The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The script scripts/antigravity_quota.py uses subprocess.run to call system utilities ps and lsof for local process and port discovery. Evidence: subprocess.run(["/bin/ps", "-ax", "-o", "pid=,command="]) and subprocess.run([lsof, "-nP", "-iTCP", ...]) in scripts/antigravity_quota.py.
[DYNAMIC_EXECUTION]: The developer utility scripts/verify_codexbar_parity.py executes a binary specified via command-line arguments for parity verification. Evidence: subprocess.run([args.codexbar_bin, "cost", ...]) in scripts/verify_codexbar_parity.py.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill scans local directories associated with AI agent harnesses (e.g., ~/.codex, ~/.claude, ~/.gemini) to read usage metadata. Evidence: SKILL.md defines default discovery paths in user home directories.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted JSON/JSONL trace logs which could contain malicious content, but it includes robust sanitization logic. Ingestion points: scripts/audit_harness_usage.py reads external logs. Boundary markers: The skill instructions emphasize never copying prompt or response content into output. Capability inventory: The skill can write files and execute local discovery commands. Sanitization: Employs safe_model for string normalization and html.escape in scripts/html_report.py.

audit-harness-usage