원클릭으로 Manus에서 모든 스킬 실행

claude-collab

스타3

포크0

업데이트2026년 6월 11일 15:01

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.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

masuP9

masuP9/codex-collab

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

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.sh from the repository root or ${CODEX_HOME:-$HOME/.codex}/skills/claude-collab/scripts/claude-helpers.sh.
If claude is 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)/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 -U3 when the diff is at most 600 changed lines and 48KB.
Treat 300 changed lines or 20KB as 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 600 changed lines or 48KB, 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 load CLAUDE.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, and Grep; 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.

이 저장소의 다른 Skills

같은 저장소

codex-collab

masuP9/codex-collab

This skill should be used when the user asks to "collaborate with Codex", "use Codex for planning", "get Codex review", "delegate to Codex", "code review", "PR review", "review this code", "review this PR", "Codexと協調", "Codexにレビュー", "Codexに計画を作成させたい", "Codexに任せる", "Codexに委任", "Codexと連携", "Codexに相談", "Codexの意見", "Claude plans", "Claude-led", "Claudeが計画", "Claude主導", "Codexに実装させる", "Codexに実装を任せる", "Claudeがレビュー", "コードレビューして", "PRをレビューして", "コードをレビューして", "このコードをレビュー", or mentions coordinating tasks between Claude Code and Codex CLI. NOTE: This is the default skill for generic code/PR review requests. Use devils-advocate only when the user explicitly asks for adversarial/red-team critique of a design or proposal.

2026-07-123

contradiction-lift

masuP9/codex-collab

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.

2026-06-213

dialectic-loop

masuP9/codex-collab

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.

2026-06-203

collaborative-planning

masuP9/codex-collab

This skill should be used when the user wants to "plan with Codex", "create a plan", "draft implementation plan", "help me plan", "plan before implementing", "refine a plan", "review my plan idea", "Codexと計画を立てる", "実装計画を作成", "計画を立てたい", "計画を一緒に考えて", "実装方針を相談", "Codexと設計を練る", "計画だけ作りたい", "実装前に計画を立てたい", "計画をレビューしてほしい", or mentions creating/refining an implementation plan as the deliverable without proceeding to implementation. NOTE: This is for PLAN-ONLY workflows that produce a plan document as the deliverable. Do NOT use when the user wants implementation (use codex-collab instead). Do NOT use for adversarial design critique (use devils-advocate instead). Do NOT use for debugging (use strong-inference instead).

2026-03-113

devils-advocate

masuP9/codex-collab

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-collab for those). Also NOT for investigating unknown bugs (use strong-inference for that).

2026-03-113

strong-inference

masuP9/codex-collab

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).

2026-02-083

name

claude-collab

description

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.sh from the repository root or ${CODEX_HOME:-$HOME/.codex}/skills/claude-collab/scripts/claude-helpers.sh.
If claude is 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)/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"

Prefer one complete unified diff so Claude can catch cross-file bugs:

Send the complete git diff --find-renames -U3 when the diff is at most 600 changed lines and 48KB.
Treat 300 changed lines or 20KB as 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 600 changed lines or 48KB, 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 load CLAUDE.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, and Grep; 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.