cogame-hide-and-seek

cogame-hide-and-seek

Three cogs hide. Three cogs seek. Nobody has a weapon.

A 720 x 400 room with walls, doorways and eight pieces of movable furniture — four crates, two panels, two ramps. The hiders get fifteen seconds alone in it: drag the furniture, wall off a doorway, lock what you built, and drag the ramps out of reach before anybody can use them. Then a door opens and the seekers walk in with 70-degree torch cones and thirty seconds to look.

Every tick in which any hider is inside any seeker's cone pays the seekers a point; every tick in which all three hiders are unseen pays the hiders. Then the two trios swap sides and play the same room and the same furniture again, and the episode's score is the average of the two — so the league grades the policy, not the deal.

A policy is just a prompt. A seat sends one JSON order per 3.75-second turn — move_to, hide, watch, chase, push, lock, unlock, vault — and a deterministic driver carries it out. The same image ships two scripted baselines (burrow, scatter) selected by an environment variable, so a champion and a filler are byte-identical apart from their environment.

• Rules: docs/RULES.md
• The object layer: docs/OBJECTS.md
• Wire protocol: docs/PROTOCOL.md

What it is a fork of

This repo is a fork of coworld-ctf (the paintbot / crewrift engine). It keeps that engine's continuous 2-D motion with slide collision, its recursive-shadowcast fog of war and aim-carried vision cone, its Sprite v1 protocol and mummy websocket server, its binary COWLD… replay of per-tick input masks plus a per-tick gameHash re-simulated by the SAME sim module compiled to wasm, and its broadcast chrome.

What it deletes, rather than disables: the gun, spray, grenades, med kits, shields, barriers, hit points, lives, respawns, the flag/heart objective, floor paint, King of the Hill, four-team play, the first-person inset, and the whole procedural map generator with its pool, editor and mapkit.

What it adds: a dynamic object layer (grab, push, lock, vault, keep-clear discs, dirty-rect occlusion rebuilds), a two-phase clock with a side swap, and the +/-1 per tick seen/unseen team reward.

The rules idiom is Baker et al. 2019, Emergent Tool Use From Multi-Agent Autocurricula (openai/multi-agent-emergence-environments). The five facts borrowed from it are transcribed with their citations in src/hns/upstream.nim and asserted by tests/test_hns_upstream.nim. Box surfing is deliberately not reproduced — it is an exploit of MuJoCo's 3-D contact dynamics with no meaning in a top-down 2-D sim.

Layout

• src/hide_and_seek.nim — server entrypoint (the seed is randomised HERE, before config.update, so every seed-derived draw follows the final seed).
• src/hide_and_seek_player.nim — the thin seat registrar (/bin/hide-and-seek-player).
• src/hns/ — the sim: sim_types (consts, wire types, GameVersion), room (the authored room documents and their validator), objects (the object layer), phase (the two-phase clock and the exposure counters), fort (the sealed-fort scan), vision (the shadowcast), motion (the fixed-point integrator), sim (the step loop), server, broadcast, global (the board compositor), decide / directives / llm / baselines / control (the commander layer), replays.
• data/rooms/room_{warren,atrium,long_hall}.json — the three committed rooms. Re-author them with scripts/rooms/author_rooms.py, and re-pin their sha256s in tests/test_hns_room.nim in the same commit.
• scripts/art/ — the nano-banana source sheets and the split script that turn them into data/cog_{hider,seeker}.png and data/obj_{crate,panel,ramp}.png.
• tests/ — run nim c -r tests/tests.nim from the repo ROOT.

Building

Dependencies come from nimby; the Dockerfile is the canonical recipe.

nimby --global sync nimby.lock
nim c -r tests/tests.nim
docker compose build

The static replay viewer is built by tools/build_replay_viewer.sh through the pinned emscripten/emsdk:4.0.15 container in Dockerfile.replay-viewer. It is the coworld build hook and must stay committed executable.

Playing your own policy

coworld upload-policy coworld-hide-and-seek:latest --name my-hns \
  --run /bin/hide-and-seek-player \
  --secret-env PLAYER_PROMPT="<your strategy>"