Release v3.25.1 — Codex AGENTS.md emits direct mmr review recipes · zigrivers/scaffold

Fixed

Codex AGENTS.md now emits direct mmr review recipes for review-code and review-pr instead of the scaffold run <step> shim. The shim works in Claude Code (slash commands re-inject the meta-prompt into the model's context) but breaks in Codex, which executes scaffold run as a shell command and treats the 45 KB meta-prompt as opaque stdout — agents reported "scaffold run review-code emitted the review wrapper again rather than an actionable verdict" and fell back to bare mmr review --staged. The codex adapter now inlines a deterministic recipe (full 7-level BASE_REF resolution ladder + empty-diff guard via git diff --quiet + multi-mode mmr review invocation split into separate per-mode code blocks + verdict-handling guidance) for these two tools, with gh pr view PR-number detection in the review-pr recipe. The generated AGENTS.md preamble now explicitly notes that scaffold run is the default unless the step provides an inline shell recipe. Non-executor steps still use scaffold run <slug> unchanged.

The Codex recipes ship 3-channel coverage (Codex CLI, Gemini CLI, Claude CLI). The 4th-channel Superpowers code-reviewer reconcile requires a harness that can dispatch agent skills, which Codex cannot do directly — the recipes point users at scaffold run review-{code,pr} from a Claude Code session for full 4-channel coverage.