OpenCode mirror: sourced from ~/.config/opencode/skills/nvidia-video-forge and mirrored for OpenCode CLI usage.

NVIDIA Video Forge

Build videos as a reproducible pipeline with explicit run state, bridge artifacts, and resumable iterations.

Pipeline Standard

1. Ingest all provided sources into asset_graph.json (image/video/audio/url/youtube/blog/product/person/overlay).

• For YouTube, fetch metadata via Google API (service account / API key fallback) and extract source visual evidence (video + keyframes) when enabled.

1. Derive brand profile (brand_profile.json) and optional ASR storyboard (asr_plan.json).
2. Build prompt plan from user prompt + source context + brand constraints.
3. Generate multiple hosted NVIDIA candidates.
4. Judge candidates:

• Default: chat_bridge (same chat session; no codex exec).
• Optional: nvidia_qa, openai_api, dual.

1. Iterate until pass or budget reached (--judge-max-iterations, --judge-max-minutes).
2. Apply optional voice stage and overlay stage.
3. Export run artifacts (manifest.json, QA files, bridge files, final/rejected video, dashboard).

Quick Start

Generate in default chat-bridge mode:

python3 scripts/video_forge.py generate \
  --prompt "Branded product reel with smooth camera motion" \
  --input-product-image /absolute/path/product.png \
  --input-person-image /absolute/path/person.jpg \
  --input-url https://example.com \
  --overlay-template world_news \
  --hosted \
  --api-key-env NVIDIA_API_KEY

Resume after judge response in same run:

python3 scripts/video_forge.py resume \
  --run-dir /absolute/path/output/nvidia-video-forge/<run-id> \
  --judge-response /absolute/path/judge_response.json

Debug ingestion only:

python3 scripts/video_forge.py ingest \
  --input-url https://example.com \
  --input-blog-url https://example.com/blog/post \
  --input-youtube-url https://www.youtube.com/watch?v=... \
  --out-dir /absolute/path/ingest-preview

Execution Rules

• Hosted NVIDIA generation is mandatory (--hosted).
• Default judge is chat_bridge; this path writes bridge artifacts and exits with code 3 while waiting for resume.
• chat_bridge does not execute external judge tasks by itself.
• For YouTube visual judging, configure Google auth via --google-service-account-json (RECOMMENDED) or Antigravity OAuth. Direct GOOGLE_API_KEY is PROHIBITED per OpenSIN mandate.
• Keep contract fallback and endpoint profile fallback enabled unless strict debugging is needed.
• For NVCF exec routes, provide reference media or use text-capable routes.
• Preserve all run artifacts for reproducibility.

Run State Machine

• running: candidate generation and QA loop in progress.
• awaiting_judge: bridge package emitted; run must be resumed with judge JSON.
• completed: final output accepted and written.
• failed: budget exhausted or all candidates rejected.

Resource Map

• scripts/video_forge.py: main CLI orchestrator (generate, resume, ingest, qa, voice, probe).
• scripts/source_adapters.py: multisource normalization and crawling.
• scripts/brand_profile.py: brand extraction heuristics.
• scripts/asr_pipeline.py: audio timeline/storyboard planning.
• scripts/overlay_engine.py: FFmpeg overlay templates + file overlays.
• scripts/video_judge_chat_bridge.py: bridge request/schema/template generation + response normalization.
• scripts/video_qa_gate_openai_compat.py: strict QA gate.
• references/cli.md: CLI catalog and examples.
• references/model-matrix.md: judge/generation model routing notes.
• references/troubleshooting.md: failure modes and fixes.