The fix for the local LLM in the latest version now causes an infinite loop during the conversation with your model.

[Ema-prowler]

Short description

The new fix introduces a bug in the inference.

What happened?

Right after I applied the fix—meaning I downloaded the new version—and used it for a few hours during my session, the infinite model loop started occurring; this is something I had never experienced with any previous version of the plugin.
I know this issue affects various models and that the problem might seem to lie with OpenCode, but that isn't the case.
In the year I’ve been using DeepSeek, I had never encountered this problem until now.
I’ve already tried clearing the cache,
closing OpenCode,
restarting my PC,
deleting old files,
forcing Dreamer,
and reinstalling the plugin.
Even so, the CPU issue is still noticeable at times.

Why do I say this is a Magic Context issue rather than an OpenCode issue—even though there could be two different root causes or they might be compounding each other?
Because I used to use other models without Magic Context and never had the infinite loop problem.
But once I started using Magic Context, the issue began appearing across several models.
I don't have a way to test it other than toggling the plugin on and off,
but this version introduces the problem when using DeepSeek v4.1 Flash.

Diagnostics

Plugin version

No response

OpenCode version

No response

Platform

No response

Client

OpenCode Desktop

Log output (optional)

[magic-alfonso]

Thanks for the report. Before I can look at this I need to know what "infinite loop" looks like on your side, because there are several different things it could be and they have different causes:

• the model keeps calling the same tool (which tool, with what arguments?),
• the model keeps producing the same text,
• a new assistant turn starts by itself after the previous one finished, without you sending anything,
• or the CPU stays busy with no output.

Please send:

1. a screenshot or a short paste of the repeating part of the conversation (a few repetitions are enough);
2. the diagnostic bundle: run npx @cortexkit/magic-context@latest doctor --issue 494 — it sanitises paths and secrets and attaches the plugin log, config and versions;
3. which version you were on before the update (the "fix for the local LLM" you mention — do you mean 0.42.6's change to the local embedding runtime?), and whether the loop also happens with the plugin enabled but dreamer disabled.

One thing worth checking on your side while you gather that: if the repeated part contains the text [dropped in a tool call, that is a known shape with weaker models and tells me exactly where to look.