The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands and custom scripts (setup-run.sh, finish-run.sh) to manage the development environment. This includes git operations such as git worktree, process monitoring using pgrep and ps, and macOS system utilities like lsregister and osascript.
[EXTERNAL_DOWNLOADS]: The setup-run.sh script executes pnpm install within the project's web directory, which downloads and installs Node.js packages from the official npm registry as part of the project setup.
[DATA_EXFILTRATION]: The skill accesses the user's local configuration file at ~/.config/vignette/settings.json to seed scratch settings for agent worktrees. While the script processes this data locally and it belongs to the same vendor ecosystem (Vignette), it represents access to sensitive user configuration.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from docs/TODOS.md to define tasks for sub-agents, which could be manipulated to inject instructions.
Ingestion points: The orchestrator reads task descriptions from docs/TODOS.md in the base repository.
Boundary markers: The templates interpolate raw text from the TODO file into sub-agent briefs without explicit delimiters or warnings to ignore embedded instructions.
Capability inventory: Sub-agents have permissions to edit code, execute build scripts, and drive application input via input.sh.
Sanitization: There is no evidence of filtering or validation for the content of the TODO items before they are processed by the Opus sub-agents.
[DYNAMIC_EXECUTION]: The skill documentation describes the ability to evaluate dynamic JavaScript within the application's internal editor via a custom URL scheme (vignette://eval?<javascript>), used for state inspection and driving the app during development tasks.

vignette-todo-run