Video localization

Use the attached video's existing download, transcript, and timestamps. Never ask the user to install Python or deploy another application. Preserve original media and subtitles.

1. Establish the requested deliverables and target language when provided. Load the installed translate skill using the native read tool and its path from the available skill catalog. Follow it in this session, passing the requested target language, any explicit subtitle layout limits, whether dubbing is requested and the requirement for a complete saved subtitle track. It owns language defaults, existing-track reuse, terminology, translation, polishing, subtitle length/width constraints, split/alignment rechecks, timing and batch saves. Its required references/subtitle-constraints.txt supplies the adapted VideoLingo policy. If that skill or policy is unavailable, report the missing dependency rather than improvising a second translation workflow.
2. Continue only after translate has verified the complete track and returned its actual sourceKey and any subtitleLayout exceptions. A saved track alone does not prove layout compliance. Report unresolved default-budget exceptions alongside usable outputs; an unresolved strict user limit remains an incomplete requirement. If requested, load export_subtitles with that key: srt/vtt exports files, mkv muxes selectable subtitles, and mp4 burns subtitles. bilingual=true includes the currently viewed original text without changing the user's source selection; target line breaks do not reflow that original track. Preserve its timing rather than copying upstream gap filling or short-cue extension.
3. For dubbing, follow translate section 6 to prepare contextual numeric readings and any necessary spoken repairs. Load manage_tts_model and search the official sherpa-onnx catalog with action=status, query, language or family; follow nextOffset for more results. The available models extend beyond the original presets. Use refresh=true for newly published models. Inspect the saved default, language hints, download size, voice defaults/ranges and installed state. Honor the requested model; otherwise use the saved default when compatible. If another model is required for the target language, choose a compatible model for this call and explain the choice. Pass its exact modelId to create_dubbed_video for this task; use action=select only when the user requests a persistent default change. Models and required vocoders/phonemizers download automatically on first use. Use action=prepare to inspect actual native voice ranges and bundled reference WAVs before choosing a specific voice. Omit voice for the model's language default; IDs are model-specific. Pocket and ZipVoice require reference audio: reference.file is a bundled WAV name returned by preparation, relative to the model directory. ZipVoice also requires reference.text to exactly transcribe that WAV; use its returned transcript or verified publisher documentation, never invent one. Pocket can use a bundled default reference without text. Load create_dubbed_video with the verified sourceKey and its sparse speechOverrides; these bind one-based saved cue indices and exact originalText to spoken text without rewriting display subtitles. Choose background=separate to preserve accompaniment while removing original vocals, or remove when the user wants dubbed voice only. Do not describe source separation as voice cloning. Keep the default maximum speed unless the user requests another limit. If a cue cannot fit, pass the actual fitting error back to translate, shorten only that cue's spoken wording without losing meaning, and retry the same sourceKey while retaining all valid overrides. Limit the whole request to 3 dubbing attempts, then report remaining failures. Verify the tool's actual modelId, voice and applied override count. Never silently cut off words or claim success after an error.
4. Deliver the requested outputs and their saved paths. Distinguish saved subtitles from a completed rendered video. A tool failure or partial translation is not a finished deliverable. Avoid a lengthy narration of intermediate steps.