Hermes Azure AI Foundry Encrypted-Reasoning Replay Can Break Follow-Up Conversations (#63257)
Bug Tracker — July 12, 2026. A reported provider-integration defect affects conversation continuity when encrypted reasoning history is replayed to Azure AI Foundry through the Responses API.
Issue Summary
Issue #63257 reports that Hermes strips the id from stored encrypted-reasoning items during Responses API history replay. Azure AI Foundry requires that identifier on subsequent turns, so the service returns HTTP 400 invalid_payload. The reported result is that Azure GPT-5.x conversations can become effectively single-turn, while resumed sessions and deployment switches can also fail.
Observed Failure Path
| Step | Reported Behavior | User-Level Consequence |
|---|---|---|
| Initial turn | Encrypted reasoning item is stored in conversation history. | Session appears to begin normally. |
| History replay | Hermes removes the item ID while reconstructing Responses API payload. | Replay no longer conforms to Azure AI Foundry expectations. |
| Follow-up / resume | Azure rejects payload with invalid_payload. | Conversation continuity is lost; retries and provider/deployment changes are disrupted. |
Operational Impact
Classification: P2. This page summarizes a publicly reported issue; behavior, scope, and remediation status may change as the upstream project investigates and merges fixes. Teams should validate their own deployment before treating any workaround as complete.
Mitigation and Verification
- Before relying on Azure GPT-5.x multi-turn workflows, test fresh, resumed, and provider-switch sessions with encrypted-reasoning history enabled.
- Capture the outbound replay payload only in a redacted, authorized test environment and verify required reasoning-item identifiers survive serialization.
- Track the linked upstream issue for a fix; after updating, rerun continuity tests instead of assuming a successful first turn proves recovery.
- For production workflows, preserve a bounded restart/recovery path so a rejected replay does not silently discard task state.
Source Evidence
Primary issue: NousResearch/hermes-agent issue #63257
For a managed-agent comparison, the relevant question is whether provider replay, error handling, token growth controls, and session recovery are observable and bounded under failure—not whether a one-off request completes in the happy path.