Feature request — play-mode visual capture (screenshots during a playtest)

Back to board
Hi, this is a feature request rather than a bug report — feel free to say "platform won't allow it," but we wanted to lay out the case with the actual incidents it would have prevented, because it's the single biggest quality gap left in our agent workflow. **What we're asking for:** any way to get real rendered pixels out of a *running* playtest — even asynchronously, even low-resolution, even if it lands as a PNG file next to the run_test report instead of coming back through MCP. **Why edit-mode screenshots don't cover it:** our game (a horror co-op) builds essentially everything at runtime. The map is generated per run, the entire UI (~36 LocalScripts) is created in PlayerGui — `StarterGui` is empty — and the look is defined by darkness, fog, the flashlight cone and post effects. `manage_camera screenshot` works great for props and static scenes, but by definition it can never show the thing the player actually sees. **Three documented incidents where every measurement was green and the PICTURE was the bug** — this is the fact base for the request: 1. *The invisible walk animation.* Our monster's server-side gait ran flawlessly — the server reported 8 of 16 bones deflected mid-stride. Players saw a gliding statue: `Bone.Transform` set on the server does not replicate, so the animation existed only server-side. Every number we could measure said "working"; only a human playtest caught it, and the diagnosis took a dedicated session with bone-angle probes on both sides. One frame of play-mode capture would have shown it instantly. 2. *The washed-out lobby UI.* A new bottom dock passed every check we have — element exists, visible, correct size, 0 contrast violations, 0 touch-target violations. The human tester reported it as "barely visible": the ScreenGui sat at DisplayOrder 0, underneath the HUD's full-screen vignette. The defect was purely visual stacking — present in any rendered frame, absent from every per-element metric. 3. *The particle stream that shot through the floor.* A faucet emitter cloned from a shower head kept speed 16–20 with lifetime ~0.4s — a ~7-stud jet for a 0.68-stud drop, firing through the basin, the cabinet and the floor. Geometry checks can't see it (particles aren't geometry), and in edit mode the emitter doesn't run. Found by a human, fixed by recomputing v0 from drop height. We've measured the cost of this class of defect: in our last UI review round, an agent captured and audited 10 runtime screens (367 text elements) via a self-built reconstruction pipeline — and still produced 2 false findings that a real frame would have settled immediately (both were renderer artifacts of the reconstruction, not real defects; each cost a correction round). The pattern is consistent: anything we can measure, we verify agent-side; anything that only shows up in the rendered frame goes through a human playtest round-trip, typically a day later. **What we've built to close the gap ourselves, and where each tool ends:** - *RayCam*: a ray-grid "camera" inside run_test, pixels printed to the test log, PNG built from the log file. Answers "what stands where" — but rays hit collision geometry only, so CanCollide=false detail is invisible, and there are no textures, no lighting, no GUIs. - *guicam*: a client probe dumps the PlayerGui tree (positions, sizes, colors, text, DisplayOrder, clipping) through a RemoteEvent; a renderer redraws it as a PNG. It found incident #2's stacking bug on its first real run — but it's a reconstruction: images, gradients and the 3D scene behind the UI are approximations (see the false findings above). - Neither can ever answer: "is this room readable in the dark?", "does the flashlight beam wash out the vignette?", "does this material read as metal under game lighting?" For a horror game those questions ARE the quality bar. **On the platform hurdle:** your screenshot tool's own description explains why play-mode capture was cut from v1 — CaptureService's temporary contentId can't be converted to an EditableImage from a non-edit-DM context. We don't know a way around that either. But the transport constraint might be softer than "return it via MCP": run_test already writes a report directory per run. If a capture could be saved as a file into that directory server-side — even fire-and-forget, even only at test end — our tooling would pick it up from disk exactly like it picks up test-log.txt today. No MCP image payload needed. And if CaptureService is a dead end entirely, a documented "not possible" would also help — we'd stop probing at it. Thanks for considering it — and no hard feelings if the platform simply says no. The 2.11/2.12 releases (preflight, validation, failures.jsonl) have all landed exactly on the pain points we reported, which is why we keep writing these up.