<system-reminder> keeps moving, causing unnecessary prompt processing in llama.cpp

[jacekpoplawski]

Description

For some reason, OpenCode keeps moving <system-reminder> around. As a result, the prompt history changes, so the cache cannot work correctly. In llama.cpp, this leads to a lot of unnecessary prompt processing time being wasted.

Would it be possible to keep <system-reminder> in the same place once it has been inserted for the first time? If something changes, you could add a new <system-reminder> instead.

details explained in the comment: ggml-org/llama.cpp#22031 (comment)

Plugins

No response

OpenCode version

1.14.19

Steps to reproduce

1. Start in "plan" mode
2. Switch to "build" mode to trigger <system-reminder>
3. Tell opencode to do something complex, like use lots of tools and write lots of files
4. Use any next prompt (like "thank you") and observe

<system-reminder> moves from the previous prompt to the next prompt and all the tokens (thousands!) between prompts must be processed again because cache is broken

Screenshot and/or share link

example:

 1184:  <|im_start|>user
 1185:  add one comment to it
 1186:  <system-reminder>
 1187:  Your operational mode has changed from plan to build.
 1188:  You are no longer in read-only mode.
 1189:  You are permitted to make file changes, run shell commands, and utilize your arsenal of tools as needed.
 1190:  </system-reminder><|im_end|>
 1191:  <|im_start|>assistant
 1192:  <think>
 1193:  The user wants me to add a comment to the `build_norm_gated` function. I should add a descriptive comment before the function.
 1194:  </think>
 1195:
 1196:  <tool_call>

then it moves after another prompt:

 1213:  <|im_start|>assistant
 1214:  Added.<|im_end|>
 1215:  <|im_start|>user
 1216:  thank you
 1217:  <system-reminder>
 1218:  Your operational mode has changed from plan to build.
 1219:  You are no longer in read-only mode.
 1220:  You are permitted to make file changes, run shell commands, and utilize your arsenal of tools as needed.
 1221:  </system-reminder><|im_end|>
 1222:  <|im_start|>assistant
 1223:  <think>

Operating System

Linux and Windows

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Queued user messages are serialized inconsistently across turns and break prompt caching #21518: Queued user messages are serialized inconsistently across turns and break prompt caching — this issue specifically describes <system-reminder> being placed inconsistently across turns, causing KV cache invalidation, which is very similar to what you're reporting.
• reasoning_content stripped from assistant messages on replay, causing KV cache invalidation on local inference #19081: reasoning_content stripped from assistant messages on replay, causing KV cache invalidation on local inference — same symptom (unnecessary prompt reprocessing in llama.cpp due to message serialization inconsistency), different root cause.

If your issue is distinct (e.g. it affects all turns, not just queued messages), please clarify that in a comment so maintainers can track it separately.

[qiujunhan]

#32246 When updating agents.md for the qwen3.6 27b model, it always triggers prompt processing at 100%.