The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The check-pm-facts.py script performs network requests to official registries and documentation endpoints (GitHub, npmjs.org, getcomposer.org, php.net, and endoflife.date) to verify the accuracy of its internal version support tables. These downloads are restricted to well-known, trusted technology services and are used solely for informational maintenance, not for executing remote code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves analyzing untrusted repository files, which presents a diagnostic surface for potential indirect instruction injection.
Ingestion points: The pm-audit.py tool reads third-party files including package.json, composer.json, Dockerfile, and CI workflow YAMLs from the user-provided project path.
Boundary markers: Monitored data is processed and output in structured formats (TSV/JSON), though it is not wrapped in specific LLM-facing boundary markers during its internal diagnostic pass.
Capability inventory: The skill uses high-privilege tools (Bash, Read, Write) to inspect local files and can execute git or docker binaries via subprocess.run calls within its scripts.
Sanitization: A robust redact() function is implemented in the Audit class within pm-audit.py to identify and mask tokens, passwords, and sensitive query parameters discovered during analysis before they reach the agent's output.
[COMMAND_EXECUTION]: Diagnostic scripts use local subprocess execution to query git tracking status and docker build contexts. These operations are performed using structured argument lists (list-based subprocess calls) rather than raw shell strings, preventing command injection vulnerabilities. Several automated alerts related to remote code execution patterns were investigated and confirmed to be false positives associated with unit test mocks in tests/facts.sh.

package-manager-ops