Fireworks · kimi-k3

Write index.html with canvas fireworks, then verify with browser tools.

Clean load, no console errors. Demo shell fires at 400ms; screenshot to check. Let me take motion or screenshot after some time, then click.

Frames all show dark sky — no visible fireworks in motion frames. Possibly the demo shell already exploded before frame capture and faded, or rendering isn't happening (rAF may be throttled in headless? but motion frames should capture). Wait — the initial reload frame 1 also showed nothing. Motion spans 1.7s. Demo launches at 400ms, shell takes ~50 frames to ascend... Should be visible in frames 2-4. But all dark.

Possibly the canvas isn't being drawn — check: I do fillRect then draw shells. Should work. Maybe rAF isn't running in the sandbox/headless? The motion frames show identical images. Let me click and screenshot to test. Or maybe rAF only fires when visible. Let me try clicking and take motion.

Click worked — shell ascending near bottom (x≈640). The earlier motion probably didn't advance rAF between frames or timing. Now capture motion to see burst.

The shell exploded and particles faded — the motion frames captured after the burst finished (particles have decay ~0.008-0.02/frame, so lifetime ~50-125 frames ≈ 1-2s; frames all dark because burst happened and faded before frame 1? Actually motion captured frames appear all empty, likely the whole sequence passed during the wait, or frames show only trailing faint residue. It seems motion capture misses transient moments — frames are identical and dark, suggesting the rAF loop may only run when the browser tool takes a snapshot (headless throttling). Actually the click frame did show the ascending shell, so rendering works; the frames in between bursts just fade quickly.

Concern: the burst itself might be too brief/dim to capture. But it likely works. To verify burst visually, click and immediately screenshot. The shell takes ~550px/speed ~13 ≈ 42 frames ≈ 0.7s to ascend. The screenshot happens at some delay after click. Timing is tricky. Alternative: make verification easier — not necessary; the click frame showed the ascent correctly targeting click location (x≈640, clicked 600,250 — shell heading toward it).

Let me try click then screenshot quickly to catch the burst. Actually screenshot is a separate action; there may be delay. Try click at higher point so burst happens sooner, then screenshot.

Burst captured — radial burst centered around (700,150) with particles spreading and falling with trails, fading. Now test Clear: click again, then Clear, then screenshot.

Clear works — clicked Clear button while a shell was mid-ascent, scene is now fully dark and empty, no new burst occurred. Done. Report.