Replicate marginaleffects avg_predictions with "by" argument by hand

I cannot replicate marginaleffects::avg_predictions by hand when using the by argument.

In the following example, the coefficient returned by marginaleffects for the nodegree==0 condition equals 8290.

How can I obtain it by hand?

library(tidyverse)

mydf <- Rdatasets::rddata("lalonde")
mod <- lm(re78 ~ treat + married + nodegree, data=mydf)
summary(mod) 

marginaleffects::avg_predictions(mod, variables="treat", by="nodegree") # 8290

mdf <- mydf
mdf$treat = 0
mdf$nodegree = 0
res1 <- predict(mod, newdata = mdf, type = "response")
mean(res1) # nope

mdf <- mydf %>% filter(nodegree==0)
mdf$treat = 0
res2 <- predict(mod, newdata = mdf, type = "response")
mean(res2) # nope

Counterfactual variables replicate the entire data set (see ?avg_predictions). So to replicate

mdf <- mdf2 <- mydf
mdf$treat = 0
mdf2$treat = 1
mdf <- rbind(mdf, mdf2)
res1 <- predict(mod, newdata = mdf, type = "response")
mean(res1[mdf$nodegree==0]) # 8290
mean(res1[mdf$nodegree==1]) # 6046