The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/evidence.py utility executes local commands to capture their output (stdout/stderr) and exit status for the audit trail. It uses process replacement (os.execvp) and descriptor-based directory binding to ensure the command runs in the intended environment.
[DYNAMIC_EXECUTION]: scripts/evidence.py uses a Python launcher script string executed via a subprocess to perform low-level process setup (fchdir, descriptor management) before executing the target payload. This is a structured approach to secure execution rather than arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data including command outputs, file contents, and conversation transcripts to generate audit logs.
Ingestion points: Command output streams, imported files, and workspace session transcripts (located via the pstack-harness skill) are used as evidence sources.
Boundary markers: Audit rows are stored in a structured TSV format.
Capability inventory: The skill possesses capabilities for file system reads and local command execution via its helper scripts.
Sanitization: The scripts/log.sh utility implements sanitization for spreadsheet software by escaping characters that could be interpreted as formulas (=, +, -, @) and stripping control characters. The scripts/evidence.py script validates path inputs to prevent directory traversal and symlink-based attacks.

show-me-your-work