六如 · 共梦图

0003: the whole picture at once, a living picture, cursors, chat and an AI companion

Status: accepted (crit 10, riff 8 pod brief). Keeps 0002's consent model and 0001's stack. Changes three of 0002's rules, listed below.

Context

The crit-9 room was quiet: you noticed other people only through what they placed, and the picture sat in a long scroll that each viewer panned and zoomed alone. The crit-10 pod brief asks for a room that feels inhabited. Everyone sees the whole picture at once and the picture breathes. You can see where everyone is pointing and talk to them. When you're alone, you can invite an AI companion to paint with you.

Rules this changes

  1. "No chat" becomes "one chat". The room has a single shared chat. Likes, rankings, streaks, scores, AI judging and autoplay audio stay out. The chat is the first free text in the app. It is plain text only, up to 300 characters, with control characters stripped. Each person may send five messages in any 15 seconds. It says on its face that messages are public and kept. There is no moderation queue.
  2. "Never simulate other players" becomes "never pass a program off as a person". The companion is allowed, but only when someone invites it. It is labelled as AI everywhere it appears: its 机 mark wears an "AI" tag and a hatched ring, and the same label appears on its cursor, in the inspector, in requests and on the archive's contributor marks.
  3. Pan, zoom, the overview strip and Fit are gone. Every viewer sees the entire scene.

Decisions

The scene size: 1800×1000 for new rounds, stored per round

The 2400×900 scroll (8:3) shows as a thin band on a 16:10 laptop. To keep 2400 units across a 1440×900 window it has to draw at 0.6×, about 540 px tall. On a portrait phone it is a sliver. We compared three shapes against the real available area at 1440×900, 1280×720, 844×390 (a phone held sideways) and 390×844. The available area is what's left after the header, tray and tool rail:

1800 is a little narrower than the old 2400, so the default motif sizes (a 560-unit water patch, a 230×330 pine) now take more of the picture. A 14–16 motif scene reads as a full landscape rather than scattered stickers, and the 48-object and twelve-per-person limits still fit without crowding. The existing scaffold (2400 wide) is centred, and the scene clips it.

Each round stores its own size (rounds.scene_w, scene_h). Rows from before this change have none, which means 2400×900. Every archive draws at its round's size, so the 16 dreams saved by the crit-9 app render byte-for-byte as they did (spec/layout.test.ts checks one of them). Positions are validated against the round's own size.

Portrait phones see the whole picture in the band between the header and the tools, at about 380 px wide. That's small but playable, and a dismissable hint suggests turning the phone sideways. We didn't make a separate portrait layout, which would break "everyone sees the same picture". We didn't force landscape either, because a phone that can't rotate would lock its owner out.

No black bars. The SVG's viewBox follows the window's own shape, so a textured paper bleed fills whatever the scene doesn't, out to every edge. The picture itself has a faint deckle line so you can see where the paper's edge is. A tap outside it doesn't place anything.

Slow panorama depended on the camera. 慢读 Slow reading replaces it. A soft pool of light moves right to left across the finished picture, the way a handscroll is read. It rests on each object and names who brought it. It is local to the viewer and steps rather than glides with motion off.

The living picture: one shared wind, from the server's clock

public/motion.js holds pure functions of the server-corrected clock, the round's lightning moment and each object's stored data (id, position, flip). They never use per-client randomness, so two screens side by side show the fish in the same place. We checked this in a real browser: both windows' transforms matched the formula for every animated object. One wind field drives everything that sways. It breathes, and a gust crosses the scene left to right every 23 seconds. The lightning (电) is the same wind, much stronger and faster. Boat bob, perching, lantern glow and the moon's reflection read the same field.

A crane that slides isn't walking, so six motifs are drawn from separate parts: fish (tail), crane (neck and both legs), deer (head and four legs), birds (each of three), lantern (light) and willow (fronds). The parts are made by scripts/split-motifs.py and stack back into the original pixel-for-pixel. A new water-glint overlay carries light along the water. Durable state, hit testing, relationships and archives all use the placed position. Only transforms and opacity change each frame, and nothing is rebuilt. prefers-reduced-motion or the Motion toggle gives a still picture. The archive stays still.

Cursors: a WebSocket, never stored

Cursors are presence, not state. They go over one WebSocket per tab (src/presence.ts, the ws package) and never touch SQLite or the event stream. They're sent in scene units, so they line up on any screen. The sender throttles to about 16 Hz and the server batches at 20 Hz, so locally a cursor arrives in under 200 ms (spec/presence.test.ts). The upgrade requires an existing visitor cookie and a same-site Origin. Messages are capped at 512 bytes and rate-limited, every field is validated and clamped, and pings drop dead sockets. Drag and ghost previews ride on the cursor message, which replaces POST /api/preview and its SSE preview event. Durable actions stay as POSTs through src/game.ts. There was no reason to move them, and that keeps idempotency, monotonic events and gap-resync exactly as they were.

Chat: one room, durable, across rounds

src/chat.ts follows src/game.ts's pattern. A message is one transaction with an idempotency id, appending a monotonic chat event that SSE carries to everyone. Each message records the round and phase it was sent in, so the history reads dream by dream. The page loads the newest 50 messages, and older ones page in on demand (capped at 100 per page). Exports in Markdown (each dream's prompt as a heading), plain text and JSON are streamed with Content-Disposition, for the whole history or one dream. A dream's conversation is shown beneath its saved picture; the archived scene is untouched. A message can point at one object ("about this"), which shows as a link that highlights the object if it's still there.

The companion: a server-side player that can only do what a person can

src/companion.ts is a Participant flagged agent = 1. Its identity is a server-only token that can never be a visitor cookie. It has its own dealt palette and a person's quotas, and it never holds one of the twelve people's marks. At most one is present per round (a partial unique index), it leaves when the round ends, and anyone can invite or dismiss it.

Every 5–30 seconds (COMPANION_MIN_SECONDS, COMPANION_MAX_SECONDS) it takes the same snapshot a browser gets and lists candidate moves. Positions come from relationship anchors, invitations and open space. Each candidate is filtered by running the real src/game.ts action inside a transaction that is always rolled back (game.wouldAccept), so "valid" means the server accepted it, not that a second copy of the rules agreed. The survivors are scored and one is sampled. Its cursor glides there, and then the move is made through the same function with an idempotency id. "Do nothing" competes with every move and usually wins. It answers requests aimed at it every couple of seconds, well inside the 20-second expiry. With nobody connected it holds no timers and makes no writes, and it never starts a round (startRound refuses an agent).

The model is a linear scorer over ten features of "this motif, here, in this scene": relationships formed, band height, overlap, gap, crowding, edge, balance, fit to the prompt and how full the picture is (src/companion-model.ts). Most weights are learned by scripts/train-companion.ts, a conditional logit fitted to the 97 motifs people actually placed in the live app's 16 public dream traces. Each real spot is set against 24 random spots for the same motif in the same scene; the real spot ranks first 21 times in 95, against a chance rate of about 4. The fit can't see two features, prompt fit and fullness, because they're the same for every spot in a group. Those two, an extra overlap penalty and the per-action priors (how often to wait, borrow, offer) are set by hand and marked as such in the weights file. That is 10 weights plus 13 priors and a temperature, under 1 KB. An OpenAI-compatible adapter is optional. It is off unless COMPANION_LLM_URL and COMPANION_LLM_MODEL are set, has a 3-second timeout, receives the top candidates, returns a validated index and falls back to the local model on any error. No service was signed up for and no key is stored.

What it costs

0002: one shared dream, temporary custody, consent to change hands

Status: accepted (crit 9, riff 8 pod brief). Supersedes the product half of 0001; the stack decision there still holds.

Context

The crit-9 brief asks for real time (one person's change reaches everyone else within about a second, no reload) and one decision about how the app behaves when several people use it at once. The pod's brief turns the old permanent wall into a seven-minute cooperative ink handscroll: everyone arranges prepared motifs on one composition, then it freezes, is saved as a dream trace, and dissolves.

The question that matters is not the transport. It is what happens when two people want the same thing on the same picture: the boat one person placed is exactly where another person's river wants it.

Decision

We share one changing composition, while direct manipulation follows temporary custody. Others can propose a new view or receive an offered object; consent and atomic handoffs prevent a shared canvas becoming a fight.

Alternatives weighed

  1. Everyone edits everything immediately. The most fluid, the fewest steps, and the closest to a shared whiteboard. But with 2–4 people and prepared motifs, the dominant move becomes rearranging other people's work, and last-write-wins means careful placement is undone silently. It also erases the thing worth noticing (someone gave you their boat). Conflict handling would be either a lock (a fight over who grabbed first) or a merge nobody can see happening.
  2. Each person's objects stay private and immutable to others. Simple, safe and conflict-free. But it makes the scroll four solo collages that happen to overlap: nobody can complete anybody else's scene, which is the core pleasure the brief names ("I brought a boat; someone else gave it a river"). The six similes about passing things on (泡, 影) would have nothing to act on.
  3. Shared composition, temporary custody, consent to change hands (chosen). Exchange is meaningful because it is visible and agreed, and no one can casually destroy another person's careful placement.

What it costs

Related behaviour this decision depends on