The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates multiple shell commands and CLI tools to perform its auditing functions. It invokes bun, git, python3, and bash for data gathering, processing, and rendering. On macOS platforms, the skill utilizes sandbox-exec with a restrictive profile to confine the execution environment of probe scripts, effectively denying network access and filesystem writes while stripping inherited environment variables. On other platforms, it executes child processes with a stripped environment and a wall-clock kill-timer as documented in the security policy.
[REMOTE_CODE_EXECUTION]: The skill implements a dynamic probe system where modular TypeScript files are loaded and executed at runtime. These probes are used to classify recurring verification shapes without model calls. In scripts/probe-loader.ts, the skill employs dynamic import() to evaluate these modules from the skill's own probes/ directory or from a local configuration path (~/.config/keel/probes/). This behavior is a core functional requirement of the tool's auditing capabilities and is supported by a documented security model that encourages user review of probes.
[EXTERNAL_DOWNLOADS]: Several automation workflows within the skill's artifacts reference external resources from well-known and trusted sources. These include cloning the author's own bstack tool from GitHub, fetching binaries from the Model Context Protocol (MCP) registry, and retrieving GPG keys from Hashicorp's official release domain. All detected network operations target established technology providers or the author's own infrastructure in accordance with the skill's intended purpose.

keel