The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and renders external metadata from Salesforce orgs. \n
Ingestion points: Metadata is retrieved in scripts/metadata_listing.py (via zip retrieve) and scripts/fetch_soql.py (via SOQL query results).\n
Boundary markers: Reporting logic in render_architecture.py utilizes fenced code blocks (e.g., text blocks for prompt content) to prevent misinterpreted output.\n
Capability inventory: The skill performs subprocess execution of the sf CLI and local file system writes to cache and data directories.\n
Sanitization: It implements _md_escape for markdown table formatting, uses shlex.quote for shell safety, and enforces strict alphanumeric regex validation for all identifiers via fs_guard.py.\n- [COMMAND_EXECUTION]: Necessary interaction with the sf CLI is performed via subprocess calls to fetch and display org metadata.\n
Evidence: scripts/sf_cli.py, scripts/resolve_bot.py, and scripts/metadata_listing.py.\n
Mitigations: The skill strictly uses list-based command arguments to avoid shell injection and validates all user-supplied and discovered parameters against a restrictive regex in scripts/_shared/fs_guard.py.\n- [SAFE]: The skill implements advanced security controls for handling sensitive authentication tokens during its operation.\n
Evidence: scripts/rest_client.py includes a custom StripAuthOnCrossHostRedirect handler to prevent token leakage on redirects and a redact_text utility to ensure tokens are scrubbed from all logging and error reporting output.\n
Subprocess execution in sf_cli.py and resolve_bot.py includes ownership and symlink checks to prevent file system attacks.

agentforce-architecture-analyze