The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several shell commands to facilitate its workflow, including git init, git commit, git clone, and node. These commands are used to manage the state of the artifact and interact with external plugins. Using shell commands with user-provided or dynamically generated arguments represents a significant capability that requires trust in the input sources.\n- [REMOTE_CODE_EXECUTION]: The skill identifies and executes a specific Node.js script located on the user's file system through dynamic path resolution. It attempts to find codex-companion.mjs in the environment or specific plugin directories. Running scripts from paths determined at runtime is a dynamic execution pattern that could be exploited if a malicious file is placed in a searched directory.\n- [EXTERNAL_DOWNLOADS]: The 'Ground Truth' step involves cloning external Git repositories. The skill provides instructions to execute git clone on URLs provided as input, which establishes network connections to arbitrary remote servers.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the artifacts it processes. Ingestion point: The ARTIFACT_PATH identifies the file or diff being reviewed. Capability inventory: The driver has the ability to edit files (apply findings), commit changes (git commit), and execute shell commands (node task). Sanitization: There is no explicit sanitization or validation of the 'fixes' suggested by the LLM reviewers before they are applied to the artifact. A malicious artifact could trick a reviewer into suggesting a 'fix' that introduces a vulnerability, which the driver would then apply.

multi-llm-convergence