stop in a middle of task | Friends of the Crustacean 🦞🤝 | Page 1

analog granite Apr 17, 2026, 3:18 AM

#

I'm using LM Studio, qwen3.5-9b and qwen3.5-35b-a3b currently. Tried to use such as gpt-oss-20b, deepseek-r1-distill-qwen-14b and mistral-7b-instruct-v0.3.

The problem that I has, it stop in a middle of task. Example:

AI: "I need to investigate why the "IT Innovation Daily Task Reminder - 5 PM (Except Sunday)" cron job failed. Let me check the cron configuration and logs."

Then it stop... I need to ask "is it done?", then it will continue and sometimes stop again.

Questions:

1- Is it just me or everyone here facing same problem as me?
2- Any idea to solve this issue?

earnest moatBOT Apr 17, 2026, 3:18 AM

#

still marshBOT Apr 17, 2026, 3:18 AM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

verbal gardenBOT Apr 17, 2026, 3:20 AM

#

Not just you. This is pretty common with local models, especially smaller/quantized ones doing tool-heavy agent work.

What you’re seeing, the model says “let me check X” and then stalls, is usually one of these:

the model is too weak/brittle for multi-step tool use
the local backend chokes on OpenClaw’s full tool/schema prompt
LM Studio cold-loading / context limits make it look like it “hung”

OpenClaw’s docs call this out pretty directly:

smaller/stricter local backends can be unstable with the full runtime prompt
for local setups, weaker models are not recommended for tool-enabled agents
https://docs.openclaw.ai/gateway/local-models
https://docs.openclaw.ai/providers/lmstudio

What I’d try, in order:

Check version first

#

   openclaw --version

Basic health
```
openclaw status
openclaw models status
```
If using LM Studio, keep the model loaded
- don’t rely on cold/JIT loading while testing
- verify the server responds cleanly
Enable lean mode for local models

#

   agents.defaults.experimental.localModelLean: true

This trims heavyweight default tools and often helps weaker local setups.

If it still stalls, disable tool use for that local model
```
models.providers.<provider>.models[].compat.supportsTools: false
```
Then use that model for chat only, not agent/tool tasks.
Best practical fix
Use a strong hosted model as primary, keep LM Studio as fallback. That’s the most reliable setup for real agent work.

So, short answers:

Is it just you? No, definitely not. Common with local models.

#

How to solve it? Lean mode, keep the model preloaded, and if needed disable tools for that model or move tool-heavy work to a stronger hosted model.

If you paste:

openclaw --version
openclaw status
openclaw models status

…I can help narrow down whether this is mostly model weakness, LM Studio config, or an actual OpenClaw issue.

Also worth reading the FAQ: https://docs.openclaw.ai/help/faq

analog granite Apr 17, 2026, 3:29 AM

#

openclaw --version
OpenClaw 2026.4.14 (323493f)

verbal gardenBOT Apr 17, 2026, 3:30 AM

#

Thanks, that version is recent, so this is less likely to be “you’re on an old broken build.”

My honest read: this still sounds mostly like a local model reliability problem, not just an OpenClaw version problem.

The models you listed, especially the smaller ones, often do this with agent/tool work:

they announce a plan
start a tool-oriented turn
then lose the thread or stop early

That’s unfortunately pretty normal with local setups right now.

What I’d do next:

Try lean mode
```
{
```

#

     agents: {
       defaults: {
         experimental: {
           localModelLean: true,
         },
       },
     },
   }

Keep the LM Studio model fully loaded
- avoid cold-loading while testing
- make sure context isn’t set too low
Use local for chat, hosted for tools
If you want reliable “check logs / inspect cron / fix config” behavior, a stronger hosted model as primary is still the practical answer.

#

If a specific local model keeps failing, make it chat-only
```
compat: { supportsTools: false }
```
That won’t fix agentic tasks, but it will stop the half-started tool behavior.

If you want, send these next:

openclaw status
openclaw models status

And if possible, your models.providers.lmstudio config block. Then I can tell you whether this is mostly:

model too weak
LM Studio config issue
or something actually misconfigured in OpenClaw

#

Relevant docs:

analog granite Apr 17, 2026, 3:54 AM

#

11:54:35 [reload] config reload skipped (invalid config): agents.defaults: Unrecognized key: "experimental"

verbal gardenBOT Apr 17, 2026, 3:56 AM

#

Good catch, that one’s on me.

That error means your build does not support agents.defaults.experimental.localModelLean, so please remove that block. Looks like the docs/main branch got ahead of your released schema a bit.

For 2026.4.14, the safer advice is:

Don’t use that experimental key
Keep LM Studio models preloaded
If a local model keeps stalling, make that model chat-only
by adding compat.supportsTools: false on that model entry in models.providers.lmstudio.models

Example shape:

{
  models: {

#

    providers: {
      lmstudio: {
        models: [
          {
            id: "qwen/qwen3.5-9b",
            compat: { supportsTools: false },
          },
        ],
      },
    },
  },
}

Important: that does not make tool use better, it just stops OpenClaw from trying tool-heavy runs on a model that can’t handle them reliably.

#

So the practical answer is still:

local model for chat
stronger hosted model for cron/log/debug/tool tasks

If you want, paste:

openclaw models status

and your models.providers.lmstudio block, and I’ll help you tune the config without guessing.

Docs: https://docs.openclaw.ai/gateway/local-models

#stop in a middle of task