The Error type should be smaller

[Techcable]

The problem / use case
Currently, processkit::Error is over 100 bytes. This bloats both direct users and types like vcs_core::Error which contain it indirectly. These large types frequently fire the clippy lints result_large_err and large_enum_variant. These clippy lints are both on by default; they aren't pedantic.

Errors are presumably on the cold path, so having large error types reduces the performance of the common case.

Proposed solution
Since the processkit::Error enum is public. It is not possible to fix this without a breaking change.
The largest variants only use stdlib types like String, Vec, and Duration.
We cannot shrink the size of these types even if we wanted to.

One option is to box these individual fields of the enum. This is a breaking change.

A more robust solution is to rename the current processkit::Error enum to something like ErrorReason, ErrorKind, or ErrorInfo. Then processkit::Error could hold a boxed reference to the kind. This is similar to the stdlib std::io::Error and std::io::ErrorKind distinction, although our "kind"/"reason"/"info" enum would hold more data than io::ErrorKind.

struct Error {
   reason: Box<ErrorReason>
}
impl Error {
  pub fn reason(&self) -> &ErrorReason {
    &self.reason
  }
}
enum ErrorReason {
    // all the same variants we currently have in Error
}

Alternatives considered
Leave as is, hurting performance and requiring clippy suppressions.

[ZelAnton]

Thanks for the detailed write-up and for pinpointing the downstream large_enum_variant/result_large_err impact on embedders like vcs_core::Error — that framing is spot on.

We agree with the approach. We'll adopt the std::io::Error/ErrorKind shape: rename the current enum to a public ErrorReason and make Error a wrapper holding Box<ErrorReason> with a reason() accessor (a single outer box rather than boxing individual fields). This drops Error to pointer size and un-bloats every Result<T, Error> and any embedding enum. The existing accessor surface (code(), program(), stdout(), is_timeout(), diagnostic(), …) will carry over unchanged, and the hardened Debug/Display (stream previews, PATH redaction, bidi/ANSI sanitization) will be preserved by delegating to the inner ErrorReason.

Since this changes the public API (direct variant matching on Error will no longer compile), it's a breaking change and the crate is currently on 2.x — so this will land in the next major release (3.0), batched with other breaking changes rather than shipped piecemeal. Tracking it internally toward that release.

[ZelAnton]

@Techcable
I will release 3 version tomorrow, with the PTY support and your fix

[Techcable]

I might suggest another change for version 3: Support for non-tokio async runtimes.