The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends a global installation of the dmux orchestration tool via npm install -g dmux. This dependency is hosted at github.com/standardagents/dmux, which is a third-party repository from a source that is not verified as trusted.
[COMMAND_EXECUTION]: The skill documentation includes examples of shell command execution for managing tmux panes and git worktrees. It also highlights a helper script scripts/orchestrate-worktrees.js that executes commands defined in a plan.json file through the launcherCommand field, enabling arbitrary code execution on the user's system.
[INDIRECT_PROMPT_INJECTION]: The orchestration helper reads and processes a plan.json file to coordinate agent tasks. If this plan is derived from untrusted repository content (such as a project being audited), it could serve as a vector for malicious task injection or command execution.
Ingestion points: The plan.json file is used as input for the orchestrate-worktrees.js script.
Boundary markers: There are no boundary markers or explicit isolation instructions provided to prevent the orchestrator from executing commands or tasks embedded within untrusted JSON data.
Capability inventory: The skill has the capability to execute shell commands via the launcherCommand property, manage git worktrees, and perform file system writes (writing task, handoff, and status files).
Sanitization: No sanitization or validation mechanisms are described for the launcherCommand or task fields within the JSON plan.

dmux-workflows