The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external AI CLI tools including claude, gemini, codex, opencode, and copilot. These executions are parameterized with user-supplied flags that are filtered for shell metacharacters.
[EXTERNAL_DOWNLOADS]: The skill uses npx to download and execute ACP adapters for Claude and Codex. These packages (@anthropic-ai/claude-code-acp and @zed-industries/codex-acp) originate from well-known and reputable organizations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external AI tool responses and local files via the --context argument, creating a surface for embedded instructions.
Ingestion points: Tool outputs (stdout), context files (via Read tool).
Boundary markers: None explicitly implemented for tool response interpolation.
Capability inventory: Command execution of CLI tools, file writing (temporary question files), and file reading.
Sanitization: Implements an extensive regex-based redaction system for API keys (Anthropic, OpenAI, Google, GitHub, AWS) and Authorization headers.
[DYNAMIC_EXECUTION]: The skill dynamically builds shell commands from templates and executes a local script (acp/run.js) to handle ACP transport for specific providers.

consult