Weather log

The dated record of shader weather and the water effects described with it: what each live check found, what was changed after it, the tunings that came before the current ones, and the status recorded when the storm work landed. It is a log, so an entry describes the code as it was on its date. How weather works today is in Shader weather.

Entries are oldest first. The first five are the live checks of the falling rain, splashes and puddles; their dates are those of the commits that recorded them. The passages are kept as they were written on the weather page, so "above" and "below" in them point at that page as it was then. New entries go at the end, under a ## heading that ends with the date.

First live check, 2026-09-26

Godot 4.5.1, Mineclonia, the warm beach at Godot (515, 4, 447), noon, /weather rain. The weather node's state was right (intensity 0.5, the mesh at the camera, one rain spawner), and nothing could be seen. Only with the streak opacity raised tenfold, the intensity at 2 and the cover map switched off together did faint streaks appear, with bright curved arcs on the right of the frame. What was changed, headless:

• Opacity. Peak streak opacity had been 0.17 on the nearest layer, drawn unshaded in the sky gradient's colour, which is small beside sunlit sand once the exposure is applied. The rain is now lit by the scene's own lights, peaks at 0.68 near and 0.34 far, and is denser; the constants live in weather.gd, reach the material from there, and a test holds the peak to at least 0.5 near and 0.25 far.
• Intensity. The references assumed two Mineclonia spawners; there is one, so ordinary rain was drawn at half strength. Rain is now 1 and thunder 1.8.
• Arcs. The rain rings on water and ground were evaluated one cell at a time but could grow past their cell, so each was cut off by a straight line; with the sun in the water those cut rings are bright partial circles. Rings now stay inside their cells, and water rings stop at 22 nodes rather than 30. That was a real fault, but not the arcs: the second check found them in open air (below).
• The cover map. No fault was found in the lookup on reading, and a test now reads the uploaded texture back the way the shader does over an open beach and finds every point a layer draws at open, while an empty or part scanned map withdraws its area so the shader treats everything as open. The C++ scan against a real map is still untested; open_share in the status is the way to check it live.

Second live check, 2026-09-26

Same place and time. Status: rain 1.0, eye_open true, open_share 0.87, cover over the eye 2.5, so the map was not hiding anything. Streaks were visible at the defaults, over sky and sand, but three faults remained, all turning with yaw: the rain gathered into one narrow bright band (straight ahead at yaw 60), the rest of the sky was speckled with sub pixel white dots like stars, and bright curved arcs hung in open air (upper right at yaw 0, which is the same direction as the band).

All three were one fault. To keep the rain still in the world when the player walked sideways, the column coordinate carried the camera's world position along each layer's tangent. That term changes round the turn at a rate of the camera's distance from world zero, about 680 nodes at the beach, against layer radii of 3 to 22. In most directions it squeezed a layer's columns far below a pixel (the stars); in the two directions square to the camera's position it spread them back to their proper size (the band); and between, where it cancelled the arc length exactly, one column was smeared across the view (the arcs). The wind's shear had the same fault, smaller, measured from world zero instead of the eye. What changed:

• The column coordinate is arc length alone. No shift keeps the pattern still in every direction at once on a cylinder round a moving camera, so there is none, and the rain moves with you sideways.
• The shear is measured from the eye, the lean is held to 0.35, and the edge on fade now ends at 60 degrees, so the column coordinate advances at no less than 0.39 of its calm rate anywhere a layer is drawn.
• Minification, for the specks the far layers could still make at low resolution: columns under 5 pixels apart fade, and so do streaks under 8 pixels long (flakes likewise by their cell size).

Third live check, 2026-09-27

The owner, in play, at night on sand (the screenshot is sand lit by a torch, wet and glossy). Rain looked good at a distance, leaning and dense. Three faults:

• "Like I have an umbrella": no rain within a node or two of the player, plain looking down at about 30 degrees. The nearest layer was 3 nodes out, so a ray more than 28 degrees down met the ground first, and the edge on fade removed everything past 45 degrees anyway. The two low layers above are the fix; the test finds 178 of 359 quarter degree steps of depression clear under the old layout and none under the new, down to a clear circle of half a node.
• No splashes or puddles visible. The flecks were too small, too brief and too grey (above), and there were no puddles.
• Lightning. The strike went down the particle path as a flat unshaded quad, culled whenever the middle of the bolt left the view, and the server's white sky was eased away before it could be seen (above).

What was changed, headless (from the commit that made the changes): two low layers, at 0.45 and 1 node, which existed only below the eye (fading in from 40 and 16 degrees down, out by 88), fainter and with thinner streaks, and the wind's shear held to 1.73 radii of height so they could not fold. The splash fleck was made larger, longer lived, denser and given a share of the sky fill as its own light, and shallow noise puddles were added on flat, open, up facing ground. A strike spawner was recognised by texture and drawn by lightning.gd with an omni light, and main.gd held a white sky out of its easing and turned it into a flash. With shader weather off, the strike kept the particle path, and the particle path's culling box was made to include the particle's own size: it was a node round the emitter, 50 nodes up the bolt, so the bolt vanished whenever its middle left the view.

Fourth live check, 2026-09-27

The owner, in play, with Godot 4.5.1 against Mineclonia. Looking down: "it's still like a pillbox, the rain rushes in underneath". Looking up: the streaks became radial lines round a clear circle overhead. "The ground does get wet as the rain persists, but I've never seen a splash or a puddle", while rain rings did show on water. What was changed, headless:

• The cylinders are gone. They are edge on above and below the eye, which is both faults, and no layout of them fixes that. The rain is now a box of drops fixed in the world (above). The test casts rays at pitches from 80 degrees down to 80 up, at three yaws, and finds drawn drops within 12 degrees of every one in at least six of eight moments.
• Splashes and puddles: the terms were in the one array shader that does not draw the ground in play (above, "Splashes and puddles in play"). Found by reading materialFor and upstream's bunching, and confirmed against Mineclonia's textures; not by a live client, which the status trace is now for. Both shaders now draw them, puddles are darker and a little more widespread (a fifth of open flat ground after a minute of rain), and flecks are brighter.

The cylinders the box replaced. The box replaced six cylinders nested round the camera, which could not work at steep angles: a cylinder is seen edge on above and below the eye, so looking up its streaks became radial lines round a clear circle, and looking down the nearest wall was a ring of rain rushing in under a dry disc. A box has no angle it looks wrong from. What was lost: rain beyond about 12 nodes. The cylinders drew a layer 22 nodes out, which is what the owner found good at a distance; the far box fades out from 10.5 nodes to its sides at 12, and the storm fog and sky are left to carry the distance. The cylinders were four to six layers of overdraw across the whole view; with the box, most pixels carry no rain at all.

Splashes and puddles in play. Neither was ever seen, through two live checks, while the ground did go wet and rings did show on water. The terms were only in nodes_array.gdshader, and in play the ground is not drawn by it. GoannaClient::materialFor puts a node array on nodes_array_scissor.gdshader if any of its layers has alpha, and the arrays are upstream's bunches (NodeVisuals::fillNodeVisuals), which group tiles by size alone, 256 at a time. 1160 of Mineclonia's 1841 16 pixel textures have some alpha (counted over every 16 pixel PNG in its mods, near enough the set its nodes use), so the chance of a bunch of 256 with none is nil: every array is a scissor array, and sand, grass and stone are all drawn by the scissor shader, which had the wetness block and nothing after it. The terms now live in one function both shaders call, and the status trace below names the shader of the ground under the eye. The same cause kept the parallax march, which is only in nodes_array.gdshader, off the ground too. That was fixed on 2026-09-27 by choosing the shader per tile, by its own layer (docs/systems/materials.md, "Which shader draws a tile"): sand, grass tops and stone now go to nodes_array.gdshader on the near mesh, while cut-outs and the far tiers stay on the scissor shader. Both still carry the rain terms.

Fifth live check, 2026-09-27

The owner, in play at night against Mineclonia, in rain on open ground lit by a torch (the Godot version was not recorded). The splashes were seen at last: "bright white giant spots ... in game it just looks like hundreds of big spots appearing and disappearing without any kind of concentric ring effect, and their colour doesn't seem well matched to the surface". The brightened flecks of the fourth round were the whole of what showed: a fifth of a node of near white albedo with light of its own, over a ring too faint beside it to read. And for puddles, the owner's idea: "it may be possible to get the soft wet puddle look by softening the reflectivity and normals, like those dips in between ridges are what would get filled in by water." What was changed, headless:

• Splashes are small rings, 7 to 15 centimetres across, about 31 a square node a second at intensity 1, each spreading over a third of a second, in the normal only, over a brief wet mark that darkens and smooths the ground's own colour. Nothing white is mixed in and no light is added; the test holds rain's effect on the albedo to a factor of at most 1.
• Standing water fills the relief's hollows, by the height in the _n map's alpha, rising with the wetness; the noise puddles are kept as basins that raise the level until the crests drown too.
• The damp film is softer: roughness toward 0.32 rather than 0.13, and the relief softened rather than left whole under a gloss.

What these replaced, after the fifth check: a splash was a fleck a fifth of a node across for a tenth of a second, its albedo mixed 80 per cent toward white with a share of the sky fill added as light "so it can be seen at night", with a ring 0.7 nodes across in the normal that the fleck drowned out. In play it was a field of big white spots coming and going with no ring, their colour nothing to do with the ground. A puddle was a patch of the same noise with a hard edge, half the albedo and a mirror, drawn over the relief whatever its shape.

The damp film before this change was a roughness of 0.13 over everything open with the relief left whole, which read as wet plastic.

Other games' weather, run live, 2026-09-27

Regional Weather was run live on 2026-09-27: Minetest Game with Climate API, Regional Weather, lightning, lighting_monoid and player_monoids as world mods, and a headless client on the software renderer. Each preset forced with /set_weather: rain, heavy rain, snow and heavy snow were all taken as weather (about a dozen bursts at once) and drew 0.52, 0.93, 0.78 and 0.70, scaled by a storm field of 0.1 to 0.2 at the spot; hail was left to its particles. Its snow and heavy snow both reach the cap of 2 before the storm field, so the two look alike. Two traps for anyone repeating it: Climate API skips a dead player, and a client that spawned underground dies there; respawn and fly up before forcing a preset. Nothing was looked at on a GPU, and the lightning mod's strike (one spawner of one sprite, lightning_lightning_N.png, the sky set plain white for a moment) has not been seen.

Snowdrift's particles never reached the client before 2026-09-27, when the batched particle packet was first read (protocol coverage).

The other three were run live the same day, each alone on Minetest Game with a headless client on the software renderer. theFox's weather: rain 1.18 and storm 1.13 under a weak storm field, snow 0.90, hail and sand as particles; standing flown up at y 45 nothing came, since the mod only rains where it finds sky light in the air above, which was not generated yet. Mymonths: rain 1.2, storm 1.14, snow 0.88 where the ground was snowy, hail and sandstorm as particles; its weather follows the nodes round the player, so a spot by sand or snow changes it. Snowdrift, a test copy with its precipitation forced on and its freeze read from a setting: rain 0.27 and snow 0.21 under a weak storm field. theFox's rain uses sprites 25 across and runs about five bursts at once, so by the drop count it is already heavy; at a storm cell's heart its rain and its storm both reach the cap and look alike.

Seen on 2026-09-27 on the software renderer: Regional Weather's sandstorm browned the distance and the sky with blowing grains, its hail drew as short white streaks. With shader weather off, 24 rain spawners took the covered material without a shader error, but particle rain did not show on the software renderer with the covered material or the plain one, so whether the cover test hides the right particles has not been seen.

Seen on 2026-09-27 only on the software renderer, in a disposable Minetest Game world with a faked snowfall: driven streaks, the far hill greying, snow settling on sand. Not yet seen on a GPU.

Ripple and splash tunings, 2026-09-30 to 2026-10-02

The ripple patch's body footprint is a disc whose soft edge (kEdge) is centred 0.42 nodes out, just outside a player's 0.3. The first versions centred both the shove and the push on a Gaussian under the body: the crest it shoved up stood a fifth of a node ahead, inside the body, and the ring it pushed out was a node across before it was steep enough to see, so the water hard by a body lay flat whatever it did, a dead zone round the player seen looking down.

Three tunings came before the one current on 2026-10-02. The first, on a swimmer at 3 nodes a second, drew nothing a player could see: wading at 1 node a second left a slope of 0.09 against the rings' 0.3. The second pressed down on the water under every body all the time and swayed that press at a fixed 1.1 Hz, which drew waves out of a player standing still, at a rate that had nothing to do with what the player did, and the press, built up over half a second, left the middle of the pattern behind a moving player. The third lifted the surface with the body's own vertical speed, in the air as well as in the water, making water to do it, and a player jumping up and down piled up broad mounds with no troughs, drawn as white smoke rings by the crest whitening that went with them. Jumping in and out for five seconds peaked at 0.27 and -0.23 with no water made under that tuning; a gentle bob, a fifth of a node up and down, makes faint rings (0.03), as it does in real water, where narrowing the push to sharpen them only made them fainter. The test prints and floors the steepest slope 0.4 to 1.5 and 1.5 to 6 nodes from a swimmer. Now, for a player: nothing holding still; moving, 0.35 to 1.2 near and 0.22 to 0.30 further out; bobbing 0.11 near and 0.024 out; 0.09 out two seconds after dropping in. Jumping in and out for five seconds peaks at 0.08 round the body. It also checks a swimmer's crest is ahead of it and its trough behind, centred on it.

The refraction offset through the ripple patch was at first the whole normal times 0.006 times the ray's length through the water, which moved a shallow bed by a pixel. Its replacement, the tilt times the depth over the distance, was still only a subtle distortion looked down on, so caustics were added in sunlight. Seen from high up even that was not enough, and the rings round a body did not show at all, so shading to the sun and froth were added on the top of the water.

The froth and shading were seen headless on software rendering: a hard kick beside a standing player froths as clumps of rimmed bubbles of mixed sizes; the shading did not show in those captures, where the patch steps slower than real time and the light was dim, and neither has been looked at on a GPU.

The drips a body throws coming out of the water at first came out of a box through the whole body, which in first person put them in front of the lens. The droplets were at first round billboards 0.07 across, which read as white bubbles coming out of the chest and hands.

Status when the storm work landed, 2026-10-10

Status: seen five times, not right yet, and rebuilt since. The first two live checks were Godot 4.5.1, Mineclonia, the warm beach at Godot (515, 4, 447), noon, /weather rain. The first showed no rain at the defaults; the second showed streaks at the defaults, but gathered in one narrow band, with the sky speckled and bright arcs, all turning with yaw. The third, by the owner in play at night on sand, found the rain good at a distance, leaning and dense, with a clear circle round the player looking down, no splashes or puddles to be seen, and lightning drawn poorly. After changes for that, the owner in play again found the rain still a "pillbox" looking down, with rain rushing in underneath, radial lines round a clear circle looking up, and still no splash or puddle on the ground, though rain rings did show on water. The live check entries above say what was changed after each. The falling rain has since been rebuilt as a box of drops, and splashes and puddles now reach the shader that draws the ground. The owner then saw the splashes, as big bright white spots with no ring, their colour not the ground's (the fifth check); they have been rebuilt as small rings in the ground's own shading, and the puddles as water filling the relief. All of that has been tested headless only, and nothing of it has been observed.

The storm work's own status, as recorded at its landing. Its section names are those of the weather page at the time.

Storms, other games' weather and snow: status

Landed on main on 2026-10-10, after a rebase onto weeks of other weather, water and sky work. What it adds, each described in its own section:

• Storm cells (below): a drifting field of storm cells scales rain, snow, hail and dust from a drizzle at a cell's edge to a downpour at its heart, and scales the storm deck and the wetness with it.
• Snowstorms: snow rises further than rain at a cell's heart and turns into a whiteout that closes the fog in and greys it and the sky.
• Settled snow: snow builds goanna_snow_cover instead of wetting the ground, and the node and leaves shaders dust up facing surfaces open to the sky, texel by texel.
• Wet grass droplets (grass_volume.gdshader): a few tilted, mirror smooth droplets on wet blades where the sky reaches, gated, like the rest of the wet look, by the tier's wet surfaces flag.
• Hail and blowing sand ("Hail, blowing sand and the rest"): drawn by the precipitation shader as kinds of their own, with dust closing and browning the fog.
• Covered particles: weather still drawn as particles is kept out from under roofs by the rain cover map.
• Other games' weather ("Other games"): Mineclonia and VoxeLibre (mcl_weather), pmb_core (aom_weather), backroomtest (br_weather), Kythen, Regional Weather (Climate API), theFox's weather, Mymonths and Snowdrift.
• storm_override on the weather node: 0 to 1 pins where in a storm cell the viewer stands, for a test or a screenshot; below 0, the default, the field decides.

How it was verified, and how not:

• On 2026-09-27, before the rebase, each supported mod was run against a local server (Minetest Game, or Mineclonia for mcl_weather) with a headless client on the software renderer, and the recognised weather and its intensity were read from the client ("Other games").
• On 2026-09-28, still before the rebase, a GPU look pass in headless gamescope found the defects listed below.
• At the landing, on 2026-10-10, only headless checks with Godot 4.5.1's dummy renderer: project/tests/weather.gd and the native tests (cmake --build build --target check), both passing, and the other headless tests that load the changed scripts: ripples, wake, water_optics, cloud_layers, procedural_grass, local_play, graphics_profiles and render_features pass; local_play_scene (3 menu failures) and portal_materials (3 failures) fail the same way on main before the landing. These draw no frame. The rebased code, which now sets its shader globals per view for local multiplayer, loads the covered particle shaders through the view's scope and gates the droplets on the tier flag, has not been run against a server or looked at on a GPU since the rebase. None of it is claimed to look right.

Known defects, open, from the 2026-09-28 GPU look pass (realistic rates, not the fake spawner):

• Blowing dust is invisible. The grains (DUST_RADIUS 0.012, alpha 0.5) are near invisible, and Mymonths' sandstorm at 0.67 shows nothing. The fog does close in and brown.
• The sky stays sunny under a storm. storm_cover reaches 0.93, but the sky keeps blue gaps and full sun lies on the ground.
• Hail is sparse: thin, sparse streaks rather than a hailstorm.
• The fake snow spawner is far heavier than real snow. weather snow fake=true (particles.gd, inject_test_spawner) is 400 a second over 16 by 16 nodes, about 39 times Mineclonia's snow by density, so it sits at the cap of 4 and a full whiteout at any storm strength. Judge snow with a spawner of 100 a second over 50 by 50 (density 1) instead. Its rain is over the cap too. Not changed here, since local_bench.gd's rain benchmark uses the same spawner.

Rain and snow at realistic rates did scale with the storm as intended in that pass.