[riclarsson]

AI PR message:

Summary

Error messages are shorter and easier to read, and ARTS_USER_ERROR / ARTS_USER_ERROR_IF are replaced by a plain function, arts::user_error.

Most of the 215 changed files are the mechanical macro conversion. The substantive changes are in debug.h/debug.cpp, mystring.h/mystring.cpp, workspace_method_class.cpp, gen_auto_py_methods.cpp, species_tags.cc and the files listed under "Message fixes".

Error layout

• ARTS_METHOD_ERROR_CATCH indents the error it passes on by two spaces, so nested functions give a readable trace instead of one flat block.
• Python method errors no longer print "Method reports the following error(s):"; the error is indented under the method call instead.
• Errors from methods in agendas list their inputs and outputs without the internal _/@ prefixes, and the line explaining those prefixes is gone. Inputs and outputs were printed swapped; that is fixed.
• Release builds show user errors as the message only; function names come from ARTS_METHOD_ERROR_CATCH. Debug and RelWithDebInfo builds still add file, function, line and column. The Fail check: <condition> line is gone in all builds.
• New indent(text, indentation = " ") in mystring.h. mystring.h/mystring.cpp move from the strings library to util, so debug.cpp can use it.

Example, an IGRF time outside the valid range, Release build. Before:

Cannot execute method:

  atm_fieldSchmidthFieldFromIGRF(atm_field : self.atm_field,
                                 time      : User-provided Time)

Method reports the following error(s):
Fail check: year < valid_range.first or year > valid_range.second

User error found in:

Filename:      .../src/core/igrf/igrf.cpp
Function Name: std::pair<Coefficients, Coefficients> IGRF::coefficients(const Time &)
Line Number:   40
Column Number: 11


Please follow these instructions to correct the error:

The IGRF coefficients are valid from the decimal year 1900 to 2030, but the time 1850-01-01 00:00:00.000000+00:00 is the decimal year 1850

After:

Cannot execute method:

  atm_fieldSchmidthFieldFromIGRF(atm_field : self.atm_field,
                                 time      : User-provided Time)

    The IGRF coefficients are valid from the decimal year 1900 to 2030, but the time 1850-01-01 00:00:00.000000+00:00 is the decimal year 1850

arts::user_error replaces the macros

arts::user_error("x has {} elements but y has {}", x.size(), y.size());

if (x.size() != y.size()) arts::user_error("x has {} elements but y has {}", x.size(), y.size());

• [[noreturn]], so it can end a function or a switch default.
• Checks are written as an if around the call. The arguments are only evaluated when the check fails, exactly as with the old macro; a passing check costs nothing beyond its condition, in every build type.
• The call location is captured by the format string's consteval constructor (std::source_location as a default argument). No macros, standard C++23.
• Checked at compile time: the format string must use every argument (std::format itself accepts unused ones), and every argument must be formattable. A violation is a compile error at the call.
• A message without arguments is used as it is, without formatting.
• 1,692 call sites converted with a script: 122 ARTS_USER_ERROR, 1,570 ARTS_USER_ERROR_IF.

Message fixes

• No more "Internal error" for user mistakes: unknown species, isotopologues and CIA partners in species tags now name the bad part, list the valid choices and point to arts-cat-data; HITRAN/JPL record parsing, surface normals and Python callbacks are reworded.
• The two Wigner initialization checks said "Must not be initialized." when the library was not initialized.
• Short messages now give the values involved: sensor element setters, path grids, polarization, DISORT subsurface sizes, surface extrapolation, gridded-field shape mismatches (was "bad field"), MT_CKD data, line-shape temperature models, Jacobian x-vector checks.
• Eight messages silently dropped their arguments (iostream-style "text", x, "more", or a missing {}); found by the new compile-time check: atm_field.cpp, xml_io_old.cc (2), lbl_hitran.cpp, rtepack_transmission.cc, test_mdspan.cc (3).

Other fixes

• extract1D read past the end of single-point grids (x.size() > 0 should be > 1).

Notes

• Debug builds no longer show the text of the failing condition; file and line point to it.
• doc/arts/dev.error.rst is rewritten for arts::user_error.
• Tested with Clang on macOS: builds without new warnings and passes the full test suite. Not tested with MSVC.

🤖 Generated with Claude Code

[olemke]

I like that the error macros are now plain functions.

I don't like that the location of the error (filename, function, line number) is now only available in debug/relwithdebinfo builds. Knowing the exact place where the error occurred without having to re-run the whole setup in a debug/relwithdebinfo build to get that information was very helpful. Especially if you try to diagnose error reports from end-users.

[riclarsson]

I agree it is useful. But the error messages looks more like C++ stacktraces than useful help to ensure people can fix what is wrong. That is not a good way to represent what is happening for python users.

[olemke]

I agree we should avoid confusing users, but we also need to make sure we can
support them efficiently. Knowing the exact error location is often critical
for that. I'd prefer to keep it readily available. Could we keep it and simply
label it as the C++ error location? For example:

Cannot execute method:

  atm_fieldSchmidthFieldFromIGRF(atm_field : self.atm_field,
                                 time      : User-provided Time)

      ARTS C++ error location:
        Filename: .../src/core/igrf/igrf.cpp
        Function Name: std::pair<Coefficients, Coefficients> IGRF::coefficients(const Time &)
        Line Number:   40

    The IGRF coefficients are valid from the decimal year 1900 to 2030, but the time 1850-01-01 00:00:00.000000+00:00 is the decimal year 1850

[riclarsson]

I agree in this example that it works.

Now imagine the method above is called in a parallel region.1 line of error information vs 4 lines of file/line information. Not just that, the error is caught and rethrown after the for loop, which will give a line number unrelated to the error itself.

Also, the user cannot fix the error. In fact, the error is in the python code. I think that the C++ should just be conditional. Perhaps compile time conditional is too much and this should be controlled by a global? One that is always off in release but on in reldeb.

[olemke]

The global seems to be a great solution to cover all the bases. Nice!

[riclarsson]

@olemke Could you check what is wrong with windows conda?

[olemke]

@olemke Could you check what is wrong with windows conda?

Must have been a temporary breakage in the miniconda chain. Started working again now without any change on our side.