The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external git repositories, which could contain malicious instructions aimed at influencing the agent's behavior. * Ingestion points: The agent reads commit messages via 'git log', file contents via 'git show', and repository-defined build/test scripts. * Boundary markers: The skill lacks explicit boundary markers or 'ignore' instructions for the content it ingests from the repository history. * Capability inventory: The skill has the ability to read and write to the local filesystem, stage files using 'git add', and execute arbitrary build/test commands defined in the project structure. * Sanitization: No sanitization or validation is performed on commit messages or file metadata before the agent interprets them to determine the 'intent' of changes.
[DYNAMIC_EXECUTION]: The skill executes repository-specific code during the verification phase. * Evidence: Step 8 ('Build and test before declaring success') instructs the agent to run the project's own build and test commands, which constitutes execution of potentially untrusted scripts found within the repository being analyzed.
[COMMAND_EXECUTION]: The skill makes extensive use of host-level shell commands to interact with the git repository. * Evidence: Instructions include various 'git' commands (log, show, diff, status, add, config, merge-base, grep) and filesystem operations ('ls', redirecting output to '/tmp'). While these are scoped to the repository, they involve direct interaction with the host's shell.

fix-merge-conflicts