The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to transmit local diagnostic data, environment details, and error logs to a public GitHub repository (celo-org/buy-skill). This behavior poses a risk of accidental sensitive data exposure, which the skill attempts to mitigate through explicit redaction instructions for private keys, SSH keys, wallet addresses, and bearer tokens.
[COMMAND_EXECUTION]: The skill utilizes multiple local system commands to gather environment data and perform security checks. This includes the use of npx to check the version of the vendor-supplied @celo/buy package, uname for system information, grep for secret scanning, and the gh CLI for interacting with GitHub.
[DYNAMIC_EXECUTION]: The skill instructions include the dynamic generation of shell scripts and temporary files to handle the submission of reports. It uses security best practices such as umask 077 for temporary file creation and quoted heredocs (<<'BODY') to prevent shell expansion of potentially malicious content within the error reports.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of tool error messages and execution outputs to populate issue templates.
Ingestion points: Untrusted data enters the context through the error, actual, and repro fields during the bug reporting process.
Boundary markers: The skill uses structured Markdown tables and block headers to delimit user-provided data from system instructions.
Capability inventory: The skill has the ability to execute shell commands (gh, grep, npx) and write to the local filesystem.
Sanitization: The skill relies on a regex-based grep backstop scan and a mandatory human-in-the-loop approval step where the user must review the final rendered issue before submission.

file-feedback