GLM uses dcp messages as input/confirmation

[itpropro]

I had multiple instances with GLM-5 where the prune and other DCP messages were processed as input. The model took them as confirmation and just executed things it was not supposed to.
I have to disable the complete plugin for now or is there any way to make loading the plugin conditional to the model?

[Tarquinen]

Not yet, there's a new opencode hook that I think will enable things like per model settings, but haven't gotten around to implementing that yet (open to PRs). I'm not really sure what you mean by the prune and DCP messages were processed as input, do you have any pics I could look at?

[itpropro]

This is an example. There was already a final output and summary above that and everything was done, then prune kicked in and it started the Thinking on the screenshot.

[Tarquinen]

This is normal, when models use tools some kind of tool output is returned and the model generates another response. Agentic loops don't end turns on tool calls. Smarter models will probably think less and just output something like "what do you want to do next" instead of all that thinking output you got from glm 5.

[itpropro]

These steps only take place with DCP, it goes one step further and sometimes this one step does something that otherwise would have needed approval.

[Tarquinen]

I don't understand what you're saying. You can't have a model use a tool and then output nothing afterwards, it's not possible. Prune is a tool just like read/edit/anything else, after it uses the tool the agentic loop continues. If the model decides to do something weird due to the output of the prune tool, which is basically just something like "8 tools pruned succesfully: 1) bash: Create winapps... etc", that's just the model behaving poorly. I'll make a per model toggle eventually (duplicate issue of #77)

[itpropro]

I know how the tool calling works, but this didn’t ever happen before and the difference is that is seems to misinterpret the prompt from the tool call and especially it’s output for something the use wrote.
I never had a tool call resulting in a confirmation of a prompt that was target towards a user. This also didn’t happen before the 2 version of the plugin, so I assume it has either something to do with the model or a changed tool description/prompt.
It’s probably a combination of GLM-5s tool calling in combination with the “automatic” nature that the plugin nudges it’s actions.

[mja00]

Here is an example of the plugin's output triggering the model to think I've confirmed its plan. The only thing that stopped it is OpenCode's plan mode being strictly read only.

[Tarquinen]

I think this is just crazy halucination from GLM, the tool call doesn't output a user message and definitely doesn't output anything like "Otherwise the rest looks good... etc". It seems like you're both also having this issue in plan mode, I know opencode injects some user messages that remind the model to stay read only, maybe this combined with DCP causes weird behavior. Not sure yet how to fix this or if it's even fixable...

[mja00]

Yeah it might just be a OpenCode harness tweak is needed for this GLM model. It's nice to use but it does like to drift pretty quick.