Carbon

Carbon is an experimental successor language to C++ from the carbon-language project. As of 2026-09-05 there is no 1.0 and no 0.1: the 0.1 milestone is the minimum language for evaluation, and the 2025 roadmap revision moved its earliest realistic date to the end of 2026 after memory-safety design was added to its scope. The only downloads are nightly toolchain tarballs tagged v0.0.0-0.nightly.YYYY.MM.DD, on limited platforms. Treat every finding below as dated to the nightly you ran.

Contract

Field	Bound contract
Trigger	A team with a large C++ code base asks whether to evaluate Carbon, wants to run the toolchain, wants to call C++ from Carbon, or asks how Carbon's safety model and readiness compare with C++.
Authority	Reversible local: writes only the toolchain checkout or extracted nightly tarball, Carbon source files, and --output artifacts in the working directory; rollback is deleting them. No remote mutation.
Side effect	A Bazel build of the toolchain compiles a pinned LLVM commit and takes substantial disk and time.
Done	The evaluation report states the nightly date or commit used, what compiled and what did not, and a decision (evaluate further or stay on C++) tied to the constraints in step 7.

Inputs

• The C++ code base under consideration and its constraints: stdlib and Boost dependence, template depth, platforms.
• A toolchain source: a nightly tarball from the GitHub releases page, or a checkout of https://github.com/carbon-language/carbon-lang with Clang 19 or newer, libc++-dev, libc++abi-dev, and lld installed. Bazel is pinned by .bazelversion (8.6.0 at grounding) and fetched by scripts/run_bazelisk.py; LLVM is pinned as an upstream commit through git_override in MODULE.bazel, not as a release number.
• The evaluation question: syntax and ergonomics, C++ interop, or readiness.

Procedure

1. Record the toolchain identity. From a nightly: extract the tarball, take the version from the carbon_toolchain-${VERSION} directory name, and run ./carbon_toolchain-${VERSION}/bin/carbon help. From source: ./scripts/run_bazelisk.py run //toolchain -- help builds and runs the driver (target //toolchain:carbon, alias //toolchain:toolchain); //toolchain/install:carbon_toolchain packages the tarball. Done when: the nightly tag or commit hash is written into the report.

2. Compile the README example to prove the toolchain works:

   import Core library "io";
   
   fn Run() {
     Core.Print(42);
   }
   

   carbon compile --output=forty_two.o forty_two.carbon then carbon link --output=forty_two forty_two.o. Type-check only with carbon compile --phase=check file.carbon (phases: lex, parse, check, lower, optimize, codegen). Done when: the binary prints 42.

3. Read the syntax against the design docs, not against tutorials. A package declaration is [impl] [package Name] [library "Name"];; there is no api keyword, and an API file is one that omits impl. The entry point is a function named Run in the default library of the Main package; its valid signatures are undecided. Variables are var x: i32 = 3;. Printing is Core.Print(...) after import Core library "io";. Done when: each construct in the evaluation has a line in docs/design/README.md behind it.

4. Test C++ interop with the documented form. Import a header into the Cpp namespace and call through it:

   import Cpp library "circle.h";
   
   fn Area(r: f64) -> f64 {
     return Cpp.circle_area(r);
   }
   

   import Cpp library "<cstdio>"; works for standard headers. Interop is semantic: C++ functions, classes, and macros map into Cpp. with explicit-cast rules for types such as Cpp.int. The design says Carbon functions and types may be marked as exported to C, comparable to extern "C", but no exported-to-C++ header generator exists in the docs; Carbon.h is not a real artifact. Done when: at least one C++ function is called from Carbon on the recorded nightly, or the failure is quoted.

5. State the safety model as designed, not as marketed. Safe and unsafe code are split by a narrow unsafe keyword, with two modes: Strict Carbon marks all unsafe code, Permissive Carbon relaxes that for C++ interop and migration. Type safety and temporal and data-race safety are meant to come from the type system at compile time, spatial safety from run-time checks. The docs state the temporal model is expected to follow Rust's direction at the highest level. None of this is implemented enough to evaluate; say so. Done when: the report cites docs/design/safety/README.md and marks the model as design-stage.

6. List the limitations that apply today: no standard library comparable to C++'s, toolchain and syntax change nightly, editor support is thin, and the platform list is short. Done when: each limitation is tied to something observed on the recorded nightly.

7. Decide. Evaluate further when the team owns a large C++ code base it intends to migrate gradually, wants C++ interop without an FFI layer, and can absorb nightly churn on an evaluation branch. Stay on C++ when production stability is needed now, when the code leans on the standard library, Boost, or deep template metaprogramming, or when the target platforms exceed the nightly's list. cpp-modules and cpp-templates cover the C++ side of that comparison. Done when: the decision names the constraint that settled it.

Failure and recovery

Failure	Cause	Fix
Bazel build fails on LLVM	Host Clang older than 19, or libc++, libc++abi, lld missing	Install the listed prerequisites; let Bazel fetch the pinned LLVM commit rather than pointing at a system LLVM
C++ import rejects a header	Unsupported C++ feature in the header	Wrap the needed calls in a plain C or simpler C++ shim header and import that
Tutorial syntax fails to parse	Pre-0.1 churn	Rewrite against docs/design/README.md at the same commit as the toolchain
Compiler Explorer and local results differ	Different nightlies	Pin both to the same date; carbon.compiler-explorer.com is the project's own instance
Needed library function does not exist	No standard library yet	Call the C++ library through import Cpp library
Link error mixing Carbon and C++ objects	Toolchain mismatch	Link with carbon link from the same nightly that compiled the objects

Output

An evaluation report naming the nightly tag or commit, the examples that compiled and ran, the C++ interop result, the safety model status as design-stage, the observed limitations, and the decision from step 7 with its deciding constraint.