The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external GitHub repositories, such as README.md files and issue comments. This surface is protected by structured data parsing that whitelists specific JSON keys and truncates long text fields, reducing the risk of embedded malicious instructions influencing the agent's behavior.
[COMMAND_EXECUTION]: The skill executes local scripts (scripts/gh-recon.ts and scripts/contribute-recon.ts) to perform its tasks. These scripts include significant safety checks: they validate their own canonical paths to prevent path traversal or symbolic link attacks and verify their own SHA-256 hashes to ensure the executable code has not been tampered with.
[EXTERNAL_DOWNLOADS]: Network interaction is conducted via the gh CLI and is restricted to the github.com domain. The skill uses a 'plan' and 'run' architecture where a set of GET endpoints is first proposed to the user for approval; the subsequent 'run' phase enforces the approved plan hash, preventing the model from dynamically changing request targets.
[CREDENTIALS_UNSAFE]: The skill proactively protects user and repository secrets through a redaction engine. It uses regular expressions to detect and mask typical credential patterns (GitHub PATs, AWS keys, private keys, and Authorization headers) from repository content before it reaches the AI agent's context.

tailrocks-contribute-recon