Local model writes out the tool call as text instead of actually calling it

Asked 16 Apr 20259,840 views3 answers

29votes

Running the generator on my own machine because the source cannot leave it:

bash

sdods agent generate -p rwa-bank --adapter ollama --model qwen2.5-coder:7b --goal "add a contact"

It does not call anything. It writes a paragraph about what it is going to do, then a line that looks like this in the answer text:

text

step_find({"project":"rwa-bank","text":"I navigate to the Contacts page"})

and stops. The structured tool-call field is empty. Same model calls tools fine in other tooling, so I do not think it is the model.

agents mcp config

asked by Tanvi Bhatt126 · 16 Apr 2025

3 answers

33votes✓
Accepted answer
It is not the model, it is the size of the menu you handed it. From the MCP guide:
SDODS exposes 33–35 tools to a role and Playwright adds about 24; a 7–8B local model starts describing tool calls in prose instead of making them somewhere past that.
That is the behaviour you are looking at, and it is remarkably consistent: past roughly twenty functions a 7–8B model stops choosing and starts narrating.
The fix is the small profile, which is supposed to be picked for you:
bash
```
sdods agent generate -p rwa-bank --adapter ollama --model qwen2.5-coder:7b --profile small --goal "add a contact"
```
Under small the generator sees five tools — step_find, feature_list, feature_read, feature_parse, feature_write — one call per turn, no browser server attached, and the system prompt carries sixty of your project's real step patterns instead of the general conventions.
Check what you actually got with --dry-run first; it prints the profile, the reason it was chosen and the tool list.
answered by Kostya Petrov7,903 · 16 Apr 2025
15votes
One detail that will change what you are seeing: under the small profile a call the model wrote as prose is read back out of the answer and executed. The run says so, in those words:
text
```
recovered a step_find call the model wrote as text
```
So the exact output you pasted is the case that profile was built for. Under full it is ignored, which is why it looks like nothing happened.
Also worth checking before you blame the profile: ollama ps and the CONTEXT column. A 4096-token context truncates an agent prompt silently on the first turn, and a truncated prompt looks exactly like a stupid model. Local models has the numbers.
answered by Reeta Nandalmaintainer41,280 · 17 Apr 2025
6votes
We ended up mixing: agents.models takes a model per role, so planning runs on a hosted model and generation stays local. Planning was the role the 7B could never finish anyway.
answered by Thom Vasseur5,470 · 23 Apr 2025

Have the same question?

Ask it with the command you ran and the output you got, and it will be answered here.

Ask a question

Local model writes out the tool call as text instead of actually calling it

3 answers

Related