Could tool output visibility be configurable?

[BerniSc]

I've been really enjoying CodeCompanion - thanks for all the work you've put into it!

I was looking at PR #3327 and the changes around displaying tool output. I can definitely see the reasoning behind streamlining it, especially when tool output gets large and the chat becomes quite noisy.

Personally though I really liked the previous behaviour, especially being able to optionally expand the folded output of a run command.

For me, seeing the actual output while the agent is working makes it much easier to follow whats going on. Not only can I see which commands its running, but also what they're producing. This sometimes lets me notice earlier, that it is going in the wrong direction before it gets too far and I can intervene with the btw feature. I also find it quite useful for learning - seeing the output gives me context about what the agent is discovering and why it is taking its next steps, which I wouldn't necessarily get from seeing the commands alone.

My understanding is that #3327 is mainly a change to whats displayed in the UI, rather than a change to what is sent back to the LLM. If thats correct I was wondering whether it might be possible to make this configurable, so that users prefering the more verbose view could opt back into seeing the tool output?

I'm also curious if there are drawbacks to doing this that I'm overlooking - besides of course making the config more complex. Maybe there's a reason why making the old behaviour configurable is less straightforward than it seems from the outside.

If its something you'd be open to, I'd be happy to help with testing, creating a MR or otherwise help out with getting it working.

Thanks again for the great plugin!

[olimorris]

Thanks for your kind words. Really appreciate it.

So to answer your question directly, "could tool output visibility be configurable?". Yes, it could. Do I want to? Not especially 😆 . I really disliked the UX of tool output previously. It was janky, and the experience of seeing tons of lines of text appear in the chat buffer, then be folded and disappear, didn't feel great.

My rationale for removing it was, that I believe there are two extremes of agentic engineer (and I'm using that term pretty loosely):

• The first one doesn't care what the agent does. All it cares about is just getting the output.
• The second one cares deeply about what the agent does and will interrogate and review every single line (that's me btw!)

I don't believe that the tool output gives enough to be able to satisfy the second persona. If someone is truly between both of them, then they could just open the debug window and see all of the tool calls and the corresponding output. For me, just having a clean chat buffer was more important.

I have considered a compelling use case where tool output could be shown in a dialog box if the user moved their mouse over a tool call in the chat buffer. However, I'm very focused on the roadmap for CodeCompanion right now. So this would be very low down in the pecking order.

But as always, more than happy to have a discussion. I always want to hear the community's opinion.

[BerniSc]

Thanks for the thoughtful reply - that clears things up a lot.

I totally get where you're coming from, and I actually agree with your persona distinction. I'm probably somewhere in the middle, depending on the Task😄. Just reading the tooloutput is indeed a bit thin for a full understanding. I do, however, like having essentially the same information available to me as the agent has, without necessarily having to explicitly interrogate it every time. Sometimes I just want to be able to glance at what happened and build a mental model of what the agent is doing, rather than only digging into it when something goes wrong.

I have looked through the debug window a few times to get the tool outputs, but I've found it a little difficult to orient myself in there, especially when trying to follow a particular tool call in the context of the conversation.

That dialog-box idea you mentioned actually sounds pretty awesome to me. Having the tool call in the chat as the clean, compact representation, but being able to hover over it and get the output when I want to dig deeper, feels like it could be a really nice middle ground between the two personas.

I might actually be trying my luck at whipping up a little prototype extension or something to see what that could look like. If I get anywhere with it, I'd be happy to share what I come up with - and I'd certainly be interested in any thougts you might have along the way😄

[BerniSc]

So, I actually went ahead and built the thing I was talking about here😄

https://github.com/BerniSc/codecompanion-toolresults.nvim

It's still very much a work in progress, but the core is at least stable enough for my workflow and I've been really enjoying it so far. It basically gives me a way to inspect tool results on demand without cluttering the chat buffer.

I still have some ideas for where to take it further - including the hover idea mentioned above (currently its vim-style-cursorbased) and potentially a larger overview of performed tool calls/results - but I figured it's at a point where it could be worth sharing.

Feedback is very welcome, especially if anyone gives it a try and has suggestions or notices anything that could be improved or done differently🙂