The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to manage tmux sessions and git worktrees for parallel agent environments. It executes commands such as dmux, git worktree add, and node scripts/orchestrate-worktrees.js.
[REMOTE_CODE_EXECUTION]: The instructions direct users to download and install an external tool dmux from github.com/standardagents/dmux. This repository is outside the scope of known trusted providers.
[EXTERNAL_DOWNLOADS]: The skill references multiple external agent harnesses including Claude Code, Codex, OpenCode, Cline, and Qwen without specifying versioning or source verification for the orchestration environment.
[DYNAMIC_EXECUTION]: The skill executes local scripts (e.g., scripts/orchestrate-worktrees.js and scripts/lib/tmux-worktree-orchestrator.js) that are referenced but not included in the provided skill files, making their logic unverifiable.
[DYNAMIC_EXECUTION]: The orchestration helper uses a plan.json file with a launcherCommand field (e.g., codex exec --cwd {worktree_path} --task-file {task_file}). This pattern of string interpolation into shell commands presents a risk of command injection if the input variables are not properly sanitized.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the task-planning data pipeline.
Ingestion points: Untrusted task descriptions are read from plan.json (specifically the task field) and passed to sub-agents.
Boundary markers: No delimiters or instructions to ignore embedded commands are specified for the task content.
Capability inventory: The skill can execute arbitrary shell commands via the launcherCommand template and manage multiple terminal panes.
Sanitization: There is no evidence of input validation or sanitization for the task strings before they are processed by the orchestration scripts.

dmux-workflows