Skip to main content
中文
Back to Codex
TOOL TUTORIAL

09 | Codex Automations: Background Tasks without Unattended Control

GlobalCoreHub Editorial Team2 min read

An Automation is a background colleague, not autopilot

The upstream guide uses this distinction deliberately. Automations can run on a schedule, heartbeat, or thread follow-up, but background work still needs ownership. The right candidates are periodic tasks with clear boundaries and output a person can triage, such as documentation-drift checks, PR-comment summaries, or read-only status inspection.

Start with read-only automation. Its output should make the next human decision obvious.

Three trigger types

  • Cron runs on a fixed schedule or interval.
  • Heartbeat checks a state on a continuing rhythm.
  • Thread follow-up triggers work related to a particular thread’s next action.

Choose based on the real signal. A follow-up to a changed PR does not require a broad periodic job; a weekly documentation check does not need to run constantly.

Write durable automation prompts

The course asks for a goal, source, read/write scope, stop condition, output format, and escalation condition:

Each week, check whether documentation has drifted from the current API definition.
Default to read-only. Report differences, source evidence, impact, and the smallest next step.
If code, external systems, or sensitive configuration would need changing, stop and request human confirmation.

This turns output into triage: ignore it, request a decision, create a small task, or treat it as urgent.

Permissions, Skills, and worktrees

An Automation can call a stable Skill so it does not rediscover the process on every run, but it needs narrower permissions than interactive work. For writing work, the upstream guide suggests considering a worktree for isolation and retaining review, tests, and merge as explicit human steps.

Automations are also different from GitHub Actions or CI. CI is part of a build/check pipeline; an Automation is a background helper around workflow analysis, follow-up, or limited action. Both need logs, recovery paths, and a way to turn them off.