The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes user input and repository source files, creating a surface for indirect prompt injection. The skill mitigates this through a multi-pass review loop (Safety, Surface, Audit) and enforces mandatory validation signals and invariant checking.
[REMOTE_CODE_EXECUTION]: Executes local Python scripts for validation and linting. The skill uses uv run for executing scripts/lint_fix_skill_contract.py and references internal system scripts for quick validation. No remote script execution from untrusted sources was identified.
[COMMAND_EXECUTION]: Leverages system tools for code searching and verification. It uses ripgrep (rg) to locate behavior tokens and evidence and executes local test suites (e.g., pytest, Makefile) to confirm fixes.
[DATA_EXFILTRATION]: Designed to operate in local-first mode. The skill explicitly instructs to choose local signals without network access and contains directives to avoid logging sensitive information or PII.
[SAFE]: Implements comprehensive safety guardrails. It includes a strict autonomy gate and mandatory self-review loop while focusing on correctness and data safety as top priorities.

fix