The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the configuration of shell command hooks through the add_hook.py script. These hooks allow the harness to execute arbitrary commands in response to specific lifecycle events such as PreToolUse or Stop.
[EXTERNAL_DOWNLOADS]: The documentation includes instructions for using curl to communicate with the Letta API at api.letta.com. These operations are used for managing server-side agent settings like name, description, and model configuration.
[PROMPT_INJECTION]: The skill's hook system introduces a surface for indirect prompt injection. Prompt-based hooks can process and interpolate untrusted data from tool inputs or session context into prompts for LLM evaluation.
Ingestion points: add_hook.py input arguments and runtime tool/session data processed by hooks.
Boundary markers: Example configurations in references/hooks.md lack explicit boundary markers or sanitization for interpolated variables like $ARGUMENTS.
Capability inventory: The skill allows defining both shell command execution and prompt-based logic that can influence agent behavior and decision-making.
Sanitization: Input provided to the configuration scripts is stored without sanitization or validation against malicious patterns.

configuring-your-harness