For the first two caveats, I actually agree that we could and should handle ambiguous input. It just hasn't been a priority because doing that properly would be quite a bit of work, and such ambiguous syntax isn't particularly common. See https://github.com/mvdan/sh/issues/686 for my current thoughts on how to tackle it.

The third caveat concerns parsing `export` and `let` as keywords rather than as builtins. Like the README says, this is to properly build the syntax tree without leaving opaque strings as expressions, but also to support `declare foo=(bar)` which wouldn't work if `declare` was treated like any other builtin simple command.

How else would you have a static parser handle these two builtins? They are in a bit of an awkward middle ground between builtin and keyword. My instinct is that giving them special treatment in the parser to allow tokens like `(`, while at the same time representing them in the syntax tree with opaque strings as expressions, would be pretty underwhelming to any users of the parser.

That said, we already have that problem with `let "foo=123"` for example, where our parser currently represents the expression as the quoted string without going any deeper. https://github.com/mvdan/sh/issues/754#issuecomment-96329574... considers doing a second parse stage in the shell interpreter to fix cases like these, though always doing a second parse could get expensive.

We _could_ leave all arithmetic expressions as input strings in the parser, and do all the actual parsing when they are evaluated. That would be more compatible with Bash and more consistent. But it would also be less useful to any parser users who don't run into any of these weird edge cases, which aren't common at all, I think.

In short, I have some ideas, but I'm not sure at all what's best :) Doing a good job for 99% of users feels better than aiming for 100% compatibility with bash syntax, particularly where bash syntax is a bit weird.

We have used it for more than a year in CI at work. We have many dozens of scripts from various coders. Nobody has ever complained that those caveats would have affected them. I was not aware of them, needed to search now where they even are (need to follow the more info link).

Have to agree with the other response. Demanding perfection is not realistic in many situations. If you can get 95% of the way there with 50% of the code/effort, you should do it. Sometimes being productive is knowing when that last percentage just isn't worth it.

I can’t say I’ve ever used the forms they don’t support and I am the person who uses all sorts of esoteric bash features at work.

Also for static typing an analysis I would absolutely give up even more syntax that is ambiguously parsed.

you could always do

  shfmt foo.sh || true

worst case it just doesn't format it?

At a previous job, my team slowly built an absolutely unmanageable bash script (hundreds of lines, including _many_ complex, multi-line jq incantations). The first step in migrating that script to go was to embed it in a Go binary and run it with gosh.

Kudos to Daniel for building such a wonderful package.

An alternative to have both `shfmt` and `shellcheck` and all pre-commit hooks managed (that is, automatically installed until no longer used and garbage collected) is to use https://devenv.sh/.

After `devenv init`, update `devenv.nix` as follows:

    { pkgs, ... }:

    {
      pre-commit.hooks = {
        shellcheck.enable = true;
        shfmt.enable = true;
      };
    }

We introduced it at work after the code base containing had been developed for 3 - 4 years with just manual reviews. The changes we needed to make were rather small and most of them were oversights, the code should not have looked like that in the first place. Of course your mileage might vary.

The only thing that disturbed me personally is no space before semicolon. Without having read what spec says I think the semicolon in shell is more of a command separator than a terminator. So I had always formatted it symmetrically, with space before and after like e.g. && or a pipe. shfmt did not support that so I had to adapt. Still having a tool and skipping review discussions outweighs this minor matter of taste.

  for i in $(seq 100)
  do if expr $i % 5 > /dev/null
     then if expr $i % 3 > /dev/null
          then echo $i
          else echo fizz
          fi
     else if expr $i % 3 > /dev/null
          then echo buzz
          else echo fizzbuzz
          fi
     fi
  done

https://github.com/mvdan/sh/blob/master/syntax/canonical.sh is one small example.