The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses pnpm or npm to install the @cursor/sdk and other development dependencies. These downloads are performed from the official NPM registry to support the runtime execution of the DAG runner.- [COMMAND_EXECUTION]: The skill uses shell scripts to locate runtime components and execute the run_dag.ts script using tsx. The runner invokes subagents that interact with the local workspace to perform decomposed tasks.- [PROMPT_INJECTION]: The skill is designed to process user-provided task descriptions and intermediate subagent outputs. It implements a boundary strategy by using markdown delimiters and specific headers to separate upstream context from current task instructions, mitigating the risk of instructions being interpreted as data.
Ingestion points: User-provided goals and the resultText from preceding tasks in the DAG.
Boundary markers: Employs --- delimiters and structured headers in the context construction logic.
Capability inventory: The orchestrated subagents can modify the filesystem and execute shell commands within the project context.
Sanitization: Data is truncated to specific character caps, but no content sanitization is applied to LLM outputs.- [DYNAMIC_EXECUTION]: The runner script dynamically generates React source code and writes it to a .canvas.tsx file to facilitate real-time UI updates within the IDE. This mechanism is a legitimate use of the platform's visualization capabilities.

dag-task-runner