hermes - 💡(How to fix) Fix [Feature]: Streamlining of review threads

Problem or Use Case

The Problem: Every 10 turns (memory) and 15 tool iterations (skills), Hermes spawns a background review thread. That thread calls run_conversation() with the entire messages list as conversation_history. The review agent then sends this full context to the LLM — triggering a complete prefill pass. This is expensive and blocks inference pipelines.

Key insight from the code: The review agent already inherits _cached_system_prompt (line 442) to hit prefix cache, but the conversation_history is still the raw full message list — no compression applied.

Proposed Solution

Use a cheaper model for review — the fork could use a smaller/faster model specifically for the review pass.

Alternatives Considered

Compress the messages snapshot before passing to review — use _compress_context or a lighter summarization pass so the review agent only gets a condensed version of the conversation.

Increase nudge intervals — 10 turns is aggressive. Bump to 20-30 for memory, 30+ for skills.

Disable background review entirely if it's more cost than value — set nudge_interval: 0 for both.

Feature Type

Performance / reliability

Scope

Medium (few files, < 300 lines)

Contribution

• I'd like to implement this myself and submit a PR

Debug Report (optional)