Skip to content

Add json_document and json_view: parse, accept, types, materialize - #5621

Closed
nlohmann wants to merge 18 commits into
json-view/08-view-builderfrom
json-view/10-view-document
Closed

nlohmann wants to merge 18 commits into
json-view/08-view-builderfrom
json-view/10-view-document

Conversation

@nlohmann

@nlohmann nlohmann commented Sep 29, 2026 •

Copy link
Copy Markdown
Owner

Part of the stack for the zero-copy view (#5295). This PR adds the new header <nlohmann/json_view.hpp> with the public classes, built on the parser from the PR below. Element access, values, dump(), and comparisons follow in the next PRs, so each one stays reviewable.

Summary

json_document parses JSON text into a compact, read-only index. json_view is a cheap handle to one value in it. Strings and numbers are not copied out of the text.

#include <nlohmann/json_view.hpp>

std::string text = receive();
auto doc = nlohmann::json_document::parse(text);   // borrows text
nlohmann::json_view root = doc.root();
if (root.is_object() && !root.empty())
{
    nlohmann::json j = root.materialize();          // the value json::parse(text) returns
}
  • basic_json_document<BasicJsonType>: parse, parse_copy, accept, read (reuses the document's memory), root, is_discarded, source, owns_source, node_count, memory_usage, shrink_to_fit.
  • basic_json_view<BasicJsonType>: type, the is_* queries, operator bool, size, empty, materialize, source_offset.
  • Aliases: json_document, json_view, ordered_json_document, ordered_json_view.
  • Input:
    • borrowed: lvalue contiguous byte containers, std::string_view, C strings, character arrays, pointer ranges, and (in C++20) other contiguous iterator ranges;
    • owned: an rvalue std::string (moved in) and anything else parse() accepts (copied or read).
  • Errors: a parse error throws the exception basic_json::parse would throw for the same input. The library parser is run on the failing input (a cold path), so message, position, and id are the same. Inputs of 4 GiB or more are rejected with the new out_of_range.416.
  • Floats are converted when read, with the library's converter from Speed up the lexer: own float parser, string scan, and \u table #5738: the value does not depend on the locale at the time of the access, and reading a float or double does not allocate.
  • materialize(): builds the value with the SAX handler parse() uses, so JSON_DIAGNOSTICS parent pointers are set as usual.

Changes

area files
header include/nlohmann/json_view.hpp, detail/view/{errors,input,materialize,number}.hpp, single_include/nlohmann/json_view.hpp (amalgamated with the external key from the first PR of the stack, so it includes json.hpp instead of copying it)
build Makefile (amalgamate, check-amalgamation, include.zip, release), cmake/ci.cmake, check_amalgamation.yml, meson.build, Bazel, the C++20 module, labeler.yml
tests unit-json_view.cpp, unit-json_view_macros.cpp, fuzzer-parse_json_view.cpp
benchmarks tests/benchmarks/src/benchmarks_view.cpp (ViewParse, ViewRead, ViewParseIndented, ViewAccept, ViewMaterialize; built only if the header exists) and Accept for json::accept
docs API pages for both classes and the aliases, features/json_view.md with a "which representation to choose" section (linked from the feature overview, last in the Features navigation), 23 examples, the exceptions page, README, and a "Node index of JSON views" section on the architecture page (the 16-byte node: fields, navigation, and a worked example), linked from the feature page and the pages that mention the 16 bytes

Tests

  • unit-json_view.cpp:
    • materialize() equals parse() for handwritten and generated documents, with duplicate keys, 100,000 levels of nesting, and JSON_DIAGNOSTICS parents;
    • parse errors equal parse()'s, message included, for malformed inputs under all option combinations;
    • every input kind, borrowed or owned;
    • reuse with read(), moves, shrink_to_fit(), and source offsets.
  • unit-json_view_macros.cpp: includes the header without JSON_TEST_KEEP_MACROS, as users do. It checks that the view does not rely on macros json.hpp undefines, and that none of its own NLOHMANN_VIEW_* macros leak.
  • Fuzzer: accept, value, and exception parity with json::parse; 637,000 runs without a finding at this PR, and 14.5 million on the complete stack.

Benchmarks

tests/benchmarks at this PR (Google Benchmark, Apple M1, median of 5):

file json::parse json_document::parse speedup json::accept json_document::accept speedup
jeopardy 151.7 ms 26.1 ms 5.8× 82.0 ms 26.1 ms 3.1×
canada 9.18 ms 1.03 ms 8.9× 6.16 ms 1.03 ms 6.0×
citm_catalog 3.13 ms 0.42 ms 7.5× 2.03 ms 0.42 ms 4.9×
twitter 1.48 ms 0.24 ms 6.1× 0.80 ms 0.24 ms 3.3×

materialize() of the whole document (ViewMaterialize) takes 80.1, 4.65, 1.06, and 0.72 ms. So even when a json value is needed in the end, json_document::parse followed by materialize() is 1.4–2.1× faster than json::parse.

x86-64 numbers are still missing.

Public API

No breaking change. This PR adds a new header, new class templates, and the exception id out_of_range.416; nothing existing changes. The C++20 module exports the new names. The label aspect: json_view used by labeler.yml has to be created in the repository.


Written by Claude Code.

🤖 Generated with Claude Code

@nlohmann
nlohmann added this pull request to stack #5636 September 29, 2026 14:20
@nlohmann nlohmann changed the title json view/10 view document Add json_document and json_view: parse, accept, types, materialize Sep 29, 2026
@nlohmann

Copy link
Copy Markdown
Owner Author

CI fixes for this PR (commits a8e9ce9 and 94cd91c, plus the fixes from #5617/#5618/#5620 merged in):

Verified with gcc:16 (linux/amd64) and the CI flags (C++11 and C++20), and with a JSON_NOEXCEPTION build run as --no-throw: unit-json_view, unit-json_view_macros and unit-json_view_builder pass. make amalgamate leaves the tree clean.

— posted by Claude Code on behalf of @nlohmann

@github-actions

Copy link
Copy Markdown

🔴 Amalgamation check failed! 🔴

The source code has not been amalgamated and/or formatted correctly, or BUILD.bazel is out of date.

@nlohmann
nlohmann removed this pull request from stack #5636 September 30, 2026 13:19
@nlohmann
nlohmann force-pushed the json-view/08-view-builder branch from ab4f646 to dc9899f Compare September 30, 2026 13:20
@nlohmann
nlohmann force-pushed the json-view/10-view-document branch from 94cd91c to a887559 Compare September 30, 2026 13:20
@nlohmann
nlohmann added this pull request to stack #5739 September 30, 2026 13:21
@nlohmann
nlohmann force-pushed the json-view/10-view-document branch from a887559 to b136677 Compare September 30, 2026 15:34
nlohmann added a commit that referenced this pull request Sep 30, 2026
The generator for BUILD.bazel only listed single_include/nlohmann/json.hpp
in the singleheader-json cc_library's hdrs. Bazel's sandboxing only exposes
declared headers, so a consumer of :singleheader-json that includes
<nlohmann/json_fwd.hpp> failed with "file not found" (verified with Bazel
9.2.0). Meson and include.zip already ship json_fwd.hpp; the Bazel target
was missing it.

Add single_include/nlohmann/json_fwd.hpp to the hdrs list and regenerate
BUILD.bazel. Verified that `bazel build //:json //:singleheader-json`
succeeds and that a consumer cc_binary depending on :singleheader-json can
now include <nlohmann/json_fwd.hpp>.

Overlaps #5621, which adds json_view.hpp to the same hdrs list in
gen_bazel_build_file.cmake; that PR will need a rebase.

#5716 item 2

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
nlohmann added a commit that referenced this pull request Sep 30, 2026
fuzz_testing and fuzz_testing_{bon8,bson,cbor,msgpack,ubjson} seeded
fuzz-testing/testcases from tests/data, which was removed in dbf1a1f
(2020) when the test data moved to the external json_test_data repo.
The find command found nothing, but the pipeline's exit status was
that of xargs, so the recipe still reported success with an empty
corpus, and the printed afl-fuzz command would refuse to start.

The recipes were also six near-identical copies with unquoted -name
patterns, used the legacy CXX=afl-clang++, and fuzzing-start/stop were
missing from both the help output and .PHONY. tests/fuzzing.md already
documents the working flow (download json_test_data, then
`make -C tests fuzzers`), so replace the six broken targets and their
help lines with a single pointer to that document instead of trying
to keep six copies of a fragile shell pipeline in sync.

This does not affect OSS-Fuzz, which builds through tests/Makefile.

Overlaps #5621, which adds a seventh copy of the same broken line for
fuzz_testing_json_view.

#5717 item 2

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
@nlohmann
nlohmann force-pushed the json-view/10-view-document branch 2 times, most recently from ccd2753 to 80763fe Compare September 30, 2026 18:06
@nlohmann
nlohmann marked this pull request as ready for review September 30, 2026 18:17
@nlohmann
nlohmann force-pushed the json-view/08-view-builder branch from c195988 to e96e298 Compare September 30, 2026 18:19
nlohmann added a commit that referenced this pull request Sep 30, 2026
…zel (#5715 item 5)

The amalgamation/format check existed three times with three different
file sets: the Makefile's pretty/check-amalgamation, the pull_request-only
check_amalgamation.yml workflow, and the ci_test_amalgamation CMake target
that also runs on direct pushes to develop/master/release/*. The CMake
target's glob was a strict subset of the workflow's (missing the
docs/mkdocs/docs/examples/*.hpp headers, tests/abi/, tests/cmake_*/project/,
tests/cuda_example/, tests/fmt_formatter/, and tests/module_cpp20/), and it
never checked BUILD.bazel at all, so a misformatted file in any of those
paths, or a stale BUILD.bazel, could reach develop through a direct push
even though the PR-only workflow would have caught it.

Make ci_test_amalgamation glob the same roots (docs/mkdocs/docs/examples,
include, tests) and extensions (*.hpp, *.cpp, *.cu) as check_amalgamation.yml,
excluding tests/thirdparty/ and tests/abi/include/nlohmann/ the same way, and
regenerate and diff BUILD.bazel next to json.hpp/json_fwd.hpp. Also add
docs/mkdocs/docs/examples/*.hpp to the Makefile's pretty/pretty_format
targets, which were missing the four custom_*_type.hpp example headers, and
drop the stale "called by Travis" comment on check-amalgamation (Travis is
gone; nothing currently calls that Makefile target from CI).

Leaves the workflow itself untouched: it deliberately runs amalgamate.py
from a fresh develop checkout so a PR cannot change the tool that checks it.

Overlaps #5610 and #5621, which each add a new amalgamated header and touch
the same INDENT_FILES/ci_test_amalgamation/check_amalgamation.yml hunks.

Verified with `make check-amalgamation` on this branch: clean, no diff.
#5715 item 5.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The public classes of the zero-copy view (#5295), in the new header
<nlohmann/json_view.hpp>:

- basic_json_document<BasicJsonType>: parse (borrowing contiguous byte
  inputs, owning rvalue strings, streams, and other inputs), parse_copy,
  accept, read, root, is_discarded, source, owns_source, node_count,
  memory_usage, shrink_to_fit
- basic_json_view<BasicJsonType>: type and the is_* queries, size, empty,
  materialize (the value parse() would produce, built by the same SAX
  handler), source_offset
- the aliases json_document, json_view, ordered_json_document, and
  ordered_json_view

A parse error throws the exception basic_json::parse would throw for the
same input: the library parser is run on the failing input, so messages,
positions, and exception ids are the same. Inputs of 4 GiB or more are
rejected with out_of_range.416.

The single header single_include/nlohmann/json_view.hpp keeps including
json.hpp; make amalgamate, check-amalgamation, include.zip, and release
handle it.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
unit-json_view.cpp: type queries, size, and empty against basic_json;
materialize() against parse() (json and ordered_json, generated documents,
duplicate keys, 100,000 levels of nesting, parent pointers with
JSON_DIAGNOSTICS); parse errors and their messages equal to parse() for
malformed inputs and all option combinations; the overflow of a float
document (1e39, 3.4028236e38) as in parse(); NUL and BOM; borrowed and
owned inputs (strings, C strings, literals, vectors, string_view, streams,
wide strings, parse_copy, and iterator ranges over pointers, vectors,
strings, and lists); reuse with read(); moves; shrink_to_fit() of the index
and of the decoded strings; source offsets.

unit-json_view_macros.cpp includes the header without
JSON_TEST_KEEP_MACROS, as users do: the view must not depend on the macros
json.hpp undefines, and must not leak its own.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The new single header goes through the same checks and install steps as
json.hpp and json_fwd.hpp:

- cmake/ci.cmake: ci_test_amalgamation regenerates, formats, and compares
  json_view.hpp as well
- check_amalgamation.yml: the pull request check does the same; it runs
  develop's tools, so it needs config_json_view.json on develop first
- meson.build: installs single_include/nlohmann/json_view.hpp
- gen_bazel_build_file.cmake, BUILD.bazel: json_view.hpp joins the
  single-header target; the glob of the other target already covers the
  new headers
- labeler.yml: an "aspect: json_view" label for the header, its
  detail/view headers, tests, and documentation

The CMake install rules for include/ and single_include/, the REUSE
catch-all, and Package.swift cover the new files without changes.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The nlohmann.json module includes json_view.hpp in its global module
fragment and exports basic_json_document, basic_json_view, and the four
aliases next to basic_json, json, and ordered_json. features/modules.md
lists them, and tests/module_cpp20 parses a document, so that a missing
export fails the ci_module_cpp20 job.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
fuzzer-parse_json_view.cpp checks for every input that
json_document::accept agrees with json::accept, that an accepted input
materializes to the value json::parse returns, and that a rejected input
makes both throw the same exception with the same message. It is built
like the other fuzzers (tests/Makefile, and the root Makefile's
fuzz_testing_json_view target, which starts from the JSON test corpus) and
listed in tests/fuzzing.md.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- API pages for basic_json_document and basic_json_view, one per member,
  and for the four aliases, each with an example
- features/json_view.md: the problem the view solves, ownership and
  lifetime, what matches basic_json::parse() and what differs, and when
  to choose json, ordered_json, SAX, or the view
- the examples show why one would use the view, not only how: borrowed
  vs. owned input, reading a few fields and materializing one subtree,
  reusing a document across many messages
- registered in the mkdocs navigation, llms.txt, the docset, the
  exceptions page (out_of_range.416), architecture.md, the integration
  page, and the README; the yyjson credit is added to the README and
  license.md

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
benchmarks_view.cpp adds ViewParse, ViewRead (a reused document),
ViewParseIndented, ViewAccept, and ViewMaterialize on the files of
ParseString, so that each row can be read against the json::parse row of
the same file; benchmarks.cpp gains Accept (json::accept) as the
counterpart of ViewAccept. The view benchmarks are built only if the
header directory has json_view.hpp, so that older versions can still be
benchmarked.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- the input dispatch takes byte ranges by const reference and reads the
  size once (which also settles a finding of the static analyzer); input
  adapters are taken by value
- the classification of inputs keeps its nested conditional operators, a
  constant expression of C++11 (NOLINT)
- the test's C arrays, fixed seed, and escaped literals are marked, as in
  the other tests

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The 4 GiB limit and the fallback for an input that parse() accepts but
the view rejects (a bug) are excluded from the coverage; shrink_to_fit()
of an empty document is tested.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
materialize.hpp includes <string> (build/include_what_you_use).

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
A new section describes the 16-byte node: its fields, how integers,
floats, and object members are stored, how views navigate without
pointers, and a worked example. The feature page and the pages of
basic_json_document and node_count link to it where they mention the
16 bytes.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
…header

- ci_test_noexceptions: the helpers that compare the exceptions of
  json_document::parse() and json::parse() catch them outside a
  CHECK_THROWS, so with JSON_NOEXCEPTION the first parse error aborted
  the test. Compile those comparisons only with exceptions, as
  unit-class_parser.cpp does.
- ci_test_gcc: -Werror=unused-result for CHECK_THROWS_AS(json_document::
  parse(...)); assign the result to a dummy document.
- ci_test_compilers_clang (3.6): `const json_view invalid;` needs a
  user-provided default constructor there (CWG 253); value-initialize it.
- ci_test_single_header: json_view.hpp now exists as a single header and
  contains the internal view headers, so unit-json_view_builder.cpp
  includes it instead of the detail headers in that mode, and the test
  is built again with the single header.
- Regenerate single_include/nlohmann/json_view.hpp for the builder change
  merged from json-view/08-view-builder.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The "check" job (Check amalgamation) runs develop's amalgamate.py and read
all configurations from the develop checkout, where config_json_view.json
does not exist until this stack lands, so it failed with
FileNotFoundError. Read that configuration from the pull request's
checkout; the tool itself stays develop's.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
@nlohmann
nlohmann force-pushed the json-view/10-view-document branch from 80763fe to 2d7024e Compare September 30, 2026 19:15
@nlohmann

Copy link
Copy Markdown
Owner Author

Restacked #5621–#5635 onto the current json-view/08-view-builder (e96e2982a), which now contains develop's json_literals.hpp (#5610). The old branches conflicted with it.

Conflicts (only in this PR, 10-view-document): develop added json_literals.hpp in the same places where this PR adds json_view.hpp. Every resolution keeps both:

  • Makefile: amalgamate/pretty/zip/check-amalgamation/release targets and shasum list
  • BUILD.bazel, meson.build, src/modules/json.cppm
  • .github/workflows/check_amalgamation.yml, cmake/ci.cmake
  • docs/mkdocs/docs/integration/index.md

The upper 14 branches rebased without conflicts; git range-diff shows no content changes beyond the resolutions above.

Checked before pushing, for every branch head:

  • Regenerating single_include/ (json.hpp, json_fwd.hpp, json_literals.hpp, json_view.hpp) and BUILD.bazel changes nothing.
  • unit-json_view*, unit-udl and unit-no_automatic_udls pass with Clang at C++11 and C++17.
  • 10 and 24 also compile with GCC 16 using CI's warning flags, at C++11 and C++20.

Pushed with --force-with-lease against the previous heads; the old heads are kept locally as backup refs.

— posted by Claude Code on behalf of @nlohmann

nlohmann added a commit that referenced this pull request Sep 30, 2026
…er (#5735)

* Remove dead doctest help entry and pretty_format target from Makefile

The top-level Makefile still carried three leftovers:

- The help text listed a "doctest" target that was removed in #4560,
  so "make doctest" fails with "No rule to make target". The example
  check now runs as "make check_output -C docs".
- "pretty_format" ran clang-format on all sources, but .clang-format
  was deleted in #4573, so the target reformatted everything in the
  default LLVM style, against the Artistic Style formatting that
  "make pretty" applies and CI enforces.
- "clean" removed benchmarks/files/numbers/*.json, a directory that no
  longer exists since the benchmarks moved to tests/benchmarks (#3462).

Only maintainer tooling changes; the library is not affected.

Part of #5717

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Document tools/macro_builder and tidy up serve_header.py

tools/macro_builder generates the NLOHMANN_JSON_EXPAND,
NLOHMANN_JSON_GET_MACRO and NLOHMANN_JSON_PASTE* macros in
macro_scope.hpp, but nothing referred to it. Add a README that explains
what it generates, how to run it and where the output goes, and which
dependent tables (NLOHMANN_JSON_DOUBLE_PASTE, NLOHMANN_JSON_TYPE_BODY)
are maintained by hand. Point to it from a comment above
NLOHMANN_JSON_EXPAND. The generator itself is unchanged; following the
README reproduces the header byte for byte.

In serve_header.py, drop the LGTM suppression (LGTM.com shut down in
2022), replace the """.""" placeholder docstrings with real ones, and
import socket and ssl at module level. DualStackServer.server_bind uses
socket, which was only imported under __main__; when the module was
imported instead, the NameError was swallowed and IPV6_V6ONLY was not
cleared.

The header change is a comment only; behavior, API and ABI are
unchanged.

Part of #5717

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Hash every release_files artifact, not a hardcoded subset

The `release` target signed and copied json_fwd.hpp into release_files
alongside json.hpp, but the shasum line that writes hashes.txt only
listed json.hpp, include.zip and json.tar.xz. Users could not verify
the published json_fwd.hpp against hashes.txt.

Hash every file in release_files except the .asc signatures instead
of naming files by hand, so a newly shipped header (such as the
json_literals.hpp that #5610 adds to this target) cannot be missed
again.

Only affects the generated hashes.txt release artifact; the library
itself is unaffected.

Overlaps #5610, which touches the same lines to add json_literals.hpp
to the release target.

#5717 item 1

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Remove the broken fuzz_testing* Makefile targets

fuzz_testing and fuzz_testing_{bon8,bson,cbor,msgpack,ubjson} seeded
fuzz-testing/testcases from tests/data, which was removed in dbf1a1f
(2020) when the test data moved to the external json_test_data repo.
The find command found nothing, but the pipeline's exit status was
that of xargs, so the recipe still reported success with an empty
corpus, and the printed afl-fuzz command would refuse to start.

The recipes were also six near-identical copies with unquoted -name
patterns, used the legacy CXX=afl-clang++, and fuzzing-start/stop were
missing from both the help output and .PHONY. tests/fuzzing.md already
documents the working flow (download json_test_data, then
`make -C tests fuzzers`), so replace the six broken targets and their
help lines with a single pointer to that document instead of trying
to keep six copies of a fragile shell pipeline in sync.

This does not affect OSS-Fuzz, which builds through tests/Makefile.

Overlaps #5621, which adds a seventh copy of the same broken line for
fuzz_testing_json_view.

#5717 item 2

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Remove stale Travis comment above check-amalgamation

check-amalgamation carried "Note: this target is called by Travis",
left over from before the project switched off Travis CI. The prior
Makefile cleanup commit removed the other stale Travis-era leftovers
(the doctest help entry, pretty_format, and the benchmarks/ path in
clean) but missed this comment.

#5717 item 4

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix stale install/usage instructions in the vendored amalgamate README

tools/amalgamate/README.md is the unmodified upstream text and no
longer matches how the tool is used here:

- It named a Bitbucket origin that no longer exists; CHANGES.md
  already tracks the GitHub mirror commit this copy is based on.
- It asked for Python 2.7, but CI and the Makefile run the script
  with python3.
- It told readers to run ./test.sh (not vendored) and install to
  /usr/local/bin; in this repository the tool runs through
  `make amalgamate`.
- Its usage synopsis showed `-v` taking no argument, but the script's
  own argparser requires `choices=["yes", "no"]`, so that form fails
  with "argument -v/--verbose: expected one argument". The Makefile
  calls it as `--verbose=yes`.
- It pointed at test/source.c.json and test/include.h.json, which are
  not vendored; the configs actually used are config_json.json and
  config_json_fwd.json.

Rewrote only the Installing and Using sections to match; left the
"Here be dragons" caveats and the rest of the vendored code untouched
to avoid diverging further from upstream.

Overlaps #5615, which edits amalgamate.py, this README and CHANGES.md.

#5717 item 6

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Derive generate_natvis.py's ABI tag list and version from abi_macros.hpp

generate_natvis.py hard-coded abi_tags = ['_diag', '_ldvcmp', '_dp',
'_bics', '_psp', '_snul'] and required --version on the command line.
The source of truth is include/nlohmann/detail/abi_macros.hpp: the
NLOHMANN_JSON_ABI_TAG_* defines, the argument order of
NLOHMANN_JSON_ABI_TAGS_CONCAT, and NLOHMANN_JSON_VERSION_MAJOR/MINOR/
PATCH. Nothing checked that the copies stayed in sync, and they have
drifted apart before: _dp was added in #4517 but missed here until
#5544, and 3.11.3 shipped json_abi_v3_11_2 namespaces (#4340).

Parse the tag list (in NLOHMANN_JSON_ABI_TAGS_CONCAT order) and the
version from abi_macros.hpp instead of hard-coding them. Make
--version optional (falling back to the parsed version) and default
the output directory to the repository root the script lives in.

Add a "natvis" Makefile target that runs the script, and extend
check-amalgamation to regenerate nlohmann_json.natvis and fail on a
diff, the same way it already does for the amalgamated headers and
BUILD.bazel. Wire the same regeneration into check_amalgamation.yml,
using the tool copy checked out from develop (as the workflow already
does for amalgamate.py) and installing jinja2 from
tools/generate_natvis/requirements.txt. Update the tool's README to
say it must be re-run after adding an ABI tag or bumping the version.

Verified: a run against develop produces no diff (with either the
default or an explicit --version 3.12.0); adding a dummy
NLOHMANN_JSON_ABI_TAG_* without a matching #define makes the script
fail loudly instead of silently omitting the tag; xmllint --noout
passes on the regenerated file; and running the script from a
directory other than the one being checked (simulating the workflow's
separate tool checkout) against this repository root also produces no
diff.

Overlaps #5600, which added _ekmo to the same hand-written abi_tags
line and regenerated the file.

#5717 item 3

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Strip the leading "./" find(1) prefix from release hashes.txt entries

bc6e7db (#5717 item 1) switched the release target's shasum line from
naming files by hand to $$(find . -type f -not -name '*.asc' | sort),
so a newly shipped header is hashed automatically. Run from inside
release_files, that find prints paths as "./json.hpp" instead of
"json.hpp", so hashes.txt lists "./json.hpp" etc. instead of the plain
filenames it always used. shasum -c still verifies "./json.hpp" fine,
but it is a needless cosmetic regression for anyone reading the file
or matching it against release notes.

Strip the "./" prefix with sed before sorting, keeping the filenames
exactly as before while still hashing every artifact automatically.

Review fix for #5717 item 1 (PR #5735).

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix check_amalgamation.yml: pass --version to generate_natvis.py

bff45f1 (#5717 item 3) made --version optional in
tools/generate_natvis/generate_natvis.py and wired the workflow's new
"Regenerate nlohmann_json.natvis" step to call it without --version,
relying on the script deriving the version from abi_macros.hpp itself.

But NATVIS_TOOL_DIR is checked out from develop, the same way TOOL_DIR
already is for amalgamate.py, precisely so an in-flight PR's tooling
changes cannot mark themselves clean. Until this PR (or an equivalent)
merges to develop, that checkout is the old generate_natvis.py, whose
--version argument is still required=True. The new step's invocation
of "generate_natvis.py $MAIN_DIR" (no --version) then fails argparse
on this PR's own CI run with "the following arguments are required:
--version", before the check ever gets to compare output.

Extract the version from $MAIN_DIR's own abi_macros.hpp in the
workflow and always pass it as --version. That satisfies the old
script's required argument and is accepted as an explicit override by
the new one, so the step behaves the same whether NATVIS_TOOL_DIR holds
the pre- or post-merge tool, and stays correct for later PRs that bump
the version.

Verified by running the workflow step's shell logic locally against
both the pre-#5717 generate_natvis.py (checked out at 633de8e) and
the new one: both produce the identical nlohmann_json.natvis as the
committed file.

Review fix for #5717 item 3 (PR #5735).

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Extend tools/macro_builder to also generate DOUBLE_PASTE and TYPE_BODY

tools/macro_builder only emitted NLOHMANN_JSON_EXPAND..PASTE64, so two
other tables that scale with the same max_args stayed hand-maintained
with nothing checking them: NLOHMANN_JSON_DOUBLE_PASTE (added by hand
in #4563 for the *_WITH_NAMES macros) and the 64-slot dispatch table of
NLOHMANN_JSON_TYPE_BODY (the #4041 zero-member/one-or-more-member
switch). Both tables pass one macro name per slot to the same
NLOHMANN_JSON_GET_MACRO dispatch as PASTE, so they can drift out of
sync with max_args exactly the way _dp did in the ABI tag list fixed
by #5544.

Extend main.cpp with build_double_paste_code() (same recursive-doubling
shape as build_paste_code(), but DOUBLE_PASTE consumes two arguments
per member, so an even slot index falls back to the next lower odd
DOUBLE_PASTE<N>) and build_type_body_table() (max_args - 1 MEMBERS
slots and one trailing EMPTY slot, 8 per line, matching how it is
written by hand today). Add a "type_body" argument that selects the
TYPE_BODY block, since it lives at a separate location in
macro_scope.hpp from the EXPAND..DOUBLE_PASTE63 block; plain invocation
is unchanged apart from covering the extended range. No longer emit
the tool's old trailing blank line, so its output is directly diffable
without post-processing.

Verified with c++ -std=c++11: running the tool (with and without
"type_body") and piping the raw output through the pinned astyle
reproduces both blocks of the current macro_scope.hpp byte for byte.
tests/src/unit-udt_macro.cpp (all NLOHMANN_DEFINE_TYPE_*/_WITH_NAMES/
zero-member variants) passes unchanged under -std=c++11 and -std=c++17
with -fsanitize=address,undefined.

Add a "macro_builder_check" Makefile target that builds main.cpp,
regenerates both blocks into a scratch directory inside the repository
(astyle's --project lookup needs the target files under the same tree
as .astylerc, unlike an external /tmp directory), and diffs them
against the corresponding ranges of macro_scope.hpp; wire it into
check-amalgamation next to the natvis check. Wire the same regeneration
into check_amalgamation.yml, splicing the (still unindented) generated
blocks back into the PR's own macro_scope.hpp before the existing
astyle/amalgamation step runs, so that step's own tree-wide astyle
pass both indents them and folds any drift into the amalgamation
patch/diff the workflow already produces.

Unlike amalgamate.py and generate_natvis.py, this step builds
tools/macro_builder/main.cpp from the pull request's own checkout
($MAIN_DIR) rather than a separate checkout of tools/ at develop: this
tool has no independent source of truth to regenerate against (its
README documents that it must reproduce macro_scope.hpp byte for
byte), so a develop-pinned copy would only reproduce the
generate_natvis.py trap fixed in a previous commit on this branch,
where a PR that teaches the tool to cover more of the file fails its
own CI until that PR merges and updates the develop copy.

Add tools/macro_builder/README.md documentation for both new tables
and the two-invocation usage, and a short pointer comment above
NLOHMANN_JSON_TYPE_BODY (the EXPAND pointer already covered the first
block; extended its wording to include DOUBLE_PASTE63).

Closes #5717 item 5 in full, completing what the documentation-only
"Document tools/macro_builder..." commit already on this branch left
open (that commit's README/pointer-comment half stands; it also covers
item 7).

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

---------

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
nlohmann added a commit that referenced this pull request Sep 30, 2026
)

* Fix the ci_cmake_flags wiring so every option is checked

The CMake 3.31.6 flag list referred to itself before it was defined,
so only JSON_BuildTests was checked with that version. The targets for
the CMake running the build ("_2") were created but never added to
ci_cmake_flags, and the three versions shared one build directory.
JSON_StrictNulHandling was not in the list at all.

Use the 3.5.0 list for 3.31.6, add JSON_StrictNulHandling, and create
one ci_cmake_flag_<flag> target per option for the running CMake with
its own build directory. Also use the function parameter in the
COMMENT, refresh the stale version comment, and let ci_clean remove
the downloaded cmake-<version> directories instead of the long-gone
cmake-3.5.0-Darwin64.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix ci_test_clang_libcxx_cxx* jobs silently building without warnings

CMake only seeds CMAKE_CXX_FLAGS from the CXXFLAGS environment variable
when the cache entry is unset, so the explicit -DCMAKE_CXX_FLAGS="-stdlib=libc++"
argument made it ignore CXXFLAGS="${CLANG_CXXFLAGS}" entirely. The six
ci_test_standards_clang (..., libcxx) jobs therefore compiled without
-Weverything/-Werror while their libstdc++ siblings did use them.

Pass -stdlib=libc++ through the same CXXFLAGS value instead of a separate
-D argument, and give the target its own build directory
(build_clang_libcxx_cxx${CXX_STANDARD}) so it no longer shares a CMake
cache with the libstdc++ variant. Suppress the resulting
-Wthread-safety-negative finding from libc++'s std::mutex annotations,
which fires on doctest's reporters in this translation unit only.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Remove the no-op AppVeyor with_win_header job

The with_win_header matrix entry patched Windows.h into
single_include/nlohmann/json.hpp before building, but JSON_MultipleHeaders
has defaulted to ON since #3532 (2022-06), so CMakeLists.txt points the
tests at include/ and the patched single header is never compiled. The
job has been a no-op VS2015 build since then.

Windows.h coverage already exists through tests/src/unit-windows_h.cpp
(#3631), which runs in every MSVC job. Delete the dead matrix entry and
its before_build steps, and cite unit-windows_h.cpp from the QA page.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Remove unused ci_oclint and ci_pvs_studio targets

No workflow invokes ci_oclint, ci_pvs_studio, or their tool discovery.
ci_oclint also had a side effect on every JSON_CI configure: it copied
the single header into src_single/all.cpp and added an add_executable()
for it without EXCLUDE_FROM_ALL, so a plain build compiled a 1.2 MB
translation unit that only that unused target consumed. ci_pvs_studio
duplicates the Makefile's pvs_studio target, which is kept.

Also drop the duplicate --check-level=exhaustive flag passed twice to
the same ci_cppcheck invocation.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Stop Dependabot from proposing astyle bumps

astyle is deliberately pinned at 3.4.13 because newer versions reformat
unrelated lines and this version defines the formatting that
check_amalgamation.yml enforces. Without an ignore rule, Dependabot
keeps opening PRs for every new astyle release (most recently #4580,
#4942, #5445, #5448), each of which fails the amalgamation check and
gets closed unmerged.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Move Linux arm64 CI from dead Cirrus CI to ubuntu-24.04-arm

Cirrus CI stopped reporting check runs on develop sometime after
d10879b (2026-05-26); every commit since has only github-actions
check runs, so .cirrus.yml silently lost its only consumer while
README.md, FILES.md and the QA page kept advertising the coverage.

Add a ci_test_arm64 job to ubuntu.yml using the same pinned
actions/checkout and lukka/get-cmake actions as the other jobs, on the
native ubuntu-24.04-arm runner, with a step that confirms uname -m
reports aarch64. Delete .cirrus.yml and its README badge and FILES.md
section, and update the QA page's arm64 row.

Part of #5715

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Make scan-build fail on findings and drop irrelevant checkers (#5715 item 4a)

ci_clang_analyze ran scan-build without --status-bugs, so the job
passed whenever the ninja build succeeded, no matter what the
analyzer found ("No bugs found" in a green run gave no signal either
way). It is also missing --use-analyzer=${CLANG_TOOL}, so scan-build
picks whichever clang happens to be first on PATH inside the
silkeh/clang:dev container instead of the one this file already
selected and versioned.

Add --status-bugs and --use-analyzer=${CLANG_TOOL} to the scan-build
invocation. While here, drop the osx.*, webkit.*, fuchsia.*, and
optin.mpi.* checkers from CLANG_ANALYZER_CHECKS: none of them apply
to this portable C++ library, and leaving them enabled only adds
noise once the job can actually fail on a finding.

The job is currently clean (0 bugs), so this alone does not surface
any new finding; it only makes the existing "no bugs found" result
authoritative. This is 4a of 3 independent steps in #5715 item 4;
4b (Infer) and 4c (IWYU) still need their existing findings triaged
before --fail-on-issue/-Xiwyu --error can be added, and are handled
in separate commits.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Deduplicate the amalgamation/format check's file set and add BUILD.bazel (#5715 item 5)

The amalgamation/format check existed three times with three different
file sets: the Makefile's pretty/check-amalgamation, the pull_request-only
check_amalgamation.yml workflow, and the ci_test_amalgamation CMake target
that also runs on direct pushes to develop/master/release/*. The CMake
target's glob was a strict subset of the workflow's (missing the
docs/mkdocs/docs/examples/*.hpp headers, tests/abi/, tests/cmake_*/project/,
tests/cuda_example/, tests/fmt_formatter/, and tests/module_cpp20/), and it
never checked BUILD.bazel at all, so a misformatted file in any of those
paths, or a stale BUILD.bazel, could reach develop through a direct push
even though the PR-only workflow would have caught it.

Make ci_test_amalgamation glob the same roots (docs/mkdocs/docs/examples,
include, tests) and extensions (*.hpp, *.cpp, *.cu) as check_amalgamation.yml,
excluding tests/thirdparty/ and tests/abi/include/nlohmann/ the same way, and
regenerate and diff BUILD.bazel next to json.hpp/json_fwd.hpp. Also add
docs/mkdocs/docs/examples/*.hpp to the Makefile's pretty/pretty_format
targets, which were missing the four custom_*_type.hpp example headers, and
drop the stale "called by Travis" comment on check-amalgamation (Travis is
gone; nothing currently calls that Makefile target from CI).

Leaves the workflow itself untouched: it deliberately runs amalgamate.py
from a fresh develop checkout so a PR cannot change the tool that checks it.

Overlaps #5610 and #5621, which each add a new amalgamated header and touch
the same INDENT_FILES/ci_test_amalgamation/check_amalgamation.yml hunks.

Verified with `make check-amalgamation` on this branch: clean, no diff.
#5715 item 5.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Download prebuilt CMake binaries on Linux x86_64 instead of building from source (#5715 item 6)

ci_get_cmake() downloaded the source tarball of CMake 3.5.0, 3.31.6, and
4.0.0 and compiled each one completely (including CMake's own test
helpers) with -DCMAKE_POLICY_VERSION_MINIMUM=3.5 as a workaround for
building old CMake with a newer one. On CI this made the
ci_cmake_options (ci_cmake_flags) job take about 11 minutes, most of it
spent building CMake itself, even though Kitware has published
ready-to-run Linux x86_64 archives for all three of these releases
since 3.20 (lowercase platform name).

On Linux x86_64, download and unpack the prebuilt
cmake-<version>-linux-x86_64.tar.gz archive instead and point the
existing ${var} output at its bin/cmake, skipping the configure/build
steps and CMAKE_POLICY_VERSION_MINIMUM entirely. Keep the previous
source build as a fallback for any other platform (macOS, Linux
aarch64), since Kitware does not publish binaries for every
CMake/platform combination this project might build on.

Verify the downloaded archive against Kitware's own published checksum
before unpacking it: download cmake-<version>-SHA-256.txt alongside the
archive and run `sha256sum -c` on the matching line. A CI job that wgets
and untars a binary from a release page with no integrity check is a
supply-chain gap; Kitware has published this file for every release
since 3.20, so checking it costs one extra download and one grep.

As a separate, mechanical change: the ci_cmake_options job's container
only needed to stay on ubuntu:focal for the source build's
libssl-dev dependency and its own aging toolchain; now that the
Linux/x86_64 path never compiles CMake, drop libssl-dev from its apt
install line and move the job to ubuntu:24.04 (Ubuntu 20.04 left
standard support in May 2025). ci_clean already removes the
cmake-3.5.0/cmake-3.31.6/cmake-4.0.0 directories from the #5715 item 3
fix, and the prebuilt path reuses those same directory names, so no
further cleanup changes are needed.

Overlaps #5598, which edits the same ci_cmake_options matrix line in
ubuntu.yml; a rebase may be needed once that lands.

Verified locally: `cmake -S . -B build -DJSON_CI=On` configures cleanly
on macOS/arm64 (source-build fallback branch) and on Linux/x86_64 in an
ubuntu:24.04 Docker container (47 `ci_cmake_flag_*` targets generated,
one built and run successfully); `.github/workflows/ubuntu.yml` still
parses as valid YAML; downloaded the real v3.31.6 Linux x86_64 archive
and SHA-256 file from Kitware and confirmed the `grep | sha256sum -c`
pipeline both accepts the genuine file and is anchored to the exact
filename (not a prefix match).

CI must confirm: the prebuilt-binary path actually runs on the
ubuntu-latest/ubuntu:24.04 x86_64 runner, all `ci_cmake_options`
entries still pass with the new container's GCC, and the job's
runtime drops from roughly 11 minutes.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Make Infer fail on findings, with a type-level baseline for the ~174 pre-existing ones (#5715 item 4b)

ci_infer ran `infer run` without --fail-on-issue, so the job passed
regardless of what Pulse found; the last recorded run (35829411620,
commit 1054b20) logged "Found 174 issues" and still went green.
report.txt was also never uploaded, so the full finding list was only
ever visible in the truncated 5-issue console excerpt.

Add a repository-root .inferconfig (auto-discovered by Infer; passing
--project-root on the `infer run` invocation makes sure it is found
even though the analysis runs from build/build_infer) that sets
fail-on-issue and disables the six PULSE issue types that made up all
174 findings in that run: PULSE_UNNECESSARY_COPY_ASSIGNMENT (129),
PULSE_UNNECESSARY_COPY (22), PULSE_UNNECESSARY_COPY_INTERMEDIATE (15),
PULSE_RESOURCE_LEAK (5), PULSE_CONST_REFABLE (2), and
PULSE_UNNECESSARY_COPY_OPTIONAL (1).

This is a deliberate, narrower fix than "triage and fix everything in
this PR": the visible sample is entirely doctest-macro copies in test
code (for example tests/src/unit-algorithms.cpp:141 and
tests/src/unit-bjdata.cpp:3706), but 169 of the 174 findings were never
uploaded anywhere and this PR cannot respectably claim to have fixed
issues it never saw, including the resource-leak and const-refable
ones that are the most likely to be genuine bugs. Disabling by issue
type is a coarser baseline than a per-finding one (Infer has no
built-in per-finding baseline short of the two-run `infer reportdiff`
workflow, which this repository does not have the CI infrastructure
for), but it has the same effect today: the job goes from always green
to green-only-when-clean-of-everything-else, so CI now fails the
moment a *new* issue type appears, and report.txt is uploaded as a
workflow artifact on every run (including failures) so the six
disabled types can be triaged and re-enabled incrementally in follow-up
PRs.

#5715 item 4b. 4a (scan-build) and 4c (IWYU) are handled in separate
commits.

Verified: .inferconfig parses as JSON, ubuntu.yml still parses as
YAML. Infer itself is not available in this environment (v1.3.0 tar.xz
requires a Linux x86_64 runner), so CI must confirm that `infer run
--project-root ... -- make` picks up .inferconfig, that fail-on-issue
takes effect, and that the six disabled types actually suppress the
existing findings without also hiding an unrelated new one.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix IWYU findings for json.hpp/json_fwd.hpp/ordered_map.hpp and make CI fail on new ones (#5715 item 4c)

ci_single_binaries ran IWYU via CMake's CXX_INCLUDE_WHAT_YOU_USE launcher
property, which only printed "Warning: include-what-you-use reported
diagnostics" without failing the build: CMake's own __run_co_compile
wrapper does not propagate the launched tool's exit code, so even
`-Xiwyu --error` could never fail `cmake --build` this way. Verified
this empirically by injecting a deliberately-unused #include and
confirming the build still exited 0.

Fix the findings from the last recorded run (issue #5715 item 4, log
35829411620):
- ordered_map.hpp: add <new> (placement new) and
  nlohmann/detail/abi_macros.hpp; drop <memory> (std::allocator is
  still visible transitively via <vector>, confirmed by full local and
  containerized test suite runs).
- json_fwd.hpp: drop <memory> (same reasoning). Keep every forward
  declaration IWYU wanted removed (adl_serializer, basic_json,
  json_pointer, ordered_map): this file's only job is to forward-declare
  them for downstream users, so "nothing in this TU uses them" is
  expected, not a real finding. Mark each with `// IWYU pragma: keep`.
- json.hpp: add <cmath>, <cstdint>, <set>, <type_traits>,
  <unordered_map>, and the detail/abi_macros.hpp, detail/input/json_sax.hpp,
  detail/meta/detected.hpp, thirdparty/hedley/hedley.hpp includes IWYU
  says it needs. Do NOT remove adl_serializer.hpp,
  detail/conversions/from_json.hpp, detail/conversions/to_json.hpp,
  detail/macro_unscope.hpp, or ordered_map.hpp as IWYU suggests: nothing
  else in include/nlohmann includes adl_serializer.hpp or
  ordered_map.hpp, so basic_json<>'s own default template arguments
  (JSONSerializer = adl_serializer, and ordered_json = basic_json<ordered_map>)
  would lose their complete type; detail/macro_unscope.hpp is what
  undoes the JSON_* macros detail/macro_scope.hpp defines earlier in
  this same file, and removing it leaks those macros into every
  translation unit that includes <nlohmann/json.hpp>. Verified by
  actually removing them in a scratch test: the header still "compiles"
  stand-alone but ordered_json and every macro-using translation unit
  break. Marked each `// IWYU pragma: keep`.

Enforce it with `iwyu_tool` (ships with IWYU, e.g. as /usr/bin/iwyu_tool
on Debian/Ubuntu) instead of relying on the launcher property: it reads
compile_commands.json (now exported project-wide under JSON_CI) and
does return a real exit code for its own analysis, independent of
CMake's wrapper. ci_single_binaries now runs it over every
src_single/*.cpp with `-Xiwyu --error`, so a *new* finding fails CI.

json.hpp itself is excluded from that hard gate: even after every fix
above, IWYU's suggestion for one remaining symbol (a container
`swap, operator!=` used somewhere via a templated comparator) is not
deterministic — repeated, otherwise-identical containerized runs
reported <set>, then <unordered_map>, then <map> as "the" header to
add/remove for the exact same source. Gating a whole CI job on a
nondeterministic suggestion would make ci_single_binaries flaky rather
than informative, so json.hpp keeps the existing informational warning
(still shown during its normal compile) without failing the build on
it. Every other one of the ~50 single-header checks is included in the
hard gate.

#5715 item 4c. 4a (scan-build) and 4b (Infer) are separate commits.

Verified: full local ctest suite (129/129) and the ci_single_binaries
target itself both green in a containerized silkeh/clang:dev run
(matching the actual CI job) after this fix; a deliberately-reintroduced
unused #include in ordered_map.hpp was confirmed to fail
`cmake --build ... --target ci_single_binaries` (exit 2) with this
change, and to pass without it, on the same container/IWYU version CI
uses. `make check-amalgamation` is clean. Compiled with Clang and GCC
at -std=c++11/14/17/20 locally with no new warnings.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

---------

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Signed-off-by: Niels Lohmann <mail@nlohmann.me>

# Conflicts:
#	Makefile
#	cmake/ci.cmake
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
json_document::read is a member function, not POSIX read(), and the C
string overload requires null-terminated input like json::parse.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Conflicts in the See also lists of nine basic_json pages, is_discarded.md, and features/index.md, where develop (#5638) and this branch both added entries: kept both. Ran make amalgamate.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
…HEAD

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
@nlohmann
nlohmann removed this pull request from stack #5739 October 6, 2026 09:26
@nlohmann

nlohmann commented Oct 6, 2026

Copy link
Copy Markdown
Owner Author

The zero-copy view stack (#5295) was rebased onto develop and regrouped from 20 PRs into 10, one squashed commit each. The changes of this PR are now in #5620; its review comments stay here for reference.


Written by Claude Code.

@nlohmann nlohmann closed this Oct 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant