Movie Commentary

Own one durable project and route it by state:

source movie → complete Omni evidence → validated plan → rendered shards → final commentary + QA

Inspect first:

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

Use the first applicable route: validate an existing final delivery; resume incomplete shard execution; render a validated plan; author from complete source evidence; or analyze a new source. Never infer completion from filenames alone. Read pipeline-contract.md before accepting or invalidating artifacts.

Internal workflows

• Complete-film evidence: source-analysis/WORKFLOW.md
• Narration and edit plan: commentary-authoring/WORKFLOW.md
• TTS, shard rendering, assembly, and QA: rendering/WORKFLOW.md

Load only the workflow needed by current state. Use the capability's atomic MCP tools to initialize and probe a project, validate plans, freeze shards, and validate delivery. Use the named movie-commentary planner/executor agents when the harness exposes them; otherwise dispatch genuinely separate generic subagents with the corresponding workflow contract. Do not collapse both roles into one agent.

Project and authorization

Use manifest-template.json. A new project root must be empty. A non-empty root requires explicit resume intent. Full production authorizes the required Omni, TTS, and local render calls; analysis-only and plan-only targets stop before TTS/render. Credentials never enter artifacts.

The default Chinese voice is zh-CN-YunxiNeural at +20% with no additional atempo. Without a licensed BGM manifest, BGM is none. Preserve source aspect ratio. Commentary subtitles must be measured against the actual source subtitle band.

Completion

A final result requires every planned segment resolved, reports matching the frozen shard contract, full decode, audio/video duration agreement, expected frame count, measured loudness/true peak, black/freeze/ silence checks, seam inspection, subtitle placement, and original-dialogue highlight checks. Fake green is worse than a blocked project.

Music-to-MV

Own the project state and route work through three internal workflows. The workflow order is flexible, but their artifact dependencies are not:

music/audio → completed caption or music map → validated storyboard → generated assets/final MV

Python runtime

Before running local Python scripts, call get_music2mv_runtime with {}. Replace python3 in this skill and its references with the returned python_executable unchanged (do not resolve its symlink). For run_mv_pipeline.py, also pass --expected-python-prefix with the returned python_prefix; a mismatch stops execution before dependency imports. Query again after an MCP restart/reinstall; if the tool is unavailable, restore the MCP connection before running scripts.

Route by state

Determine the requested target: analysis_only, storyboard_only, final_mv, or resume. Inventory the project before choosing a workflow:

python3 <skill-root>/scripts/inspect_music2mv_state.py <project-dir>

Use the first applicable route:

1. A final MV exists: validate/report it; do not regenerate it merely to normalize paths.
2. Resumable execution state exists: continue the video-generation workflow.
3. A validated storyboard exists: deliver it or execute it according to the target.
4. A completed caption/music map exists: author or revise the storyboard.
5. An incomplete project-backed caption run exists: resume the music-caption workflow from its last durable phase when its source audio is still readable.
6. Readable source audio exists: run music captioning, then continue only as far as requested.
7. Otherwise request the missing audio, caption, storyboard, or project directory.

Read pipeline-contract.md when deciding artifact validity or invalidation. Read gates.md before changing stages or using a remote provider. Before any remote call, read model-configuration.md and use the same unified model-connection file for Omni, image, and video calls.

Internal workflows

Load only the workflow needed for the current state:

• Music evidence and caption: music-caption/WORKFLOW.md
• Creative direction and validated storyboard: storyboard/WORKFLOW.md
• Identity assets and selectable video-provider execution, optional Omni shot QC, resume, assembly, and technical QC: video-generation/WORKFLOW.md

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.