← LeaderboardThe Skill Leaderboard

coding-backend-codex

Runtime prompt layer for Codex coding agent runs.

Published by ahmadrosid ↗ahmadrosid/nakama

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

name: coding-backend-codexdescription: Runtime prompt layer for Codex coding agent runs.disable-model-invocation: true

Use Codex for coding tasks when the Coding Agent Harness context selects it. Follow that context's command template and run it through Nakama's bash tool with codingAgent: true, cwd set to the target project, and a task-appropriate timeoutMs.

Prerequisites

  • Install Codex on the Nakama host if needed: npm install -g @openai/codex.
  • Authenticate with codex login or configure OPENAI_API_KEY. A missing API key does not imply that CLI login is missing.
  • Check codex --version and codex login status when installation or auth is uncertain.

One-shot coding

codex exec is non-interactive and works with Nakama's pipe-based bash tool. Do not request a PTY or use an interactive codex session there.

codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --color never 'Make the requested change, run targeted checks, and summarize the result.'

Use --dangerously-bypass-approvals-and-sandbox only when the operator has authorized full host access, as in this setup. Do not use codex exec --full-auto: Codex rejects that option. Do not use codex exec -- --full-auto either: -- makes --full-auto prompt text, not an option. If a run fails with unexpected argument '--full-auto', correct the command and retry the task.

--skip-git-repo-check permits scratch work outside a git repository; it does not replace setting cwd to the right project for repo changes. Put the task prompt in one shell argument and quote it safely.

Reviews

From the target repository, use codex review --base origin/main for branch changes or codex review --uncommitted for local edits. Keep review work read-only unless the user asked for fixes.

After the run

Check the bash exit code and timeout result. Verify changed files and targeted tests before reporting success. For implementation tasks, follow the coding-agent skill's branch, commit, push, and PR instructions unless the user requested local-only work.