bug: Mux suggests compaction but always fails to do so

[f0ssel]

My prompt got really big (multiple plans I guess) and I ran into this flow. The input actually suggested the -t flag, I'm not sure if that's real.

[ammario]

FYI the -t controls number of output tokens, not input, but I see how it's confusing.

There's a couple things we should do better here:

1. Automatically retry compaction with 1m context enabled if the model supports it on context exceeded failure
2. Advise the user to configure a model with a greater context window if no current model can handle it

I'll handle.

[ammario]

Pt. 1 of this is done here: #2366.

That change alone should solve your specific case.