Hermes: retry a single step without re-running the whole workflow

A nine-step workflow that fails on step seven used to cost you steps one through six again. Now it costs you step seven.

Long workflows fail late. Until this build, recovering meant restarting from the top, which on anything involving a slow export turned a small correction into a twenty-minute wait — and re-ran side effects that had already happened.

Runs now checkpoint after each step. A failed or unsatisfactory step can be retried on its own, optionally with an edited input, and the steps after it resume from there. Steps already completed are not re-executed.

It is the change preview users asked for most, and the one that made the longer workflows worth writing at all.

What changed

• Per-step checkpointing on every run.
• Retry one step, optionally with an edited input.
• Downstream steps resume; completed steps are never re-executed.
• Checkpoints are kept with the run history and expire with it.

The product

• Hermes — read the case