30 replies

Running OpenClaw on a local model. Running into issues with tool calling and compaction.

✅Solved

Currently trying to run OpenClaw on a Google Cloud VPS with a local model running through a GPU instance on Lambda.ai.

The model being run is Qwen3-32B.
Agent responds both through the dashboard and Slack but I am running into a few issues.

First of all, tool calls don't seem to be functioning. For example, with the prompt

"How is the weather in Tokyo?"

"How is the weather in Tokyo?"

the agent is replying in the dashboard chat what I assume is the tool call JSON

<tool_call>
{"name": "weather", "arguments": {"location": "Tokyo"}}
</tool_call>

<tool_call>
{"name": "weather", "arguments": {"location": "Tokyo"}}
</tool_call>

Second issue I am running into is that after every prompt, except the first one, compaction is getting called:

Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

since I haven't gone anywhere deep enough in any session so far, the reply from the assistant has almost always been

NO_REPLY

NO_REPLY

Solution

  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml

  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml

qwen3_xml

qwen3_xml

isn’t a valid option in your vLLM build, run:

vllm serve --help | grep -i tool

vllm serve --help | grep -i tool

…and pick a parser that exists (common alternates people try with Qwen are

qwen3_coder

qwen3_coder

hermes

hermes

, depending on the model/template).

Success condition: OpenClaw should stop showing

<tool_call>

<tool_call>

as text, and instead you’ll see “real” tool execution (tool call → tool result → final answer).

---

Jump to solution

FotC�

Friends of the Crustacean 🦞🤝•2d ago•

30 replies

Andy

Running OpenClaw on a local model. Running into issues with tool calling and compaction.

✅Solved

"How is the weather in Tokyo?"

"How is the weather in Tokyo?"

the agent is replying in the dashboard chat what I assume is the tool call JSON

<tool_call>
{"name": "weather", "arguments": {"location": "Tokyo"}}
</tool_call>

<tool_call>
{"name": "weather", "arguments": {"location": "Tokyo"}}
</tool_call>

Second issue I am running into is that after every prompt, except the first one, compaction is getting called:

Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

since I haven't gone anywhere deep enough in any session so far, the reply from the assistant has almost always been

NO_REPLY

NO_REPLY

Solution

  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml

  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml

qwen3_xml

qwen3_xml

isn’t a valid option in your vLLM build, run:

vllm serve --help | grep -i tool

vllm serve --help | grep -i tool

…and pick a parser that exists (common alternates people try with Qwen are

qwen3_coder

qwen3_coder

hermes

hermes

, depending on the model/template).

Success condition: OpenClaw should stop showing

<tool_call>

<tool_call>

as text, and instead you’ll see “real” tool execution (tool call → tool result → final answer).

---

Jump to solution

Running OpenClaw on a local model. Running into issues with tool calling and compaction.

Running OpenClaw on a local model. Running into issues with tool calling and compaction.

Similar Threads

Similar Threads

Similar Threads