The Agent Skills Directory

[PROMPT_INJECTION]: The skill follows a structured orchestration logic and does not contain instructions designed to bypass safety filters or override system constraints. It does present a surface for indirect prompt injection through its input mechanisms. \n
Ingestion points: The [feature or area] command-line argument and the contents of the production/review-mode.txt file. \n
Boundary markers: None; the skill interpolates these inputs directly into the context provided to subagents. \n
Capability inventory: The orchestrator can perform file system modifications, task delegation, and shell execution via subagents. \n
Sanitization: No explicit validation or filtering of external data was observed. \n
This ingestion surface is evaluated as safe because the skill is intended for interactive development where all subagent proposals must be reviewed and selected by the user via AskUserQuestion.\n- [DATA_EXFILTRATION]: No network activity, sensitive file access (such as credentials or environment variables), or data exfiltration patterns were identified.\n- [REMOTE_CODE_EXECUTION]: The skill does not download or execute code from remote sources. All tasks are delegated to internal subagent types. \n- [COMMAND_EXECUTION]: The skill employs powerful tools such as Bash and Task for development-specific purposes like performance profiling and hardening. These operations are conducted within the scope of the target feature and are gated by user approval at each phase transition.

team-polish