The Agent Skills Directory

[SAFE]: The skill uses a strict 'mutation via verb' architecture. All file modifications (creating catalogs, adding entries) are delegated to specific MCP tools (invariants_scaffold, invariants_add). The agent is explicitly instructed never to hand-write or append YAML manually.
[SAFE]: The tool architecture enforces declarative-only enforcement logic. The skill documentation and reference files (check-mode.md) specify that the enforcement DSL is restricted to a sandbox-free combinator tree (grep, structural, heuristic signals). It explicitly forbids executable escape hatches like script, exec, or code, which are rejected at the schema validation level.
[SAFE]: The skill implements a robust verification flow. Every mutation requires a mandatory dryRun step where the agent must show the rendered diff to the user and obtain explicit confirmation before proceeding to a real write operation.
[SAFE]: The skill handles sensitive namespaces securely. It distinguishes between user (project-owned) and dev (tool-substrate) namespaces, preventing collision or spoofing of core architectural rules by rejecting unauthorized tier assignments outside the tool's own repository.

invariants