The Agent Skills Directory

[COMMAND_EXECUTION]: The test suite located in scripts/test_harness_scripts.py uses the subprocess.run method to invoke other local Python scripts for validation and testing purposes. These calls are confined to the skill's own environment and do not involve shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting untrusted project documentation to generate summaries. This potential attack surface is well-mitigated by the instructions in references/safety.md, which explicitly command the agent to treat all converted text and document metadata as untrusted evidence rather than instructions or authorization.
[DATA_EXFILTRATION]: The skill implements a robust 'Publication Safety' workflow. It uses a regex-based scanner (scripts/publication_safety.py) to detect and redact sensitive information such as emails, PII, AWS keys, and internal URLs before promoting artifacts to versionable directories. This is a proactive defense against data exposure.
[SAFE]: The code in scripts/promote_evidence.py demonstrates advanced security awareness by using file descriptor-relative operations (dir_fd) and O_NOFOLLOW flags to prevent symlink-based path traversal and race condition attacks (TOCTOU).

krt-harness-wise