[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.