The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several local shell commands to facilitate the bundling process and verify the output. This includes ls, wc, and grep for file verification, and git ls-remote to identify the specific commit of a remote repository.
[REMOTE_CODE_EXECUTION]: The skill uses bunx repomix to dynamically download and execute the Repomix utility from the npm registry. While Repomix is a common community tool for preparing codebases for LLMs, the use of bunx involves executing code from a remote repository at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface for indirect prompt injection because it processes untrusted data from external repositories to write system instructions and define the core concepts of a Gemini Gem.
Ingestion points: The agent reads the bundled knowledge file (gem-<slug>/knowledge/<slug>-knowledge.txt) containing the full text of the source repository in Step 3.
Boundary markers: While Repomix adds ## File: <path> headers to the bundled content, there are no explicit instructions or delimiters used when the agent reads this file to prevent it from following instructions embedded within the source repository.
Capability inventory: The skill possesses the capability to execute shell commands (bunx, git, ls, grep, wc) and perform file system writes to create the Gem folder and its contents.
Sanitization: The skill does not implement sanitization or filtering of the repository content before the agent analyzes it to generate the Gem's instructions.

repo-to-gem