GitHub Repo
Hermes Agent v0.21.7 switches source installs to release tags and fixes reasoning-content continuation
The latest Hermes Agent patch changes the default update target for official source installs to stable release tags, making deployed versions easier to track. It also fixes session replay for signed reasoning content, though switching update channels still has version rollback details to watch.
Nous Research released Hermes Agent v0.21.7. The change with the greatest deployment impact is that official source installs now default to tracking stable releases. The release notes are dated October 11, while the GitHub page shows it was published on October 12, within this review window. The patch also includes a fix for continuing reasoning content. Release notes
Previously, updates followed commits on main. Now, hermes update and the desktop update button for source installs select the latest published vX.Y.Z stable release and check out its exact commit, excluding drafts, previews, and canary releases. This makes it easier for engineering teams to keep multiple agent hosts on the same code version and trace behavior changes by release tag. It improves deployment reproducibility, though stability still needs to be validated against each workload. Update documentation
There is an important distinction during migration: installs that have never selected a channel and are already ahead of the latest stable release will keep their current commit and wait for a newer release to catch up. However, explicitly running hermes update --set-channel stable pins the install to the stable release commit, even if that commit is older. Those who need to keep testing the development branch can save that choice with --set-channel main, or override the channel for a single update with --channel main. Forks and mirrors continue to track main, while Store, MSIX, and packaged desktop builds keep their own update mechanisms. Channel rules
Another fix concerns the model protocol: the release notes say that signed thinking content can now be replayed when resuming a session in the CLI and when the gateway starts the next turn of a conversation. Bedrock Converse reasoning models are no longer rejected during continuation. Hermes already stores messages, tool calls, and results in SQLite, so preserving a conversation and correctly reconstructing a model request are separate steps that need to be validated independently. Patch details, session documentation
Deployers should next test recovery of existing long-running tasks and tool round trips, and record the commits before and after updating. Stable tags provide a clear comparison point, but this announcement includes no measurements of continuation success rates or performance. A complete change summary is also slated for v0.22.0.