The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands for orchestration and system monitoring. The templates/driver.template.sh script executes the claude CLI to run sub-sessions, and adapters/pi/guard.ts uses execSync with the df command to verify available disk space before allowing training operations. These are documented, functional uses of command execution.
[INDIRECT_PROMPT_INJECTION]: The design involves reading state files (state.json, progress.log, commands.log) and interpolating their content into prompts for subsequent agent sessions. While this creates a surface for indirect prompt injection if external data influences these files, the skill mitigates this by implementing a comprehensive guard.ts that validates all tool calls before they are executed or recorded. Additionally, the system uses clear boundary markers like ===== STATE ===== and explicit termination tokens (STAGE_DONE) to maintain control over the agent's context.
[SAFE]: The skill implements significant security controls. The adapters/pi/guard.ts script includes an environment variable allowlist to prevent the leakage of provider keys or shell startup hooks to subprocesses. It also contains a robust SECRET_PATH regex that blocks access to sensitive files such as .ssh, .aws, .env, and /proc/self/environ. The driver script utilizes a turn budget (PI_KIT_TURN_BUDGET) and a loop breaker to prevent excessive token consumption from repetitive or failing commands.

tao-token-efficient-execution