The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions in SKILL.md and SKILL.ko.md recommend a bootstrap method that pipes a remote shell script from GitHub directly into the shell (e.g., curl -fsSL ... | sh). This is a high-risk pattern as it executes unverified code from a remote source.
[EXTERNAL_DOWNLOADS]: The scripts/install-binary.sh script downloads compiled binary archives from the jazz1x/rallish repository on GitHub.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its data processing logic.
Ingestion points: Turn history and user/agent notes are read from ~/.rallish/sessions/<id>/log.jsonl.
Boundary markers: There are no markers or instructions to ignore embedded commands within the ingested log data.
Capability inventory: The skill uses Bash to execute commands and Read for file system access across all scripts.
Sanitization: The skill does not perform sanitization or validation on the "notes" received from other rally participants before incorporating them into its response logic.
[COMMAND_EXECUTION]: The skill executes various system commands, including the rallish CLI for session management, make build for local compilation, and rm -f for deleting socket files during crash recovery.

rallish-operator