[KaiwenLi]

Summary

Ships the scene-by-scene breakdown capability of the existing visual-analysis skill to every hosted plugin. Companion to ChatCut-Inc/chatcut#5434.

• codex and claude: visual-analysis/SKILL.md now points to references/scene-breakdown.md for whole-video and reference breakdowns.
• grok and kimi: adds the visual-analysis skill (entrypoint, scene reference and the concat-images helper). Each host's verification skill (grok/skills/verification, kimi/skills/chatcut-verification-guide) links to it.
• The scene reference is byte-identical to the monorepo copy. A breakdown reads each scene's spoken words from the existing ChatCut transcript (transcriptRangesMs) and never transcribes again. A missing transcript is reported as a gap, and audio is used only for tone, music, effects or sync.
• Version bumps so installed copies update: claude 1.10.24, codex 1.10.22, grok 1.10.18, kimi 1.10.18. Manifests and MCP headers are kept in sync.

Reference material (in the companion PR)

Higgsfield's backend analysis prompt is not visible. Only the tool descriptions and the app's routing skill are.

• Higgsfield tool instructions · routing skill
• Screenshots: 1 · 2 · 3 · 4
• JSON samples: Higgsfield excerpt · ChatCut benchmark excerpt
• Benchmark summary: docs/features/visual-analysis.md

Verification

• node --test tests/*.test.mjs (Node 24, CODEX_SQLITE_HOME unset): 144/144. The new tests/visual-analysis.test.mjs checks:
  • one entrypoint per host;
  • an identical scene reference on all four hosts;
  • the seven scene fields with audio_basis: chatcut_transcript;
  • the Grok and Kimi verification routes.
• Not verified: a live run through each hosted client after installing the plugin.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}