Agent notes

August 9, 2026 ยท View on GitHub

Omni-Rewriter is maintained as a typed, model-extensible prompt-expansion framework. H3 video and the initial image dialects are profiles, not the framework boundary. Before changing prompts, profiles, runtimes, or models:

  1. Read docs/architecture.md, the relevant profile guide, and .cursor/skills/omni-rewriter-h3-pe/SKILL.md. For a new model family, also follow .cursor/skills/omni-rewriter-model-contribution/SKILL.md. For promo films / RAW-vs-PE montages, follow .cursor/skills/omni-rewriter-promo-pipeline/SKILL.md and docs/promo/README.md.
  2. Preserve public contracts in src/omni_rewriter/models/.
  3. Add/adjust tests under tests/.
  4. Update docs if task routing or render formats change.
  5. Keep expand separate from generation; document adapter/runtime evidence and label untested compatibility as unverified.

Do not commit secrets, full videos, or private vendor dumps beyond the sanitized skill archives already under docs/references/. Use public evidence to bridge demos and deployable workflows; do not claim to reproduce undisclosed closed-source behavior.