The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PR comments and reviews from GitHub (ingestion points in fetch_pr_comments.py). It includes a dedicated _untrusted_doc.py script that wraps this external content in <<<UNTRUSTED_EXTERNAL_CONTENT>>> delimiters (boundary markers) to prevent the agent from following malicious instructions hidden in the data. Capabilities like command execution (discover_and_verify) and API writing (post_thread_replies) are only performed after these safety measures.
[COMMAND_EXECUTION]: The discover_and_verify tool executes build and test commands (like npm test or pytest) identified within the local repository. The skill uses a secure run_command helper in _skill_cli.py which executes these via subprocess.run with shell=False, effectively preventing shell injection attacks.
[EXTERNAL_DOWNLOADS]: The skill communicates with GitHub's servers using the gh CLI to fetch and post PR data. These network operations are standard for the skill's purpose and target a well-known service domain.
[SAFE]: The skill follows security best practices by implementing hard gates for merges (assert_no_merge) and ensuring local changes are pushed (assert_fixes_pushed) before communication with the remote CI bot. No suspicious obfuscation, persistence mechanisms, or privilege escalation patterns were detected.

heyeddi-ci-respond