Skip to content
vavio

What happens to my scheduled Hermes jobs when my model plan hits its usage limit?

When your model plan or key hits its usage limit, a scheduled Hermes job fails unless Hermes can continue on another key for the same provider or, for a job without its own model pin, on a fallback provider you have configured. The schedule keeps going: the job is tried again at its next run. Hermes alerts you once for a repeated error rather than on every run. For OpenAI Codex, which states how long its limit lasts, Hermes holds the job until the limit resets.

This follows Hermes Agent v0.23.1, commit 3b42157, checked on 16 November 2026. We reviewed the official documentation and source; we have not exhausted a plan or watched a held job on a Vavio computer. The commands below are upstream Hermes commands, not Vavio CLI commands.

What does Hermes try before the run fails?

Hermes tries these in order during the run:

Step When it applies What it changes
Retry the same key A brief rate limit without a stated plan limit (once, when you have a pool) Nothing; the same key is tried again.
Next key in a credential pool You have added more than one key or sign-in for the same provider Usage moves to the next key’s account. A plan limit such as ChatGPT’s “usage limit reached” rotates immediately, without a retry.
Fallback provider Every key is exhausted, the job has no model pin of its own, and you have set fallback_providers The job continues on a different provider and model, billed by that provider.
Fail the run None of the above can continue The run is recorded as failed with its error.

The model pin matters most. A job with its own provider, model or endpoint never falls back to your global chain: if that route fails, the run fails. Pool rotation for the same provider still applies to it. To keep fallback for a job, leave it unpinned and set its model through cron.model instead. Can a scheduled Hermes job use a cheaper model covers pinning.

A fallback keeps the job alive, but on a different model. Its output and its bill can differ from what you expected, and the first request on the new model gets no prompt-cache discount.

Does Hermes keep retrying while the limit lasts?

For most providers, no special wait applies. A failed recurring job runs again at its next scheduled time, and if the limit is still in force, that run fails too. An hourly job against a plan that resets in five hours can therefore fail about five times. Each attempt may still reach the provider.

There is one exception in the pinned version. When the OpenAI Codex provider reports that its quota is exhausted with a “retry after” time, and no fallback is available, Hermes holds the job. Usually the first run to hit the limit fails normally; the hold starts on the next run, when Hermes sees the limit before calling the model:

  • The one failure alert says the provider’s window is closed and that the job is held until a stated time.
  • No run fires and no alert is sent before that time. The held time is stored on the job as quota_hold_until.
  • Missed runs inside the window are combined. A job on a cron expression resumes at its first scheduled time after the window; an interval job resumes when the window ends.
  • For a scheduled job on a cron expression, if the window ends well before its next scheduled time (at least half its period earlier), the blocked run is retried once when the window ends. Manual runs are not retried this way.
  • One-shot jobs and paused jobs are not held. Any later run that reaches the model clears the hold.

The source starts a hold only from that Codex error. A 429 that a model API returns partway through a run is not held this way; it is retried on the job’s normal schedule.

Network and DNS failures before any model call are a different case. Hermes re-runs those after 5, 15 and 30 minutes, as described in why a job ran but no result arrived.

How will I know a job is failing on its limit?

A recurring job that keeps failing with the same error alerts you once. Hermes records the failure as an incident and withholds further alerts for that error. While the job keeps failing, it reminds you again every six hours by default (cron.failure_repeat_alert_hours). A different error alerts at once, and a successful run resolves the incident.

Every failed run is still recorded, so the job’s record shows the limit even when your chat is quiet. If you can reach the Hermes command line:

hermes cron list
hermes cron runs <job_id> --limit 20
hermes cron incidents --state alerted

Replace <job_id> with the job’s ID. The list shows the job’s last error, and the run history shows each failed attempt. Acknowledging an incident with hermes cron incidents ack <id> silences that exact error for good, including reminders, so use it only once you have decided to wait out the limit.

You can also ask Hermes directly:

List my scheduled jobs that failed today and show each one’s last error. Tell me which failures mention a rate limit, quota or usage limit, and whether any job is held until a provider reset.

What can I change so jobs keep running?

  • Add a fallback provider with hermes fallback add, and keep jobs that should use it unpinned.
  • Add another key or sign-in for the same provider to its credential pool, if your provider’s terms allow it.
  • Move frequent jobs to a cheaper or separate model with cron.model, so scheduled work does not use up the plan your chats rely on.
  • Run less often while the limit is tight. A job that only needs a check can use a script that skips the model when nothing has changed.

Each option moves usage to an account you pay for. Check that provider’s current price and terms before relying on it for unattended work.

Vavio supplies a managed remote Hermes computer for $14.99 per month, with model usage billed separately by your model provider. Hosting keeps the Hermes runtime off your laptop; your model plan’s limits still apply to every job it runs. See the current hosting offer, or apply the same settings to a Hermes server you maintain yourself.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source. This is a source review, not a live usage-limit trial.

More from the Vavio blog: Hermes Agent guides