Continue agent mode returns raw JSON for tool calls instead of executing them
In Continue agent mode with the Ollama provider, tool calls stop executing and the assistant only outputs raw JSON like {"name": "builtin_search_web", "arguments": {"query": "cast ai documentation"}}. It happened with the AUTODETECT model on Ollama and also with qwen2.5-coder:7b, qwen3:8b, and a vLLM provider. Is this a model problem or a Continue problem, and is there a workaround?
Originally asked on GitHub: https://github.com/continuedev/continue/issues/6178
6 answers
Status: closed, with the maintainers explaining the cause.
A Continue contributor said this is a common issue with smaller-parameter models that advertise tool support but aren't reliable at invoking tools; the client receives model output it can't execute. The suggested workaround at the time was to enable User settings -> Experimental settings -> "Only use system message tools" on a prerelease build. A maintainer later closed the issue noting that system message tools are now the default for models without explicit tool support.
Other findings from the thread:
- One commenter attributed the problem to newer Ollama releases and suggested downgrading to Ollama v0.7.1, but the thread doesn't confirm this fixed anything.
- A user compared Ollama 0.9.2 with LM Studio using identical qwen2.5-coder:1.5b config, roles, and prompt ("list all files in directory ."): LM Studio executed the tool calls correctly while Ollama failed. The thread doesn't settle whether the root cause is the provider or the model.
General advice: if you hit this, testing the same model through a different provider can help you tell whether the issue is model-side or provider-side, and check that the model config lists
tool_useundercapabilities.Did this work for you?What @RomneyDa tried on GitHub (Outcome unclear)
A maintainer suggested enabling 'Only use system message tools' under User settings -> Experimental settings, explaining that smaller-parameter models that support tools often fail at using them properly. This later became the default for models without explicit tool support.
Source: https://github.com/continuedev/continue/issues/6178#issuecomment-3139836069 (opens external site)
Did this work for you?What @ghdoergeloh tried on GitHub (Didn't work)
A user compared Ollama 0.9.2 and LM Studio with identical qwen2.5-coder:1.5b config and the same prompt: LM Studio executed the tool calls correctly, while Ollama failed. No conclusion about the root cause was drawn in the thread.
Source: https://github.com/continuedev/continue/issues/6178#issuecomment-2991893818 (opens external site)
Did this work for you?What @steve-fraser tried on GitHub (Outcome unclear)
A commenter suggested downgrading Ollama to v0.7.1, saying newer Ollama versions seem to cause the problem.
Source: https://github.com/continuedev/continue/issues/6178#issuecomment-2985195560 (opens external site)
Did this work for you?What @neo-rooney tried on GitHub (Didn't work)
A user tested qwen2.5-coder:7b, which responded in JSON format, and qwen3:8b, which showed a toggle button but neither executed the tool nor gave a text response.
Source: https://github.com/continuedev/continue/issues/6178#issuecomment-2983350123 (opens external site)
Did this work for you?What a GitHub user tried on GitHub (Didn't work)
The reporter used Continue agent mode with the prompt 'search web for docs for cast ai' against the Ollama AUTODETECT model; the assistant only emitted the JSON tool call and never executed it.
Source: https://github.com/continuedev/continue/issues/6178 (opens external site)
Did this work for you?