A heartbeat automation polls. It wakes on a fixed recurrence and asks GitHub what is true now [2]. RuntimeWire's read of the default prompt has Codex inspect the latest PR state with `gh`, including mergeability and current checks, and re-check live GitHub state rather than relying on earlier turns [4]. That instruction is the good engineering here. Long-running agents fail by acting on a snapshot they took half an hour ago, and telling the child task to distrust its own history is cheaper than keeping a cache coherent with a repo two other people are pushing to.
Polling has a price in latency. A 10-minute recurrence gives six opportunities an hour and up to 144 a day while the task stays unblocked [1]. The prompt already anticipates pipelines slower than the interval, telling Codex to revisit pending checks on the next heartbeat [4]. On a slow pipeline, then, you are paying for turns that mostly observe.
The judgment call sits in one phrase: repair failing checks caused by the PR [4]. Attribution is the hard part of CI triage. A flaky integration test and a real regression look the same in the logs and annotations the prompt tells Codex to read before editing [4]. Get the attribution wrong and the loop edits code to satisfy a failure the code did not cause, then commits and pushes it to the PR branch [4].
Then the gate. Mutations happen inside an isolated git worktree [4], and with auto-merge enabled the embedded instructions merge once required checks pass and the PR is mergeable, and on a failed merge diagnose, update the branch, retry, and continue until the PR is merged or closed [5]. Read that as a state machine. Every internal transition is retry. The only stop condition outside the agent is whatever your repository demands before a merge is allowed. If that is required status checks and nothing else, required status checks are now your review policy in full, and a stored preference called `git-pr-watch-auto-merge` decides when it applies [7].
RuntimeWire frames the target as the costly last mile of keeping a PR green and mergeable while CI and the base branch keep changing [11]. That is a fair description of the repair half. The merge half automates the step where a person used to be in the path, which is a policy decision that lives in your branch protection rules, not in a desktop settings pane.
On evidence: this is static analysis of a shipped package. RuntimeWire extracted the Electron and bundled frontend code, traced the user-facing settings into stored configuration keys, followed the automatic-fix module through automation creation, heartbeat scheduling, worktree and push behaviour and optional merge logic, and published a sha256 for `app.asar` [1][7][10]. It rates the finding partially reproduced and says it has not yet run the feature end-to-end against a live GitHub pull request [9]. The prompt text is therefore well attested; the runtime behaviour of the loop against real checks is not yet observed by anyone outside OpenAI. RuntimeWire presents the work as original reporting based on reverse engineering, and says its method included searching OpenAI's public product documentation, release material and current Codex descriptions for the same feature [8].
