Using Mistral 3.5 with high reasoning is barely usable

[jsobiecki]

Description

I'm quite happy with Mistral Medium 3.5 model running on default variant. Not sure if it's more about problem of Model or if it's related to opencode itself, so I'm reporting it here

Plugins

none

OpenCode version

1.14.43

Steps to reproduce

1. Set mistral connection using "vibe api key" from mistral studio
2. Select Mistral Medium 3.5
3. Select variant "high"
4. In plan mode, request task (eg. plan approach to single task)
5. Model stop working at random places, without finishing task, hig

Screenshot and/or share link

Fragment of processing...

....
<thinking>
Thinking: The user wants me to create the test file. I've successfully written it. Now I need to make it executable and verify it was created correctly. Let me also check if there's a runner script that needs to be updated to include this new test.
</thinking>

bash{"command": "chmod +x /home/[REDACTED]/scripts/e2e/generate-infrastructure-files/test-auto-deploy-values-syntax.sh", "description": "Make test file executable"}read{"filePath": "/home/[REDACTED]/scripts/e2e/run-generate-infrastructure-files-tests.sh"}

AND THAT's IT.  No error, just ended output. At variant "default" it executes all commands as expected.

### Operating System

Dockerized ubuntu 26.04

### Terminal

Ghostty

[MyChaOS87]

i see the same ass soon as variant high is selected

Let me try a different search - look at the Oh My Posh website directly.
webfetch{"url": "https://ohmyposh.dev/docs/segments", "format": "text"} webfetch{"url": "https://ohmyposh.dev/docs", "format": "text"}

[Johannbr]

I've got the same issue. Cannot use mistral-medium-3.5 with reasoning on latest opencode (it works well with official CLI : vibe).
The current version of "@ai-sdk/mistral" is "3.0.27" on opencode
I've found several issues on this SDK : vercel/ai#15194
Not sure this is the root cause, but the ai-sdk is two month old.

I've tried v3.0.37 of @ai-sdk/mistral, no success, still a lot of tools error when thinking is enabled :

The arguments provided to the tool are invalid: Model tried to call unavailable tool '{"filePath": "/Users/ME/Dev/opencode-bundle/opencode/packages/opencode/src/cli/cmd/tui/context/thinking.ts"}'. Available tools: bash, context7_query-docs, context7_resolve-library-id, glob, grep, invalid, read, skill, task, webfetch, worktree_create, worktree_delete.

NB: I've made the comments there : ElioNeto/teamcode#851 , I didn't realize it was another repo.

[dbosk]

I see the same. On high it stops constantly, I have to type "Please continue." to make it resume. If I switch out of high it seems to keep working.

Seems like it's thinking output tokens are just normal output instead of thinking as for other models.

[pijyoi]

Version 1.18.5 seems to have fixed this