The Agent Skills Directory

[COMMAND_EXECUTION]: The script extensively uses Bash commands and Git to manage project worktrees and branches. Key operations include git worktree add, git branch, and git commit to synchronize task states with the project's repository. These commands are executed locally within the project scope.
[DYNAMIC_EXECUTION]: The environment command parses and executes shell commands from external markdown files referred to as 'recipes'. Specifically, it extracts code blocks under headings like ## Tokens, ## Bring up, ## Address, and ## Tear down and executes them using a subshell. This functionality is the primary purpose of the environment management feature and includes a safety mechanism that refuses execution in autonomous mode (exiting with code 70), thereby requiring manual user approval for these actions.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided content such as task goals and summaries, which are stored in task.md and task.json. While the skill extracts and compares sections of these markdown files (e.g., during repair and split actions), the text is not directly executed as code. However, the reliance on content within task.md creates a minor ingestion surface for untrusted data that could influence agent behavior if the agent treats the goal text as instructions.
[PRIVILEGE_ESCALATION]: The skill utilizes the EnterWorktree tool to change the agent's working directory scope to paths outside the default .claude/worktrees/ directory. The instructions explicitly state that this tool will prompt the user for approval, maintaining transparency and preventing silent scope escalation.

task