claude-code

# Claude Code Use this skill when the user wants Codex to help them work with Claude Code. ## Workflow 1. Clarify the user goal only when the target repo, command, or desired handoff is ambiguous. 2. If local setup matters, inspect the environment before giving advice: - Check whether `claude` is available with `claude --version`. - If this plugin's MCP tools are available, call `claude_code_status` first. - Check the current repository state with `git status --short` when operating inside a repo. - Read project docs such as `README.md`, `CLAUDE.md`, `AGENTS.md`, or package scripts before suggesting commands. 3. When preparing a handoff prompt for Claude Code, include: - The concrete objective. - Relevant files or directories. - Constraints from the user. - Verification steps to run after edits. 4. When troubleshooting, separate setup problems from project problems: - Setup: missing CLI, authentication, shell path, permissions, or network access. - Project: failing install, failing tests, missing env vars, or unclear repo instructions. 5. Prefer precise terminal commands the user can run directly, and explain what each command proves. ## Account Access This plugin never stores Claude credentials. It accesses a Claude Code account only through the local `claude` CLI after the user has authenticated with Anthropic. Use this setup sequence: 1. Check whether the CLI is installed with `claude_code_status` or `claude --version`. 2. If it is missing on Windows, install it with one of Anthropic's supported commands: - `irm https://claude.ai/install.ps1 | iex` - `npm install -g @anthropic-ai/claude-code` 3. Call `claude_code_login` to open the official browser login and complete the confirmation. 4. Re-run `claude_code_status`. 5. Use `claude_code_prompt` only after the CLI is installed and authenticated. ## Verification Runs `claude_code_prompt` runs only in the current Codex task workspace, always starts Claude Code with `--safe-mode`, and defaults to `manual` permissions. This prevents repository configuration, hooks, plugins, and MCP configuration from being loaded implicitly. Use `manual` when each action should ask for approval, or `plan` when the task must remain read-only. Autonomous modes, including `dontAsk`, `auto`, and `acceptEdits`, are intentionally unavailable through this plugin. Do not use bypass-permissions modes from this plugin. To switch accounts or refresh auth, call `claude_code_login`, or run `claude` and use `/login` or `/logout` in the Claude Code session. ## Common Workflows ### Review or second opinion Use `claude_code_prompt` with `permissionMode: "plan"`. Include the review target, relevant files, constraints, and expected verification. Ask Claude Code to report findings before suggestions. ### Planned implementation Use `plan` when the user wants analysis without edits. Use `manual` only when the user explicitly wants an interactive implementation where every action still requires approval. ### Setup and repair Run `claude_code_status`, use `claude_code_login` only when authentication needs attention, then run status again. Keep installation problems separate from project-specific failures. ## Handoff Prompt Template ```text Goal: <one concrete outcome> Context: - Repo: <path or project name> - Relevant files: <files/directories> - Constraints: <user constraints> Task: <specific implementation or investigation request> Verification: - <command or manual check> - <command or manual check> ``` ## Guardrails - Do not claim Claude Code has been installed or authenticated unless the local command confirms it. - Do not invent Claude Code flags. If exact CLI syntax matters and the installed CLI help is unavailable, say what needs checking. - Keep handoff prompts compact enough to paste into a terminal session.