The Agent Skills Directory

[COMMAND_EXECUTION]: The nemoclaw_run function in assets/nemoclaw/bashrc-snippet.sh uses eval to execute shell commands within the sandbox environment to facilitate logging and environment persistence.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration templates from the official iblai GitHub repository using curl as part of the setup process.
[PERSISTENCE_MECHANISMS]: The installation instructions for the NemoClaw sandbox involve appending NODE_OPTIONS and helper functions to /sandbox/.bashrc to maintain the hardened environment across sessions.
[DYNAMIC_EXECUTION]: The .nemoclaw-preload.js script uses dynamic require() calls to load security guard modules from the /tmp directory at Node.js startup, based on paths provided by the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill manages agent workspace files including SOUL.md and IDENTITY.md, which creates an attack surface where instructions embedded in these data files could influence agent behavior.
Ingestion points: The AgentConfigPrompts component and mentors/{mentor_unique_id}/agent-config/ API endpoint handle the prompt data (SKILL.md).
Boundary markers: None explicitly defined in the provided instruction files for the prompt editing surface.
Capability inventory: Command execution via bash -ic and nemoclaw_run, and file writing via openclaw CLI tools.
Sanitization: Not detailed in the static instructions for the prompt field content.

iblai-vibe-agent-sandbox