crosscheck

Files (2)

• SKILL.md
• codex_ask.sh · scripts/

SKILL.md (first lines)

# Claude-hosted Crosscheck

This compatibility entrypoint delegates to the shared procedure in [verify](../verify/SKILL.md). Read that file and follow its routing, permission, evidence, wrapper, strategy, stopping, and reporting contracts. Do not maintain a second workflow here.

- Claude is the host; Codex is the read-only counterpart. The host owns implementation. In Codex, use the shared entrypoint's model-family routing instead.
- Choose the strategy automatically, within task-scoped reviewer permission; do not ask which method to use. Respect an explicitly selected method.
- Infer the stopping stage from intent: opinions → `analyze`, plans → `plan`, completed work → `verify`, authorized implementation → `full`. Advice does not default to implementation.
- For advice and planning, read the shared [strategy reference](../verify/references/crosscheck-strategies.md). Broad improvement exploration normally uses independent proposals plus mutual review; a concrete draft uses critique; a narrow exploration uses independent proposals.
- Say when an idea is poor value, recommend keeping the current approach when justified, and preserve evidence-based dissent. No mandatory praise, criticism, or consensus.
- Every Codex call goes through this entrypoint's `scripts/codex_ask.sh`, which links to the shared implementation. Keep its read-only sandbox and exact returned session ID; never call `codex exec` directly or recursively invoke crosschecking.
- Preserve user model/effort settings. Local edits, Git changes, remote mutations, and deployment retain their existing approval boundaries.
- Reply in the user's language. Report actual reviewer responses and checks only; no claimed crosscheck without a collected response.

From moveju112/skill_verify, MIT.

More from moveju112/skill_verify