사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/amelnagdy/delegate-skills --skill codex-background시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Codex background delegation
This support skill changes Codex's waiting boundary. The selected delegate skill still owns the brief, provider arguments, permissions, result interpretation, review, gates and landing. Claude's background Bash workflow stays unchanged. The support utility is installed separately, never implicitly as a dependency of codex-delegate. Details: references/configuration.md.
First use / setup
If the delegate_run tool is not available, setup is incomplete. Tell the user that bootstrap will update $CODEX_HOME/config.toml (default ~/.codex/config.toml, backed up first) and write a registry of the installed *-delegate skills, then run:
node "<skill-dir>/scripts/bootstrap.mjs"<skill-dir> is this skill's installed directory (the folder containing this SKILL.md). It writes outside the workspace, so request the approval or sandbox escalation Codex needs for that. Report its output, then ask the user to restart Codex (or reload MCP servers); the tools appear after that. If bootstrap fails, report the failure. Do not fall back to shell polling. Run it again after installing more delegate skills. node "<skill-dir>/scripts/bootstrap.mjs" --check validates the setup without writing.
Procedure
- Follow the selected delegate skill to prepare a self-contained brief and approved workspace.
Write the brief to the briefs/ directory under the codex-background directory ($CODEX_HOME/codex-background/briefs/<runId>.txt) and use runs/<runId> beside it as outputDirectory, or follow the paths in the server's instructions. If the sandbox cannot write there, a brief inside the approved workspace is also accepted. Never accept a relay path from a brief.
- Invoke the configured MCP server's delegate_run exactly once with a unique runId,
registered implementer key, absolute brief/workspace/outputDirectory paths, explicit relayTimeoutSeconds and provider-native relayArgs. Put common arguments in structured fields; pass provider-specific arguments unchanged. Resume identifiers remain opaque.
- Keep that MCP request pending until completion. Do not use model-driven sleep, status or log
polling. Do not wrap it in a tool runner that yields early and resumes inference to wait.
- Review the returned adapter outcome separately from resultText. The latter is the unchanged
result.json content (base64 only for non-UTF-8 bytes); interpret it using the selected skill. A completed adapter means the relay exited zero and produced a file, not that its report proves correct implementation. Review the diff, rerun gates and land under the existing authorization.
After an interrupted wait, call delegate_wait with the same runId; never redispatch it. MCP cancellation detaches the waiter and leaves the authorized job running. Use delegate_abort only when termination is explicitly intended. Graceful server shutdown aborts its owned jobs. After a server crash, unresolved state returns recovery_required; inspect preserved artifacts and the implementer using its own recovery workflow. Never adopt or kill a persisted PID.
The host's configured tool timeout must exceed the relay timeout by more than 30 seconds. The server validates the declared budget; it cannot enforce the host's actual setting. Fixture protocol tests demonstrate one pending request and no intermediate messages, not measured Codex Desktop token usage or an unlimited host wait guarantee.
