Back Home

GitHub Repo

Hermes Agent Reports Delegation Model Configuration Mismatch; Existing Conversations May Continue Using Parent Agent Model

On November 13, the Hermes Agent community reported that long-running conversations may continue using the parent agent’s model after delegation settings are changed, without issuing a warning. A contributor has proposed a warning fix, but it remains unclear whether it will be included in an official release.

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

On November 13, the Hermes Agent community reported a delegation routing issue: in a long-running Discord gateway conversation, a user changed the profile’s delegation backend to model_map and specified a different set of models and providers, but all four background delegations used the parent agent’s model. The issue remains open, and the project has labeled it a P2 bug with a workaround. Issue report

The key is that the profile and the conversation retain two different states. The official documentation explains that the delegation backend, along with the model-visible tool parameters, is fixed for the conversation; changes made directly to delegation.backend are only meant to take effect in a new conversation. This keeps the tool structure consistent with the actual routing, but also means existing conversations continue using the old backend. Official delegation documentation

The old backend in the report was profile_routes. When a task does not specify a profile, it falls back to the parent agent’s route, so the delegation model and provider set in the profile do not have the intended effect. The issue is therefore one of observability: the settings appear to have changed, but the execution results do not clearly indicate that the old backend is still in use. In workflows that rely on cross-review by different models, this could mean the expected independent review does not happen; cost and quality assessments also need to account for the actual route rather than just the profile settings. The latter points are engineering implications inferred from this mechanism.

A contributor has submitted a fix on a branch that adds logging and a route.warning result field, and corrects the tool description. The proposal preserves the existing behavior of pinning the backend for a conversation rather than switching models automatically. The commit also adds a regression test for a mismatch between the profile and conversation backends, but a branch commit does not mean the fix has been released. Fix commit

Engineering teams can follow the documentation by starting a new conversation with /new, or use /delegation <backend> --now to rebuild the tools for the current conversation on the next turn. After switching, they should still check which model and provider the subagent actually used. Track whether the warning fix reaches the main branch and an official release. Current evidence comes from a specific v0.23.1 environment and a public reproduction; it does not establish that all delegation configurations are affected. Usage instructions

Sources

  1. Hermes Agent issue #154936:對話固定的委派後端使模型設定靜默失效
  2. Subagent Delegation
  3. fix(delegation): expose stale session backend when config changes