Roadmap

Current Omni-Rewriter ships as a general prompt-expansion framework: transport-neutral requests, typed profile outputs, deterministic validation, dialect rendering, and bounded LLM repairs. H3, Seedance, and LTX-2.5 video PE, plus Seedream and Qwen-Image image PE, are its first profiles.

The project helps the community bridge polished demos and the explicit prompt contracts required by public/open generators. It relies on public evidence rather than attempting to reproduce undisclosed closed-source behavior. Expansion and generation remain separate. Community contributions are welcome — see CONTRIBUTING.md.

Near-term (maintainers)

• Ingest public H3 skill contracts (references/jahnson-h3-skill-*.txt) into PE rules, harness docs, and Cursor skills.
• Ship low-res RAW vs PE gallery media on the GitHub homepage.
• Add CONTRIBUTING / PR / issue templates / CODE_OF_CONDUCT / SECURITY.
• Publish Video / Image / Unified contribution boards, a model-contribution AI skill, and deterministic PR contract checks.
• Define a stable profile/renderer extension interface for model families beyond H3 and the initial image dialects.
• Add opt-in OpenAI-compatible Qwen image generation, Hunyuan custom-vLLM, Omni video/WAN, and independent LingBot adapters without coupling them to expand.
• Add mock contract tests for image/video submit, polling, base64/URL decoding, download limits, WAN mapping, Hunyuan extension fields, and LingBot subprocess/rewriter behavior.
• Publish H3 RAW vs PE demo surfaces under docs/ (landing, showcase, homepage gallery).
• Ship LTX-2.5 video PE (public paragraph dialect + optional official CLI runner; live generate unverified).
• Run live SGLang/vLLM-Omni WAN compatibility matrices across pinned runtime releases.
• Strengthen image validators for quote-language consistency (Chinese “” vs English "").

Community / research TODOs

• Supervised fine-tune a smaller writer on validated Omni-Rewriter traces (video H3 + Seedream/Qwen image PE), reducing repair rate and latency.
• RL / preference optimization (DPO / GRPO / RLOO) against downstream scores: H3 structural conformance, lip-sync / cut adherence judges, image aesthetic + instruction following judges.
• Add a bounded VLM-guided PE loop: generate candidates, judge full temporal/audio evidence, select or revise prompts, and record every iteration separately from deterministic repair.
• Implement the explicit Video, Image, and Unified profile targets in community-models.md, one evidence-scoped PR at a time.
• Multilingual eval suites and public leaderboard hooks.
• Streaming expand API and batch JSONL expand CLI.
• Video reconstruct / v2pe (v1): observe a local short clip (ffmpeg evidence pack → VideoObservation) then reuse draft/validate/repair to emit H3 t2va PE. Expand ≠ generate; do not inline the source mp4 into expand. CLI omni-rewriter reconstruct, HTTP observation JSON at POST /v1/reconstruct. Design: design/video-reconstruct-pe.md. Optional generate and fl2va/i2va remain later.
• Safer media sandboxing for untrusted reference images in shared deployments.

Contributions that keep the public contracts (RewriteRequest, ImageRewrite, BaseRewrite, Ref2VARewrite) stable are preferred.