The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local repository (code changes and history) to generate natural language commit messages.
Ingestion points: git diff, git diff --staged, and git log output are used in Step 1 and Step 2 to understand the intent of changes.
Boundary markers: The skill does not explicitly define delimiters to separate file content from instructions in the prompt, relying on the model's context understanding.
Capability inventory: The skill possesses capabilities to modify the repository state via git reset, git add, and git commit.
Sanitization: The instructions mitigate this by requiring a confirmation step (Step 4) where users review the generated message and the list of files before execution, though this can be bypassed in 'Quick Commit Mode' unless sensitive files are detected.
[COMMAND_EXECUTION]: The skill executes shell commands constructed from strings generated by the AI model.
Evidence: Step 5 involves executing git commit -m "<message>" using text generated based on code analysis.
Mitigation: The skill includes explicit security logic for shell escaping, mandating the use of single quotes for arguments or backslash-escaping special characters ($, `, , ") when double quotes are required, which prevents arbitrary command injection into the host shell.

yy-commit