The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill relies on dynamic downloads of numerous analysis tools from public registries (NPM, PyPI, Go) using npx, uvx, and go run. Examples include jscpd, aislop, zizmor, complexipy, fallow, and staticcheck. The use of unpinned versions with @latest tags introduces a potential supply chain risk, as the environment's security depends on the integrity of these public packages at the moment of execution.\n- [DYNAMIC_EXECUTION]: When the --deep flag is used, the skill is instructed to execute the target repository's own build processes, such as cmake or meson for C/C++ projects and dotnet build for C# projects. This execution of arbitrary build logic from an untrusted repository is a high-privilege operation. The skill identifies this as a 'deep profile' capability and implements a footprint comparison check to detect unauthorized filesystem changes.\n- [COMMAND_EXECUTION]: The skill makes extensive use of shell commands to invoke a wide array of static analysis tools, compilers, and formatters. This includes a 'Fix' mode where the agent is authorized to apply autofixes and perform commit-sized changes to the codebase based on tool recommendations.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code from the user's repository by passing it through various static analysis tools. The reports (JSON, SARIF, TSV) are interpreted by the agent and subagents to produce grades and remediation plans. A malicious repository could attempt to poison tool outputs to influence the agent's logic or the resulting 'fix plan'. The skill mitigates this through structured templates and human review checkpoints.\n- [CREDENTIALS_UNSAFE]: The skill implements a robust secret detection process using gitleaks and specific regex filters to redact sensitive data from reports. However, the analysis process requires handling repository files that may contain secrets in temporary directories (scratch/). The skill uses a 'Secrets never quoted' contract and performs a final directory-wide grep to ensure no credentials are leaked in the final output.

crank-health