ChatGPT tasks can now wake up when a pull request changes
Maestro Brief · Published by Maestro Mojo
2026-08-26
Maestro’s take
ChatGPT scheduled tasks just learned to listen.
That matters more than another automation screen. A task can now react when a pull request changes, a Slack channel gets a message, or an email arrives. The agent does not have to check every 15 minutes and hope something happened.
For developers, this is useful today as a triage loop. It is not yet a local coding-agent loop.
TL;DR
OpenAI says eligible ChatGPT plans can trigger scheduled tasks from supported GitHub, Slack, and Gmail events.
GitHub triggers can watch pull-request reviews, comments, commit updates, or merges. Slack and Gmail triggers can watch selected messages. Several matching events may be grouped into one run.
The catch is important. Event-triggered tasks run in ChatGPT on the web or mobile. They are not available in the ChatGPT desktop app, Codex CLI, or the IDE extension. Web tasks also cannot work directly in a folder on your computer.
Use them to read, summarize, classify, and prepare the next move. Do not treat them as a silent replacement for your local build-and-test loop.
Why Maestro users should care
A timer asks, “Is there work yet?” An event says, “The work changed.”
That removes pointless polling. It can also make an agent feel much more responsive.
A useful first workflow is simple:
- A reviewer comments on a pull request.
- ChatGPT groups the feedback.
- It separates requested changes from questions.
- It drafts a short action plan with links to the relevant review activity.
- A person decides what happens next.
The agent wakes up at the right moment. The human still owns the decision.
Do this
Do: Start with a read-only PR-review assistant.
Ask it to summarize new review comments, identify blockers, and prepare a checklist. Keep the output in Scheduled for review.
Why: You can test whether the trigger and instructions are useful without giving the task permission to change code or communicate on your behalf.
Do not do this
Do not: Let one unattended task interpret a comment, change code, approve the result, and merge it.
Why: Event triggers improve timing. They do not prove the agent understood the request, ran the right tests, or respected every trust boundary.
OpenAI also says event-triggered tasks cannot combine an app event with a time schedule. If you need both, use separate tasks. Connected apps must be authorized, and the GitHub app must be allowed to access the repository.
One thing to try
Create one task for a low-risk repository:
When a new pull-request review or comment arrives, summarize the requested changes. Separate blockers, questions, and optional suggestions. Link to the relevant activity. Do not change code, post comments, approve, merge, or close anything.
Review the first few runs. Tighten the prompt if it invents urgency or misses context.
The simple rule: use events to wake the agent; keep a human at the gate.
Published August 26, 2026. Maestro’s opinions and summary are AI-generated.