hermes - 💡(How to fix) Fix Gateway mode doesn't invoke pre_llm_call / post_llm_call hooks — LCM and other context plugins fail silently in platform sessions [1 comments, 2 participants]

Problem

The pre_llm_call and post_llm_call plugin hooks are invoked in CLI mode (run_agent.py) but not in Gateway mode (gateway/run.py). This causes context management plugins like Hermes-LCM to silently fail for all platform sessions (Feishu, Telegram, Discord, etc.).

Root Cause

In run_agent.py, the LLM call lifecycle is:

pre_llm_call hook → [LLM call + tool loop] → post_llm_call hook

Both hooks receive conversation_history=list(messages) and in LCM's case, post_llm_call triggers the ingest into lcm.db.

In gateway/run.py, the message handling path does NOT invoke these hooks. It only calls:

• on_session_finalize
• on_session_reset

Impact

• LCM threshold_tokens stays at 0 in Gateway mode (set only via update_from_response() in CLI loop)
• should_compress_preflight() with its ingest logic is never called
• All platform sessions (Feishu, Telegram, WhatsApp, etc.) bypass LCM entirely
• Only CLI sessions work correctly with LCM

Expected Fix

In gateway/run.py, add pre_llm_call / post_llm_call invocations in the message processing path, matching what run_agent.py does. Specifically around where the agent is called to produce a response.

Alternatively, add a new hook like on_message_processed that fires after each message exchange with the full conversation_history.

Evidence

hook 调用分布：

• CLI mode (run_agent.py:8962): ✅ pre_llm_call with conversation_history
• CLI mode (run_agent.py:11868): ✅ post_llm_call with conversation_history
• Gateway mode (gateway/run.py): ❌ neither hook is called

Workaround

Manual /compress command works because it explicitly calls compress(), but this must be done manually per session and doesn't auto-trigger.

This issue affects all plugin authors building context/memory systems that rely on these hooks.