#Ollama - wont finish tasks.

1 messages · Page 1 of 1 (latest)

ember furnace
#

Im running open claw with GLM4.7 Flash, via Ollama (locally) the model itself is not bad. but it just randomly stops mid task and doesnt try to continue or give any explanation. Just gives up with no reply.

I have GLM4.7-Flash 30B Q4_K model with 128k context, reasoning off, which should be enough for what i think are basic tasks.. its a bit lazy on the follow through

glad sandBOT
#

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

ember furnace
#

@short hatch

short hatchBOT
# ember furnace <@1457407575476801641>

⚠️ Agent failed before reply: No API key found for provider "anthropic". Auth store: /srv/clawdbot/state/agents/main/agent/auth-profiles.json (agentDir: /srv/clawdbot/state/agents/main/agent). Configure auth for this agent (clawdbot agents add <id>) or copy auth-profiles.json from the main agentDir.
Logs: clawdbot logs --follow

burnt tundra
ember furnace
#

@short hatch ollama models are not finishing tasks ... they will stop half way and sit idle.. whats going on

#

@short hatch brotha u there

short hatchBOT
# ember furnace <@1457407575476801641> ollama models are not finishing tasks ... they will stop ...

This usually isn’t OpenClaw “getting lazy” — it’s almost always the run aborting (timeout / Ollama disconnect / tool-call parse error), and Discord just doesn’t show the reason. The reason will be in the gateway logs.

Can you paste these (they’re the fastest way to pinpoint it):

  1. Versions
openclaw --version
ollama --version
  1. A redacted debug bundle (includes log tail):
openclaw status --all
  1. Or reproduce it while tailing logs and grab the first error after it “goes idle”:
# 
openclaw logs --follow

Common fixes depending on what the logs say:

  • Timeouts (local models can be slow): bump the agent reply timeout (default is 30 min):
    https://docs.openclaw.ai/gateway/troubleshooting#agent-timing-out
    (key is reply.timeoutSeconds)

  • Tool-call / JSON issues: some Ollama models claim tool support but still output malformed tool calls occasionally → the agent run can crash mid-loop. Logs will show parsing/schema errors. In that case, try a different tool-capable Ollama model (docs list a few) or use a hosted fallback for tool-heavy tasks:
    https://docs.openclaw.ai/providers/ollama

  • Ollama disconnect / OOM / model unload: logs will show ECONNRESET, “connection closed”, or similar — then it’s Ollama/runtime stability (GPU memory pressure is a common culprit).

Drop the status --all + the log snippet around the stall and I’ll tell you exactly which bucket you’re in.

#

Yep — I’m here. When it “stops mid task and sits idle”, that almost always means the run errored/aborted and Discord just didn’t show the reason. The reason will be in the gateway logs.

Can you paste these (redacted is fine):

openclaw --version
ollama --version
openclaw status --all

And if you can reproduce once more, run this in a terminal while you do, then paste the first error that appears right when it stalls:

openclaw logs --follow

Most common culprits are: reply timeout, Ollama disconnect/OOM, or a tool-call JSON parse/schema error from the model. Once I see the log line, I can tell you which it is + the exact config tweak.

ember furnace
#

there are no errors it just stops mid way through a task like its finished.

short hatchBOT
# ember furnace there are no errors it just stops mid way through a task like its finished.

If there’s truly no error and it just “decides it’s done”, there are basically 3 common reasons with Ollama/local models:

  1. It hit a generation cap (OpenClaw maxTokens or Ollama num_predict) so the model stops outputting and returns “done”. That looks like “it gave up”, but it’s just truncated.
  • Check what OpenClaw thinks the model limits are:
openclaw models status
  • Check Ollama-side limits:
ollama show <your-model-name>

Look for num_predict / num_ctx / params.

If you’re using an explicit model entry in OpenClaw, make sure maxTokens isn’t tiny, or set it via model params (supported key is maxTokens):
https://docs.openclaw.ai/gateway/configuration (see agents.defaults.models[...].params)

  1. The model is accidentally returning the silent token NO_REPLY (yes, some weaker/quirky tool models do this). Then Discord shows nothing and it feels like it stopped. You’ll only see that in gateway logs.
#
  1. It aborted/timed out/disconnected but you’re not looking at the gateway logs. Discord won’t surface it.

Even if you think there are no errors, can you run this right after it stalls and paste the last chunk?

openclaw logs --limit 200

If you paste that snippet, I can tell you immediately whether it’s (a) maxTokens/num_predict truncation, (b) NO_REPLY, or (c) an abort/disconnect.