#Token counting not working.

1 messages · Page 1 of 1 (latest)

raw merlin
#

When I add a local model to openclaw through the openclaw config or openclaw onboard for some reason it defaults to 16000 tokens. When i talk to the bot it does count the tokens then almost immediately starts compacting. If I go into openclaw.json and change the context window to 262144 like it is set in LM Studio, it stops counting tokens in /status so I have to guess how many tokens I have used. I don't understand why it works when it is set to the default 16000 but doesn't count tokens once I set it to the proper value.

magic rapidsBOT
median domeBOT
#

This is probably two separate things:

  1. Why it defaults to 16k

    • If you added the model through the generic custom/OpenAI-compatible path, OpenClaw uses a conservative fallback context window instead of LM Studio’s real one.
    • So 16000 is usually not “LM Studio says 16k”, it is “OpenClaw couldn’t reliably infer the real limit on that path”.

  2. Why /status stops showing token usage at 262144

    • /status token counting depends on usage metadata returned by the backend (or transcript fallback).
    • A lot of local OpenAI-compatible backends do not return reliable usage counters, so OpenClaw can run the model fine but still show token usage as unknown.

  3. Why it compacts almost immediately at 16k

    • 16k is just too small for OpenClaw’s normal agent overhead, tools, system prompt, etc.
    • So with that tiny effective window, auto-compaction kicks in very fast.

A few practical notes:

  • If you’re using LM Studio, prefer the built-in lmstudio provider over a generic custom provider.
#
  • If you stay on the custom-provider path, setting the real contextWindow manually is the right move, but /status may still be inaccurate if the backend doesn’t emit usage.

Please paste these so I can tell you if this is expected behavior or a version-specific bug:

openclaw --version
openclaw status --all

Also, if you can, paste the model block from openclaw.json and say whether you configured it as:

  • lmstudio/...
  • or a generic custom/OpenAI-compatible provider

Docs:

raw merlin
#

Ok I am an idiot. I didn't realize there was a LM Studio provider choice let me try actually setting it up proper

raw merlin
#

Thank you that fixed it!

median domeBOT
#

Glad that helped! 🎉

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution.
That helps Answer Overflow index it for the next person who hits the same weird token-counting issue.

If the local-model /status counters get weird again, drop the version + model block here and I can take another look.