-
Before submitting your bug report
Relevant environment info- OS: MacOS
- Continue version: 1.1.14
- IDE version: vscode 1.98
- Model: deepseek v3
- config:
OR link to assistant in Continue hub: DescriptionDescriptionI am using Continue (VSCode extension) v1.1.14 and have configured the DeepSeek-V3 model with the following capabilities: capabilities:
- tool_use However, despite enabling tool_use, the model does not seem to leverage the built-in tools to fulfill my requests. Additionally, I have tested Claude-3.5 in agent mode, but its performance in handling tool usage was also suboptimal. In comparison, Cline performs significantly better in this aspect. I am unsure whether this is due to a misconfiguration on my end or if the agent mode in Continue is not fully utilizing the available tools as proactively as Cline does. QuestionsIs there a specific configuration required to ensure that DeepSeek-V3 (or other models) actively use tools when tool_use is enabled? If agent mode is activated, does it fully leverage the tools as autonomously as Cline does? Are there any best practices or additional parameters I should adjust to improve tool usage? I appreciate any guidance on this matter. Thanks for your support! To reproduceNo response Log output |
Beta Was this translation helpful? Give feedback.
Replies: 11 comments
-
Hi, do you find that the model does not call tools at all or just is rarely doing so? Or can it sometimes call tools? I have tested with QwQ and claude, but in the OUTPUT - Continue - LLM prompt/Completion I do not see any requests with parameters of tools, which can be reproduced on 2 different Linux machines. This only happened after I updated continue to 1.1.14 pre-release with the new ui, it didn't have this problem back in 1.1.13. But I also noticed that the toolcall parameter has disappeared. Edit: I was using config.json and had put capabilities: {
"tools": true
} in every LLM block in my config file |
Beta Was this translation helpful? Give feedback.
-
as same with you. |
Beta Was this translation helpful? Give feedback.
-
I have used Cline, and I must say it works really well. I believe this is something worth considering for Continue—learning from Cline’s approach and potentially incorporating similar enhancements. Since Cline is also open-source, it could serve as a valuable reference. |
Beta Was this translation helpful? Give feedback.
-
Hi @luxixing, I have figured out how to call tools before I was just about to blame the devs for not testing this before releasing it🤣. If you are also experiencing that models telling you that they do not have the capability of calling tools or something. I am not exactly sure if this problem has anything to do with the version of continue, I am using 1.1.15 now, in this version I can switch to agent on the model and let it do tool calls, their docs probably needs to be updated though. Maybe the devs these days are working on something else and is too busy to respond. But it is really frustrating to have problems like this. |
Beta Was this translation helpful? Give feedback.
-
Following the upgrade to version 1.1.16, I conducted tests using both DeepSeek v3 and Gemini 2.5 Pro in agent mode. My findings are as follows:
Observed Behavior:
Observed Behavior: Strong Recommendation: I strongly urge the Continue development team to thoroughly study and implement Cline's agent capabilities. |
Beta Was this translation helpful? Give feedback.
-
cc @RomneyDa here |
Beta Was this translation helpful? Give feedback.
-
Cline does not do the tool calling in the openai compatible api way. It provides the tools directly in chat and does the parsing it self, which surprisingly, works more reliably. Maybe the problems are with the model providers. |
Beta Was this translation helpful? Give feedback.
-
Even in Cursor and RooCode, it is known that DeepSeek is just not great with tool use, so I have a few questions:
|
Beta Was this translation helpful? Give feedback.
-
It might be the reason for the lost tool call state here: Lines 908 to 920 in 38342fa Also see: |
Beta Was this translation helpful? Give feedback.
-
This is correct, @RomneyDa is working on a PR that will enable "virtual" tool calling, eg defining the tools directly in the system message, that will hopefully perform better with current DeepSeek models. |
Beta Was this translation helpful? Give feedback.
-
Happy to share this was merged recently and now in release. Documentation went live today. https://docs.continue.dev/customize/deep-dives/model-capabilities |
Beta Was this translation helpful? Give feedback.
Happy to share this was merged recently and now in release.
#6723
Documentation went live today.
https://docs.continue.dev/customize/deep-dives/model-capabilities