verifylisted
Install: claude install-skill moveju112/skill_verify
# Verify
One shared entrypoint for Claude Code, Codex, and pi. The current agent remains the host and owns implementation; the reviewer only reads and reviews.
## Route and permission
- Ordinary verification: inspect and test locally; never call an external LLM.
- Crosscheck: require explicit task-scoped permission to involve the selected external reviewer. Loading this skill, installing it, or asking for generic verification is not permission.
- Claude host → Codex reviewer. Codex host → Claude reviewer; never invoke Codex recursively from Codex.
- In pi, inspect `PI_PROVIDER` and `PI_MODEL` using the shell tool. Claude/Anthropic model → Codex reviewer; GPT/OpenAI model → Claude reviewer. For another model family, unknown metadata, or an ambiguous router, ask which reviewer to use. Pi is a runtime, not a model family.
- Respect an explicitly requested reviewer, but do not describe same-family review as independent cross-model verification; clarify that conflict before calling.
- Infer `analyze`, `plan`, `full`, or `verify`. `analyze` stops after analysis, `plan` after planning, `verify` only reviews existing work. Only `full` implements, and only within existing change permissions.
- Review-only requests stay read-only. Ask separately before DB, network, remote, destructive, deployment, Git mutation, or service-changing operations beyond the authorized reviewer call.
## Local verification
1. Read applicable project rules; inspect `git status --short` and the relevant di