Codex Claude Pm

Codex Claude PM

Use this skill to make Codex the planner, reviewer, and final owner while Claude Code performs the implementation work. Treat the workflow as an engineering ticket system, not an open-ended chat.

Core Rule

Codex owns correctness. Claude Code may write code, but Codex must still inspect scope, compare against acceptance criteria, and independently verify the result before accepting it.

Workflow

1. Orient in the repository.

• Check the current working directory, git status, relevant files, package scripts, and existing tests.
• Prefer rg and targeted file reads. Avoid broad rewrites before understanding local patterns.
• If the worktree is dirty, note user changes and avoid overwriting unrelated edits.

1. Decide whether to dispatch.

• Dispatch to Claude Code when the request is mostly implementation and can be specified with concrete files, constraints, and acceptance criteria.
• Keep Codex-local work when the task is mainly product judgment, architecture choice, security review, small surgical edit, or high-risk secret/credential handling.
• Split work before dispatch if it touches more than 3 unrelated modules, exceeds about 300 expected changed lines, or has separable dependencies.
• Read references/dependencies.md when the task depends on external tools, credentials, project services, generated assets, or multi-step task ordering.

1. Write a Task Package.

• Use references/task-package-template.md.
• List precise files and boundaries.
• Convert vague intent into numbered executable requirements.
• Define mechanical acceptance criteria and exact test commands.
• Include a required Completion Report format from references/completion-report-template.md.

1. Quality-check the Task Package before sending.

• Read references/review-checklist.md when the task has meaningful risk.
• Ensure Claude Code can act without follow-up questions.
• Prefer smaller, verifiable tasks over one broad prompt.
• Include a dependency map when execution order, tooling, runtime services, or credentials affect success.

1. Dispatch to Claude Code.

• If the local CLI is available, use scripts/Invoke-ClaudeDispatch.ps1.
• If the CLI is unavailable or blocked by login, output the Task Package for manual paste into Claude Code.
• Prefer non-interactive Claude Code mode: claude -p.
• Use foreground for small bounded tasks; use -Background for long-running implementation, broad investigation, or unclear runtime.
• Preserve Claude Code stdout as the transport result. Do not rewrite the Completion Report before Codex review.

1. Verify the result.

• Read Claude Code's Completion Report.
• For background work, use scripts/Get-ClaudeDispatchRun.ps1 to check status and fetch the result.
• Inspect git diff --stat and targeted git diff for modified files.
• Run the requested tests yourself when feasible.
• Check that all acceptance criteria are actually satisfied; do not trust PASS labels alone.

1. Decide.

• ACCEPT only when requirements, constraints, and tests are satisfied or deviations are explicitly acceptable.
• Issue a Change Request when a fix is needed. Use references/change-request-template.md.
• If Claude Code exposes unrelated failures, record them separately instead of silently expanding scope.
• Read references/evaluation.md when judging whether this skill itself worked well on the task.

CLI Dispatch

Use the bundled script from the repository root or target project root:

powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Invoke-ClaudeDispatch.ps1 `
  -TaskPackagePath .\.codex-claude-pm\tasks\TASK-APP-001.md `
  -WorkingDirectory . `
  -PermissionMode acceptEdits `
  -OutputFormat text

Useful modes:

# Preview command and write logs without calling Claude Code
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Invoke-ClaudeDispatch.ps1 `
  -TaskPackagePath .\.codex-claude-pm\tasks\TASK-APP-001.md `
  -DryRun

# More autonomous local execution in a trusted sandbox or disposable worktree
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Invoke-ClaudeDispatch.ps1 `
  -TaskPackagePath .\.codex-claude-pm\tasks\TASK-APP-001.md `
  -PermissionMode auto

# Constrain Claude Code to specific tools
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Invoke-ClaudeDispatch.ps1 `
  -TaskPackagePath .\.codex-claude-pm\tasks\TASK-APP-001.md `
  -AllowedTools "Read","Edit","Bash(npm test*)","Bash(git diff*)"

The script writes each run under .codex-claude-pm/runs// with the task package, command metadata, stdout, and stderr. Each run now also writes status.json and has a stable Job ID.

Background execution:

powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Invoke-ClaudeDispatch.ps1 `
  -TaskPackagePath .\.codex-claude-pm\tasks\TASK-APP-001.md `
  -WorkingDirectory . `
  -Background

Inspect runs:

# List recent runs
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Get-ClaudeDispatchRun.ps1 -WorkingDirectory .

# Show one run result
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Get-ClaudeDispatchRun.ps1 `
  -WorkingDirectory . `
  -JobId cc-20260706-120000-abc123 `
  -Result

# Best-effort cancellation of a background run
powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Stop-ClaudeDispatchRun.ps1 `
  -WorkingDirectory . `
  -JobId cc-20260706-120000-abc123

This mirrors the useful part of openai/codex-plugin-cc: setup/status/result/cancel are first-class operations, while Codex remains responsible for the final review.

Dependency Check

Before the first real dispatch in a project, run:

powershell -ExecutionPolicy Bypass -File .\codex-claude-pm\scripts\Test-ClaudeDispatchPrereqs.ps1 `
  -WorkingDirectory .

This checks whether claude is available, whether the working directory and git state are readable, and whether non-interactive Claude Code can be invoked for basic CLI metadata.

Accuracy Practices

• Make every requirement observable: name the file, function, UI state, API response, or test behavior.
• Put "do not change" constraints in the Task Package, especially public APIs, unrelated files, migrations, lockfiles, generated files, and user data.
• Include negative cases and boundary cases in acceptance criteria.
• Require Claude Code to report exact files changed and test outputs.
• Run independent verification from Codex after Claude Code returns.
• Prefer acceptEdits or auto over bypassPermissions unless the project is isolated and trusted.
• Use a fresh branch or worktree for large tasks.
• If a completion report says tests failed because of pre-existing issues, verify that claim with targeted commands before accepting.

Evaluation

Use references/evaluation.md to score the skill after real tasks. A good run produces a specific Task Package, a bounded diff, a structured Completion Report, independently verified tests, and either a clear ACCEPT or focused Change Request.

Read references/codex-plugin-cc-study.md when changing the dispatcher architecture.

Manual Handoff Format

When automatic CLI dispatch is not appropriate, respond with:

1. The Task Package to paste into Claude Code.
2. The exact target working directory.
3. The expected Completion Report format.
4. The verification commands Codex will run after the report returns.

References

• references/task-package-template.md - canonical Task Package format.
• references/completion-report-template.md - required Claude Code response format.
• references/change-request-template.md - format for follow-up fixes.
• references/review-checklist.md - pre-dispatch and post-dispatch quality gates.
• references/dependencies.md - dependency taxonomy and dependency map format.
• references/evaluation.md - how to judge whether the skill is useful and accurate.
• references/codex-plugin-cc-study.md - implementation lessons adopted from openai/codex-plugin-cc.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ZhouYinLong-lab
• Source: ZhouYinLong-lab/Dog-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.