The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code and logs during its analysis and rewriting phases. This is a documented vulnerability surface common to all migration tools.
Ingestion points: git clone of user-provided repositories and log file ingestion (SKILL.md, Phase C2).
Boundary markers: Subagents are controlled via structured context blocks and file path references rather than inlining untrusted content.
Capability inventory: The skill uses git, aws, and uv commands and performs network requests to provider APIs.
Sanitization: Relies on the agents' ability to distinguish application code structure from instructions.
[DYNAMIC_EXECUTION]: In the baseline evaluation phase, the skill generates specialized Python scripts from predefined templates in its reference files (e.g., resolve_source_model.py and source_baseline.py) to interact with provider APIs. This execution occurs locally within a pinned virtual environment using the uv toolchain.
[COMMAND_EXECUTION]: The skill makes extensive use of local shell commands, including git for repository management, aws CLI for identity and model access verification, and uv for managing its internal Python dependencies.
[EXTERNAL_DOWNLOADS]: The skill clones user-specified repositories via Git and installs required Python dependencies from official registries (PyPI). All external provider API interactions target well-known and established domains.

llm-to-bedrock