The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from a journal file which may contain content from various sources or previous agent interactions.\n
Ingestion points: The scripts/journal.py read command outputs content processed by the agent in SKILL.md during the 'Monthly review' phase.\n
Boundary markers: Entries use structured date headings (## YYYY-MM-DD kind: title), but the content within individual fields is not strictly sanitized for instructions.\n
Capability inventory: The skill has file-write capabilities restricted to append-only mode via the bundled script, with no network or general shell access.\n
Sanitization: No explicit instruction filtering is performed on ingested journal text before agent processing.\n- [COMMAND_EXECUTION]: The test utility scripts/test_journal.py executes the local journaling script to verify its behavior.\n
Evidence: The script uses subprocess.run([sys.executable, str(SCRIPT), *args], ...) to invoke journal.py.\n
Note: This is a standard testing pattern and follows best practices by using a list of arguments to avoid shell injection vulnerabilities.\n- [SAFE]: The core journaling functionality in scripts/journal.py is implemented using the Python standard library with an append-only file opening mode ('a'), which ensures that existing user data is never overwritten and minimizes the risk of accidental or malicious data modification.

credo-journal