How Megaparsec combines error information

The source code appears to be in Text/Megaparsec/Internal.hs, although I haven't gotten my head around it yet.

I tried a simple example and it seems to work the way I expected.

To understand the behavior of megaparsec's <|> operator it is useful to know about "consuming" and "non-consuming" (or "empty") parses. To illustrate that concept I'll compare a literal string parser to a parser that parses each character separately, watch:

> let p = string "abc"
> let q = sequence [char 'a', char 'b', char 'c']
> parseMaybe (p <|> string "abd") "abd")
Just "abd"
> parseMaybe (q <|> string "abd") "abd"
Nothing

So, what happened? Well, when string "abc" tries to parse the string "abd" it fails without consuming any input. Or you can think of it as backtracking back to the beginning of the string. In contrast, the parser sequence [char 'a', char 'b', char 'c'] does consume the 'a' and 'b' characters even if it fails. In this case, <|> will not even try to use the string "abd" parser.

You can manually force the parser to backtrack by using the try function as follows:

> parseMaybe (try q <|> string "abd") "abd"
Just "abd"

But note that this can cause exponential running time, so try to avoid it.

To answer your question given this information: the error information will be combined but only if both arguments of <|> are failing without consuming any input. If either consumes input, then only the error information from that branch is used.