The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface because it relies on untrusted data from external websites to guide the modification of local source code. A malicious website could potentially craft its content to influence the resulting code patches.
Ingestion points: Untrusted data is ingested from DOM snapshots, network logs, and trace artifacts (e.g., summary.md, state/, and network.jsonl) generated during the diagnostic phase.
Boundary markers: The instructions do not specify the use of delimiters or sanitization techniques when the agent interprets data from external pages.
Capability inventory: The agent has access to the Edit and Write tools to modify Javascript files and the Bash tool to execute the webcmd command.
Sanitization: There are no explicit requirements to validate or escape content retrieved from websites before incorporating it into code fixes.
[COMMAND_EXECUTION]: The skill automates a workflow that involves runtime code modification and execution. It patches local Javascript adapter files and executes them using the webcmd tool to verify repairs, which is a form of dynamic execution based on external inputs.
[DATA_EXFILTRATION]: The skill provides functionality to report fixed adapters by creating issues on the vendor's GitHub repository. This process transmits diagnostic trace artifacts to an external service. While this is a documented feature and requires user confirmation before using the gh tool, it facilitates the movement of local diagnostic data to an external repository.

webcmd-autofix