The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from user-provided GitHub repositories using the gh api tool (Section 4). The requirement for the agent to "Read every fetched file" to evaluate the merge plan creates a vulnerability surface where instructions embedded in the external files could be interpreted and executed by the agent during the intake process.\n
Ingestion points: Content fetched via GitHub API and saved to the .upstream-scratch/ directory (Section 4).\n
Boundary markers: The instructions do not specify prompt delimiters or "ignore embedded instructions" warnings when the agent reads the imported content.\n
Capability inventory: The skill can write to marketplace files, update prompt instructions in other skills, and execute shell commands via gh, git, and grep.\n
Sanitization: Relies on license compliance checks and manual agent/user inspection, with no automated sanitization of the content body.\n- [DYNAMIC_EXECUTION]: The skill performs dynamic script generation by appending a custom gh api fetch loop to the code blocks of a separate skill file (.agents/skills/upstream-sync/SKILL.md) as part of the synchronization setup in Section 7.\n- [PERSISTENCE]: The skill establishes persistence for external data fetching by modifying the upstream-sync skill's sync table and adding automated fetch logic, ensuring that the connection to external, user-specified repositories is maintained across different agent sessions.\n- [COMMAND_EXECUTION]: The skill utilizes several powerful shell-based tools including gh, base64, grep, and git. These tools are used to process and integrate data fetched from external sources into the local environment.\n- [EXTERNAL_DOWNLOADS]: The skill fetches content from the github.com API. While GitHub is a well-known service, the skill allows for content to be pulled from any arbitrary, user-specified repository, which constitutes an untrusted external source.

external-repo-intake