The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing untrusted data (e.g., PR titles, commit messages, or external filenames) that an agent may record in the decision log.
Ingestion points: Untrusted data enters the skill via arguments passed to the scripts/log.sh command (specifically the decision, why, and evidence fields).
Boundary markers: The skill documentation mandates the use of a specific helper script to ensure rows remain well-formed.
Capability inventory: The skill possesses the capability to append content to local files on the disk via standard shell redirection.
Sanitization: The skill implements strong sanitization in scripts/log.sh. The clean() function stripping tabs and newlines to prevent log structure corruption and prefixes sensitive characters (=, +, -, @) with a single quote to prevent Spreadsheet Formula Injection (CSV Injection).
[DATA_EXPOSURE]: The skill involves writing to local files such as decisions.tsv or .audit/<task-slug>.tsv. The analysis confirmed that these are used only for operational logging. No hardcoded credentials, sensitive system path access, or network exfiltration patterns were identified.
[COMMAND_EXECUTION]: The skill utilizes a bash script (scripts/log.sh) for file operations. The script uses safe shell practices, including set -euo pipefail and proper variable quoting, to prevent command injection during log rotation or directory creation. It does not execute remote code or download external scripts.

decision-log