The Agent Skills Directory

[PROMPT_INJECTION]: The skill analyzes untrusted repository content, creating a vulnerability surface for indirect prompt injection where malicious instructions could attempt to subvert the agent's audit decisions.\n
Ingestion points: The audit process involves reading source code and documentation across the entire repository (SKILL.md, Phase 2).\n
Boundary markers: The skill does not provide delimiters or specific instructions to treat code comments or strings as inert data.\n
Capability inventory: The agent has permissions to delete files, execute shell commands, and configure automation hooks or CI pipelines (SKILL.md, Phase 5 and 6).\n
Sanitization: No sanitization or validation of the codebase content is described before it is processed by the agent.\n- [COMMAND_EXECUTION]: The workflow involves executing command-line utilities to scan the codebase and verify the project state.\n
Evidence: Instructions include running rg for searches, npx knip for unused dependency scanning, and tsc --noEmit for validation (SKILL.md, Phases 1, 3, and 5).\n- [REMOTE_CODE_EXECUTION]: The skill executes code from a remote registry through the npx utility.\n
Evidence: The baseline measurement phase calls for running npx knip (SKILL.md, Phase 1). As this targets a well-known development tool from a standard registry, it is noted as safe behavior.\n- [EXTERNAL_DOWNLOADS]: The skill methodology utilizes external packages and references.\n
Evidence: It specifies the use of the knip package and methodology from the DeepSeek Harness repository (SKILL.md).

subtraction-audit