diff --git a/docs/plans/next-session.md b/docs/plans/next-session.md index be20f7b..212fe92 100644 --- a/docs/plans/next-session.md +++ b/docs/plans/next-session.md @@ -12,33 +12,21 @@ planned or known, not built. ## Next features -3. **Wave B starts with a shared session clock.** Proposed on 2026-11-07, not decided yet: it changes how every - multiplayer game plays. Today each player's game runs in their own browser at their own speed (pause, 0.5× to 8×), so - players in one session are on different days. The proposal: the server keeps the time and every game in the session - follows it. The simulation stays in the browser. - - **Server:** a session gets a pace when it is created: one game day every 30 s to 5 min (agent mode's list). Never the - normal 2.5 s, which is 1,440 game days an hour. The day only moves while at least one player was seen recently (the - roster already keeps `seen`), so nobody comes back to years of unattended play. Every roster answer carries the - server's time and the session day. - - **Browser:** the clock is locked to the session. `UI.clockLocked` and `HOOKS.speed()` (`50-state.js`) already hide - the speed buttons and refuse pause, and `save()` already leaves `uiSpeed` alone while locked. The code goes in - `src/game/`, because `src/agent/` does not ship (ADR 0004), so it needs its own copy of agent mode's worker ticker - (`src/agent/85-ticker.js`) to keep time in hidden tabs. Agent mode also sets `HOOKS.speed`: the session wins, and - agent mode is off in a multiplayer game. - - **Catching up:** a game behind the session day runs days until it matches, a few per frame, with its progress - shown. It needs its own loop: with the route worker, `runClock` holds one day at a time. A day takes about 30 ms on - the big test network, so 400 missed days take about 12 s on a desktop and longer on phones. Late joiners start on - the session's current day (to check: what in the world depends on the day a game started). - - **Nobody can pause or speed up.** That is what "the same day" means. - - **Later, optional: the server plays absent players.** It opens an offline player's save in a headless game (jsdom, - as `scripts/lib/mcp-game.js` already does with its realtime clock) and runs their days with the intern. The - browser takes the save back on return through phase 1's save numbers, so the server is just one more device. Up to - about 0.5 GB of RAM per absent player. Without it, other players see an absent network stand still. - - **Not proposed: a server that runs every game** with the browser only showing it. Every click would be a round - trip, and the page reads `S` and `R` directly everywhere, so it would be a rewrite. Its one gain over the above is - verified cards (today the server trusts them). - - **Phase 2 gets simpler:** its wall-clock timing (tenders that close after 10 real minutes, `idle` ≥ 300 s, notice on - each player's own clock, the soak test's "own speed") can move to game days. +3. **The server runs every game of a multiplayer build** (decided 2026-11-07, being designed and built). In a page built + with `MP_API`, each network's days run on the session server, not in the browser, so every player of a session is on + the same day at the same speed. A single-player game in that build is a session of one and runs there too; old local + saves keep working in the browser, and the plain build stays exactly as it is (static, no server, no requests). + - **Speed:** 1× by default (one day per `DAY_MS`, 2.5 s). Players vote for another speed, pause included; alone in a + session your vote decides at once. The world pauses while no player is connected. + - **Absent players:** while anyone plays, the networks of those away keep running on the server (their intern's + standing orders help), so nobody falls behind. + - **Why:** the same day for everyone, which phase 2's trades need, and less work for phones at higher speeds (at 1× the + day is about 10–15% of a phone's CPU on the big test network, at 8× about half; map and panels stay in the browser). + - **What stays in the browser:** map drawing and panels. The flow tables (`R.ef`/`R.fl`, about 3 MB per routing on the + big network) cannot be shipped every day, so what reads them (a selected link on the map, the link panel, PNI value, + puzzles) has to ask for them. + - **Phase 2 gets simpler:** with every world in one server process, its wall-clock timing (tenders that close after 10 + real minutes, `idle` ≥ 300 s, notice on each player's own clock) can move to game days. 4. **Wave B: #97 co-op companies** (several players run one network with roles, budgets and approvals), host-authoritative over the multiplayer server, and **multiplayer phase 2** (players as carriers in each other's simulations: transit, peering, PNIs, shared fiber): design in `docs/plans/multiplayer-phase2.md`. @@ -75,7 +63,5 @@ planned or known, not built. ## Housekeeping -- `npm run perf -- --browser` needs Chromium (`sudo pacman -S chromium`, then `PERF_CHROME=/usr/bin/chromium`); only the - jsdom benchmark has been run. -- Upstream release tags are not on `vanvik` yet: `git push vanvik --tags` if wanted. +- `npm run perf -- --browser` runs with `PERF_CHROME=/usr/bin/chromium`. - The merged worktrees under `.claude/worktrees/` and their `worktree-wf_*` branches can be removed.