translation-novel
Pass
Audited by Gen Agent Trust Hub on Aug 12, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on internal library scripts (e.g.,
scene_chunker.py,novel_prompts.py,character_extractor.py) to perform its narrative processing tasks. These scripts are part of the local distribution and handle logic such as scene boundary detection and character state management. - [INDIRECT_PROMPT_INJECTION]: The skill processes external novelist prose, which presents an indirect injection surface. It addresses this through a mandatory evidence chain:
- Ingestion points: Chapter text is ingested in Layer 2 (Preparation) as a Markdown payload.
- Boundary markers: The DRAFT and REFLECT prompts use
<TRANSLATE_THIS>tags to explicitly separate content to be translated from surrounding context. - Capability inventory: The skill utilizes file reading for manifest processing and platform-integrated web search tools for L3 glossary resolution.
- Sanitization: The
protect-pass-spec.mdprotocol defines a system that masks formatting, code, and URLs into opaque sentinels (⟦P:NN⟧) before processing, preventing the model from obeying instructions hidden within the source text. - [EXTERNAL_DOWNLOADS]: The glossary resolution logic (L3) utilizes web search capabilities to verify terminology. This is documented as an optional feature and is implemented using standard agent tool abstractions without bypassing platform controls.
Audit Metadata