The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The 'Webhook Router' implementation pattern described in components.md is vulnerable to command injection. It uses an unsanitized path from an incoming request to dynamically construct a directory path and execute a script.
Evidence: The script extracts $WORKFLOW from $request via jq and then executes (cd "workflows/flows/$WORKFLOW" && ./run.sh). An attacker could use path traversal or shell metacharacters in the path to execute arbitrary commands.
[COMMAND_EXECUTION]: The skill's primary function is to generate and execute shell scripts (run.sh) that use various CLI tools like curl, jq, yq, and flock to perform automated tasks.
Evidence: Multiple files (data-flow.md, lifecycle.md, errors.md) provide templates for scripts that execute arbitrary shell commands and process local files.
[CREDENTIALS_UNSAFE]: The documentation encourages a practice of storing full database connection strings, which often include plaintext usernames and passwords, inside markdown files.
Evidence: components.md provides a format for 'Database Connections' that includes a 'Secret' field meant to store a full postgresql://user:pass@host:port/database string in a markdown file.
[EXTERNAL_DOWNLOADS]: The skill's architecture is built around continuous network interaction, using curl to fetch data from remote APIs and receive triggers via webhooks.
Evidence: data-flow.md and lifecycle.md demonstrate fetching JSON data from external URLs and posting results to third-party services like Pushover.

Workflow