The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF files to extract text segments for translation. In 'Handoff' mode, these segments are passed directly to the AI agent. A malicious PDF could contain instructions designed to override agent behavior during the translation process.
Ingestion points: scripts/translate_pdf.py (and the underlying pdf2zh core) reads and parses local PDF files supplied by the user.
Boundary markers: Absent. There are no explicit delimiters or system instructions to ignore potentially malicious commands embedded within the text segments being translated.
Capability inventory: The skill has access to subprocess execution (opening files, self-updating), local file system writes (saving translated PDFs, logging), and network access (Google Translate API, GitHub update checks).
Sanitization: Absent. Text segments extracted from the source PDF are interpolated directly into the translation workflow without filtering.
[COMMAND_EXECUTION]: Several components use subprocesses for legitimate utility purposes.
app/gui.py uses subprocess.Popen to open translated files or directories using system commands like open or xdg-open.
app/update.py generates and executes a PowerShell script (updater.ps1) to perform a self-update by swapping application directories.
[DYNAMIC_EXECUTION]: The script scripts/translate_pdf.py uses importlib.import_module to dynamically load the core translation logic and check for the presence of native dependencies at runtime.
[EXTERNAL_DOWNLOADS]: The skill downloads various assets and updates from external sources.
scripts/fetch_assets.py downloads layout models and fonts required for PDF rendering.
app/update.py and the Android UpdateChecker.kt fetch release information and binaries from the author's GitHub repository (breslee1707/VI-Translate).
These downloads target the vendor's own repository or well-known infrastructure.

pdf-translate