The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local development commands, including unit tests, linting, and type checking, to verify the integrity of code changes during the 'apply' phase.
[COMMAND_EXECUTION]: A local utility script named validate-yaml.py is invoked to ensure that the generated refactoring checklists are correctly formatted.
[SAFE]: The skill implements a 'Refusal Routing' mechanism that explicitly directs the agent to halt and suggest alternative workflows if certain safety criteria are met, such as attempting to modify third-party code or refactoring modules without existing test coverage.
[SAFE]: The workflow requires mandatory human approval at multiple stages: scan selection, design review, and after every individual code modification step. This significantly limits the risk of unintended or autonomous malicious behavior.
[SAFE]: The skill possesses an inherent surface for indirect prompt injection because it reads and processes local source code which could contain malicious instructions.
Ingestion points: The skill scans project source code and imported modules as defined in the 'scan' phase (SKILL.md).
Boundary markers: While the skill uses structured Markdown and YAML for its internal state, it does not explicitly use delimiters to ignore instructions embedded within the code it reads.
Capability inventory: The skill can modify local files and execute shell commands to run project-specific tests and linters (SKILL.md, reference/refusal-routing.md).
Sanitization: This risk is mitigated by the 'Refusal Routing' rules and the strict human-in-the-loop approval process required for every proposed change.

cs-refactor