kre-video-translator（kre 视频翻译器）

Kre Video Translator

Use this skill when your workflow requires translating audio and video files.

KreTrans provides very high translation accuracy. Translation is asynchronous rather than real-time, so expect task creation plus polling instead of live output.

Setup

Before the first run, read {baseDir}/references/setup.md.

Preflight

1. Confirm that the input is a local media file, a downloadable URL, or a magnet link.
2. Confirm that KRETRANS_API_KEY is available.
3. Confirm that ffmpeg is available in PATH.
4. Confirm that one Python launcher is available.
5. Confirm that the Python requests package is installed.
6. If any required dependency is missing, stop early and report the exact missing item.

Workflow

1. Resolve the input type: local media path, URL, or magnet.
2. If the input is a URL, download it first.

- Prefer yt-dlp --no-playlist -o "<output_dir>/%(title).120s-%(id)s.%(ext)s" "<url>" when yt-dlp exists.

1. If the input is a magnet link, download it first.

- Prefer aria2c --dir "<output_dir>" --seed-time=0 "<magnet>" when aria2c exists.

1. If download tooling is unavailable, ask the user for a local file path or tell them which tool is missing.
2. Call {baseDir}/scripts/translate.py with the final local media path.
3. Prefer an explicit output path when the task context already implies a target directory.
4. Return the generated .srt path and summarize any API or task failure.

Execution Rules

• Use this skill only for local-file translation output. translate.py does not accept URL or magnet input directly.
• Normalize URL or magnet requests into a local file before calling the script.
• Treat translation as an async job flow. Do not describe it as real-time or imply immediate streaming results.
• Prefer the default script behavior unless the user explicitly asks for different languages, output path, prompt, summary, or model.
• Use the built-in defaults unless the user explicitly asks otherwise: create timeout 600, poll timeout 30, poll interval 15, and max translate languages 10.
• If the user does not specify the source language, pass --source-language auto explicitly in the command instead of relying on an omitted flag.
• Never infer the source language from the target language.
• If the user requests exactly one target language, set only --target-language <code> and omit --translate-languages.
• Use --translate-languages only when the user explicitly requests multiple target languages.
• Preserve the actual generated .srt file path and report that exact path back to the user.
• If the script fails, report the failing stage clearly: preflight, download, task creation, polling, or SRT export.
• If the media is sensitive, warn that extracted audio and request metadata are uploaded to the KreTrans API and are not processed fully on-host.

Default Parameter Strategy

Use these defaults unless the user explicitly asks otherwise:

• --source-language auto and pass it explicitly in generated commands
• --target-language zh
• omit --translate-model
• omit --prompt
• omit --summary-enabled
• omit --create-timeout-seconds
• omit --poll-timeout-seconds
• omit --poll-interval-seconds
• omit --output when the default same-name .srt next to the input file is acceptable

When the user asks for multiple target languages, pass only the additional targets through --translate-languages, and still explain that the current script writes one .srt using target_language.

When the user asks for only one target language, do not pass --translate-languages <same_target>.

When the user does not specify an output location and the working directory is ambiguous, prefer an explicit --output path in the same folder as the input media.

Language Argument Rules

Use these command construction rules consistently:

• User says "translate to German" and does not specify the source language:

- use --source-language auto --target-language de - do not use --source-language de - do not add --translate-languages de

• User specifies both source and target:

- use both explicit values, for example --source-language en --target-language de

• User asks for multiple target languages:

- keep the primary output language in --target-language - put only the additional target languages in --translate-languages

Preferred single-target example:

python3 "{baseDir}/scripts/translate.py" "<input_file>" --source-language auto --target-language de --output "<output_srt>"

Runtime Requirements

• Required:

- KRETRANS_API_KEY from https://kretrans.com/console#api-management - ffmpeg - Python package requests - python, python3, or py

• Optional:

- yt-dlp - aria2c

translate.py only accepts local file input. The script sends requests to https://api.kretrans.com/v1/api. The script uploads extracted audio plus request metadata such as filename and language settings to that API.

Python Launcher

Pick the first available command:

• Windows: python, then py -3
• macOS/Linux: python3, then python

Response Contract

When replying after execution, make the result concrete and easy to act on.

• On success, include:

- the final .srt path - the input file path that was actually translated - the target language - whether the source started as a local file, URL, or magnet - any important warning such as fallback behavior or auto-created output path

• On preflight failure, include:

- the exact missing requirement - one concrete next step - whether the user can continue by supplying a local file instead

• On API or task failure, include:

- the failing stage - task_id or request_id when available - the main error message - retryability when the script reports it

Do not answer with a vague success message such as "done" or "completed". Always include the output path.

Reply Patterns

Success pattern:

Generated subtitle file: <output_srt_path>
Translated source: <input_media_path>
Target language: <target_language>
Source type: <local|url|magnet>
Notes: <warnings_or_none>

Blocked pattern:

Unable to continue: <missing_dependency_or_missing_input>
Next step: <single concrete action>

Task failure pattern:

Translation failed at: <stage>
Source: <input_media_path>
Task info: <task_id/request_id/or_none>
Error: <primary_error_message>
Retryable: <true|false|unknown>

Error-To-Reply Mapping

Convert raw tool output into a direct user-facing status.

• If the error says the API key is missing:

- say that KRETRANS_API_KEY is not set - tell the user to get it from https://kretrans.com/console#api-management

• If the error says ffmpeg is missing:

- say that local media extraction cannot start - report that ffmpeg must be installed and available in PATH

• If the error says the input file does not exist:

- report the exact input path - ask for a valid local media path

• If URL or magnet download tooling is missing:

- name the missing tool: yt-dlp or aria2c - tell the user they can instead provide a local file

• If task creation fails:

- say that the request reached the API stage but the task was not accepted - include request_id when available

• If polling fails with task status failed:

- say that the remote translation task failed after creation - include task_id, error, error_code, and retryable when available

• If SRT export fails after a successful task:

- say that translation completed but subtitle export failed locally - include the intended output path

Prefer short, concrete wording over raw traceback dumps. Preserve task identifiers and the main error fields when they exist.

Examples

PowerShell:

$env:KRETRANS_API_KEY = "<KRETRANS_API_KEY>"
python "{baseDir}/scripts/translate.py" "D:/videos/demo.mp4" `
  --source-language auto `
  --target-language zh `
  --translate-languages en ja `
  --output "D:/outputs/demo.srt"

Bash:

export KRETRANS_API_KEY="<KRETRANS_API_KEY>"
python3 "{baseDir}/scripts/translate.py" "/tmp/demo.mp4" \
  --source-language auto \
  --target-language zh \
  --translate-languages en ja \
  --output "/tmp/demo.srt"