Codex Cross Review

Prerequisites

1. Codex CLI installed: codex --version must succeed
2. OpenAI API key configured: Codex must be authenticated
3. Phase artifacts exist: the phase being reviewed must have produced deliverables

Configuration

The skill reads ~/.codex/config.toml for model and effort settings. Users configure their preferred model and reasoning effort there — the skill never overrides these.

Display the current config to the user at the start of every invocation:

Codex Config: model={model}, reasoning_effort={effort}

Execution

Phase Auto-Detection

If no phase number is provided, detect the current phase from:

1. Most recently modified docs/phase-N-*/ directory
2. Ask user if ambiguous

Spawn Cross-Reviewer Agent

Task(subagent_type="rtl-agent-team:codex-cross-reviewer",
     prompt="Run cross-review for Phase {N}.
             Phase intent: {brief description from phase docs}.
             Target artifacts: {list of docs/code paths}.
             Changed files: {from git diff}.")

Integration Protocol (for Phase Orchestrators)

Phase orchestrators (P1-P6) should invoke cross-review as their penultimate step (before declaring phase complete):

# In orchestrator agent prompt, before final step:
Step N-1: Cross-Review Gate
  Task(subagent_type="rtl-agent-team:codex-cross-reviewer",
       prompt="Cross-review Phase {N}. Artifacts: {paths}.")
  # MANDATORY explicit verdict check:
  Read(".rat/cross-review/phase-{N}/cross-review-report.md")
  # Parse verdict field — must be CONSENSUS or USER_DECIDED
  # If verdict != CONSENSUS and user did not approve → do NOT proceed to completion

Orchestrators that integrate cross-review:

• p1-research-orchestrator — review spec analysis quality
• p2-arch-orchestrator — review architecture decisions + ref model
• p3-uarch-orchestrator — review uArch design + BFM
• p4-implement-orchestrator — review RTL implementation
• p4-rtl-sanity-orchestrator — review rapid implementation
• p5-verify-orchestrator — review verification completeness
• p6-review-orchestrator — review design note quality

Review Loop Summary

Round 1: Claude → phase summary → Codex reviews → findings JSON
Round 2+: Claude → fixes + rebuttals + agreement ledger → Codex re-reviews → updated findings
Stability: 2+ consecutive rounds with verdict==APPROVE + no new critical/major + no still_disagree + no oscillation = CONSENSUS
Max 5 rounds → AskUserQuestion escalation to user

Convergence Improvements (v0.7.2+)

• Agreement Ledger: settled items are injected into each follow-up prompt to prevent context-reset re-raises
• Anti-Oscillation: re-raised settled items without new evidence are rejected automatically
• Stability Criterion: consensus requires 2+ consecutive rounds with verdict==APPROVE, not single-round APPROVE
• Oscillation Detection: tracked in resolution-state.json for diagnostics

Artifacts

Cross-review artifacts are phase-scoped for traceability:

• .rat/cross-review/review-schema.json — shared JSON schema
• .rat/cross-review/phase-{N}/phase-summary.md — phase context sent to Codex
• .rat/cross-review/phase-{N}/prompt-round-R.txt — exact prompt sent each round
• .rat/cross-review/phase-{N}/round-R.json — Codex's structured response each round
• .rat/cross-review/phase-{N}/resolution-state.json — running resolution tracker
• .rat/cross-review/phase-{N}/cross-review-report.md — final summary report
• .rat/cross-review/phase-{N}/escalation-summary.md — generated if user escalation needed

Completion marker: .rat/state/cross-review-phase-${N}-done (e.g., cross-review-phase-2-done)