Problem

When using a local llama.cpp server (via openai-completions provider), the KV cache is fully invalidated on every message because the system prompt changes between turns.

Root Cause

buildInboundMetaSystemPrompt() injects per-message metadata into the system prompt (via extraSystemPrompt), including message_id which changes every message:

{
  "schema": "openclaw.inbound_meta.v1",
  "message_id": "2959",
  "sender_id": "414434471",
  "chat_id": "telegram:414434471",
  ...
}

For Anthropic/OpenAI this is fine — they handle caching server-side. But for llama.cpp, the KV cache works by matching the token prefix byte-for-byte. Any change in the system prompt (first message in the conversation) shifts the entire token sequence and forces a full reprocessing of the entire context.

Additionally, buildInboundUserContextPrefix() prepends changing metadata (including message_id) to each user message.

Impact

With a 196K context window model like MiniMax-M2.5, every message requires reprocessing the full conversation history from scratch, making the agent progressively slower and wasting compute.

Suggested Fix

One or more of:

1. Strip volatile fields (message_id) from inbound meta for openai-completions providers
2. Move per-message metadata out of the system prompt and into the last user message only, so the system prefix stays stable and cacheable
3. Add a provider-level config like "stableSystemPrompt": true that suppresses per-message metadata injection into the system prompt
4. Move the inbound meta block entirely to user messages instead of appending to the system prompt

Option 2 or 4 would be the cleanest — the system prompt stays byte-stable across turns, and llama.cpp can reuse the KV cache for the entire prefix.

Environment

• OpenClaw 2026.2.17
• Provider: openai-completions pointing to llama.cpp server
• Model: MiniMaxAI/MiniMax-M2.5 (196K context)

Relevant Code

• src/auto-reply/reply/inbound-meta.ts → buildInboundMetaSystemPrompt()
• src/auto-reply/reply/get-reply-run.ts → where extraSystemPrompt is assembled