Use when changing the in-world CCTV / "Truman Show" surveillance cameras in Tiny World Builder — render-to-texture security feeds, the black-and-white CRT/VHS monitor shader, camera placement (lobby side-cams, pumpkincam, treecams), subject tracking, or the lobby screen cutting to live feeds.
Installs into .claude/skills of the current project.
Are you the author of Tinyworld Cctv Truman?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/jasonkneen-tinyworld-cctv-truman)
---
name: tinyworld-cctv-truman
description: Use when changing the in-world CCTV / "Truman Show" surveillance cameras in Tiny World Builder — render-to-texture security feeds, the black-and-white CRT/VHS monitor shader, camera placement (lobby side-cams, pumpkincam, treecams), subject tracking, or the lobby screen cutting to live feeds.
---
# Tiny World CCTV / Truman Cameras
Low-res black-and-white security cameras that watch the lobby and feed both
physical monitors and the big lobby presentation screen. They idle-sweep, then
pan to look at whoever is MOVING nearby — like the hidden cameras in The Truman
Show. Do not mount CCTV or the big screen in non-lobby worlds.
## Files
- `engine/world/62-cctv-truman.js` — the core system. IIFE exposing
`window.__tinyworldCCTV`. **4-space body indent on purpose** so the
duplicate-declaration guard in `tools/check.js` (which only scans 2-space
top-level decls) ignores its locals.
- `engine/world/63-cctv-placement.js` — mounts cams + monitors on lobby room
`enter`, tears down on non-lobby `enter` or `leave`. Exposes
`window.__tinyworldCCTVPlacement`. Lobby detection is
`window.__TW_LOBBY_WORLD_SLUG || 'tidewater-bay'`.
- `engine/world/58-lobby-presentation.js` — the lobby-only big screen; its
`tick()` cuts between slides and the hottest live feed.
- `scripts/landing-feed.js` + `styles/landing.css` — the public landing-page
live-worlds panel. World rows are buttons: click once to expand an island
CCTV preview (2D canvas from `/api/worlds.preview.cells`), click again to
collapse. This is a lightweight marketing/front-door CCTV treatment, not a
Three.js render target.
- Tick wiring: `engine/world/25-animation-loop-schema.js` calls
`window.__tinyworldCCTV.tick(t,dt)` then `window.__tinyworldLobby.tick(t,dt)`
**before** `renderScene()` so feeds captured this frame appear this frame.
## How it works
- Each camera owns a `PerspectiveCamera` + a small `WebGLRenderTarget`
(`FEED_W×FEED_H`, 4:3). `tick()` round-robins captures (`CAPTURES_PER_FRAME`,
capped at `FEED_FPS`) using `renderer.setRenderTarget(rt); renderer.render(scene, cam)`
— **always save/restore the previous render target** (`getRenderTarget()`),
and bail if `renderer`/`scene` aren't ready.
- The monitor material is a `ShaderMaterial` (CRT_FRAG): luminance B&W, scanlines,
a rolling interference bar, hash static, vignette, and `uSignal` dropout that
dissolves toward static. A second canvas texture (`tCaption`) bakes the camera
name + live date/time + blinking REC and is composited in the shader. The
material is tagged `userData.windowLightEffect = true` + `lightVisual = true`
so `prepareFadeable()` does NOT swap it for a fade material (skill:
tinyworld-render-performance).
- Must end the fragment with `#include <colorspace_fragment>` (r185 output color
space), like every other ShaderMaterial here.
- **Subjects**: `setSubjectsProvider(() => [{pos:Vector3, name}])`. The room
feeds `WS.subjects()` (self + peer avatar sprite positions). Cameras + monitors
live under `WS.avatarParent()` — the SAME local frame as the avatars and lobby
screen — so subject positions need no conversion.
- **Truman tracking**: `aim()` scores each in-range subject by `prox*0.6 +
moved*1.0` (movement dominates) using per-feed last-position memory; it lerps
`curLook` toward the winner (`TRACK_LERP`) or idle-sweeps (`IDLE_LERP`). A
per-feed `activity` score spikes on motion and decays (`ACTIVITY_DECAY`).
- **Hot feed**: `activeFeed(minActivity)` / `feedsByActivity()` let the lobby
screen auto-cut to whichever camera has something happening. `glitch(id,amt)`
drops `uSignal` briefly to sell a cut.
## Placement (63)
On lobby `enter` only (after a 350ms delay so cells + lobby screen exist) it
mounts:
- `lobby-l` / `lobby-r` — flank both sides of the presentation screen, angled at
the crowd (toward +z).
- `pumpkincam` — over the biggest `kind:'pumpkin'` cell (scans `world[][]`,
sorts by floors).
- `treecam-1/2` — over the tallest `kind:'tree'` cells.
Monitors stack up the sides of the lobby screen. `window.__tinyworldCCTVFeeds`
lists mounted feed ids. On non-lobby `enter` or `leave`, clear feeds, disable
capture, and remove monitors.
## Lobby cutting (58)
`build()` stashes `screenMesh` + `slideMat`. The screen is shown only in the
configured lobby world, not every island. The state machine auto-advances slides
(`AUTO_ADVANCE`), then after `SLIDE_DWELL` swaps the screen material to a
`monitorMaterialFor(hotFeed)` for `FEED_DWELL`, then back. Manual presenter
`go()` and `hide()` snap back to slides. New API: `tick, setCycle, showSlides,
showFeed, liveFeed`.
## QA
- `?cctv=demo` (or `=1`) drops 4 monitors around origin watching a bobbing test
subject and self-drives the tick — verifies the CRT look + tracking without a
multiplayer room.
- Landing page QA for `scripts/landing-feed.js`: if local `/api/worlds` is not
available, mock `window.fetch('/api/worlds')`, append a fresh copy of the
script, click `.hero-feed-link`, and assert `.hero-feed.is-expanded`, one
`.hero-feed-cctv-canvas`, and the `.hero-feed-cctv-meta` status/link render.
The panel should anchor below the nav (`.hero-feed.is-expanded`) and scroll
internally instead of overflowing the hero.
- Headless sanity: eval module 62 under a THREE stub, `addCamera` + `tick` a few
frames with a moving subject, assert `activeFeed()` returns the feed and
`activity > 0`.
- `npm run check` (duplicate-decl + i18n), `npm run smoke`, `./publish.sh`.
## Pitfalls
- Don't capture every feed every frame — round-robin, or the extra
`renderer.render` passes tank FPS (the renderer contract is single-pass per
feed; this is the only sanctioned extra-RT path here besides pixelation).
- Don't forget to restore the render target after capturing, or the main scene
renders into the last feed's RT (black screen).
- Keep monitor materials off the fade pipeline via the `windowLightEffect` tag.
- Cameras + monitors MUST be added under `avatarParent()`, not `scene`, or they
won't inherit the tinyverse scale/offset and subject tracking will be wrong.