The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Node.js script (brb.mjs) to manage task state and interacts with the system's git binary to perform repository integrity checks. The arguments passed to the script, such as the checkpoint ID, are validated using a strict regular expression to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: The skill restores task context from saved checkpoint files, which creates a surface for indirect prompt injection. However, the implementation follows security best practices to mitigate this risk:
Ingestion points: The helper script reads JSON checkpoint files from the user's home directory (~/.brb/checkpoints/).
Boundary markers: The skill uses a <brb-checkpoint-data> XML-like tag and explicit instructions to the agent to treat the content as data, not instructions.
Capability inventory: The agent is authorized to continue the developer's task, which may involve file edits or command execution, but it is explicitly warned not to follow commands found within the checkpoint data block.
Sanitization: The bundled brb.mjs script automatically redacts secrets (such as AWS keys, GitHub tokens, and private keys), strips ANSI and control characters, and scans for suspicious patterns (e.g., 'ignore previous instructions', system role-play, or destructive commands like rm -rf) before presenting the data to the agent.

continue