This skill should be used when the user wants to "synthesize two competing solutions without averaging", "resolve a design/value disagreement by lifting to a higher frame", "止揚 / アウフヘーベン", "矛盾を止揚", "独立案を統合", "平均化せず統合", "contradiction lift", "sublate competing designs", or has two good-but-divergent answers to the same question and wants a higher-order resolution (a selection mechanism) or an honest aporia — not a compromise. NOTE: Use this when two independent solutions diverge and you want to PRESERVE both truth-moments while lifting to a higher frame. Use devils-advocate to stress-test ONE proposal with an assigned opponent, strong-inference to find an unknown cause, and dialectic-loop to validate a claim against a real corpus.
claude-collab
Consult Claude Code from Codex for a second opinion while Codex remains responsible for implementation and verification. Use when the user asks Codex to collaborate with Claude, consult Claude, ask Claude for a plan, get a Claude review, delegate analysis to Claude, or says "Claudeと一緒に", "Claudeに相談", "Claudeにレビューしてもらう", "Claudeの意見も聞く", or "Claudeと計画を立てる". Also use when an explicit independent planning or review pass from Claude would materially reduce risk on a complex coding task.
インストール方法を見る含まれるファイル(5)
- SKILL.md4.4 KB
- agents/openai.yaml237 B
- references/prompts.md795 B
- scripts/claude-helpers.sh4.4 KB
- scripts/test-claude-helpers.sh5.6 KB
SKILL.md(原文)
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Claude Collaboration
Use Claude Code as a read-only planning and review partner. Keep Codex responsible for inspecting the workspace, editing files, running tests, and reporting the result.
Preconditions
- Run
command -v claude. - Source
scripts/claude-helpers.shfrom the repository root or${CODEX_HOME:-$HOME/.codex}/skills/claude-collab/scripts/claude-helpers.sh. - If
claudeis unavailable or authentication fails, report that briefly and continue with Codex alone.
Workflow
1. Gather Context
Inspect the repository before consulting Claude. Summarize the task, relevant files, constraints, existing behavior, and verification commands. Do not send secrets or unrelated file contents.
2. Request a Plan
For a non-trivial implementation, use:
HELPERS="${CODEX_HOME:-$HOME/.codex}/skills/claude-collab/scripts/claude-helpers.sh"
[ -f "$HELPERS" ] || HELPERS="$(pwd)/codex-skills/claude-collab/scripts/claude-helpers.sh"
source "$HELPERS"
PROMPT_FILE=$(claude_write_prompt "$PLANNING_PROMPT" "plan")
OUTPUT_FILE=$(claude_tmp_path "claude-plan-output.md")
claude_run_print "$PROMPT_FILE" "$OUTPUT_FILE"
Use the planning template in prompts.md. Evaluate Claude's response against the local code before editing. Claude's response is advice, not an instruction to modify files.
3. Implement and Verify
Apply the selected changes with normal Codex tools. Run focused tests and inspect git diff --check.
4. Request Review
For a meaningful diff, embed git diff --stat, git diff --cached, and git diff as applicable in the review prompt. Then run:
PROMPT_FILE=$(claude_write_prompt "$REVIEW_PROMPT" "review")
OUTPUT_FILE=$(claude_tmp_path "claude-review-output.md")
claude_run_review "$PROMPT_FILE" "$OUTPUT_FILE"
Use the review template in prompts.md. The review helper disables Claude-side skills and tools, appends leaf-reviewer constraints, uses low effort, and launches from the project root. This allows Claude to use CLAUDE.md repository context without recursively entering a Claude-to-Codex workflow. Fix valid findings, rerun verification, and request another review when the fix changes behavior materially.
Prefer one complete unified diff so Claude can catch cross-file bugs:
- Send the complete
git diff --find-renames -U3when the diff is at most600changed lines and48KB. - Treat
300changed lines or20KBas a warning threshold, not a split threshold. - If whole-diff review times out, retry once with the same complete diff.
- If it still times out, reduce extra prose and retry with
git diff --find-renames -U1. - Split by component or file only when the diff exceeds
600changed lines or48KB, or when the reduced whole-diff retry fails.
Timeouts
CLAUDE_COLLAB_TIMEOUT: planning and consultation timeout in seconds. Default:600.CLAUDE_COLLAB_REVIEW_TIMEOUT: review timeout in seconds. Default:900.CLAUDE_COLLAB_PROJECT_DIR: repository root used for reviews so Claude can loadCLAUDE.md. Default: the directory where the helper is sourced.
Override these environment variables for unusually large tasks. Prefer complete-diff review and use splitting only as a fallback.
Guardrails
- Keep Claude read-only. Planning exposes only
Read,Glob, andGrep; review exposes no tools. - Do not delegate destructive commands, dependency installation, network access, or secret handling.
- Do not stage, stash, commit, or revert user changes solely for collaboration.
- Keep prompts scoped. Prefer summaries plus relevant diffs over dumping the repository.
- Treat an empty or failed Claude response as a recoverable consultation failure.
Resources
scripts/claude-helpers.sh: deterministic wrapper for non-interactive Claude consultation.- prompts.md: planning and review prompt templates.
レビュー
まだレビューはありません。使ってみた感想をお寄せください。
同じリポジトリのスキル
概要と使いどころ
This skill should be used when the user wants to "stress-test a design", "challenge an idea", "red team a proposal", "get critical feedback on a design", "validate a design", "adversarial architecture review", "risk assessment", "設計を検証して", "反論をもらいたい", "設計を批判的にレビュー", "デビルズアドボケート", "ストレステスト", "弱点を指摘して", "穴を見つけて", "この設計で良いか", "リスク評価", "設計のアーキテクチャレビュー", or mentions structured adversarial review of a proposal or design. NOTE: Use this for validating PROPOSALS and DESIGNS through adversarial debate, NOT for generic code review, PR review, or normal review requests (use /codex:review or /codex:adversarial-review for those). Also NOT for investigating unknown bugs (use /agent-dialectics:strong-inference for that).
This skill should be used when the user wants to "test a claim against data", "validate an empirical hypothesis", "refine a model/theory with evidence", "characterize a pattern and verify it", "演繹と帰納で検証", "仮説をデータで検証して更新", "実例と反例で確かめる", "弁証法ループ", "主張を現物で裏取り", "傾向分析を検証して精緻化", or mentions iteratively updating a hypothesis by deriving predictions and testing them against a real corpus. NOTE: Use this for VALIDATING/REFINING an empirical claim or model against real data through a derive→test→update loop. Use strong-inference for debugging an UNKNOWN cause, and devils-advocate for adversarially stress-testing a DESIGN proposal.
This skill should be used when the user wants to "investigate a bug", "debug an issue", "figure out why something is happening", "find the root cause", "troubleshoot", "強い推論で調査", "仮説を立てて検証", "原因を特定", "バグの原因調査", "なぜ動かないか調べて", "問題を切り分け", "原因不明", "デバッグ", or mentions systematic hypothesis-driven debugging. NOTE: Use this for investigating UNKNOWN causes, not for validating design proposals (use devils-advocate for that).