hermes - 💡(How to fix) Fix Automatic context compaction can hide or drop just-completed assistant response [1 pull requests]

Bug description

A completed assistant response disappeared from Hermes Workspace immediately after automatic context compaction triggered. This was observed in a long, tool-heavy session. The Workspace UI may be involved, but the failure is tightly coupled to the agent/session compaction event, so this should be tracked on the Hermes Agent side as well.

Related Workspace issue: https://github.com/outsourc-e/hermes-workspace/issues/505

User-visible impact

The user loses the final deliverable right after a long run completes. This is especially damaging because compaction tends to happen after high-context, high-effort responses.

Observed flow

1. User works in Hermes Workspace on localhost:3000.
2. Assistant completes a long, tool-heavy response.
3. Automatic context compaction triggers immediately after completion.
4. The just-completed assistant response disappears from the visible conversation.
5. User can still see/recover some information from another interface/session, but the Workspace thread lost the answer.

Expected behavior

Automatic compaction should preserve all finalized messages durably and should never remove or overwrite the most recent assistant response.

Backend/session invariants that should hold:

• Final assistant message is persisted before compaction starts.
• Compaction summary/reference message is appended or stored separately, not used as a replacement transcript snapshot.
• Any session refetch after compaction includes the just-finalized assistant response.
• Message IDs remain stable across compaction.
• Stream finalization and compaction cannot race such that the final assistant message is dropped.

Actual behavior

The final assistant message disappeared immediately after auto-compaction.

Suspected areas

• Stream completion vs. compaction race condition
• Session DB/message persistence ordering
• Compaction replacing recent transcript slice with a stale summary snapshot
• Workspace receiving an incomplete post-compaction transcript from Hermes Agent APIs
• Last assistant message not flushed/committed before compaction mutation

Suggested investigation

1. Audit the order of operations when an agent run completes near the compaction threshold.
2. Confirm message persistence happens before auto-compaction mutates session state.
3. Add a regression test:
   • create long session near compaction threshold
   • stream/finalize assistant response
   • trigger auto-compaction
   • reload session/messages
   • assert the final assistant message still exists
4. If the backend is already correct, document the expected client-side merge contract so Workspace can fix its reconciliation.

Environment

• Host OS: macOS
• Frontend: Hermes Workspace, localhost:3000
• Backend: local Hermes Agent gateway/session stack
• Session: long tool-heavy conversation with automatic compaction

Severity

High. Causes apparent loss of completed agent work.