Codex Process Jobs v0.1.0
Initial public beta of Codex Process Jobs: a dependency-free Codex plugin for running ordinary macOS and Linux commands as durable detached process jobs instead of holding an agent turn open.
Highlights
- Durable detached execution with process identity, bounded stdout/stderr, terminal status, and safe process-group cancellation.
- Namespaced start, status, tail, result, and cancel skills for CMake builds, long tests, inference runs, data processing, and repair utilities.
- Best-effort conversational completion across Codex App, CLI, VS Code, and remote/mobile-driven hosts, backed by durable status and result retrieval.
- Consent-gated
PostToolUse,Stop, andUserPromptSubmithooks; the installer never trusts hooks automatically. - Critical-job safeguards, explicit Goal integration, completion batching, proactive bounded result inspection, and preservation of prior cache generations for open tasks.
- Two-phase installer with an explicit choice of global, project, or no
AGENTS.mdadoption policy. - 148 tests plus a macOS/Ubuntu, Node.js 18/22 CI matrix.
Install with Codex
Tell Codex:
Install
joelfarthing/codex-process-jobsfrom GitHub. Run the installer's read-only preview first and describe every local change. Then ask whether I want the optionalAGENTS.mdpolicy globally, in one project, or not at all. Do not apply the installation until I approve the preview and policy scope.
After installation, restart the Codex client and explicitly review and approve the plugin's hook hashes in /hooks.
Important limitations
- This is an independent community beta, not an official OpenAI plugin.
- Automatic conversational wake uses experimental local Codex transports and is best-effort; durable status and result retrieval remain authoritative.
- Commands receive no interactive stdin and must remain finite foreground processes. Servers, watchers, daemonized commands, and external fire-and-exit launchers need another lifecycle mechanism.
- Windows is not currently supported.
- CPJ primarily improves workflow quality by releasing the conversation. Token usage can be lower, neutral, or higher depending on the foreground baseline and result-inspection behavior; no universal savings percentage is claimed.
See the README, full changelog, and security model for details.