The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/audit.py invokes the git binary to retrieve repository information.
Evidence: subprocess.run(["git", "-C", str(repo), *args], ...) is used in the git utility function.
Context: The calls are made using an argument list rather than a shell string, which prevents command injection via the --repo argument.
[INDIRECT_PROMPT_INJECTION]: The skill reads data from project manifest and memory files, which are then included in the audit report provided to the agent.
Ingestion points: File reads in scripts/audit.py for files such as manifest.json, memory/topics/*.json, and policy.json.
Boundary markers: Absent. The audit report is generated as a structured JSON object but does not contain explicit instructions for the agent to ignore content within the data fields.
Capability inventory: File system writes via atomic_write (to rebuild catalog.jsonl) and command execution via git subprocess calls.
Sanitization: Absent. Content retrieved from JSON files is parsed and re-serialized into the final audit output without validation or escaping of string values.
[DATA_EXPOSURE]: The skill accesses and reports on project identities and session metadata stored in the user's home directory.
Evidence: The audit_project function in scripts/audit.py reads from the path defined by $HARNESS_HOME (defaulting to ~/.harness/projects).

harness-audit