Replace the /agents hero background with a layered signal field

[ksorathia]

What

Rebuilds the /agents hero: a new canvas background, and a pass over the copy that sits on top of it.

The background (AnimatedAgentRuns, a new component) replaces the node-cloud dot grid with a layered signal field. Runners travel a lattice across three depth planes, fork (mass halves, the child sinks a plane), and merge (mass sums, the survivor surfaces). Only a runner heavy enough to have merged can claim one of four endpoint pads — arrival is earned.

Depth renders as one camera's worth of cues: perspective scale, resample defocus, atmospheric fog, and color cooling on the back planes, with bloom reserved for the front. All tuning lives in a documented SETTINGS constant at the top of the component.

AnimatedDotGrid stays where it is — KnockAiBanner still uses it. This PR adds a component and points /agents at it rather than rewriting the old one.

Details

The field

• Entry: the field draws itself in rather than switching on. Density, simulation speed, and ink start at floors and ease (cubic out) to full over 4.2s of live time. First paint presims 0.5s, not the full four — the point of the entrance is watching the traces extend, and a fully presimmed field has already done that offscreen.
• Pads: placement is random per load among curated slots (one pad per region, so the composition can't go lopsided). A pad completes after 2 earned claims, plays a done cascade, fades out, and a successor fades in elsewhere.
• Pointer: runners within 150px bias toward the cursor, the planes take a parallax push, and a click in the hero spawns a boosted burst on the front plane.
• Reduced motion: renders a single pre-simulated composed frame at full density and brightness, never animates.
• Carried over from the old component: theme palette via useTheme(), IntersectionObserver pause offscreen, debounced ResizeObserver, one-time fade-in. The 55% bottom mask moved onto the canvas itself, so the page no longer wraps it in a masking Box. Fog resolves --tgph-surface-1 at runtime so back planes fade toward the real page surface.

The copy on top

• One entrance timeline: the canvas announces its first drawn frame through onReady, which opens a gate the heading, veil, and CTA animate off. A 300ms timer is the ceiling, so a slow or failed canvas can't hold the copy hostage; <noscript> releases the animations outright.
• Copy CTA: both glyphs and both labels stay mounted and stacked, so the button cross-fades on copy instead of cutting and never reflows. The visible stacks are aria-hidden, so the button carries an aria-label and the confirmation lives in its own polite live region.
• Coding tools: the bare hover-dimmed wordmark row became an "Open in" launch bar — one ghost icon button per tool, each deeplinking with the setup prompt loaded.
• lib/debounce.ts: debounced functions now expose .cancel(), so the resize handler can drop a pending call on unmount.

Verification

• Dark/light × 1280px/390px at 2s and 25s: zero console errors, steady 120fps in the prototype harness this was ported from.
• Motion burst frames confirm depth cross-fades settle with no doubled traces.
• Reduced-motion screenshots are byte-identical 1.5s apart.
• yarn type-check and yarn format.check pass; CI green.

Known issue

AnimatedAgentRuns keys cell occupancy per plane, but setPlane doesn't clear a mover's old-plane stamps — so for up to mergeWindow, a stale id can pull mass across depths despite the same-plane-only collision rule. Real but low severity: a rare, sub-second depth hop in a background animation. Shipping as-is; the thread records why the obvious same-plane guard is the wrong fix.

Review notes

Best reviewed running (yarn dev, /agents) — the entry, pointer response, and pad completions are the point. Hard-refresh replays the entry; pad placement varies per load.

[cursor]

Risk HIGH: Replaces the /agents hero background by rewriting AnimatedDotGrid.tsx with a new layered signal-field canvas animation (~1,400 lines added).

Reasons

1. Modifies a .tsx file under components/ — a HIGH-risk trigger per classification rules.
2. Large component rewrite (1,407 additions / 265 deletions) with new simulation, rendering, and interaction logic.
3. Affects a shared UI component used on /agents; regressions could impact page performance, theme behavior, or accessibility.
4. Introduces runtime canvas work (runner simulation, pointer bias, pad completion cascades) that warrants manual visual and performance review.
5. Not a content-only or OpenAPI spec change — scope is front-end animation infrastructure.

Notes

• Run yarn dev and review /agents in dark/light mode at desktop and mobile widths; the PR author notes entry animation, pointer response, and pad completions are best verified live.
• Confirm reduced-motion behavior renders a static frame and does not animate.
• Watch for console errors, frame-rate drops offscreen (IntersectionObserver pause), and theme palette integration via useTheme().
• Pad placement is random per load — spot-check several hard refreshes for layout balance.
• Author ksorathia is not on the repository contributor list; no auto-approval applied.

  

_{Sent by Cursor Automation: Docs PR classifier}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 568c195. Configure here.}

[cursor]

Stale occupancy allows cross-plane merges

Low Severity

Merge resolution looks up a cell owner by id but never checks that the owner still shares the incoming runner’s plane. After a survivor surfaces via setPlane, its old-plane scent can still trigger merges for about mergeWindow, so mass can jump across depths despite the same-plane-only collision rule.

Additional Locations (2)

• components/ui/AnimatedAgentRuns.tsx#L562-L568
• components/ui/AnimatedAgentRuns.tsx#L158-L162

 

^{Reviewed by Cursor Bugbot for commit 568c195. Configure here.}

[ksorathia]

Correct — occOwner is keyed per plane, but setPlane doesn't clear the mover's old-plane stamps, so a stale id can survive a merge for mergeWindow and pull mass across depths. Accepting it for this PR: it needs a stale-owner-means-empty-cell path in advanceCell (the naive same-plane guard is worse — it falls through to the annihilation branch and would kill a live runner on a phantom stamp), and the visible result is a rare, sub-second depth hop on a background animation. Tracking it as follow-up rather than reopening simulation logic on an approved PR.