The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill describes a three-phase architecture that spawns claude -p subprocesses to handle complex code fixes. This design delegates execution to a new agent instance using runtime-generated JSON data, creating a dynamic execution flow controlled by the workspace state.\n- [INDIRECT_PROMPT_INJECTION]: The skill establishes a vulnerability surface where malicious code in workspace files can influence the behavior of remediation subprocesses. \n
Ingestion points: Workspace files are processed by linters in the multi_linter.sh hook, generating violation summaries. \n
Boundary markers: There are no described delimiters or instructions to ensure the subprocess ignores instructions embedded within the violation data. \n
Capability inventory: The secondary claude -p instance has the capability to write and modify code in the repository. \n
Sanitization: The skill does not document any sanitization or validation logic for the linter output before it is interpolated into the subprocess prompt.\n- [COMMAND_EXECUTION]: The skill employs PreToolUse hooks to intercept and block specific shell commands. It restricts legacy package managers such as pip, npm, and poetry to enforce the use of modern tools like uv or bun.\n- [EXTERNAL_DOWNLOADS]: The setup process involves installing several third-party utilities and linters from public registries, including jaq, ruff, uv, and biome. These are well-known developer tools used for static analysis and formatting.

plankton-code-quality