The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage files and state, including git diff, git restore, mkdir, rm, curl, and playwright-cli.
[REMOTE_CODE_EXECUTION]: The skill follows a dynamic execution pattern (Category 10) by generating a temporary TypeScript file (e2e/tmp-screenshots.e2e.ts) derived from the $ARGUMENTS parameter and executing it via playwright-cli test. This involves running agent-generated code at runtime.
[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8) because it ingests untrusted data from $ARGUMENTS and uses it to construct executable test scripts.
Ingestion points: Custom instructions provided in the $ARGUMENTS parameter (SKILL.md).
Boundary markers: Absent; the instructions are directly translated into test logic without delimiters or warnings.
Capability inventory: The skill can write files, execute shell commands, and run Playwright tests which have full access to the browser and potentially local network resources.
Sanitization: No sanitization or validation logic is defined for the content of the generated test script.
[CREDENTIALS_UNSAFE]: The instructions (Steps 6a and 8b) direct the agent to analyze the project's authentication patterns and 'reuse the same approach'. This creates a risk of sensitive tokens, cookies, or credentials being extracted from the environment and placed into the temporary test files in plain text.

screenshots