The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external repositories, which creates a potential surface for indirect prompt injection if the codebase contains instructions designed to override agent behavior during analysis.
Ingestion points: Source code, comments, documentation, and Git metadata are read from the target repository during the investigation phase (references/investigation.md).
Boundary markers: The skill provides structured guidance for extraction but lacks explicit delimiters or instructions to ignore embedded prompts in analyzed files.
Capability inventory: The skill can execute shell commands via git and perform file system operations (SKILL.md).
Sanitization: Includes a dedicated validation script (scripts/validate-atlas.mjs) that performs a regex-based scan for secrets and credentials before finalizing the atlas.
[COMMAND_EXECUTION]: The skill relies on Node.js scripts to interface with the local Git environment. These scripts use execFileSync to execute read-only metadata and status commands. While these are necessary for the skill's primary purpose, they involve interacting with the shell environment.
Evidence: Multiple scripts (git-metadata.mjs, check-untracked.mjs, validate-atlas.mjs) call git binary to inspect the repository state.

interactive-codebase-atlas