ollama and oss-gpt-20b can't call tools

[goobeyond]

i've configured continue to use ollama as provider and gpt-oss-20b model , but whenever i ask it to analyze the code base, it can't access my files. theres a step to use 'repo_browser.print_tree' but then nothing happens afterwards. is this a problem with ollama or with continue ?

[KetAveryH]

Hi, was in a similar boat and found this post here:

#7035

seems like it should be merged in already, but if you are up to date on this newer version and it doesn't work then someone else can hopefully take it from here!

[ShaunaGordon]

That workaround seemed to work at first, but in recent updates, I've found that it dies after the first tool call with no apparent errors. The best I get now is:

And I know it's not a model issue, because I can run the same setup in Zed and it works fine.

[bdougie]

I found some existing items that might be related. Let me know if these are helpful:

#	Title	Status
#7061	Continue to call ollama's GPT-oss: 20b model problem	🟢 Open
#7023	gpt-oss interactions seem incomplete	🟢 Open
#7248	Gpt-oss-120B with Vllm chat and other functions work but agent/plan mode doesn't	✅ Answered
#7724	Local ollama models not being used/overridden by Claude or GPT	🟢 Open

[bdougie]

Based on the list above. The best we can provide is troubleshooting tips as the model is new and there are rough edges still depending on the provider, in the case of ollama: @ShaunaGordon gave some guidance there.

#7061 (comment)

[ShaunaGordon]

That's mostly just for when there are issues with memory usage and resource splitting, which came from Ollama's memory handling, and wasn't model-specific. #7035 is a more direct workaround, but gpt-oss still dies after the first tool call for me in Continue.

I've been using it with success in other tools, so I'm fairly certain it's something with Continue, but it's hard telling what, exactly. I'm also not sure I'd call the model "new" at this point, given that it's been out for over two months at this point.

[kcupr]

I am facing same issue - first call to the tools seems to be ok, following trying to invoke non-existing tools with name starting with repo_browser. Symptoms suggest that there is no tool list in the context for subsequent conversation, I tried this change in Ollama.ts then it works somehow - subsequent tool calls work then. I don't know if it is close to the solution or not, use it as a reference:

$ git diff
diff --git a/core/llm/llms/Ollama.ts b/core/llm/llms/Ollama.ts
index 0a239fce9..6f52a975a 100644
--- a/core/llm/llms/Ollama.ts
+++ b/core/llm/llms/Ollama.ts
@@ -404,7 +404,7 @@ class Ollama extends BaseLLM implements ModelInstaller {
       // format: options.format, // Not currently in base completion options
     };
     // This logic is because tools can ONLY be included with user message for ollama
-    if (options.tools?.length && ollamaMessages.at(-1)?.role === "user") {
+    if (options.tools?.length && (ollamaMessages.at(-1)?.role === "user" || ollamaMessages.at(-1)?.role === "tool")) {
       chatOptions.tools = options.tools.map((tool) => ({
         type: "function",
         function: {

continue/core/llm/llms/Ollama.ts

Line 407 in 8fe0288

if (options.tools?.length && ollamaMessages.at(-1)?.role === "user") {

(so it is continue issue, ollama served llms can't work with tools properly)