hermes - 💡(How to fix) Fix [Bug]: Compression exhaustion loop — session auto-reset re-loads bloated compressed context, causing immediate re-failure [1 pull requests]

Summary

After compression exhaustion triggers a session auto-reset, the new session immediately re-fails with the same context-overflow error, creating an infinite reset loop. The agent can respond to queries but every response is followed by "Skipping transcript persistence for context-overflow failure" + auto-reset.

Steps to Reproduce

1. Run Hermes Gateway with a profile using deepseek-v4-pro (or any model with limited context window)
2. Have a long-running conversation that eventually triggers context compression
3. Compression exhausts -> session auto-resets
4. Send any new message (even /reset or hello)
5. Agent responds correctly
6. Immediately: context-overflow failure
7. Immediately: session auto-reset
8. Loop repeats for every subsequent message

Logs

response ready: api_calls=1 response=67 chars Skipping transcript persistence for context-overflow failure in session X to prevent session growth loop. Auto-resetting session X after compression exhaustion.

After /reset, same loop continues with new session ID.

Expected Behavior

After a session auto-reset or explicit /reset, the new session should start with a clean compressed context. The old bloated compression summary should not be carried forward.

Actual Behavior

The compressed context from the exhausted session persists into new sessions, causing every new query to immediately re-trigger compression exhaustion. Only recovery is a full gateway restart.

Environment

• Model: deepseek-v4-pro
• Config: compression.enabled=true, hygiene_hard_message_limit=400
• Platform: Feishu (Lark) DM

Possible Root Cause

The compression summary/context from the exhausted session is persisted and re-loaded for new sessions. When the summary is already near the context window limit, even a single new message overflows it.