Read and follow the repository's shared Sokra-mode guidance before working on the requested task. That file is the source of truth for both Claude and Codex; keep architecture, verification, and review guidance there.
Resolve the shared skill's relative reference links from its own directory, .claude/skills/sokra-mode/. Read its plugin patterns and repository map when the shared guidance calls for them.
Invoke this skill with $sokra-mode. Preserve the existing opt-in behavior through agents/openai.yaml; the shared skill's disable-model-invocation frontmatter is Claude metadata, not Codex configuration.