The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill metadata recommends the installation of a third-party binary, codexbar, via a Homebrew tap (steipete/tap/codexbar) which is not recognized as a trusted organization.- [COMMAND_EXECUTION]: The script scripts/model_usage.py executes the codexbar CLI using subprocess.check_output. The command is constructed using a structured argument list with inputs restricted to specific allowed values (codex or claude), which prevents shell injection.- [DATA_EXPOSURE]: The skill documentation notes that the required codexbar tool reads local session logs from paths such as ~/.codex/sessions/ and ~/.config/claude/projects/. The script processes this output locally and does not perform any network operations.- [INDIRECT_PROMPT_INJECTION]: The script processes external JSON data from the CLI or user-provided files, representing a potential surface for indirect prompt injection. 1. Ingestion points: Data is ingested via sys.stdin or file paths in scripts/model_usage.py. 2. Boundary markers: The output lacks explicit markers to distinguish data from instructions. 3. Capability inventory: The script is capable of executing the local codexbar binary. 4. Sanitization: The script performs JSON schema validation and specific field extraction, which provides structural validation of the input.

model-usage