Back Home

GitHub Repo

Hermes Agent Reports Replies Not Delivered During Updates as Restart Wait and Client Reconnection Fall Out of Sync

A November 15 report on Hermes Agent found that during an update, the gateway may disconnect clients before waiting for running tasks to finish, leaving saved replies undelivered. A related restart change has been merged, but the update path for timed-out pause requests still needs attention.

Gerald L. Nino · Public domain · Image source
zh-Hant

On November 15, the Hermes Agent community filed an issue report describing how, in a Linux system service deployment, running hermes update while a conversation still had a task in progress could cause the local client to stop displaying output—even if the backend eventually completed the reply normally. The report covers one actual upgrade from v0.23.0 to v0.23.1 and includes gateway logs, update logs, and database results. It does not establish that all deployments are affected. Issue report

The logs show that after receiving a restart request, the gateway decided to wait for the current task to finish, but closed the client WebSocket about a second later. The original turn finished about two minutes after that, and the answer was written to state.db, but the old client did not reconnect successfully. This left three states out of sync: computation completed, the result was persisted, and the answer reached the UI. To the user, it looked as though the model was stuck. Incident timeline

This issue is directly related to Hermes’s service architecture. The official documentation explains that the traditional CLI is now a gateway client that attaches to conversations over a local WebSocket. Closing the terminal only detaches the connection; the gateway continues running work it has already accepted. Engineering therefore needs to protect not only backend tasks, but also ensure that clients can survive the restart wait and retrieve results afterward. CLI architecture documentation

The project had already merged a related change on November 14. It allows the gateway to continue accepting work while idle and changes some wait limits to warnings, preventing a timer from directly terminating a task. However, a follow-up comment on November 16 says that if the update process does not receive a response to its pause request within two seconds, it still stops the service immediately. The merged change therefore should not be taken as evidence that this end-to-end issue is resolved. Merged change, follow-up

The reporter’s workaround is to let the conversation turn that started the update finish as soon as possible, avoiding a prolonged wait caused by continuously polling the update log. This still does not guarantee delivery of the answer. Deployers should watch for fixes to the pause-request timeout, reconnection budget, and result redelivery together. Verification should also check the database completion status separately from whether the client actually received the content. Workaround and limitations

Sources

  1. After the user denies an edit, the agent makes the same change through another tool — Issue #154419
  2. Hermes Agent Security
  3. Hermes Agent 更新期間 WebSocket 中斷與回覆未送達回報 #156018
  4. 閘道等待閒置再重啟的已合併修改 #155532
  5. Hermes Agent CLI Interface:閘道客戶端與工作階段架構