The Agent Skills Directory

[COMMAND_EXECUTION]: The skill identifies and modifies local adapter source files (JavaScript) and then executes them to verify fixes. This automated code generation and execution loop poses a security risk if the logic is compromised by malicious input.
[PROMPT_INJECTION]: The skill processes external, untrusted data from websites to generate code patches, creating an indirect injection surface where a website could influence the agent to write malicious code.
Ingestion points: Website DOM state and network logs are collected via webcmd browser state and webcmd browser network tools.
Boundary markers: No explicit boundary markers or "ignore instructions" warnings are present to protect the agent from processing instructions embedded within the target website's content.
Capability inventory: The skill uses Edit and Write tools to modify local JavaScript files at the adapterSourcePath, and the Bash(webcmd:*) tool to execute the modified code.
Sanitization: The instructions lack specific sanitization or validation steps for content ingested from the browser before it is used to generate code patches.
[DATA_EXFILTRATION]: The skill facilitates the creation of GitHub issues via the gh tool, which involves sending diagnostic data to a remote repository. While the skill instructs the agent to sanitize this data, it still represents a mechanism for exporting information collected during the debugging process. The default target is the vendor repository agentrhq/webcmd.

webcmd-autofix