AgentSkillsCN

codex-claude-loop

在双 AI 工程闭环中，Claude CLI 以 `--dangerously-skip-permissions` 模式运行，用于规划与实施；而 Codex 则负责验证与审查输出。适用于用户提出 Claude + Codex 协作需求、交叉审核、双 AI 循环，或明确要求以跳过权限验证的模式运行 Claude CLI 时使用（“claude”、“dangerously-skip-permission”、“dangerously-skip-permissions”、“codex claude”、“dual AI”、“교차 검증”、“claude 협업”）。

SKILL.md
--- frontmatter
name: codex-claude-loop
description: Orchestrates a dual-AI engineering loop where Claude CLI runs with `--dangerously-skip-permissions` for planning/implementation and Codex validates/reviews outputs. Use when users ask for Claude+Codex collaboration, cross-review, dual AI loop, or explicit Claude CLI execution with skip-permission mode ("claude", "dangerously-skip-permission", "dangerously-skip-permissions", "codex claude", "dual AI", "교차 검증", "claude 협업").

Codex-Claude Loop

Claude CLI + Codex 협업 루프를 실행한다.

Preconditions

  • •Run --dangerously-skip-permissions only when the user explicitly requested it.
  • •Prefer isolated environment: disposable branch/worktree, no production credentials.
  • •Default Codex verification to --sandbox read-only.

Core Loop

code
Plan (Claude) -> Validate (Codex) -> Feedback ->
Implement (Claude) -> Review (Codex) ->
Fix (Claude) -> Re-validate (Codex)

Quick Commands

bash
# 1) Claude non-interactive output
claude --dangerously-skip-permissions -p "Write an implementation plan for: <task>"

# 2) Codex validation (read-only)
echo "<claude_output>" \
| codex exec -m gpt-5.2-codex --config model_reasoning_effort="high" --sandbox read-only

# 3) Claude interactive implementation session
claude --dangerously-skip-permissions

Scripted Commands

Use the bundled helper script for repeatable runs:

bash
# Claude one-shot
bash .codex/skills/codex-claude-loop/scripts/claude-codex-loop.sh claude-print "Plan this task: <task>"

# Codex one-shot review
bash .codex/skills/codex-claude-loop/scripts/claude-codex-loop.sh codex-review "Review this plan: <plan>"

# One roundtrip: Claude plan -> Codex review
bash .codex/skills/codex-claude-loop/scripts/claude-codex-loop.sh roundtrip "Implement feature X"

Operating Rules

  • •Keep handoff explicit: include assumptions, constraints, and acceptance criteria in every prompt.
  • •Treat Claude output as draft input. Use Codex to challenge edge cases, regressions, and test gaps.
  • •If either CLI returns non-zero exit, stop and summarize the failure before retrying.
  • •For high-impact edits (schema, auth, deployment), ask the user before applying changes.

Recovery

  • •Resume Claude session: claude -c
  • •Resume Codex session: echo "continue" | codex exec resume --last
  • •If context drifts, re-anchor with a short summary and current objective.