Video Translation

Own one durable project and route it by state:

source video → Agent-reconciled Omni/VAD transcript → Agent-authored dubbing plan → dubbed render → Agent listening review

Inspect first:

python3 <skill-root>/scripts/inspect_video_translation_state.py <project-dir>

Use the first applicable route: review or validate an existing delivery; resume a plan/render; translate an
accepted transcript; or analyze a new source. Read pipeline-contract.md before
accepting artifacts and dubbing-service.md when configuring the external
GPU service. Never infer success from filenames alone.

Internal workflows

• Source speech, subtitles, speakers, and timing: source-analysis/WORKFLOW.md
• Translation and duration-aware adaptation: translation-authoring/WORKFLOW.md
• Voice references, TTS, mix, remux, and QA: dubbing-rendering/WORKFLOW.md

Load only the current workflow. Reuse the bundled omni_call for audio-video understanding. Use the
video-translation MCP tools for the external dubbing service and deterministic local media execution.

Authorization and configuration

Analysis and translation-only targets stop before TTS. A full or resumed dubbing request authorizes calls
to the configured external dubbing service. The service URL comes from QWEN_MM_DUBBING_SERVER_URL. Never
write it to artifacts.

Do not use server-local paths as portable outputs. Tools upload local inputs and save returned audio into
the project. Preserve the source video stream during final remux. A delivery is complete only after the
final file fully decodes and the measured QA contract passes.

User-facing completion

After a successful render, surface the returned summary directly to the user. Report the segment count,
the number of automatic risk flags, and each segment that needs manual listening review with its time range,
translated text, and reason. Link full/translation_summary.md for a readable report and
full/translation_diagnostics.json for complete machine-readable details. Even when no automatic flags are
found, instruct the user to perform one end-to-end check for ordering, overlaps, pronunciation, and mix balance.