Go to file
dejvino 9ae00dc978 Sharper steering, solid traffic, and a map that is mostly war
Steering. The falloff with speed was linear, so the car had already lost a third
of its lock by 20km/h — exactly the speed you take a right-angle junction at, on
a map made of right-angle junctions. It is quadratic now, which leaves low
speeds nearly untouched and still calms things down at pace, plus a faster rack
and a little more lock. Pinned with a test that drives an actual quarter turn and
measures its radius. A first attempt asserted that fast corners take longer,
which is simply false: a fast car swings through ninety degrees quicker, just
across far more tarmac. Radius is what a corner costs you.

Contact. Units were points that ignored everything, so there was nothing to hit.
Vehicles now carry kinematic bodies and ramming one is a real collision that
damages them and shoves you. People get no collider on purpose — a capsule means
snagging on pedestrians or launching them — so you drive through them and they go
down, and running over a civilian is counted for later.

Territory. Liberated ground covered more than half the map because the player
started in the middle of it, so most of the world was safe by geometry. The
starting base is now at the friendly edge and the bands are anchored to give
roughly 15/25/27/33 across liberated, contested, occupied and frontier. Measured
in the running game rather than guessed, since the axis runs diagonally across a
square and area is not linear in depth.

Also fixes a fragile test that compared reward rates across whichever targets a
board happened to offer, and so mostly measured route length rather than the
novelty premium it claimed to check.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-07 21:42:17 +02:00
.claude Phase 0: drivable car with persistent wear 2026-08-07 07:39:05 +02:00
src Sharper steering, solid traffic, and a map that is mostly war 2026-08-07 21:42:17 +02:00
.gitignore Make the running game inspectable, and stop the beacons swallowing the view 2026-08-07 15:38:27 +02:00
index.html Phase 0: drivable car with persistent wear 2026-08-07 07:39:05 +02:00
package-lock.json Phase 0: drivable car with persistent wear 2026-08-07 07:39:05 +02:00
package.json Phase 0: drivable car with persistent wear 2026-08-07 07:39:05 +02:00
README.md Sharper steering, solid traffic, and a map that is mostly war 2026-08-07 21:42:17 +02:00
tsconfig.json Phase 0: drivable car with persistent wear 2026-08-07 07:39:05 +02:00
vite.config.ts Make the running game inspectable, and stop the beacons swallowing the view 2026-08-07 15:38:27 +02:00

Drive Between the Lines — prototype

Endless driving survival game.

  • Phase 0 — a drivable car with persistent wear that is felt through the wheel.
  • Phase 1 — a road network whose roads remember being driven, and escalate.
  • Phase 2 — bases, a quest board, and known-vs-new target selection.
  • Phase 3 — regional control: behind your own lines, nobody is watching.
  • Phase 4 — the front line moves, from your work and on its own.
  • Phase 5 — four distinct jobs, radio texture, and work found in the field.
  • Phase 6 — an inhabited world: traffic, people, patrols, and a live war.
  • Phase 7 — sound, synthesised from the simulation.

Running

npm run dev

Then open http://localhost:5173. Append ?seed=anything to regenerate the world — seeds may be numbers or words.

W / S throttle / brake-reverse
A / D steer
Space handbrake (rear wheels only)
R respawn the car — does not repair it
13 accept a mission, when parked at a base
M sound on/off
hold R wipe the campaign and start over (5 seconds)

Progress saves itself every 20 seconds, and immediately whenever you take or finish a job. Saves are per seed; ?seed=x is a separate campaign.

Drive to a green beacon (a base) and stop. Take a job, drive to the amber beacon, stop and hold position for three seconds, then return to any base — the amber beacon goes out once the job at the target is done, since the target is no longer anywhere you need to be.

npm test        # sim + headless physics
npm run build   # typecheck + production bundle

Road heat

Every metre driven on or alongside a road adds heat to that segment; every road everywhere sheds it slowly. Cross a threshold and the road escalates:

Clear → Patrol → Barricade → Turret

  • Patrol parks a vehicle on the verge. The road narrows.
  • Barricade puts concrete across it, leaving a gap you have to slow for.
  • Turret adds a tower overlooking the checkpoint.

Roughly four traversals take a road from clear to turret; an untouched road cools off in about four minutes. Both numbers are in sim/heat.ts and both are guesses meant to be tuned.

The road hierarchy

Roads are not all the same road:

Class Tarmac Speed Attention
Track (single car) 4.5 m ×0.55 ×0.55
Road (two lanes) 9 m ×1 ×1
Trunk (four lanes) 15 m ×1.35 ×1.35

That trade is the network doing real work rather than being the floor you drive on: a trunk is fast and direct and exactly where anyone looking for you would look; a lane is slow and awkward and nobody is watching it. Trunks run the whole width of the map and are never thinned away, so there is a shape to learn — get onto the trunk, then work down the lanes — instead of a uniform mesh where every road is as good as every other. Route previews are costed by travel time, not distance, so the board's estimate matches the way a driver would actually go.

The route corridor

Heat is credited to a road for anything within its corridor — half the tarmac plus a 7.5 m verge, so it scales with the class — not just the tarmac itself. Otherwise the dominant strategy is to drive next to every road and never accrue heat at all.

Two consequences follow, and both are deliberate:

  • The cutoff is hard, not a taper. A taper leaves a gradient to optimise along — sit at 80% of the corridor, take 20% of the heat. A cliff edge means you either use the route or you actually leave it.
  • Barricades span the full corridor, well past the kerb. A checkpoint that stopped at the tarmac would be free to round on the verge at speed.

The 7.5 m verge is about two car widths, so hugging the shoulder gains nothing, while leaving most of the map genuinely off-route. Widening it backfires: make the corridor too generous and heat becomes unavoidable everywhere, and "take a different route" stops being a choice at all.

Buildings, and why they are dense

Off-road used to be open country, which made every barricade optional — you just swung wide around it. Buildings are now laid out on a jittered lattice (random scatter cannot pack boxes and saturated at a third of the target), which takes bypass lanes around a checkpoint from none obstructed to about two thirds. You can still get around one; it means threading a maze at walking pace instead of a free detour at speed.

Other deliberate choices

  • No AI yet. The props are stationary hazards. Escalation currently means the road gets slower and more expensive to get wrong, which is enough to test the phase's real question.
  • Changes never happen on the road you are on. They are deferred until you leave. Otherwise a barricade would spawn a static collider inside your car — and you are supposed to discover escalation by coming back, not by watching it assemble behind you.
  • A checkpoint's position is seeded from its segment, so a given stretch of road always fortifies in the same place. Recognising it is part of learning the map.

The road network is a jittered grid with roughly a quarter of its edges removed, keeping the whole thing connected. The loops are the point: "take a different road this time" is not a decision unless alternative routes exist.

Missions and intel

The board at each base offers a mix of known targets and new ones, and describes the route to each in words rather than numbers — "barricaded · last word 6 min ago", "no one has been down that way".

The important part is where those words come from. The board never reads heat directly. It reads sim/intel.ts, which is a snapshot of what the player has actually observed and when. Drive a road and you write down what was on it; leave it alone and your note ages while the road keeps escalating without you. So a known route is genuinely knowable, genuinely stale, and never a guarantee — which is the tension the brief asks for.

Recon missions are the tool for converting a new target into a known one: completing one surveys every road within 130 m, filling in the map without having to drive each one.

Four jobs run end to end, and they differ in pacing and risk, not just in the noun on the board — otherwise four mission types is one mission type with four labels:

Job At the target The risk
Supply run 3s unloading the drive out
Retrieval 5s loading the drive home, with goods that break
Intel drop 9s transmitting a long wait, stationary, in the open
Survey 2s cheapest on site; the payoff is the map

Retrieval is the one that carries something home, and impacts damage it. It pays the most, but it pays for what arrives, not what you set off with. Hand-in works at any base, not the issuing one, because the brief is explicit that a blocked road home must never strand the player.

Work you find rather than work you are given

Past the line, and only when you have nothing in hand, someone occasionally turns up: a wounded stranger, a contact flashing headlights, a wreck worth stripping. They are deliberately not on a board — you cannot plan a route to one, compare it against anything, or shop between them. You take the detour in front of you or you drive on.

Radio

Chatter that frames who you are — an undercover driver being talked at by people busy elsewhere. Every line is triggered by something actually happening in the sim: crossing a border, a road you have made notorious, a car past saving, or the front genuinely drifting your way, which is the brief's "the enemy has other priorities" said out loud without ever showing a number.

Rewards, and why repairs exist

New targets pay NOVELTY_BONUS (1.8×) more than known ones. That multiplier is the phase's tuning dial: if you never take the new target it is too low, and if you never take the known one it is too high.

Payment is in parts, which repair the car — the only thing in the game that pushes condition back up. Repairs go to the worst subsystem first and stop at its ceiling, which drops permanently with every bit of damage taken (PERMANENT_SHARE, currently 30%). So a car can be patched back to what it is still capable of, but never to what it was. The HUD bars show this directly: is current condition, · is what repairs could still reach, × is gone.

This is what gives the board's reward a reason to matter. Without an economy the choice between a safe route and a risky one has no stakes; with one, the risky job is what keeps the car alive a while longer.

A world with people in it

Up to Phase 5 the world was scenery. Heat poured concrete on a road and that was the whole of the enemy. sim/units.ts is the layer that makes it inhabited.

Enemy ground is patrolled because it is enemy ground. There are standing patrols on the roads wherever you are past the line, and more of them the deeper in you go — none in liberated territory, one in contested, three in occupied, five on the frontier. Nobody summoned these; they are what hostile territory looks like.

On top of that, nothing appears — things are sent. When a road turns notorious a patrol is dispatched: it starts at a junction a few hundred metres away, drives there, sweeps up and down for about a minute, and while it works the road it brings the heat back down. Then it leaves. The enemy is not punishing you; they are re-securing a route and going home.

Everything on a road is built, including the concrete. At barricade an engineer is dispatched and the blocks go up over about twenty seconds of someone standing there; at turret another is sent and the tower takes forty. Arrive mid-build and you find a half-finished stump with nobody on it. Only once the tower stands does a gunner man it, up on top so he shoots over his own barricade. Let the road cool and the whole thing is abandoned.

Traffic and people. Civilian cars route across the network with somewhere to be; people wander on foot near the buildings. Both scatter when shooting starts.

The war you are driving through

Skirmishes break out between the two armies on contested and occupied ground, whether or not you are anywhere near. They are not a set piece for the player — you are not invited.

Rounds are real objects in flight, not instant hits resolved between two combatants. That is the whole point: if shots resolved instantly, driving through a battle would cost nothing, and "the only road to the target runs through a firefight" would not be a decision. A stray round does not check whose war it is, so it will hit a bystander, a civilian, or you.

Buildings stop bullets, which is what makes cover worth anything. Hits are routed through the same wear model as a crash, so being shot costs you ceiling too — it is permanent in the same way everything else is.

The enemy does not shoot at you by default. You are an undercover driver, not a target. They open up only past the line and on a road they have already checkpointed — that is, a road you personally made notorious.

Sound

Everything is synthesised — there are no audio files. The build has to stay self-contained, so the engine is detuned oscillators, the tyres and gunfire are shaped noise, and the ambience is a filtered drone. That constraint turned out to be an advantage: sound generated from the simulation can carry the simulation's state rather than just decorating it.

  • The engine has gears, so the note climbs and drops rather than rising forever with the speedometer.
  • A worn engine sounds wrong, not just slow: roughness rises and the top of the rev range falls away as condition drops. The car's decline is audible.
  • Bald tyres squeal sooner than good ones, off the same lateral slip the physics actually reports.
  • Gunfire is positional — it falls off with the square of distance and pans as you turn, so a firefight two streets away is a direction, not just a noise.
  • Territory has a tone. The drone drops and thickens the further past the line you are, matching what the palette does visually.

The split is deliberate: audio/mix.ts holds every decision as plain numbers and is unit-tested; audio/audio.ts is only Web Audio wiring, which cannot be tested outside a browser. Browsers refuse to make noise before the user has interacted, so the context is resumed on the first keypress or click.

Regional control

The map is cut by a single front line into four bands, running from your own territory out past the edge of the known world:

Liberated → Contested → Occupied → Frontier

Band Heat accrual Heat decay
Liberated ×0 ×4
Contested ×0.55 ×1.4
Occupied ×1 ×1
Frontier ×1.3 ×0.8

Liberated accrual is exactly zero, not merely small. A trickle would eventually fortify your own back garden, which is nonsense, and it would mean the safe area quietly punished you for using it. Behind your own lines, nobody is writing down which roads you take.

Decay is judged by whoever holds the road, not whoever holds the ground the car is standing on — your own side dismantles checkpoints on its own roads whether or not you are there to watch.

The line moves

Two forces, and they are independent by design.

Your work. Completing a mission pushes the front forward, weighted by what the job was worth — a long haul into unknown country moves the war more than a local errand, and supply runs move it more than surveys. Ground won is not held forever: the player's contribution decays over about seven minutes, so the enemy presses back on its own.

The war elsewhere. The line also drifts with nobody's help, in both directions. This is not background flavour: the brief is explicit that ambient drift is what makes the setting a war rather than a level, and that it is the only mechanism for the enemy's attention being pulled elsewhere — there is no separate "baseline threat" system alongside it.

Drift is a sum of slow waves rather than a random walk. A walk either wanders off the map or, once reined in, moves so little the war reads as static. Waves at incommensurate periods are unpredictable in practice, always bounded, and reproducible from the seed — so "the region changed while you were away" is something a test can assert. Most of the movement is shared by every boundary, because a front that advances advances as a whole; each band then wobbles a little on its own so the map is not one rigid object sliding about.

In practice a boundary swings about ±45 m and a spot near a border changes hands several times over a quarter of an hour.

Nothing announces any of this. There is no notification, no readout, no number. The minimap shows who held each patch of ground when you last stood in it — so after a while away, your map is simply wrong, and the only way to find out is to go back and look.

Your own ground is a pocket, not the map: roughly 15% liberated, 25% contested, 27% occupied, 33% frontier. The starting base sits at the friendly edge rather than in the middle, because starting central put half the world harmlessly behind your own line where nothing watches and nothing happens. Those splits are measured, not guessed — the bands are strips across a square with the axis running diagonally, so area is nowhere near linear in depth.

You are never told which band you are in by the world itself. Crossing a border eases the fog, the light and the palette: your own ground is open and green, contested goes flat and hazy, occupied turns dusty and closes in, the frontier is cold and short-sighted. The HUD names the band for now, but that line is debug scaffolding for the same reason the heat numbers are — Phase 3's gate is telling them apart without being told.

The map

A sketch map, top right, drawn from Intel rather than from the world:

  • The road network's shape is always visible, as faint dashed lines. You are an operative with a map, not an amnesiac.
  • Roads you have actually seen are drawn solid and coloured by the heat you remember on them — which is what you last saw, not what is there now.
  • Ground you have driven through is filled in, tinted by who held it when you were last there. A survey mission fills in a whole area at once.
  • The border is drawn between adjacent remembered cells that you recall as belonging to different sides — never from the front's true position, which would hand you a live readout of a line that is supposed to move behind your back.
  • Old notes fade. An observation drawn faint is one you should not trust.

Nothing on the map reads live world state. Heat is what you last saw on a road; territory is who held it when you last stood there. Both keep changing after you leave, and the map has no way of knowing.

The HUD also carries an eight-point compass arrow to the active target, since the map alone does not tell you which way to point the car.

Layout

The one rule worth keeping: src/sim/ imports neither three.js nor Rapier. Phases 14 (heat, regions, front lines, quests) are all simulation, and keeping them engine-free is what makes them unit-testable and fast-forwardable — you can run a hundred simulated days in milliseconds to tune an escalation curve without ever opening a browser.

src/
  sim/          pure model, no engine imports:
                  world, roads, heat      — the map and its memory
                  radio, opportunities    — texture: chatter and field work
                  units, combat           — traffic, patrols, soldiers, rounds
                  save                    — the campaign as JSON
                  regions                 — the front line and who holds what
                  intel, routing, quests  — what the player knows and is asked to do
                  bases, car              — where missions come from, and decline
  physics/      Rapier world, raycast vehicle, input → wheel forces
  render/       three.js scene, roads, markers, chase camera
  core/         fixed-timestep loop, seeded RNG, keyboard
  ui/           debug HUD, quest board, sketch minimap
  audio/        mix (pure: what it should sound like), audio (Web Audio wiring)
  persistence.ts  integration layer: saves in localStorage
  debug.ts        ?debug=1 handle: step, render and capture frames to shots/
  carSpec.ts    shared car dimensions, so body and mesh cannot drift apart
  heatProps.ts  integration layer: heat levels → colliders + meshes

heatProps.ts sits at the top level on purpose — it is the one module allowed to touch both Rapier and three.js, because it owns objects that must exist in both or neither. Its tests run headless: three.js scene graphs work fine in Node, so "the barricade's collider was removed along with its mesh" is a unit test.

Data flows one way: simphysicsrender. Condition reaches the physics layer already digested into a Handling by deriveHandling, so there is exactly one place where "how broken the car is" turns into "how it drives".

Notes on the state of it

  • The car handles like a placeholder. Suspension, grip, and engine numbers in physics/physics.ts and sim/car.ts are first guesses that pass a smoke test, not something tuned by feel. That tuning is Phase 0's verification gate.
  • Wear rates are deliberately aggressive so decline is visible in minutes rather than hours. Turn them down in applyWear once the curve reads right.
  • No interpolation between physics steps. Fine at 60 Hz; revisit if the step rate changes.
  • Nothing is persisted across reloads — condition, heat, intel and completed runs all reset. IndexedDB comes with the campaign layer, and sim/intel.ts plus sim/heat.ts are the two things that most need it.
  • Repairs are automatic on hand-in. Parts get spent the moment you report back. The brief wants repair to be a physical, on-foot act with scavenged parts; that is Phase 6.
  • The board's route preview assumes the shortest path. If you drive a different way, the intel you were shown described a route you did not take. findRoute already accepts a cost function, so a "safest route" preview is a small change when it is wanted.
  • Junctions have no heat of their own. A point near a junction is credited to whichever segment is nearest, so a heavily used crossroads never fortifies as a crossroads. Known gap, deferred on purpose.
  • Barricades are visually crude. Blocking the whole 24 m corridor means two long concrete slabs, which reads more like a wall than a checkpoint. The gameplay shape is right; the presentation wants berms, wire, or wreckage on the outer sections. They can also clip scenery placed on the verge — both are static bodies, so physics is unaffected, but it looks wrong up close.
  • Heat has no diegetic signal at a distance. You learn a road is hot by arriving at the checkpoint. The brief wants it readable from patrol density and wreckage before you commit — that needs the enemy presence Phase 5 brings.
  • The heat HUD lines are debug scaffolding. The brief is explicit that no numeric heat meter ships; they exist to tune the curve and should come out.
  • Bundle: ~2.7 MB raw / ~945 KB gzipped, dominated by Rapier's WASM, which rapier3d-compat inlines as base64. Switching to the non-compat @dimforge/rapier3d package serves the WASM as a separate file (~570 KB gzipped, compiled in parallel with the JS) at the cost of extra Vite plugin config. Worth doing before anyone but you plays it; not worth doing now.

Next

Two gates are open at once, and both need a person, not a test.

Phase 1: you catch yourself avoiding a road because of its history, not its distance. If it fails, check METRES_PER_HEAT (escalation too slow to matter in one session), DECAY_PER_SECOND (nothing accumulates), or whether the props are actually inconvenient enough to route around.

Phase 2: you take the new target a meaningful fraction of the time, for reasons other than curiosity — because the known one has genuinely got expensive. If the known option is always strictly better, raise NOVELTY_BONUS. If you never take a known target, lower it.

Phase 3: driving into an occupied region, you can tell it from a liberated one within a couple of minutes, from tone alone, without reading the HUD line that names it. If you cannot, the palettes in render/scene.ts are too close together — push them apart before trusting the band boundaries.

Phase 5: someone coming to this cold can describe what kind of game it is and who they are playing after one session, without the premise being explained first. If the four jobs still feel like one job with four names, the dial is MISSION_SHAPE — push the work times and the cargo risk further apart before adding anything new.

Phase 4: returning to a region after time away, you are genuinely unsure what you will find until you are back on the ground — and when it has changed, it reads as the world having moved rather than as a notification you were given. If the world instead feels arbitrary, the drift is too fast: lengthen the wave periods in sim/regions.ts before shrinking their amplitude, since it is the rate of change that reads as noise, not the size of it.

Phase 6: the world should feel lived in before it feels dangerous — traffic and people you have no reason to interact with, and a war that is obviously not about you. If firefights read as encounters staged for the player, the spacing and spawn rules in sim/units.ts are the dial.

Still deferred: a deeper repair economy with scavenged parts and on-foot work, car identity and disguise feeding into heat, and expanded combat. Nothing there should be pulled forward until the loop below it is proven.

The most useful thing before any of that is taking the debug lines out of the HUD. Several gates now depend on reading the world rather than a readout, and the scaffolding actively works against you when you sit down to judge them.

Watch for one failure mode in particular: if you find yourself taking whichever job is nearest and ignoring the intel line entirely, the board is decorative and the phase has not landed, whatever the reward numbers say.