The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes standard development tools such as git, build tools, and test runners. These are scoped to the project environment and used for 'safe probes' (e.g., local builds, isolated test data, and falsification probes) to verify code behavior. The skill explicitly limits these actions to local, non-persistent side effects unless remediation (fix mode) is authorized by the user.- [DATA_EXPOSURE]: The skill processes repository-local source code, configuration, and JSON audit results. It includes explicit instructions to treat prior findings JSON as untrusted input, requiring schema validation and checking for path traversal or repository mismatches before parsing.- [PROMPT_INJECTION]: Instructions for sub-agents are strictly defined and scoped. The coordinator agent maintains control over the verdict, severity, and final reporting, preventing delegated 'lanes' from overriding the main agent's safety or logic boundaries. There are no attempts to bypass safety filters or ignore system instructions.- [REMOTE_CODE_EXECUTION]: While the skill mentions network access for 'pull-request context' or 'third-party documentation', it does not contain patterns for piping remote content into shells or executing unverifiable external scripts. Dependency checks are limited to the project's own declared lockfiles and requirements.- [SAFE]: The skill demonstrates high security maturity by defining strict boundaries for remediation, requiring byte-for-byte tree verification after mutation probes, and providing detailed formatting for JSON output that avoids markdown injection or unauthorized file overwriting.

code-review