The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the AI agent to execute shell commands to maintain a mandatory drift cache. Specific instructions include running sha256sum <file_path> | cut -d' ' -f1 to compute content hashes and grep to validate the cache file.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) because it extracts instructions (file paths and anchors) from untrusted data sources.
Ingestion points: The agent parses Architecture Decision Records (ADRs) and BDD feature files, specifically looking for tags like @bdd: and markdown links.
Boundary markers: The skill does not define delimiters or provide instructions to ignore embedded commands within the processed ADR/BDD content.
Capability inventory: The skill uses bash command execution (sha256sum, grep, cut) and file read/write operations for the .drift_cache.json file.
Sanitization: There is no mention of path sanitization or validation to ensure the extracted paths remain within the intended documentation directories.
[DATA_EXPOSURE]: The three-phase detection algorithm instructs the agent to resolve extracted references to absolute file paths and read their content. Without strict path validation, this could be exploited via path traversal (e.g., @bdd: ../../../.ssh/id_rsa) to cause the agent to reveal the existence or hash of sensitive system files.

doc-adr-reviewer