The Agent Skills Directory

Dynamic Automation Generation: The skill is designed to have the agent author new TypeScript and JavaScript scenario runners into a local scratchpad directory and then execute them using Node.js. Running dynamically generated code is a significant capability that requires oversight, even when used for performance testing.
Command Execution and Process Spawning: The scripts utilize child_process.spawn to launch VS Code (Code OSS) and execute development lifecycle commands such as pnpm run transpile-client. This provides the agent with broad access to execute shell commands to drive performance workflows.
Management of Sensitive User Profiles: The skill includes utilities for copying and reusing VS Code user data profiles. These profiles may contain authentication secrets or personal configuration data. While the instructions advise storing these in ignored directories (.build/), the handling of credential-bearing data is a notable security consideration.
Indirect Prompt Injection Surface: The skill processes external workflow descriptions and feeds prompts into an interactive Chat interface. This creates a potential surface where malicious instructions embedded in the processed data could influence the agent's behavior. The skill mitigates this by recommending the use of throwaway workspaces and read-only commands.
Ingestion points: Workflow descriptions and Chat prompts defined in SKILL.md and user requests.
Boundary markers: The skill advises using isolated workspaces and safe commands but lacks automated input sanitization.
Capability inventory: File system writing and command execution via spawn are present in chat-memory-smoke.mts and workbench-css-performance.mts.
Sanitization: No explicit sanitization of prompt templates or user-provided messages is implemented before they are passed to the automation scripts.

auto-perf-optimize