openclaw - 💡(How to fix) Fix [Bug]: token counting broken with local LLM, only user prompts counted

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

After upgrading from 2026.5.12 to .5.18, context tokens-used status is missing output and too low

Steps to reproduce

1. Start OpenClaw 2026.5.18 with local LLM
2. Take some turns in the control UI. Observe the context bubble is indicating very low, like hundreds of tokens used instead of the correct 13K+ (or however big the system prompt is). /status similarly indicates too low.

Expected behavior

The context use bubble and output of /status should show correct context usage.

Actual behavior

It appears only the user prompts are counted, i.e. my typing; nothing else is counted.

OpenClaw version

2026.5.20

Operating system

Ubuntu 24.04

Install method

npm global

Model

Qwen3.6-35B-A3B

Provider / routing chain

openclaw -> llama.cpp openai-completions API

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

Affected: not sure - at least local LLM users with llama.cpp and similar Severity: initially cosmetic, but I haven't driven it to where it should do compaction yet. I.e. it may undercount to the point compaction only occurs when the API returns a context-too-large message.

Additional information

Accurate context usage was last seen in 2026.5.12. It has been broken for me since 2026.5.18, and is still that way in 2026.5.20.