a following wind, a wind shift, a squall, withdrawal and a crew-priority choice. The surviving flagship ends with its fires out. Starting damage belongs to the authored scenario; all subsequent changes come from the real simulation. New recordings use ruleset .4. Frozen .1, .2 and .3 files keep their original conditions, including resets and fleet reconfiguration. Weather, fire, crew positions and stocks save and resume together.

All 308 automated tests and the production build pass. Six browser gates cover paused render limits, real rain suppression, upwind/downwind movement, exact saves, WebGL/WebGPU, phone layout and the full Node-generated storm recording in Chromium. Three two-process co-op gates verify shared host weather, rejected guest weather injection, allowed crew commands and identical named-save/reinvite state. The older replay and damage-control suites and twenty-seven construction/scale checks pass too. Public evidence binds all thirty-three runtime files to the tested build. Fresh visual review caught an off-camera comparison shot; refocusing the ship fixed the capture. Sustained active-storm profiling, physical phone/older-Mac/headset performance and repeated human balance tests remain open.

Rain, warm deck fires and lanterns around a 69-metre Sovereign in WebGPU
A full-size Sovereign in a stormy night battle. Rain, smoke, lamps and people use metre coordinates.
Weather controls showing storm, wind speed and compass bearing, and rain intensity
Wind blows toward the chosen bearing. Host changes affect both players and are included in recordings.
The Sovereign deck after rain quenches the fire, with damaged hull health remaining
Rain quenches the deck fire; hull damage remains. No fire party or free repair was used in this comparison.
A rainy fleet resumed by a guest through a fresh multiplayer invitation
Public co-op resumes the same wind, rain, heat, hulls, repair stock and walking crew coordinates.

v0.3 Spreading fire and crew damage control · shipped

Permalink ↗

All hands to the fire

Alex wanted the night fires to become decisions. New naval battles now track heat in three deck sections on a cutter, six on a Sovereign and ten on a Leviathan. Some cannon hits ignite a section. Unattended heat grows, spreads along the deck, burns real hull health and can sink the ship. Severe fires can throw embers onto another hull across a gap of up to eight metres, including friendly ships. Fleet spacing now matters for fire as well as maneuvering.

Select ships and choose Damage control: Guns first, Balanced or Save the ship. Balanced sends a small support party; Save the ship draws more hands away from their normal posts. These are the existing named people, with their original roles retained. A cutter may have to give up its only gunner or rower. The ironclad can use its larger support crew, at the cost of reload support. Enemy captains also respond to severe fires.

Crew walk to the work at 1.8 metres per second, carry small buckets and throw visible water after arriving. Hot sections can injure them. Once the flames are out, assigned hands patch hull damage with small hammers; carpenters work faster. Each hull starts with repair capacity equal to thirty percent of its maximum health, and each patch consumes stock. The panel shows fire intensity, assigned firefighters and repairers, and remaining stores. The crew manifest names the jobs and lets you inspect the person doing them.

The first visual pass caught a path through the ironclad’s stern deckhouse. Deck routes now clear the deckhouse and open machinery well, and the crew camera frames the worker from outside the ship. People stay 1.717 metres tall, the ironclad remains 128 metres overall, and buckets, hammers, walking speed and ember distances use metre coordinates. The bounded light and particle pools remain. This is a compartment fire model, without fluid combustion or a complete internal ship-navigation simulation.

Fire on the Water is a new paused scenario with twelve ships and an authored emergency aboard a few damaged vessels. Start with Guns first to see the danger grow, then send fire parties and move the burning hull away from the fleet. Choose Lighting, then Night. Recordings also offers Containing the deck fire: eighty seconds in 4,641 bytes. Its four legal orders switch priorities and withdraw the ironclad while the crew extinguish fires and repair damage. Starting fires belong to the authored scenario; no health, position or outcome is injected during playback.

Saving now retains section heat, spread timers, repair stock, priorities, crew jobs and their walking coordinates. Co-op partners restore the same emergency, and a guest can change player-ship priorities through the host. New recordings use ruleset .3. Frozen .1 and .2 full-match recordings retain their earlier combat path, including fleet resizing, capital/ocean changes and both reset routes. Older naval saves keep their previous rules until Enable fire simulation is chosen; harbor economy boats still use the earlier visual damage fire. Playback never upgrades an older recording.

All 297 automated tests and the production build pass. Six browser gates verify damage and spread, real water throws, exact mid-fire saves, WebGL/WebGPU, narrow controls, finite-stock repairs and the full Node-generated fire recording in Chromium. Three two-process co-op gates verify guest priorities and identical fire, crew, stock and hull state through saving, disconnecting and inviting again. The older six-gate replay suite and twenty-seven construction/scale checks pass too. Public verification and runtime-byte binding are linked below. Fresh visual review passes after widening the worker camera. Repeated human balance tests, sustained night-battle profiling and physical phone, older-Mac and headset acceptance remain open.

Spreading fires along the deck of a 128-metre ironclad with Guns first selected
The authored emergency grows under Guns first. Heat and burning hull damage are simulation state.
A named carpenter with a bucket throwing water onto a deck fire
A real fire-party member at work. Water throws require arrival at the burning section.
Crew patching the ironclad hull with a small hammer after extinguishing the fire
After quenching the flames, crew repair actual hull health using finite onboard stores.
A shared burning fleet restored through a new multiplayer invitation
Public two-browser co-op resumes the same fire heat, jobs, walking positions, stocks and crew identities.

v0.3 Night lanterns and hull fires · shipped

Permalink ↗

Lanterns over dark water

Alex wanted naval battles to come alive after dark: torches on buildings, lanterns on ships, and fire when hulls take damage. The new warm light gives the blue night some contrast and gives scale another familiar reference. A lantern cage is twenty-two centimetres wide and thirty-four centimetres tall beside the existing 1.717-metre people. The giant ships carry more fixtures across their decks, with posts reaching down to the timber.

Completed village buildings, the dock and the completed Great Hall gain wall lights and torches. Ship lanterns follow their hull’s position and heading. At dusk the lamps brighten, warm light falls across nearby walls and crew, and small stylized glints catch the water. At noon their light fades. Harbor sources respect current visibility and construction completion, so a lamp does not reveal an unexplored building. The ocean chart also dims after dark.

A living ship below eighty percent hull health begins to show damage fire. More damage brings stronger flames; a capital ship can burn at several places along its deck. Flames, bright cores, embers and rising smoke use metre coordinates. Repair above the threshold clears the fire, and sinking removes it. The burning Sovereign in these captures was damaged by the real cannon simulation. Fire currently communicates damage visually: it does not spread, add damage over time or give crew a new firefighting order.

The effects derive from hull health and simulation time. Pause freezes the flame poses, saves reconstruct them after reload, and replay seeking rebuilds them at the selected moment. No fire timers, particles or extra save/network fields need to be stored. Co-op partners still share the same damaged ships and crews; their local lighting preferences can differ. The current .2 recorder and the retained .1 full-match compatibility fixture stay unchanged.

Large fleets need limits. Fixtures and particles use fixed instanced pools, with at most 512 lanterns emitted. The nearest weighted sources share six shadow-free point lights on High and two on Balanced or Low power. Flame/core pools hold 1,536 instances each, smoke and embers 768 each, and glows and water glints 1,024 each; excess geometry is omitted. These are stylized lights and effects, without volumetric smoke, ray-traced illumination or physically simulated water reflections.

All 285 automated tests and the production build pass. Five public browser gates cover village lighting, real damage, pause, exact saved hull/fire reconstruction, WebGL, WebGPU, narrow layout and the 128-metre ironclad. Three real two-process RTC gates restore identical ships, crew, health and destinations on the larger ocean. Six production replay gates and twenty-seven construction/scale checks pass, and all twenty-nine public runtime files match the tested build. Fresh visual and README review passes. The evidence retains the actual timing readouts; short renderer-switch captures include warm-up, and a warmed paused view is not a new sustained battle benchmark. Physical phone, older-Mac and headset performance remain unmeasured.

Village hall torches casting warm light on the wall after dark
The working harbor at night. Human-sized fixtures bring warm light to completed, currently visible buildings.
An undamaged 69-metre Sovereign with deck lanterns and visible crew
The 69 m Sovereign before battle. Small lamps light its deck while the hull and people keep their scale.
A damaged Sovereign warship with deck fires, smoke, embers and visible crew
A 69 m Sovereign takes real cannon damage. Fire illuminates the deck and reconstructs from saved hull health and time.
Distributed lanterns and visible people along the deck of a 128-metre ironclad
The 128 m ironclad carries ten small lanterns. The people and fixtures help reveal how much deck there is.

v0.3 Command UI, wider oceans and changing light · shipped

Permalink ↗

The command desk and the open sea

Alex wanted the in-game controls to receive the same care as the opening screen. Then came three more requests: cannons and explosions, larger maps, and time of day. This pass brings those together with a navy, brass and parchment command desk, a new ocean scenario and light that changes over the living world.

The selection dock now has original engraved role portraits, clickable member tiles, combat health strips and type filters for mixed groups. Click a portrait to select that person; Shift-click removes them from the group. Build opens an illustrated catalogue of the ten existing buildings, grouped by purpose with era locks, material shortages, actual costs and footprints in metres. The supplies counter shows spendable stock, with dock reserves in its tooltip.

The technology chart makes the eight existing discoveries and their prerequisites visible, with a three-age journey, availability reasons and real progress. A research button pays its cost and starts the existing queue. Narrow screens can pan the chart; refreshing resource counts preserves its position and keyboard focus, and arrow-key chart navigation leaves the world camera still. The full tutorial still completes through actual gathering, construction, boarding and delivery.

Fleet commands are grouped into Orders, Formation and stance, and Inspect and camera. Ship tiles, combined hull health, living crew, gunner counts, reload time and current orders describe the actual fleet. Cutter boarding, numbered groups, queued destinations, patrol and escort remain available. The phone command dock scrolls, and the village controls stay hidden during naval battles.

Bluewater Armada begins paused with forty-eight ships and 266 crew on a 1.28 by 1.28 kilometre ocean. Its ironclads are still 128 metres overall and its cutters eight metres. Fleet & FPS also offers a 2.56 kilometre square ocean. Water bounds, formations, navigation, minimap orders, camera, saves and co-op all use those larger coordinates. Hull dimensions, speed and weapon ranges keep their metre values. Changing ocean size restarts the fleet; fixed beacon missions retain their authored arena. Harbor terrain and editor bounds remain unchanged.

The existing cannon simulation now drives flashes at gun stations, projectile trails, smoke, impact bursts, wooden debris and water splashes. Effects follow combat events and simulation time, so pausing freezes them and seeking reconstructs them. Fixed instanced pools cap smoke at 768 particles, flashes at 256, debris and splashes at 512 each, and trails at 1,536. These are stylized visual effects; debris adds no new damage and smoke is not a fluid simulation.

Open Lighting to let dawn, day, golden hour, dusk and night unfold, or choose a fixed time with the dial and presets. A day takes twenty minutes of unpaused play by default, with five- and sixty-minute options. Automatic light follows saved simulation elapsed time; manual choices and cycle preferences stay local to the browser. Sun direction, sky, fill light and fog display change while exploration, sight and combat rules remain the same. Nights keep readable fill, and the harbor, naval WebGPU and six-thousand-person army views all use the controls.

All 282 automated tests pass. Nine production and public browser gates exercise the new controls, both oceans, WebGPU, lighting and the army view. Three two-process RTC gates share the larger ocean, accept a guest’s distant order and restore identical ships, crew, health and destinations through a named save and fresh invitation. The six-gate public replay pass and all twenty-nine runtime byte matches pass too. Twenty-seven construction/scale checks, the complete tutorial and the existing 256-ship shared save/resume flow pass on the new build. Older .1 recordings retain a frozen full-match compatibility fixture alongside the new .2 writer. A fresh visual review caught phone overlaps; those were corrected and verified before publication. Physical phone, older-Mac and headset performance and repeated human balance tests remain open.

Selected guard with an engraved portrait, health strip and command buttons
Selection represents the actual person or unit. Member tiles and type filters change the real command selection.
Parchment building catalogue with brass illustrations, costs and metre footprints
Ten existing buildings become easier to compare: purpose, era, materials and physical footprint before choosing a site.
Technology prerequisite chart and three-age progression panel
Eight discoveries, explicit requirements and the real research queue. The chart can pan without moving the world camera.
Large crewed ironclad selected on the larger Bluewater Armada battlefield
A 128 m ironclad and its working crew on the 1.28 km ocean. Larger arenas change the room around the ship.
A cutter firing its cannon under moonlight with grouped fleet controls
A real cannon event supplies the flash, smoke and debris. Night changes the light while hulls, people and combat rules keep their scale.
Time-of-day dial with live cycle, day length and dawn-to-night presets
Lighting is a local display choice. The automatic cycle follows the shared simulation clock when players use the same settings.
Harbor village under dawn lighting with the redesigned in-game interface
Dawn colors the village and fog display. Construction, exploration and resource state remain unchanged.

v0.3 Recordings, Strait battle and cinematic opening · shipped

Permalink ↗

A whole battle in four kilobytes

Alex wanted a small game recording that could be watched from another angle. Harborfolk now has a Recordings archive in the main menu and game toolbars. Record the current game, stop and save, then watch, export or import its .hfr file. The browser keeps up to eight recordings. Automatic recording of new games is optional and starts off.

The first complete demonstration lasts 114.8 seconds and weighs 3,667 bytes: twelve starting ships, forty-six people, nine legal commands and an Azure victory with one captured cutter. An automated demonstration captain issues orders through the real simulation; no mid-match damage, resources, positions or outcome are injected. Choose Watch: Battle for the Strait in Recordings, then Play.

A replay deserves decisions worth revisiting. The new Battle for the Strait scenario has three channel beacons. Staffed ships claim a neutral beacon after twelve uncontested seconds; ownership persists when vacated. Reach 240 points or eliminate the opposing fleet. Enemy captains distribute ships among vulnerable zones. The demonstration splits its fleet between channels, holds a narrow score advantage and commits two cutter crews to a western boarding attack. A passive fleet loses the same scenario in an automated simulation test. Human playtesting will tell us how these openings hold up after repeated games.

The opening now gives that fleet a proper stage. A fresh visit arrives at an illustrated dawn harbor with a large title, slow artwork motion and an Enter the harbor button. The original score starts after your gesture and respects saved audio settings. Enter or Escape reveals the captain’s menu; saved worlds skip this arrival. Watch the battle begins the real recording directly, while Take command opens the playable scenario. Loading shows the fourteen models actually completed, with recovery controls preserved. The phone menu scrolls through every choice, and reduced-motion settings stop the title animations.

Playback has pause, rewind, a seek slider, half speed through sixteen times speed, and commentary bookmarks. Orbit, zoom, follow a ship or inspect someone at their crew station. Cutters remain eight metres long, flagships sixty-nine metres overall and people 1.717 metres tall. Spectator controls replace the attack buttons. Exit restores the live harbor, and playback cannot overwrite its autosave, edit a scenario or connect a shared game.

The file contains the initial compact world and timestamped orders, reconstructing thirty simulation steps per second. It stores no video or per-frame actor transforms. Verification every ten simulated seconds and at the finish stops playback if the simulation drifts. Ordinary floating values are compared to six decimals to tolerate microscopic runtime math differences; IDs and packed army data remain checked. This command-log design differs from Unreal's replicated-state replay stream.

The host records accepted orders from both co-op players. Two independent Chromium processes verified a guest keyboard order reaching the host, then the match replaying the same villagers and resources after disconnecting. Both players can keep a copy of the file. Watching is local; synchronized watch parties are future work.

All 273 automated tests pass. The production build and published game pass complete-match verification, seeking, import/export, storage, narrow layout, autosave protection and co-op capture. WebGL and WebGPU draw the same recorded match. All twenty-nine public runtime files match the tested build. The earlier overnight soak completed five hours and forty-one minutes across seven worlds, with eleven exact save/reloads and no browser runtime errors. Six initial fixture/environment failures remain in the raw report, with passing corrections for every affected suite.

Recordings require a matching simulation ruleset. Export important files before updates; old-engine playback and migration are not built yet. Seeking re-simulates from the beginning in yielding slices, so huge battles can take time and sixteen times speed is a requested rate. Caps are six simulated hours, twenty thousand orders and eight MiB expanded. Recording adds CPU work; this release makes no new ten-thousand-unit performance or physical-device claim.

Naval replay with three channel objectives, minimap and playback controls
The complete legal-order match reconstructs in 3D from a 3,667-byte file. The camera remains yours.
Replay camera inspecting an individual crew member on a flagship
Inspect the people doing the work. Ships and crews keep their metre dimensions throughout playback.
Verified replay finish showing Azure victory with a captured cutter
The western boarding maneuver ends with a captured prize. Rewinding reconstructs the same victory.
Harborfolk cinematic title screen with an immense ship at an illustrated dawn harbor
The first arrival: original illustrated cover art, a direct entry button and a route into the actual recorded battle. This is key art; gameplay remains the 3D world shown above.
Captain’s menu with tutorial, saves, multiplayer, map editor and a featured naval battle
The captain’s menu keeps the classic RTS choices and adds direct Watch the battle and Take command routes. Saved worlds skip the arrival gate.

v0.3 Cheat codes and ridiculous units · shipped

Permalink ↗

Six metres of bad attitude

Alex asked for the kind of cheats that turn a serious RTS into a family joke. Age of Empires’ car was the reference. Harborfolk now has its own little forbidden toy box: press Enter or click Cheats in a harbor, enter a phrase and activate it. Case and extra spaces do not matter. An optional spoiler codebook reveals the list without sending any cheat text into multiplayer chat.

MID LIFE CRUISER calls an original orange open-top roadster, 4.4 metres bumper to bumper, with a visible seated driver, spinning wheels, 600 health and a speed of ten metres per second. Its combat trick is a bumper attack. The classic HOW DO YOU TURN THIS ON phrase works as an Easter egg too. HONK IF HUGE produces an approximately six-metre fantasy goose with nine hundred health, a walking leg cycle, a bobbing neck and powerful pecks. Arrival horns are synthesized; no Age of Empires models or audio were imported.

The joke still obeys the project’s scale rule. The goose is deliberately enormous geometry in the same metre world. Nearby villagers and cottages keep their dimensions. These are playable units: select them, drive or waddle to a right-click destination, Stop, assign a group, queue orders, patrol or escort. There are at most eight living toys and the existing bounded unit storage remains in force. Large toys reject paths too close to buildings or water; the existing foot grid can reject a narrow route even when a wider detour exists. Car steering physics and dynamic actor collision are future work.

TIMBER TANTRUM, LUNCH MONEY and ROCK SOLID add a thousand wood, supplies or stone. FULL PANTRY adds a thousand of all three. Supplies are this harbor’s food and trading currency; there is no separate gold system. PATCH IT UP heals living friendly harbor defenders, ships and buildings without resurrecting casualties or completing construction. MAP PLEASE explores terrain while enemy detection still requires sight. The console’s Find newest toy button selects and focuses the latest living toy.

Using a code leaves a persistent Cheats star and saved activation history. Keep a clean named copy first if you want to return to an ordinary game. The host controls cheats in co-op; the network rejects guest cheat commands, while both players can command the shared toys. Exact identities, positions, health, paths and routines survive saves and reconnects. This first toy box is for harbor worlds and authored harbor scenarios. The separate mass-army and Epic Naval Battle modes retain their own rosters.

All 262 automated tests pass, including actual goose damage, fixed art bounds, resource limits, save validation and host command authority. The new browser pass checks actual console entry, car picking and right-click movement, wheel animation, narrow layout, exact reload and two independent Chromium processes sharing toys and commanding the goose over real RTC. It passes again on the published game, whose twenty-seven runtime files match the tested build. Six additional browser suites cover RTS/audio controls, the full tutorial, shared queues and rallies, scenario editing, exact fifty-thousand-person solo saves and loading recovery. This is functional QA; no new frame-rate or physical-phone performance claim comes with the toys.

Open-top roadster with a visible driver next to unchanged villagers
The original Midlife Cruiser measures 4.4 metres bumper to bumper. Its driver remains visible as it drives.
Huge white goose looming over the roadster beside a harbor cottage
An approximately six-metre fantasy goose. Buildings and people retain their original metre dimensions.
Cheat console with resource and toy codes at a narrow viewport
Enter opens the console. The optional spoiler codebook fills the input; Activate confirms it.
Saved car and goose shared through real RTC multiplayer
The host enters cheats; either player can move, patrol or escort with the resulting toys.

v0.3 RTS controls, queues, rallies and tutorial · shipped

Permalink ↗

Learning to command

A large army is more fun when commanding it feels familiar. This release adds visible drag selection to the land battle, consistent Shift selection, type selection and ten local control groups. Double-tapping a group brings the camera to it. Harbor shortcuts cycle idle villagers, open work assignments and focus the dock. F1 opens one guide for harbor, land and naval commands. Text fields and menus keep those shortcuts from accidentally moving the world.

The Workforce panel shows actual timber, food, stone, trade and construction jobs. Add an idle villager, remove one worker or assign the selected people. Workers still walk to the worksite and earn resources through the existing simulation. A carrier finishes delivering a load rather than losing its logs. Buildings and missing jobs determine which assignments are available.

Hold Shift when issuing an order to queue it. A normal order replaces that actor’s remaining queue; Stop clears it. Harvest and work jobs finish a real cycle before continuing. Grouped movement keeps formation offsets, and a queued construction party builds one shared foundation and pays once. The queue is bounded to eight pending orders per actor. Persistent patrol and escort duties keep running until replaced or stopped.

Production rallies now remember a destination or job. Select a dock and right-click a farm, worksite or destination; newly recruited villagers go there. Select a military production building and use Y to set a rally, including a living friendly troop to escort. Queues and rallies are part of saved and shared world state. Control groups stay local to each player. Land selection operates on companies, which can contain up to one hundred individually simulated troops; individual land-troop selection remains future work.

The start menu now offers Learn to command. Its twelve-lesson Command School begins with eight villagers and no timber, then follows selection, delivered wood, farm construction, an arrived farmer, queued work, workforce assignment, a completed dock, a saved job rally, recruitment, a control group, walked crew boarding and an actual cargo delivery. It checks the world as you play. No lesson grants resources or teleports people. Save halfway, reload and continue; collapse or hide the coach when you want room to explore. This first course teaches the harbor economy, while battle scenarios retain their own briefings.

Battles now move and sound more alive. Close views get a bounded pool of tiny impact sparks, dust, launch tracers and falling human silhouettes. The largest pool is 160 cues; reduced motion lowers it. These cues add no damage or extra combat actors. Original synthesized horns and clashes accompany calm, march, engagement, victory and defeat score moods. Battle transitions use a horn cooldown, loading primes the sound history, and master mute still silences the real output.

StarCraft’s control groups, queued commands and rally-point documentation, Age of Empires’ learning goals, and Banished’s building and profession workflow informed the pass. The links below are the primary references. Art and audio remain original or part of the project’s established asset library. Every person, ship, building and new physical effect retains the existing metre scale.

The release passes 254 automated tests and ten production-browser suites, including both graphics backends, exact fifty-thousand-person solo saves, loading recovery, scenario editing and the full tutorial. Two independent browser processes use real RTC, save queued orders and rallies mid-task, reload, exchange a fresh invitation and verify recruits following the same work and army rallies. The published game’s twenty-seven runtime files match the tested production build byte for byte, and its tutorial and shared-order checks pass again.

Performance remains a measured limit. A fresh thirty-second synthetic Low power test averages 28.8 FPS, with p99 at 66.7 milliseconds and a worst frame of 117.3 milliseconds. All ten thousand people remain living and drawn and no simulation time is dropped. This is not a steady thirty-FPS pass. FORTRESS has an RTX 5090; a phone-shaped viewport and fourfold CPU throttling do not reproduce a phone GPU, older MacBook or heat. The single run is not a paired comparison with the previous diary’s longer samples. Raw evidence keeps every frame interval, release provenance and the unsuccessful target.

The battlefield, measured

FORTRESS / RTX 5090 Chromium · CPU throttle 4× · touch viewport 390 × 844 / DPR 3 · 65% framebuffer · WebGPU unavailable · Low power / 30 FPS cap · all 10,000 living and drawn · 8 s warm-up excluded · autosaves on. One 30 s synthetic run; not a physical phone, paired comparison or steady-30 pass.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL2 fallback · RTS commands30 s28.866.7 ms117.3 ms452
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Drag selection across full-size companies
Land battles select companies; each company still contains individually simulated people.
Workforce panel assigning real villagers
Plus and minus change real workers. Gathering still needs a journey and a completed delivery.
A dock job rally and the live command-school coach
A new villager follows the dock’s saved job rally to the actual farm.
A close battle with bounded combat effects on WebGPU
Physical impact, launch and fall cues run in a pool capped at 160. Moving closer preserves the original unit dimensions.
A second real browser sharing the restored rally and orders
Two browser processes exchange commands over RTC, save mid-order, reload and reconnect.
Command school completed after a real cargo delivery
Twelve lessons end with people aboard and twenty supplies actually delivered.

v0.3 Lossless save and snapshot packing · shipped

Permalink ↗

Same battle, smaller saves

Ten thousand people need to survive saving as well as fighting. This pass concentrates on the work between the simulation and its saved or shared snapshot. Positions, health, targets, attack timers, generations, orders and world dimensions stay intact. There is no new position rounding or smaller population hidden behind the performance result.

Fog and army data now pack directly from the live state before one JSON serialization. The older path built expanded JSON, parsed it, packed the fog and serialized it again. Restore can read the compact fog object directly. Earlier expanded autosaves and existing packed exports remain readable. Named saves retain validation and rollback when storage fails.

The established army layout is unchanged: twenty-three bytes per unit, with the same little-endian fields. Compatible hosts copy those array bytes in bulk; the portable field-by-field path remains available. Modern browsers can use native byte-array base64 when available. A bounded ASCII fallback covers browsers without those methods, and both paths retain input checks.

A fixed ten-thousand-person Node fixture measures sixty calls after five warm-ups. Mean army encoding falls from 8.33 to 0.63 milliseconds; snapshot preparation from 10.40 to 0.90; transport preparation from 10.59 to 0.90; receive and restore from 12.25 to 3.64. The packed army hashes and network bytes match exactly. Autosave JSON is about forty-five percent smaller in this fixture. These are CPU timings on this computer, not a tenfold browser-FPS claim.

Commands now perform one host save instead of two. Four-second autosaves, blur and page-hide writes remain active. A congested co-op link skips snapshot preparation before doing the work, and large snapshots go straight to the existing ordered chunks. Snapshot frequency and payload bytes are unchanged; a delta-based network protocol is still future work.

The fresh thirty-second synthetic comparison moves from 26.7 FPS on published v30 to 28.0 with packing alone, with all ten thousand troops living and drawn. A longer Balanced sample exposes later stalls and averages 26.1 FPS over two minutes. A late CPU profile still points mainly to collision avoidance. A tighter cell-bound experiment gave only a small, inconsistent gain and was discarded. The final rendering cleanup skips empty instance uploads and sphere tests for centers already known to be visible.

The final two-minute Low power test averages 29.6 FPS, with p99 at 50.1 milliseconds and a worst frame of 216.8 milliseconds. All ten thousand soldiers remain drawn, the simulation loses no time, and thirty autosaves succeed; the slowest save takes 22.1 milliseconds. This is not a steady thirty-FPS pass. The raw data keeps the long frames, the slower Balanced experiment and the different sample durations. FORTRESS has an RTX 5090; fourfold CPU throttling and a phone-shaped viewport cannot reproduce a phone GPU, older MacBook or sustained heat. No physical-phone ten-thousand-unit success is claimed.

Army & FPS now offers thirty-, sixty- and one-hundred-twenty-second measurements. Exports include actual save counts and cost, native codec availability, frame intervals, living and drawn populations, and simulation progress. Use those measurements to find a comfortable army size on your own machine. The full-size close models return as the camera approaches.

The release passes 240 automated tests plus browser checks for exact fifty-thousand-person solo saves, scenario editing, recovery, harbor menus, naval logistics and both graphics backends. Two real co-op browsers save a damaged ten-thousand-person battle, reload the same identities, health, targets and cooldowns, then share and resume it with a fresh invitation. One guest uses Low power while the host keeps its own graphics choice. Further work belongs in the remaining avoidance cost and smaller shared updates, followed by real older-device testing.

The battlefield, measured

FORTRESS / RTX 5090 Chromium · main-thread CPU throttle 4× · touch viewport 390 × 844 / DPR 3 · 253 × 548 framebuffer · WebGPU unavailable · Low power / 30 FPS cap · all 10,000 living and drawn · 8 s warm-up excluded · autosaves on. First two rows: 30 s samples. Final row: 120 s, with rendering cleanup. Single synthetic runs, not a phone result or steady-30 pass.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL2 fallback · v30 · 30 s30 s26.766.8 ms83.4 ms452
10,000WebGL2 fallback · Packing only · 30 s30 s28.066.6 ms100.1 ms468
10,000WebGL2 fallback · Final · 120 s120 s29.650.1 ms216.8 ms1,779
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Ten thousand full-scale soldiers in the final two-minute synthetic pressure test
Actual game capture after the two-minute CPU-pressure run. The models keep their metre scale; all ten thousand remain present.
Human-height view of the restored ten-thousand-person shared battle
The saved co-op battle resumes with the same unit identities and combat state. Close views still use the original full-detail models.
Two-minute device measurement and export controls
29.6 FPS on average, with long frames retained. This is phone-shaped Chromium under CPU pressure on an RTX 5090, not a physical phone.

v0.3 CPU and distant detail · shipped

Permalink ↗

Less work for ten thousand people

The next portability pass keeps every soldier and changes how much work the computer does for them. The CPU profile still points first to avoidance, followed by targeting and the main simulation loop. An early shared candidate-list cache preserved battle state but did not make the fixed-work test faster, so it was discarded.

The useful version remembers which nearby grid cells contain people. A small bit mask skips empty ground in the original query order, and changes only when cells gain or lose occupants. Cached troop traits avoid repeated catalog lookups. An enemy search can stop early when its complete rectangular search area contains no enemy cells. Movement lengths also use a direct square root in this bounded world. Candidate limits, thirty-Hz combat, physical dimensions and the number of people stay the same.

Three fixed-work CPU trials each warm ten thousand troops for three hundred ticks, then measure six hundred ticks with the same seed and commands. Mean tick cost falls from roughly 4.25 milliseconds to 2.88, about a third less work. All resulting packed-state hashes match. The accepted legacy, six-thousand, ten-thousand and fifty-thousand trajectory fixtures still match too, including combat and candidate-saturation counters. These are Node CPU measurements, separate from browser FPS.

Balanced and Low power now have a third drawing level for the thirty-type army. At eight projected pixels in Balanced, or twelve in Low power, each authored body and equipment part uses three intersecting physical planes with its original bounds, tint and animation channels. A simpler material avoids detailed surface lighting. Moving closer restores the original models; High keeps its earlier two levels. People, mounts and long weapons retain their metre dimensions. This extra representation uses memory of its own, so the saving here is drawing cost.

The paused six-thousand-person overview falls from 2,214,119 to 961,719 submitted triangles on both WebGL2 and WebGPU. All six thousand soldiers remain in the camera, and the saved battle stays exactly the same. At ten thousand, the distant view submits about 1.49 million triangles instead of 3.44 million. Close views and the model inspector continue to show the original art.

For the harder test, this session first remeasured the published v29 game under fourfold CPU throttling, with a phone-shaped viewport, unavailable WebGPU and every one of ten thousand troops drawn. That fresh baseline averages 21.0 FPS. The optimized candidate reaches 28.3 FPS, with the worst frame falling from 183.5 to 83.5 milliseconds. Both samples keep the simulation clock advancing without dropped time. It is closer, but the remaining hitch means this is not a steady thirty-FPS pass. The earlier 6.6-FPS diagnostic stays in the previous entry; today’s faster baseline is the comparator used here.

A subsequent check of the published optimized game averages 28.0 FPS under the same synthetic pressure, again with all ten thousand drawn and no dropped time. These are single synthetic runs on FORTRESS and its RTX 5090, not phone or 2019 MacBook measurements. CPU throttling cannot reproduce their GPUs, memory limits or sustained heat. The downloadable evidence retains every frame interval, living and drawn counts, and both fixed-work results. Warmed High-profile desktop samples remain near sixty FPS on both renderers, though WebGPU records one 33.4-millisecond frame; that hitch stays in the data too. Autosaves remain enabled.

Shared-battle checks also pass with a High host and Low power guest, each drawing ten thousand soldiers. Named mid-battle saves restore the same unit identities, health, targets and cooldowns, and a fresh co-op invitation resumes that restored battle. The graphics choice stays local to each player.

The next gate is still a real older laptop or phone: warm it, start smaller, then grow the army while measuring frame times and simulation progress. More work remains in save and snapshot packing, cheaper network updates and potentially simulation workers. Complete harbor, naval and editor touch interaction also remains open. No physical-phone ten-thousand-unit success is claimed.

The battlefield, measured

One 30 s sample per version · FORTRESS / RTX 5090 Chromium · main-thread CPU throttle 4× · touch viewport 390 × 844 / DPR 3 · actual framebuffer 253 × 548 · WebGPU deliberately unavailable · Low power / 30 FPS cap · 10,000 minimum living and drawn troops · warmed, autosaves on. Synthetic comparison, not a phone result.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL2 fallback · v2930 s21.0100.2 ms183.5 ms473
10,000WebGL2 fallback · optimized30 s28.366.7 ms83.5 ms463
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Six thousand full-scale troops drawn with cheaper distant silhouettes
A paused army overview with all 6,000 soldiers drawn. Geometry detail changes; dimensions and battle state stay fixed.
Original full-detail swordsmen restored when the camera approaches
Approaching the same army restores the original models. Soldiers remain 1.717 metres tall.
Ten-thousand-troop synthetic CPU pressure report
Phone-shaped Chromium on the RTX 5090 under CPU throttling. This subsequent public recheck averages 28.0 FPS and preserves the remaining long frames. It is not a physical phone capture.

v0.3 Portable rendering · first pass shipped

Permalink ↗

Packing the army for a phone

The RTX 5090 has been a useful place to build a ridiculous army. The next question is what it would take to play on a phone, or a MacBook from 2019. That changes the question to how little work a smaller computer can do while keeping the same people and world.

Graphics quality now offers High, Balanced and Low power in the main Options menu and Army & FPS. The choice stays in your browser and applies to harbor, naval and land rendering. Low power reduces pixel resolution and distant detail, turns off harbor shadows and targets thirty FPS. Touch-first browsers start there unless you choose otherwise. The cap limits unnecessary drawing, but it is a target rather than a frame-rate floor. The XR session bypasses that desktop cap.

None of those choices changes population, combat rules or physical dimensions. In Balanced and Low power, troops outside the camera keep moving and fighting while their models skip drawing. A close-camera check reduced drawing from six thousand soldiers to seventy nearby models, while preserving the exact paused battle and all thirty troop types. High keeps the earlier full-population submission path for comparisons.

Reusing indexed vertices keeps the same triangles, metre anatomy, equipment and animation channels. Across thirty models and both detail levels, stored vertices fall from 84,090 to 49,074. Static geometry plus index memory falls from 11.44 MB to 6.84 MB, about forty percent. That is a geometry-memory result, not a forty-percent FPS claim or the entire browser memory budget. Both-renderer galleries, motion captures and exact saves still pass.

The army also has a first touch control pass. One finger pans; two fingers pinch and rotate. Tap a soldier to select a company, then Move or Attack-move and tap a destination. A pan, pinch or cancelled gesture cannot issue an order. Portrait overview preserves the horizontal field, while close inspection keeps its normal field of view. A sticky Return to battle button keeps settings dismissible. Harbor interaction, naval tactics and the editor still need a complete touch pass.

Measure this battle records thirty seconds without restarting the current battle. Export keeps frame intervals, the actual graphics backend and framebuffer, quality, simulated time, dropped time, and minimum living and drawn populations. Menus, pauses, hidden tabs or changed settings invalidate a continuous active sample. Drawing an empty field is not proof that a thousand visible soldiers run quickly, so those counts stay separate.

The phone-shaped test uses a 390 by 844 viewport, touch, a device-pixel ratio of three, unavailable WebGPU and fourfold CPU throttling. It is still Chromium on the RTX 5090. It verifies gestures, fallback, state preservation and reporting; it cannot reproduce a phone GPU, memory pressure or sustained heat. Its thousand-person sample really draws and simulates all thousand troops, averages about thirty FPS at the cap and drops no simulation time, but includes a longer frame. The downloadable data keeps that hitch visible. No real-phone thirty-FPS pass is claimed.

A harder ten-thousand-person sample exposed the CPU limit: about 3.6 FPS under the fourfold main-thread throttle, with simulation falling behind. A CPU profile pointed mainly to avoidance and neighbour queries. Cached radii and conservative cell bounds remove impossible collision checks while preserving the original search budgets. Complete saved-state hashes still match the accepted simulation for legacy troops and six-thousand, ten-thousand and fifty-thousand-person cases. A repeated pressure sample reached about 6.6 FPS and still missed the target. These single synthetic samples are diagnostic, not a general speedup or a phone result. Guest snapshots also build their simulation grids lazily, so a rendering-only guest avoids that repeated allocation.

WebGPU is an option rather than an entrance requirement. WebKit documents it in Safari 26 on Apple platforms, and Chrome documents support for macOS and supported Android configurations. Capability detection and WebGL2 fallback stay in place. A 2019 MacBook is also a wide target: the Air has Intel UHD Graphics 617, while larger Pro models are different machines. Record the browser, operating system and GPU alongside any result.

The next physical-device gate is to warm the game, start with a thousand troops, then increase while recording frame times and the simulation clock. The complete thirty-type scenario starts at six thousand. Sustained battery and thermal testing, real Safari acceptance and the child’s computer remain open. For much larger phone battles, a cheaper distant representation, simulation workers and smaller co-op updates are candidates to measure. The existing PC can host while a guest chooses a lighter local profile; the guest still has network and rendering costs.

The battlefield, measured

Final optimized High profile · FORTRESS / RTX 5090 · hardware Chromium · 10,000 living/drawn troops across thirty types · 1440 × 900 · default distant detail · 8 s warm-up excluded · saves on. These are desktop measurements, not phone results.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL230 s60.016.8 ms17.0 ms0
10,000WebGPU30 s60.016.8 ms17.0 ms0
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Full-scale troops with conservative off-camera drawing
Actual game rendering: seventy nearby models drawn while six thousand soldiers and all thirty types remain simulated.
Harborfolk army controls in a synthetic phone viewport
Phone-shaped Chromium QA on the RTX 5090, with touch and CPU throttling. This is not a hardware phone performance test.
Current-device report with separate living and drawn troop counts
The report records the actual sample and its conditions. A thirty-FPS cap does not guarantee every frame meets the budget.

v0.3 Thirty Banners · shipped

Permalink ↗

Thirty ways to hold the field

The next request was to make the big battle into a scenario, then give it at least thirty different units with their own art and animations. Thirty Banners opens with three thousand Azure troops facing three thousand Crimson, split into thirty regiments on each side. The briefing waits while you look around. Reinforcements are off: defeat the Crimson army while keeping someone alive.

The roster runs from swords, pikes, shields and polearms through bows, crossbows, slings and early firearms to mounted scouts, lancers, armored cavalry, horse archers, a camel rider and a dragoon. A banner captain, duelist, staff fighter and torch raider complete the thirty. Each has original procedural geometry, defining equipment and individually authored movement and attack poses. Speed, armor, reach, range, damage and attack cadence give the models different combat profiles. This is still simplified combat rather than thirty finished special-ability systems.

Unit roster lets you meet them one at a time. Search for a name or weapon, orbit and zoom, and switch between Idle, Walk and Attack. These are the actual game models and shader motions, rather than artwork prepared for a separate gallery. Command this type selects that troop’s living Azure regiments. The inspector stays local to your camera, leaves a connected battle running, and hides the battle FPS counter so one person on an empty field cannot masquerade as a mass-battle result.

The size rule keeps applying. A standing person has 1.717 metres of anatomy, with helmets adding their own height. Riders keep full-length limbs, and the horses and camel occupy metre-scale footprints. A long pike remains long at both geometry detail levels. Pulling the camera back reduces detail, never the people’s world scale.

A long finite-combat soak found an enemy behavior problem: rear Crimson regiments could stop with nobody nearby. They now choose living Azure regiments to pursue, while your company orders remain yours. Left to its opening orders, this seeded scenario reaches a natural Crimson victory. Moving and combining the Azure regiments is part of the challenge. A full-population simulation test also proves the scenario can be won using ordinary repeated Charge orders, without altering troop health or enemy behavior. The result screen freezes the final pose and offers a rematch.

The first full-roster WebGPU check also caught a real portability failure. The shader used seventeen vertex-attribute locations, one beyond the default limit, although individual inspection models rendered. That build was rejected. Packing tint and team data together brought the complete layout to sixteen. The automated geometry checks now count actual attribute locations as well as buffer bindings. Rendered populations, shader errors and the real graphics backend are checked before accepting FPS evidence.

The same exact saves and direct two-person co-op carry this roster. A resumed battle retains its individual positions, health, targets, attack timers, replacement generations, regiments and random state. The original two-role Field of Ten Thousand and the harbor and naval scenarios remain available. The new army can also be restarted at other counts, finite or endless; the fifty-thousand slider ceiling remains an experiment.

Browser QA renders all thirty units on both WebGL2 and WebGPU, records their movement and attack poses over time, exercises type selection, restores exact six-thousand and fifty-thousand-person saves, and checks the narrow-screen inspector, finite result and rematch. The automated suite contains 229 tests. Two independent browser processes also exercise actual co-op commands, saved combat and reconnecting. The measurements below apply to this richer thirty-type roster; the previous entry’s simpler army has its own results.

Both real renderers passed the richer ten-thousand-person army for three minutes on FORTRESS with its RTX 5090: about sixty FPS, no frames over the thirty-FPS budget and no dropped simulation time. The worst frames were 17.1 ms on WebGL2 and 17.4 ms on WebGPU. Every frame submitted all ten thousand people across all thirty types, about 3.44 million triangles, with the complete army in view and autosaves enabled. This is a high-end desktop result, not a promised limit for a laptop or headset.

The art is deliberately low-poly and the simulation remains an open-field prototype. Ranged attacks use simplified damage, rather than full projectile flight; there is no siege pathfinding or dismounting system here. The ordinary harbor’s trainable units are a separate roster. Physical headset acceptance and performance on my kid’s computer still need their own tests.

The battlefield, measured

Thirty-type roster · 10,000 living troops with reinforcements · FORTRESS / RTX 5090 · Chromium 156 · 1440 × 900 framebuffer · default distant geometry · all army centers in view · 8 s warm-up excluded · autosaves on.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL2180 s60.016.8 ms17.1 ms0
10,000WebGPU180 s60.016.8 ms17.4 ms0
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Thirty Banners infantry and cavalry fighting at full metre scale
Actual published Thirty Banners battle in live play. Mounted and foot troops share the same scale. This on-screen counter includes first-use setup; the table measures separate warmed runs.
Full-size lancer and horse in the Harborfolk unit roster inspector
The inspector uses the same full-detail model, material and animation data as the army.
Camel rider with full-size rider anatomy in Harborfolk
Different mount, different footprint and gait. Camera distance changes; the person and animal stay at world scale.

v0.3 Mass-battle laboratory · shipped

Permalink ↗

Ten thousand little people

My kid had a new question: could we have a Lord of the Rings sized battle with ten thousand active characters, without dropping below thirty frames per second? That is a useful challenge for Harborfolk. Adding more people only counts if they still move, fight, save and share a world at the same scale.

The Field of Ten Thousand is a separate land scenario. Its default is five thousand Azure soldiers against five thousand Crimson soldiers on a field measuring 1,024 metres across. Army & FPS offers one thousand to fifty thousand characters, a WebGL2/WebGPU comparison, resolution controls and a distant-detail switch. The upper slider limit is an experiment, rather than a promised frame rate. Both model detail levels keep a person 1.717 metres tall.

Each soldier has a position, heading, health, target, attack timer and replacement generation. Companies of up to one hundred share movement orders. Click a soldier to choose a company, Shift adds companies, and right-click sends them somewhere. Attack-move, Charge and Hold work on those actual people. The Crimson army runs its own simple combat behavior. Reinforcement mode keeps the selected population alive by bringing new soldiers from the rear; fallen soldiers and replacements are counted separately.

The architecture work, including an Astra review, concentrated on keeping repeated work small. Typed arrays store the people. A spatial grid and bounded neighborhood queries replace all-pairs searches. Shared meshes animate limbs in the vertex shader instead of maintaining a separate scene object for every limb. Companies share destinations, while individual soldiers still move and take damage. The crowd-pathfinding literature informed that division; this open field does not implement a complete tiled flow-field pathfinder.

The first WebGPU smoke check demonstrated why frame rate alone is weak evidence: a shader exceeded the portable vertex-buffer limit, the army failed to draw, and the browser still reported fast frames. That result was rejected. Interleaving the static attributes fixed the actual rendering failure. The measurements now check the real backend, living and drawn counts, camera coverage, skipped renders, frame intervals and simulation time. Autosaves stay on during a benchmark.

Named saves and exports preserve the exact authoritative soldier arrays, company orders, combat totals and random seed. The existing invitation-and-answer flow supports one other commander. The host runs combat and accepts the guest’s company orders; a local presentation layer smooths the guest between snapshots without inventing another battle. The large full-state snapshots use substantial bandwidth, so this is a direct two-person prototype rather than a claim about a public multiplayer service.

The source passes 212 automated tests. They include finite victories saved after a multi-tick update, exact codec restoration, malformed-data rejection and bounded transport at the maximum army size. Twenty existing browser suites passed 260 named checks plus networking and co-op integration reports; the new army mode adds seven solo and nine two-client checks. They cover real pointer commands, renderer switching, exact 50,000-person named saves, rejoining and the earlier harbor and naval modes. The original eight-metre cutters, 128-metre ironclads and their visible working crews keep their own rules.

On FORTRESS with an RTX 5090, the default ten-thousand-person battle held about sixty FPS for three minutes on each real renderer. The worst recorded frames were 17.2 ms on WebGL2 and 17.1 ms on WebGPU, with zero frames over the 33.3 ms budget and no dropped simulation time. At twenty thousand, WebGPU also passed its one-minute sample; WebGL2 had three borderline 33.4 ms frames. Fifty thousand averaged roughly forty-five to forty-six FPS, but dense combat produced substantial dips. It does not meet the thirty-FPS floor. The table keeps those failures visible.

Disabling distant simplification submitted 2.48 million triangles for ten thousand full-detail people. Both one-minute samples averaged about sixty FPS; WebGL2 recorded one 83.3 ms frame, while WebGPU recorded no frames over budget and a 17.3 ms worst frame. That is a separate test from the default-detail table. Two independent co-op clients on this same machine also shared thirty seconds of warmed ten-thousand-person combat near sixty FPS, with zero over-budget frames or dropped simulation time; it is a short local two-client diagnostic, not a wide-area network guarantee.

These are stylized low-poly people and simplified individual combat on open ground. A high-end RTX 5090 result does not establish the same limit on a laptop. Siege navigation, full soldier physics, richer troop artwork and large public multiplayer remain future work. The XR tabletop currently belongs to harbor scenarios and still awaits physical headset acceptance; these battle scenarios use the desktop view.

The battlefield, measured

FORTRESS · NVIDIA RTX 5090 · Chromium 156 · 1440 × 900 framebuffer · 100% resolution · default distant geometry · all living soldiers submitted and all army centers in view. Warm-up excluded; automatic saves enabled.
SoldiersRendererDurationMean FPSp99Worst frameFrames > 33.3 ms
10,000WebGL2180 s60.016.8 ms17.2 ms0
10,000WebGPU180 s60.016.8 ms17.1 ms0
20,000WebGL260 s59.916.8 ms33.4 ms3
50,000WebGL260 s44.7116.7 ms166.8 ms250
20,000WebGPU60 s60.016.8 ms17.1 ms0
50,000WebGPU60 s45.7100.0 ms116.8 ms306
Download the recorded frame measurements ↗

Display-paced browser measurements, with saves enabled. A 33.3 ms frame is the 30 FPS budget. These results describe the stated hardware and test; the army-size slider lets you compare your own computer.

Thousands of full-size Azure and Crimson soldiers meeting on the battlefield
Actual published ten-thousand-person battle in Clash view. This gallery counter includes startup; the table samples reset after warm-up.
The entire ten-thousand-person battlefield and its live WebGPU performance counters
The complete army remains submitted and simulated in overview. The displayed public verification sample is separate from the three-minute sustained tests.
One full-size animated soldier seen close to person height
A person stays 1.717 metres tall when you move from the whole army to Human view. This paused close view changes only the camera.

v0.3 Boarding Drill · shipped

Permalink ↗

Across the plank

The cutters can now take a prize. Boarding Drill starts five Azure cutters opposite five anchored Crimson cutters, with Azure holding fire so there is time to practice. Select cutters, press B and click a red hull. The attackers take stations alongside it at real hull clearances and lay visible planks.

The whole three-person crew leaves its rowing, helm and gun stations to cross. A ship with its crew away cannot keep sailing or firing. Defenders leave their work too. Everyone remains 1.717 metres tall and the cutters remain eight metres long; the boarding gap is actual space between their hulls.

Casualties are permanent. Outnumbered defenders can strike their colors, changing the cutter’s ownership with surviving sailors aboard. Boarders fill missing crew duties and their companions return home. If every defender dies the hull is scuttled instead. Shots already in flight at surrender are removed so a newly captured prize is not immediately hit by its captors.

Plank progress, people crossing, deck fighting, casualties, captured ownership and crew rosters save with the battle. Browser checks verified a mid-fight reload and two independent commanders sharing the same boarding action. This remains cutter-to-cutter boarding; the enormous capital ships retain their much larger clearances and cannot use this tool.

Full-size sailors crossing a plank between two eight-metre cutters
Actual published gameplay: crews leave their work stations and cross the boarding plank.
Cutters alongside one another with crew fighting on the defended deck
Deck fighting, casualties and captured ownership belong to the saved shared battle.

v0.3 Fleet commanders & Leviathans · shipped

Permalink ↗

The leviathans have jobs

The next overnight milestone puts something much bigger in the strait: two 128-metre ironclads, each with a 17.8-metre beam and 64 named playable crew. Ironclad Showdown is a new scenario; the original 69-metre Sovereign battle remains available. Cutters are still eight metres and people are still 1.717 metres. No zoom trick changes those relationships.

HMS Warrior supplied a useful scale reference. The National Museum of the Royal Navy procurement specification states 128 metres overall and 17.8 metres beam; Royal Museums Greenwich preserves the gun-deck plan. Our mesh is a stylized Warrior-inspired prototype, with a modeled 115.8-metre hull and an open machinery well. It is not a museum reconstruction, and 64 playable work stations are not the full historical complement.

Every crew member has a saved name, health and duty. Engineers and stokers power steam movement; gunners serve the two batteries. Deckhands help reload and marines reduce crew injuries. Helm, lookout, sail trimming, rigging, carpentry and quartermaster duties support steering, range, propulsion, repairs and resupply. Losing a required role affects the ship rather than leaving a decorative empty body. The crew manifest lets you click a person and inspect their station at human scale.

Fleet tactics now include line abreast, column, wedge and screen formations. Destination planning uses the actual hull clearance in metres, translates whole layouts at map edges and rejects impossible arrangements. A huge leader therefore needs a genuinely large amount of sea room. Ships still use conservative navigation circles rather than full rigid-body hull physics.

Attack-move engages along a route. Patrol alternates between two points. Shift-right-click queues up to eight further waypoints; right-clicking a friendly ship gives an escort duty. Escorts reserve stations around the actual leader and switch to guard when it sinks. Guard holds position while firing in range, and Engage, Defensive and Hold fire stances control pursuit and fire discipline. Selected hull health, route lines, steam and wakes add useful feedback.

Those tactics live in the simulation, not just in a menu. A guest can issue them to the shared fleet. Two independent browser processes verified 256 ships and 890 starting crew, guest formation/stance/queued orders, real combat damage, and exact named mid-battle saving. After a fresh invitation, hulls, named crew, duties, casualties, orders and projectiles matched the host again. Co-op remains one guest with manual signaling and local named saves/export, rather than a hosted public lobby.

A follow-up adds two enemy doctrines that the host can change during the same battle. Crimson can hunt the player capital ship or assign real defensive escorts around its own. Switching changes enemy orders, not hull identity, crew or damage; the doctrine and resulting screen synchronize and save. The local Follow ship camera keeps a moving hull in view while preserving the current angle and zoom, and Overview or manual panning releases it. Both features passed actual browser and co-op checks before publication.

The source passed 185 automated tests plus TypeScript and production build checks. Browser tests exercised actual pointer tactics, close crew views, save/reload, renderer switching, narrow-screen crew controls, existing menus, save/editor flows and deliberate loading recovery. The public build repeated the tactics, both maximum-fleet co-op presets, menus and recovery checks. These screenshots were captured from the published game.

The finished ironclad preset was measured after renderer warm-up at 48, 192 and 256 ships on actual WebGL2 and WebGPU. Both held about 60 FPS on the RTX 5090 at 1440 by 900, with up to 890 starting crew. That is a display-paced result, not a maximum throughput ceiling or a claim about another computer. Boarding battles, complete naval engineering, a full historical complement and physical XR acceptance remain future work.

A full-size ironclad deck with dozens of small working crew members
128 metres overall, 64 named playable crew. The ship stays huge when the camera moves.
Close view of the helmsman on the ironclad with the named crew manifest
Click a name to meet the person at their work station. People stay 1.717 metres tall.
Engineers and stokers in the ironclad open machinery well
The open machinery well makes the steam crew visible; losing those roles affects movement.
Full-size ships with wedge formation route lines and tactical command controls
Attack-move, formations, stances and queued routes are saved orders for actual hulls.

v0.3 RTS front end & original score · shipped

Permalink ↗

The captain’s desk

The next overnight request was for Harborfolk to feel more like an RTS from the moment it opens. We now have an original illustrated harbor cover, a loading screen with actual model progress, a vertical start menu, scenario selection, new-harbor setup, options and an in-game menu. Continue returns to the saved world rather than quietly starting another settlement.

The reference pass looked at Age of Empires II, StarCraft, Warcraft III, Command & Conquer and Rise of Nations. Their interfaces have different personalities, but several useful patterns recur: stable resource locations, commands that follow the selection, readable key hints, a minimap, and clear separation between starting, loading, configuring and playing. Those conventions informed the captain’s desk; their artwork and branding were not copied.

The game HUD now groups timber, supplies, stone, population capacity, age and time into a compact strip. A bottom dock gives the selection an identity and keeps the six primary actions in a consistent command grid beside the coast chart. Objectives open when wanted. The centre stays available for the people, construction and huge vessels that make this project interesting.

The naval battle gains a strait chart and local fleet control groups. Click the chart to move the camera, or right-click to order selected ships. Ctrl+1–4 saves a selection; 1–4 recalls its surviving ships. Those are commands for the actual existing hulls, not substitutes for their scale. The ship slider, crew duties, combat saving and WebGL2/WebGPU comparison remain intact.

Menu transitions needed careful treatment. A solo world waits without changing its saved pause flag or camera. In connected co-op, browsing a menu does not pause the shared battle for the other player. The title menu also leaves an existing autosave unchanged until the player continues. Native save dialogs retain their Escape behavior, and new world starts preserve the previous recovery snapshot.

There is also an original procedural nautical score: distinct menu, harbor, battle and victory phrases, layered soft synthesis, low percussion, compression and subtle reverb. It is synthesized orchestral-style music, not a recording of an orchestra. Menu music begins after a gesture, master mute is genuinely silent, and independent music/effects volumes persist. Opening the menu suppresses hidden chopping and battle-effect noise.

The cover is original illustration, not a gameplay screenshot or a new promise about historical engineering. The runtime still uses the same metre dimensions. These images are actual captures of the published interface; the command-desk views show the game itself. Desktop, narrow-screen and reduced-motion layouts were checked.

The shipped source passed 171 automated tests. Real browser checks covered normal boot, menu navigation, a saved-world round trip, onboarding, measured audio output and mute, loading recovery, existing editor flows and maximum-fleet co-op. That co-op check also opened the guest’s menu while the host’s simulation continued. The refreshed production UI held about 60 FPS in display-paced RTX 5090 tests at 48, 192 and 256 starting ships on both renderers; this still does not establish an uncapped performance ceiling. Family feedback on the score and controls remains the next human check.

Harborfolk title menu over original dawn harbor illustration with a huge sailing ship
Original cover illustration, a familiar start menu, and a world that waits for Continue.
Harborfolk scenario selection showing the naval battle and developed civilizations
Choose a first harbor, a developed world or a staffed naval battle.
Harborfolk village with compact resource bar, minimap, selection card and command grid
The actual game: readable resources and commands around a clear playfield.
Harborfolk options with independent audio sliders and keyboard control hints
An original synthesized score with separate persistent music and effects gains.

v0.3 epic naval battle · shipped

Permalink ↗

A battle worth coming back to

The Azure Strait is our first fleet-sized battle: blue and red ships, two enormous flagships, visible sailors, physical shots and actual casualties. The Fleet & FPS slider chooses 6–128 ships per side: as many as 256 hulls and 778 people at the start. It adds complete ships and people to the simulation. It does not scale them down to fit the screen.

The little cutters are eight metres long. The two stylized great ships retain the existing 69-metre overall length, 11.3-metre beam, 18-metre yards and 52.4-metre keel-to-mainmast height. Every person is 1.717 metres tall. Up close, a sailor looks wonderfully small beside the hull. The large vessels use eight prototype work stations; they do not claim to reproduce the much larger historical crews or complete naval engineering.

Crew jobs matter. Without a living helmsman and rower, a cutter cannot move; without its gunner, it cannot fire. On a flagship, helm and sail handling permit movement, the two gun stations fire, the lookout extends range, the rigger improves speed, the carpenter repairs a living hull and the quartermaster improves reloading. Shots lead a moving target but must still physically reach it. Sinking leaves living survivors afloat, and repairs never bring back lost people.

Parent and child can command the same blue fleet through Co-op. A two-process browser test connected the maximum fleet, issued a guest order through a real right-click, fought, paused, and saved a named mid-battle moment. After disconnecting, the host reloaded it and created a fresh invitation. Ship positions, damage, velocities, orders, crew identities, jobs, casualties and in-flight projectiles matched exactly in the guest browser. Resumed combat then reached both players.

That larger state needed a transport change: snapshots now travel in bounded ordered pieces rather than assuming one enormous WebRTC message will arrive. Saves remain local to the browser profile, with JSON export for a portable backup. The connection still supports one guest with manual invitation and answer exchange, and there is no relay service. Those boundaries are worth understanding before planning a remote session.

The battle also has an actual WebGPU renderer beside WebGL2. Switching keeps the paused combat state and camera; each player chooses their own backend. Shared instanced geometry batches small hulls, limbs and shots. The original harbor keeps its established WebGL fog pipeline, and combat AI still runs on the CPU. The counter reports measured browser FPS, frame milliseconds and draw calls, and identifies a fallback honestly.

On Fort’s RTX 5090, production browser runs at 1440 × 900 held approximately 60 FPS with 48, 192 and 256 starting ships on both real hardware backends. The runs included moving ships, shots and casualties after warm-up. They were display-paced near 60 FPS, so they do not establish a throughput ceiling or a WebGPU speed advantage. Performance on the child’s computer is still a family playtest question.

The shipped build passed 169 automated tests, the maximum-fleet co-op save/resume round trip, backend-switch preservation and the existing eight save/editor browser checks. These images are captures of the public game. The earlier loading complaint resolved when Chrome was restarted; that human result does not establish a technical root cause. Naval headset rendering remains future work, while harbor scenarios retain the existing tabletop prototype.

Harborfolk naval fleet slider set to 128 ships per side with a real WebGPU FPS counter
Up to 256 starting ships and 778 crew. Combat changes the live counts; the hulls keep their metre dimensions.
Small sailors at their stations on a huge full-scale Harborfolk flagship
A 1.717-metre person on a 69-metre overall ship. Eight prototype jobs are visible and functional.
The guest browser showing a restored naval battle with surviving ships and crews
A named save restores the actual combat roster through a fresh multiplayer invitation.

v0.3 saves, scenarios and editor · shipped

Permalink ↗

A world to return to

The next request was less about another building and more about keeping a world: build a civilization together, save it, and return to it. The Saves library now keeps named harbors beside the automatic browser save. Jobs, crews, trade routes, paid construction, technology and explored terrain return with the people who were doing the work.

Scenarios gives you three starting points: a busy working harbor, a defended Citadel, and a maritime city with the enormous Hall and Great Warship. These are authored starting civilizations, with deliberately supplied resources and completed projects. Their farmers, merchants and crews run the actual simulation. They are useful places to explore what the prototype can become, rather than claims that every screenshot was earned from an empty map.

The Map Editor opens an isolated, paused copy. Paint land and water, place buildings and people, move objects by their coordinates, staff a vessel, and choose starting resources, age and fog. Terrain uses an eight-metre authoring grid; navigation, ordinary city construction and the rendered coastline use the same data. The town can grow beyond the starter harbor and onto an authored island when builders can reach it. Buildings, ships and villagers retain their metre dimensions.

Undo and Redo keep the recent edit history. Validate checks the draft, and Play test runs a disposable copy with ordinary game controls. Back to editor restores the draft. Return to harbor restores the original settlement and camera, leaving its autosave intact. The fixed starter harbor and infrastructure corridors remain protected, and the existing population and fleet caps still apply.

A custom scenario or saved civilization can be exported as a JSON file and imported in another browser. For co-op, the host loads that world first and creates a fresh invitation; the other player joins its restored state. You can save the shared progress and repeat that process next time. The file preserves the civilization, while the direct connection and local camera are recreated. There is still one guest, manual invitation exchange and no relay service.

A reported endless loading screen also prompted recovery work. Asset progress now shows what is opening, and bounded loading offers retry and export of the retained save. A separate guard covers the case where the main application script never finishes opening. A fresh public browser loaded successfully during investigation; that did not by itself reproduce the reported failure.

The screenshots are actual captures of the published build. It passed 152 automated tests, browser checks for edited terrain, save files, developed worlds and return-from-play-test recovery, plus three deliberate loading-failure checks. Two independent browsers also joined an edited civilization, issued a shared order, saved the job in progress, reloaded and reconnected to finish it. Physical headset comfort, performance on the child’s computer and the next family playtest remain open checks.

A huge Hall and launched warship overlook a developed Harborfolk town
An authored civilization, with the same human and building scale as a new settlement.
Harborfolk save library showing named saved worlds and export controls
Named moments, portable files and a previous harbor kept for recovery.
A full-size cottage selected in the Harborfolk map editor with coordinate controls
The editor moves a footprint through the world; it does not shrink the building.
Narrow Harborfolk screen with the map editor occupying its lower half
A compact editor leaves the upper playfield available for painting and placement.

v0.3 full-scale landmarks · shipped

Permalink ↗

Wait… it’s that big?

The next rule was simple: keep everything at scale, even when it gets inconvenient. The playable map now spans 640 by 480 metres. The camera can inspect a doorway or pull back over the coast. Fog follows that larger geography, and a Great Warship needs open sea instead of squeezing through the little cargo channel.

The Great Hall takes its roof plan from Westminster Hall: 73.2 by 20.7 metres. Villagers stay about 1.72 metres tall. You choose a surveyed, clear site and pay its actual material cost. Stone courses, thirteen timber trusses, roof bays and coverings appear as separate full-size pieces. The total height and exterior treatment are prototype choices; this is a stylized reference building rather than a complete historical reconstruction.

Construction depends on people reaching the site. Builders walk to work stations and climb to the stage they are working on. Stopping sends them safely down; resuming keeps the paid materials and existing parts. Small city buildings now assemble by courses and bays too. Dock production uses arrived shipwrights, with the keel, ribs and hull appearing as work progresses.

The Great Warship takes its main dimensions from Vasa: 69 metres overall, an 11.3-metre hull beam, and 52.4 metres from keel to mainmast top. Its broader yards need more space than the hull alone. Building starts with a bottom shell before inserted frames, then planking, decks, masts, rigging and sails—a simplified nod to the documented Dutch construction method. A finished ship launches once, lowers into the water and joins the fleet.

Eight actual villagers occupy playable duties: helmsman, lookout, two gunners, sail trimmer, rigger, carpenter and quartermaster. Vasa was designed for roughly 145 sailors and 300 soldiers. Those eight stations are a gameplay crew, not a claim to reproduce that entire complement. No extra decorative people are spawned to suggest a staffed ship.

The Scale panel fits a project in view, takes you close to the Hall doorway, and offers stone and sailcloth colours. A cutaway reveals the timber structure without changing its size. Cosmetic changes stay local and cost nothing. Two co-op players can place and build the same Hall while keeping their own camera views.

The build passed 119 automated tests, plus browser checks for earned materials, surveyed footprints, stopped and saved construction, launch, occupied duties and repeated convoy trips. Two independent browsers also exercised guest placement, safe descent and resumed Hall construction over the real direct connection. The pictures are gameplay captures with the actual simulation advanced for testing. Physical headset comfort and performance on the child’s computer remain unmeasured.

Harborfolk villagers on scaffold ladders while separate full-size Great Hall roof trusses appear
Reached work stations, stone courses and separate trusses. The people keep their height as the Hall takes shape.
A completed enormous Great Hall with several small human-sized villagers beside it
A long roof and tiny-looking people, sharing the same metre scale.
A close view of the Great Hall entrance steps with two villagers beside its fixed-size door
Getting close brings the doorway and people back into view. Their dimensions stay fixed.
A Vasa-inspired Great Warship under construction with masts and real assigned builders
The hull, decks and masts assemble as separate parts, with actual builders at their stations.
A narrow browser view of the Great Hall with its roof covering hidden to reveal the timber structure
Peek inside changes visibility, leaving the structure’s scale intact.
A launched full-size Great Warship with towering masts and villagers visible at occupied deck stations
The finished hull joins the fleet with eight occupied roles. The spars and rigging tower over the crew.

v0.3 convoys · shipped

Permalink ↗

Ships with jobs, and company

Up to six ships can share the harbor. The dock builds additional cargo boats and warships, each with its own berth and crew. Cargo boats need two people; warships need a rower, helmsman, and gunner. People walk to their assigned ship, and a hull waits until its crew is aboard.

A repeat trade route spends twenty actual supplies. Crew members carry exports aboard, sail to the outpost, and take them ashore. They then collect imports and bring them home. Timber and stone become spendable only when those imports are unloaded at the harbor. The route repeats while supplies and crew are available; a low stock pauses it rather than conjuring another load.

Select a crewed warship and right-click a cargo boat to escort it, or choose that boat in the ship panel. Its escort follows a reachable position with room between the hulls and engages the raider when the convoy is threatened. Multiple arrivals queue for the outpost berth. Stop disables repetition, while people finish handling a paid manifest safely.

This draws from Anno’s automatic trade routes, stock safeguards and separate piers, Homeworld’s guard and support orders, and 0 A.D.’s crewed trading ships. The research keeps official features, developer plans, qualitative requests and dated mod attention separate. These are useful references, rather than a ranking of every RTS player’s wishes.

The metre rule is still the center of the project. These boats retain their eight-metre hulls, human-sized people and individually visible crates. A much larger geography and evidence-backed Great Hall and Vasa-inspired ship are under development. Their roof, hull and mast dimensions will dictate the space they need. They are not yet part of this published convoy milestone.

The integrated build passed 104 automated tests. The browser checks exercise normal resource gathering, repeated paid trips, actual cargo handling, escort combat, reload and two-browser guest commands. Dock ship production still uses a timed assembly queue; arrived shipbuilders and staged construction are the next milestone. The screenshots below are real gameplay captures with the simulation advanced for testing.

Villagers carrying individual export crates along a Harborfolk pier to their cargo boat
Loading starts with two people collecting real supplies and carrying them aboard.
Two Harborfolk crew members unloading imported cargo at their home pier
A completed trade trip becomes spendable timber and stone as the crew unloads it.
Multiple individually crewed cargo boats and a warship at separate Harborfolk piers
Each hull has its own berth, cargo, orders, and people.

v0.3 plans and patrols · shipped

Permalink ↗

Small plans, and people to follow

Plans turns the sandbox into three small optional adventures. First crossing guides a dock, crew, and twenty-supply outpost delivery. Woodland Supply follows a farm and an extra cart through its first timber run. Coastal Watch asks for a scout, a mainland discovery, two guards, and a raid victory. The next goal and a practical hint stay in view.

Each plan keeps its checklist, elapsed simulation time, and completion badge with the saved harbor. Existing accomplishments count, and returning to free play preserves the village. The opening customization screen can choose a plan. Two co-op players share the same progress, so one can build while the other explores.

Units also have duties that continue after a single move. Select a guard, scout, archer, or medic, press P, and right-click the far end of a patrol route. Press E and right-click a villager, cart, or friendly unit to escort it. A selected unit shows a dashed route or connecting line. Guards and archers respond to nearby threats and resume their duty; scouts stay out of combat until ordered, and medics still heal nearby wounded allies.

Escorts follow actual positions and use land paths. They wait and retry if a route is blocked, rather than jumping through buildings or water. Stop or an ordinary move or attack clears the routine. Saved routines and guest orders use the same validated state as the rest of the game. Headset-native patrol and escort buttons remain future work; the target resolver is shared with the controller adapter.

The build passed 76 automated tests, a browser adventure using real gathering, construction, crewing and cargo delivery, and real two-browser co-op checks for shared plans and guest patrol and escort orders. The captures below come from gameplay, with the simulation advanced for testing. Physical headset comfort and performance on the child's computer remain unmeasured.

Harborfolk's Plans panel after a real cargo delivery completed the First crossing adventure
A first crossing, completed through gathering, building, boarding, and an outpost delivery.
A selected Harborfolk scout following a dashed patrol route
A scout patrols between its starting position and the chosen endpoint.
A Harborfolk scout escorting a moving villager with a visible connecting line
The escort follows a person who is actually walking, with a land route and room to stop nearby.

v0.3 Polyfork expansion · shipped

Permalink ↗

More places, more useful people

A Scout Lodge, Harbor Market, and Infirmary give the next group of people useful jobs. Scouts move faster and reveal more coast, with an attack order needed before they engage. A market trader has to reach a stall before exchanging food for timber or stone. Medics train at the infirmary and heal nearby living allies after a fight.

The expansion uses six additional free Polyfork models: a tool shed and trail-map kiosk for the lodge, an oven and drying-net rack for the market, a hand axe for the scout, and potion bottles for the medic and clinic. Existing stalls, barrels, sacks, and the Beach Hut generator supply the remaining bases. The animated people remain the original humanoids, with Polyfork equipment attached. Everything is placed in meters; doors, props, and people keep their scale.

For a possible Polyfork case study, the build retains each source page, free-status check, license check, file hash, measured dimensions, and geometry budget. The six new source models total 2,777 triangles, with one material and no textures each. Those are asset figures, not a claim about the game's frame rate. Models ship inside the game. This is an independent prototype; no Polyfork partnership or endorsement is claimed.

The RTS research points toward readability, guidance, co-op scenarios, and distinctive rosters. The Age of Empires team highlights tree and wall visibility, idle-worker warnings, and guided build orders among its popular mods. Forgotten Empires reports over a million official downloads before becoming official game content, while Mental Omega pairs a distinctive roster with extensive solo and co-op missions. These are directional signals from named projects and platforms, rather than a representative survey of all RTS players.

This batch applies the smaller lessons first: an Idle button and I shortcut, local selection groups on Ctrl plus 1–4, and optional see-through trees without shrinking them. Decorative trees yield to building footprints, while timber-bearing trees still have to be cleared normally. Short shared scenarios, patrol and escort orders, and regional rosters are promising next experiments. Deeper naval logistics also fit Harborfolk, though the direct popularity evidence is weaker.

The build passed 62 automated tests, the existing 29-check city playtest, and eighteen new browser checks covering real construction, scouting, staffed trade, actual raid damage and medic healing, saved reload, and narrow controls. Two separate browsers also built a Scout Lodge and trained a scout through the real co-op connection. The images below are gameplay captures. Human feedback, target-device frame rates, and physical headset comfort remain future checks.

Harborfolk's Scout Lodge built from a Polyfork shed and trail-map kiosk, with a newly trained scout
A compact lodge assembled from real-meter models, with the first scout outside.
A working Harbor Market with a trader, stalls, bread oven and drying-net rack
Asha has arrived at a market stall. Trading becomes available when the assigned person is there.
The Harborfolk Infirmary with a medic and villager outside
A one-bay Beach Hut becomes the clinic, with access stairs and a medic carrying a potion bottle.
A Harborfolk medic treating a surviving guard after a real land raid
A living guard was wounded in battle. The medic walked over and restored health within range.

v0.3 automatic carts · shipped

Permalink ↗

A cart that knows the way home

The next request was to let the cart finish its own work loop. Moving it to a woodland loading spot now assigns a dedicated operator and enables automatic hauling. When twenty-four logs are aboard, that person pulls it home, unloads it one log at a time, and brings the empty cart back to the same spot.

The operator waits at the handles while gatherers fill it again. Each cart has an automatic-hauling toggle and a Stop button. Deliver timber can send a partial load home too. The timber stays aboard when a journey is stopped, and saved routes retain their loading spot and operator.

More carts can be built through Build, the storehouse, or a cart panel. Each costs twenty wood. A villager walks to the assembly site and builds its parts over eight seconds; interrupted work can resume without paying again. Up to six carts have their own loads, routes, and people. Gatherers choose the nearest useful parked cart.

This build passed fifty-seven automated tests, the fourteen manual-cart browser checks, and eighteen new checks covering repeated automatic deliveries, visible assembly, a second cart receiving real timber, saved loads, narrow controls, and reset cleanup. In two separate browsers, the guest also built and positioned a second cart through the shared game connection. The browser tests advance the actual simulation and use normal orders and harvested resources. Human playtesting is still ahead.

A full Harborfolk timber cart with its assigned operator at the handles
The load is full. Its operator is starting the trip home, then will return to this loading spot.
Mara assembling the wooden platform of an additional Harborfolk cart
A second cart begins with a frame. Its villager has arrived and is building the remaining parts.

v0.3 cart and chat · shipped

Permalink ↗

Bring the cart to the woods

The cart can now go where the work is. Click it and right-click clear land near the woods. An available villager walks to its handles and pulls it along a land route. Nearby gatherers use the cart when it is closer than the storehouse.

Each deposited log appears on the cart, up to a stack of twenty-four. That timber stays on board until the operator brings it home and unloads it. Deliver timber sends it to the storehouse; each unloaded log becomes spendable. Redirecting or stopping the cart preserves its load. The original supply voyage still works, with crates appearing as the cart is loaded.

Connected co-op players also have a collapsible chat window and an unread badge. Typing stays separate from game shortcuts. Messages travel over the same direct connection as orders, with the latest fifty kept in the current page. This is chat with one invited partner; joining still uses the invitation and answer codes.

The combined build passed fifty-one automated tests, fourteen cart browser checks, and a real co-op playtest in two separate browsers. That playtest sent cart orders from the guest, exchanged messages in both directions, checked narrow controls and disconnect behavior, and confirmed that message text renders literally. The cart captures use actual gathering and delivery orders, with the real simulation advanced for testing. Human playtesting is the next check.

A Harborfolk cart parked near the woods with a stack of twenty-four individually rendered logs
A full cart beside the woods. Its visible logs match the timber waiting to be delivered.
Two connected Harborfolk players coordinating timber gathering through the chat panel
An actual two-browser co-op conversation, carried by the game's direct connection.

v0.3 controls · shipped

Permalink ↗

Click a person, give them a job

The next request was the familiar strategy-game gesture: click a person, then right-click the thing they should work on. Trees mean gathering; a farm means food; a quarry means stone; an unfinished building means construction. A boat takes the selected people as its crew, and the cart takes an operator. Empty ground still means move.

Farm and quarry workers can switch jobs directly. Small construction and income crews now stand at offset work positions, making individual villagers easier to select. Resuming a building does not charge for its materials again. Co-op uses those same orders; a tree-click coordinate also needed a fix to pass the shared command validation.

You choose sites for all seven settlement building types through Build. A footprint checks explored land, water, obstacles, and people before the villagers walk there. Right-click or Escape cancels placement. The original dock and workshop still use their fixed locations.

For WebXR, Little Cities is a reference for a calmer tabletop interface. The controller prototype now has a labeled building bubble palette inside the headset, age locks, selection and resource feedback, and buttons to scale or rotate the table. Trigger selects or places; squeeze gives a contextual order or cancels the active placement. A green or red footprint previews the site.

Automated checks cover controller callbacks and restoring the desktop camera. The browser playtests use actual mouse selection, farm and quarry orders, construction placement, and boarding a cutter with three occupied roles. Physical headset comfort and controller feel remain untested; this update does not add hand tracking.

Selected villagers working at separated stations around a Harborfolk farm
A working farm from the mouse playtest, with people selected and assigned in the world.

v0.3 · shipped

Permalink ↗

From a dock to a working town

The harbor now has a small city economy. Villagers walk to building sites, raise their structures, tend farms, and quarry stone. Cottages add room for more people. A warehouse trades food for timber and stone, so the settlement can keep growing after the nearby trees are gone.

Settlement, Harbor, and Citadel ages give that growth a direction. Eight technologies form a prerequisite tree, with useful changes to harvests, construction, ships, and defense. An academy and barracks are buildings you need to finish, rather than menu unlocks that appear by themselves.

Guards, archers, and watchtowers add the first land battles. Peaceful mode lets you choose when to try a raid. Adventure mode allows automatic raids but gives damaged buildings a way to recover; Hard mode allows destruction. The opening screen also lets you name the harbor, choose its banner, and change the landscape.

Two browsers can now command the same settlement through direct co-op. One browser runs the simulation; the other sends orders and receives updates, with its own camera and selection. Moving people and ships ease between those updates. Joining currently means exchanging invitation and answer codes, and some networks will need a relay we have not added yet.

There is also a WebXR tabletop view for compatible VR browsers, with a scale control and controller selection. Automated checks cover entering, leaving, and restoring the desktop view. Headset comfort and controller feel still need a physical playtest.

This build passed 43 automated tests, the original cargo and naval browser playtests, a complete city progression playtest, and a real two-browser co-op test. The city captures below come from ordinary game orders with the actual simulation advanced for testing. No resources or units were injected. The next verdict belongs to my kid.

2026-10-11 adjustment: the boats now turn more quickly in response to a new heading. The cargo boat, cutter, and raider use the same faster response, and their crew stations follow the hull.

Villagers tending a newly constructed Harborfolk farm
Food comes from assigned people who have reached the farm.
Harborfolk's Citadel Age and technology research panel
Three ages and a technology tree with real prerequisites.
Guards defending the Harborfolk settlement from land raiders
The first land raid, with selected defenders and visible opponents.
A working Harborfolk town with placed buildings
A settlement built through play, ready for the next human playtest.

v0.2 · shipped

Permalink ↗

An upgrade you can watch being built

The next experiment was a workshop that changes while villagers work on it. Polyfork supplies a parameterized building: the upgrade adds another bay and steepens the roof in visible stages. Doors stay people-sized as the building grows.

Paint, roof color, and weathering are separate controls. They are there to play with, without spending the resources needed for a functional upgrade. The finished workshop earns fishing supplies.

The same update adds unexplored coast, a dock production queue, and a cutter with three occupied stations: rower, helmsman, and gunner. A raider pursues and fights the crewed ship. Music, footsteps, tools, and battle sounds are synthesized in the browser.

This release passed 14 simulation tests and 34 automated browser checks. The captures below come from that working build. Camera orbit, close inspection, and a wide overview all use the same 3D world. The next playtest with my kid is still ahead.

Harborfolk workshop before its upgrade
Before: one bay, with people-sized doors.
Villagers working on an expanding workshop
During: villagers walk to the site and build while its parameters change.
Completed two-bay workshop with cosmetic controls
After: a second bay, a steeper roof, and cosmetic controls.
A crewed cutter confronting a raider at sea
The first naval encounter, with people at their stations.

v0 · shipped

Permalink ↗

Start with a boat that needs people

My kid loves Age of Empires and keeps asking for a version where people, ships, and buildings share a believable scale. He also wants to see who is actually operating the vehicles. Harborfolk starts with that question.

The first small loop has eight villagers, a wooded shoreline, and a boat. Villagers chop timber and carry it home, build a dock, load supplies, and deliver them across the water. The cargo boat needs two people, visibly rowing and steering. A cart has an operator too.

The world uses one meter per scene unit. Its inhabitants stand about 1.7 meters tall and the first boat is about eight meters long. It is a stylized little settlement, with scale treated as a rule for the assets rather than adjusted separately to fit an icon.

Harborfolk's original 3D village and shoreline
The first playable settlement. These are screenshots, not concept art.
Three visible villagers operating a cutter
A later close look at the crew: each occupied station has a job.