The Agent Skills Directory

[COMMAND_EXECUTION]: The provided templates request extensive shell access through the Bash tool. Template 1 allows for npm run * and pnpm run *, Template 7 and 8 allow for node *, and Template 2 and 3 allow various git and gh (GitHub CLI) commands. These permissions provide a path for the execution of arbitrary scripts or binaries within the user's environment.\n- [EXTERNAL_DOWNLOADS]: Template 3 ('Dependency Update') automates fetching and installing external software packages using npm update and npm install. This introduces supply chain risks if external registries or the packages themselves are compromised.\n- [DATA_EXFILTRATION]: Template 8 ('Monitoring Setup') utilizes the curl command to interact with network endpoints. While the intended use is for health checks on specified service URLs, the tool capability allows for the exfiltration of local data to remote servers.\n- [CREDENTIALS_UNSAFE]: Template 4 ('Security Scan') and Template 2 ('Batch Review') are configured to search the entire project (Read(**)) for sensitive secrets, including API keys, authentication tokens, and private keys, using regex patterns like sk-, [A-Za-z0-9]{32,}, and -----BEGIN.*KEY-----. This involves high-privilege access to sensitive local data.\n- [PROMPT_INJECTION]: Template 2 ('Batch Review') and Template 4 ('Security Scan') are susceptible to indirect prompt injection because they process untrusted external data.\n
Ingestion points: These templates read data from git log, git show, and gh pr view (pull request contents).\n
Boundary markers: The task prompts lack explicit boundary markers or isolation instructions to prevent the agent from obeying instructions embedded within the code or commit messages being reviewed.\n
Capability inventory: The templates grant the agent the ability to write to files and perform GitHub actions (gh pr review), which could be leveraged by a malicious PR to perform unauthorized actions.\n
Sanitization: There is no logic provided to sanitize or filter the input data before it is processed by the AI model.

dispatch-templates