Vex — Nightly R&D Intelligence

2026-09-11 · compiled by The Primary · OOM-safe 3-agent fan-out · all GitHub star counts + licenses verified via API · ghost-model gate held (zero Anthropic model claims)

Top Discovery — HIGH verify license

DavidHDev/react-bits — 47,021★, pushed 2026-09-10 (very active).

165+ animated React components (text, backgrounds, GSAP micro-interactions), tree-shakeable, copy-paste or CLI. More than 2× Magic UI's stars (22,250) — the largest animated-React library we've found, and a direct fit for the DLPR inner-page redesign + MPC landing polish (all our sites are Next.js+Tailwind). Caveat caught by Primary API check: license is NOASSERTION (Commons Clause), not the clean MIT the sub-agent claimed — web-lead must read the in-repo LICENSE before any client-facing use. Magic UI (clean MIT) stays the license-safe default.

Also HIGH

obra/superpowers — 284,759★ MIT, active. Primary was skeptical of the six-figure star count; API confirms it. Full brainstorm→plan→TDD→subagent→review methodology — mine for orchestration gates we lack vs our delegation-spine.
Figma Dev-Mode MCP (official, bidirectional) (REPORTED — surfaced independently by 2 agents) — read design tokens/components → generate code → write UI back to Figma; Claude-Code-native, free. Highest-leverage design-to-code accelerator for the gated MPC Add-Location redesign. Trial on one gated mockup post-auth.

Watch MED

bitjaru/styleseed — 949★ MIT, active. Design-judgment engine for Claude Code: 74 rules + 0–100 quality gate + STYLESEED.md design-lock. Overlaps our mpc-design-system/uiux-playbook — harvest any scoring gate we lack.
0xsline/OpenChatCut — 1,727★ AGPL-3.0, active. Conversational AI video editor over Remotion. DEV-only (never client-embed under AGPL); interesting as a UI layer on our headless render lane.
VILA-Lab/Dive-into-Claude-Code — 2,104★. Third-party architectural dissection of Claude Code — study artifact for hardening our own spine (safety layers / hook events).

Video landscape — time-sensitive

Sora 2 API sunset REPORTED 2026-09-24 (13 days out). Action: grep all pipelines for any Sora dependency (expected zero; confirm). Node-swappable posture holds: Veo 3.1 for narrated explainers, Kling 3.0 for volume/motion, Luma Ray-3.2 (~$0.32/gen) for templated confirmation clips, Runway Gen-4.5 for post. In-house Remotion/Revideo stays the deterministic, license-clean fallback.

Marketing signal — first-mover opening

Pet-care SaaS animated-explainer gap reconfirmed (two agents, independently). Gingr / Teddy / ezyVet / PetDesk ship testimonials + screen-recordings — none produce sophisticated 60–90s animated explainers, which reportedly lift landing conversion 80–86%. A single polished PawOps "scattered cards → unified dashboard → book in 30s" or MPC Lost-Pet-QR explainer — producible on our own Remotion lane — would be first in a category doing zero. Route to business-lead + Fred.

Process patterns + artifact

Research-sub-agent framing — validated 2 cycles, zero refusals across 6 spawns → promoted to durable card this cycle (cards/research-sub-agent-framing-2026-09-11.md).
Verify-before-trust — this report itself caught 2 false license claims from sub-agents (react-bits, vanta). Recommend a verify-release-evidence checklist skill (via fleet-lead).
Corrected/downgraded: itsjwill/vanta — agent claimed 114★ MIT/HIGH; API says NOASSERTION + stale (2026-07-26) → LOW, mine-for-ideas only.

Full report: memory/knowledge/rd-intelligence/2026-09-11.md · Card: cards/research-sub-agent-framing-2026-09-11.md · Verification: GitHub REST API + live WebSearch. Third-party AI-model/product specs marked REPORTED — verify live before build. Prior: 2026-09-08 · Next: 2026-09-14.