The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to interact with the system keychain and retrieve sensitive credentials.
Evidence: scripts/coordinator-validate.py and references/miner.md use security find-internet-password -s github.com -w to fetch GitHub tokens in plaintext.
[COMMAND_EXECUTION]: The skill provides instructions that reduce security boundaries for external tools.
Evidence: references/codex-review-cli-patterns.md and references/codex-review-invocation.md recommend the use of the --dangerously-bypass-approvals-and-sandbox flag with the Codex CLI tool.
[DYNAMIC_EXECUTION]: The skill uses dynamic module loading to verify environment prerequisites.
Evidence: scripts/validate.py utilizes the __import__ function to check for the presence of the github package at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill has a potential surface for indirect prompt injection by processing untrusted data from GitHub to drive system actions.
Ingestion points: references/feedback.md and references/fix.md fetch comments from GitHub PR reviews and issues using the gh API.
Capability inventory: The skill has capabilities to edit local files, perform git commits, and push changes to remote repositories.
Boundary markers: The skill suggests using temp files and --body-file flags to prevent shell mangling, but lacks explicit prompt boundary delimiters for the ingested text.
Sanitization: A manual validation phase is included where the agent is instructed to verify claims against the code before applying fixes, acting as a human-in-the-loop mitigation.
[EXTERNAL_DOWNLOADS]: The skill depends on external tools and libraries that are not bundled with the skill.
Evidence: Instructions in references/miner.md and scripts/validate.py reference the installation and use of @openai/codex and PyGithub.

pr-workflow