Paying for agent runs with a subscription token trades a metered bill for a silent cron

A prepaid API balance and a rotating subscription token both stop a daily job, but only one of them is forecastable. Metered credit drains against usage you can read off your own account, and when it empties, the first Claude call in the run is where the job dies ^[2][3]. The replacement credential comes from `claude setup-token`, which mints a short-lived token intended for a single CLI session rather than a service account ^[5], and it stops working when the provider rotates it ^[6]. Two of the three things that needed hardening after this pipeline went live were about how it pays for Claude ^[1][20]. The second was the worse of the two, because nothing about it produced output.

The part of the fix that does the actual work is not the issue itself. It is that GitHub's default notification settings turn a newly opened issue into an email ^[9], which is the output channel a cron job otherwise lacks. That required adding `issues: write` to a permissions block that previously carried only `contents: write` and `id-token: write` ^[10]. Cheap, and easy to omit until the day it matters.

The alert step fires on any step failure in the job, not just the auth step ^[11], which reads like blanket coverage. It is not. The Self-Refine critique pass runs with `continue-on-error: true`, so a critique call that fails or times out publishes the un-refined first draft rather than blocking the run ^[12][13]. A step that cannot fail the job also cannot satisfy a condition that tests for job failure, so that particular degradation arrives with no issue and no email ^[14].

Which puts the weight on the one gate that makes no API call at all. `slop-check.mjs` strips frontmatter, code fences and markdown, splits the remainder into sentences, and scores three signals ^[15]. The headline signal is burstiness, the standard deviation of sentence length over the mean, annotated in the script as roughly 0.6 to 1.2 for human writing and under 0.4 for machine slop ^[16]. The hard floor it enforces is 0.32 ^[17]. That is 80 percent of the script's own slop threshold and 53 percent of its human floor, so any draft scoring between 0.32 and 0.40 passes a gate whose own comment classifies it as slop ^[19]. The em dash counter and the banned-phrase list catch different things ^[18], but the flagship metric ships with an 0.08-wide band of self-declared failure.

None of this argues for going back to per-token billing. It argues that the credential you chose determines which alarm you owe. Metered credit fails on a schedule you control and can watch in a balance; a session token fails on someone else's, at which point the only thing standing between you and four quiet mornings is a `failure()` block you wrote before you needed it.