ask-peer

# Ask Peer One request, one opposite-model reply. This is the phone-friendly headless path: use it for a focused second opinion, review, or bounded work pass. Use `herdr-pair` for a visible multi-turn dialogue. Detect the current harness, then read the matching reference in full: | Current harness | Peer | Reference | |---|---|---| | Codex | Fable | `references/ask-fable.md` | | Claude Code | Codex | `references/ask-codex.md` | If the harness is ambiguous, inspect the environment and available CLI before choosing. Never ask the same model through its own CLI. ## Shared contract - Default to read-only. Grant a write pass only when the user's request requires edits, with an explicit lease: target files, forbidden changes, validation, and stop point. - Include the workspace root, the exact ask, relevant repo instructions, and the desired response format. Keep secrets out of the prompt. - Model selection belongs to each peer's own surface: call Fable explicitly; leave Codex's model unset so `~/.codex/config.toml` or the CLI default owns it. - Treat the reply as evidence, not authority. Inspect any diff and rerun the named validation before relying on a write pass. - No commit, push, merge, deploy, or credential mutation unless the user explicitly authorized that action.