codex/third_party/v8

rusty_v8 Consumer Artifacts

This directory wires the v8 crate to exact-version Bazel inputs. Bazel consumers build V8 archives from source and generate matching Rust bindings on Darwin, GNU/musl Linux, and Windows MSVC/GNU. MODULE.bazel wires RUSTY_V8_ARCHIVE and RUSTY_V8_SRC_BINDING_PATH to the @v8_targets copy of these consumer-facing selectors:

• //third_party/v8:rusty_v8_archive_for_target
• //third_party/v8:rusty_v8_binding_for_target

The Bazel v8 crate and its native source targets enable V8's in-process sandbox and pointer compression.

Local Cargo builds still use upstream prebuilt rusty_v8 archives by default. Selected Cargo CI, release, and package builds override RUSTY_V8_ARCHIVE/RUSTY_V8_SRC_BINDING_PATH with Codex release assets.

Source targets enforce the workspace runtime and sandbox/pointer-compression ABI. Release-pair targets enforce either the legacy or sandbox ABI plus Chromium's pinned runtime. Target CPU follows the named artifact. These V8 settings are internal; callers do not pass separate V8 feature or libc++ flags.

Current pinned versions:

• Rust crate: v8 = =152.2.0
• Embedded upstream V8 source for Bazel builds: 15.2.124.1

Updating to a new v8 release

Use this as the maintainer flow for a version bump:

1. Bump the v8 crate version and refresh codex-rs/Cargo.lock.
2. Update the Bazel versioned inputs in MODULE.bazel and the source pins in third_party/v8/dependencies.bzl and abseil_repository.bzl using the source dependency update procedure. Verify the imported source groups by querying the source archive and generating Torque and snapshots for each supported platform. If V8 changes its BUILD section layout, the importer fails instead of silently using an old inventory; update that boundary and review its fixed feature and CPU selections. Then refresh the matching checksum manifest and generated checksums below.
3. Publish a release-candidate PR and validate that v8-canary passes.
4. If the canary is green, publish the release tag and release build.
5. Independently verify the published Codex-built checksum manifests and record their SHA-256 digests in third_party/v8/rusty_v8_<version>_release_manifests.sha256.
6. Once the release build completes, rerun the build on the candidate branch and verify that the final artifact builds and tests pass.

When changing the prebuilt rusty_v8 http_file inputs, keep the checked-in checksum manifest and MODULE.bazel in sync:

python3 .github/scripts/rusty_v8_bazel.py update-module-bazel
python3 .github/scripts/rusty_v8_bazel.py check-module-bazel

For all eight published pairs, verify each published rusty_v8_ptrcomp_sandbox_release_<target>.sha256 against the committed rusty_v8_<version>_release_manifests.sha256 first. Copy the verified archive and binding checksums into rusty_v8_<version>.sha256, then run these commands. They validate every matching http_file entry, and CI blocks checksum drift.

Published release assets are expected at the tag:

• rusty-v8-v<crate_version>

with these raw asset names:

• librusty_v8_release_<target>.a.gz
• src_binding_release_<target>.rs

During the sandbox rollout, sandbox-enabled assets are published alongside those current assets on the same tag, with the Rust crate's sandbox feature suffix in their raw names:

• librusty_v8_ptrcomp_sandbox_release_<target>.a.gz
• rusty_v8_ptrcomp_sandbox_release_<target>.lib.gz on Windows MSVC
• src_binding_ptrcomp_sandbox_release_<target>.rs

The dedicated publishing workflow is .github/workflows/rusty-v8-release.yml. Tagged runs build release artifacts from the Bazel graph itself:

• //third_party/v8:rusty_v8_release_pair_x86_64_apple_darwin
• //third_party/v8:rusty_v8_release_pair_aarch64_apple_darwin
• //third_party/v8:rusty_v8_release_pair_x86_64_unknown_linux_gnu
• //third_party/v8:rusty_v8_release_pair_aarch64_unknown_linux_gnu
• //third_party/v8:rusty_v8_release_pair_x86_64_unknown_linux_musl
• //third_party/v8:rusty_v8_release_pair_aarch64_unknown_linux_musl

The same run also builds the matching sandbox pair targets:

• //third_party/v8:rusty_v8_sandbox_release_pair_x86_64_apple_darwin
• //third_party/v8:rusty_v8_sandbox_release_pair_aarch64_apple_darwin
• //third_party/v8:rusty_v8_sandbox_release_pair_x86_64_unknown_linux_gnu
• //third_party/v8:rusty_v8_sandbox_release_pair_aarch64_unknown_linux_gnu
• //third_party/v8:rusty_v8_sandbox_release_pair_x86_64_unknown_linux_musl
• //third_party/v8:rusty_v8_sandbox_release_pair_aarch64_unknown_linux_musl

The workflow also builds sandbox-enabled x86_64-pc-windows-msvc and aarch64-pc-windows-msvc archive/binding pairs from upstream rusty_v8 source. Those ABI-specific outputs cannot be produced by Codex's Bazel Windows GNU toolchain.

The Bazel graph pins the same libc++, libc++abi, and llvm-libc source revisions used by rusty_v8 v152.2.0. The named release and sandbox-release pair targets select those headers and fold the matching runtime objects into the final static archive so consumers can link it with the v8 crate's default use_custom_libcxx feature. The profile keeps the object files and the bundled runtime on Chromium's std::__Cr ABI namespace instead of mixing those objects with the toolchain libc++ default namespace. Selected Cargo release and package builds use these published archives. The explicit Bazel prebuilt archive targets remain available; on Linux, they decompress to a private output and weaken the ten shared std::logic_error and std::runtime_error constructors and assignment entry points listed in gnu_libcxx_shared_exception_symbols.txt. libc++ keeps these exception functions in std:: across inline ABI namespaces; the toolchain's definitions then take precedence if both runtimes are linked. Constructors accepting Chromium's std::__Cr::string remain strong in the V8 archive. The verified input and the producer's archive are never mutated. This depends on libc++'s shared exception-object ABI; updates to either libc++ revision must keep the native GNU link/runtime checks passing. The Bazel pair targets above point directly to source targets so a new release or canary never depends on an older published copy.

Windows x64 and ARM64 sandbox archives and matching Rust bindings use the selected hermetic @llvm clang-cl toolchain and its declared Microsoft SDK: //third_party/v8:v8_152_2_0_x86_64_pc_windows_msvc_bazel and //third_party/v8:v8_152_2_0_aarch64_pc_windows_msvc_bazel. The producer selectors //third_party/v8:rusty_v8_native_windows_archive_for_target and //third_party/v8:rusty_v8_native_windows_binding_for_target select source outputs directly. Both use the same sandbox, pointer compression, target CPU, and Chromium libc++ profile. The bindings use V8's compiled headers and defines so their sizes follow the archive, including CppGC members.

The Windows distribution archive uses the published MSVC wide-stdio mode (_CRT_STDIO_ISO_WIDE_SPECIFIERS=0) for V8, Chromium libc++ and their native dependencies. Native Windows CI checks the COFF directives and links and runs a stock MSVC C++ consumer that includes <cstdio> and calls v8__V8__IsSandboxEnabled() on both x64 and ARM64. Cargo runs the V8 runtime tests against that staged archive and its generated bindings.

In-tree Bazel consumers link a separate source-built archive with the hermetic toolchain's ISO stdio mode (1) and matching bindings. Both pairs use the same V8 source, sandbox configuration and pinned Chromium libc++.

Binding generation uses the published hermeticbuild bindgen v0.0.4 executables through the standard rules_rs toolchain. Their embedded LLVM 23.1.2 libclang matches the selected LLVM resource headers; Codex does not build bindgen or libclang from source.

Binding generation reuses V8_COPTS, including native /std:c++20 on MSVC. The clang-cl bindgen action passes toolchain and caller arguments through in CL driver mode instead of translating their header-search or warning semantics.

The owned engine policy preserves the GN MSVC defaults for GC-managed microtask queues, random Wasm module generation, disabled imminent-deprecation warnings, and external TypedArray backing stores. Other source targets retain the published Unix policy, including a 64-byte inline TypedArray threshold. Generated Rust bindings use that same threshold; never pair a native 64-byte threshold with GN's size-zero binding constant. These are fixed platform choices, not additional developer build switches.

Private libc++ profiles set @llvm//config:nostdincxx and explicitly supply Chromium's pinned libc++ headers and archive. The Windows MSVC profile also sets @llvm//config:nostdlibxx. These settings are target-scoped: unrelated execution tools retain their default C++ runtime. V8's own generator binaries explicitly reapply the private-runtime policy after the execution transition. Neither flag suppresses C headers, the SDK, CRT, compiler-rt, or the independent unwinder.

Build with Linux remote execution and the static CRT:

bazel build --config=buildbuddy-generic-rbe --host_platform=//:rbe \
  --platforms=@llvm//platforms:windows_x86_64_msvc \
  --extra_execution_platforms=//:rbe \
  --repo_env=BAZEL_MSVC_RUNTIME_VISUAL_STUDIO_EULA=1 \
  --repo_env=BAZEL_WINDOWS_SDK_EULA=1 \
  --features=-dynamic_link_msvcrt --features=static_link_msvcrt \
  --features=cargo-rebase-json-vfs-overlays \
  //third_party/v8:rusty_v8_native_windows_archive_for_target \
  //third_party/v8:rusty_v8_native_windows_binding_for_target

Torque and other ordinary generators execute on Linux. The MSVC snapshot generator is a Windows payload compiled with the same @llvm toolchain. rules_wine selects direct execution on Windows or its pinned Wine runtime on Linux. Linux ARM64 workers use the runtime's bundled FEX to run the x64 generator; both x64 and ARM64 targets use the existing //:rbe platform. This keeps V8's Windows-host unwind writer enabled, including generated .pdata/.xdata. Select @llvm//platforms:windows_aarch64_msvc for an ARM64 build.

Native Windows ARM64 uses an ARM64 generator for ARM64 snapshots. Other MSVC combinations use the x64 generator, including V8's x64-hosted ARM64 generator. The existing GNU snapshot action and its Windows-local CI execution remain unchanged. The existing MSVC assembly-to-C++ conversion is also retained.

Release and CI Cargo builds for Darwin and Linux use RUSTY_V8_ARCHIVE plus a downloaded RUSTY_V8_SRC_BINDING_PATH to point at those openai/codex release assets directly. We do not use RUSTY_V8_MIRROR because the upstream v8 crate hardcodes a v<crate_version> tag layout, while our artifacts are published under rusty-v8-v<crate_version>.

Do not mix artifacts across crate versions. The archive and binding must match the exact resolved v8 crate version in codex-rs/Cargo.lock.

Source dependency ownership

The Darwin, Linux GNU/musl and Windows source targets consume Codex-owned Abseil, Highway, simdutf, ICU, fast_float, dragonbox and fp16 repositories. dependencies.bzl declares the pinned imports; the *.BUILD.bazel overlays here declare the targets V8 uses. Abseil, Highway and simdutf apply their C++ runtime settings through dependency_config.bzl; ICU defines its settings in its own overlay.

engine_repository.bzl fetches the V8 source pin and installs Codex's engine BUILD into @v8. Source filegroups are read from the pinned upstream BUILD each fetch, including added, removed or reordered groups. The importer rejects missing boundaries, malformed or empty sections, and new compiled targets. Only the two known Dragonbox and FP16 declarations are skipped, since the dependency overlays already supply them. v8_binary is valid V8 syntax, but compiled targets must be reviewed against the owned engine BUILD on an upgrade. The upstream bazel/config platform and CPU package is reused; only the CPU setting changes from a CLI flag to an internal setting. v8_defs.patch records that change and the Codex differences to upstream's bazel/defs.bzl: internal ABI/runtime settings, their visibility, libc++ headers/runtime, Windows compiler and linker settings, and snapshot generation. All engine rules, including Torque and V8's test configuration helpers, load from that file. The engine BUILD owns feature policy and dependency targets. ABI, runtime and CPU settings are selected together by artifact_profile.bzl for source and release-pair targets.

V8's private Abseil is built from the pinned 20250814.1 sources and uses codex_v8_lts_20250814 for its supported inline-namespace option. Protobuf keeps the regular shared Abseil module and its lts_20250814 namespace. Both can link into one binary without interposing functions built for different C++ runtimes. No V8 dependency needs a hook in shared rules_cc libraries.

Updating source dependencies

These are reviewed with each V8 source upgrade. There is no source-pin autoupdater: the prebuilt checksum commands above and their CI check validate published archive/binding pairs, not these source revisions. Bazel verifies the declared download hashes and Git revisions; it does not choose them for us.

Use the V8 tag in the root MODULE.bazel engine_repository(name = "v8") (currently 15.2.124.1) as the input. Read that tag's unpatched DEPS and MODULE.bazel; @v8 deliberately does not include upstream's MODULE. On an upgrade, also review v8_defs.patch against the new bazel/defs.bzl and run the V8 source canary. A cleanly applying patch does not by itself verify new upstream defaults or code-generation behavior.

v8_tag=15.2.124.1 # Replace with the V8 source tag selected for this upgrade.
gh api "repos/v8/v8/contents/DEPS?ref=$v8_tag" \
  -H 'Accept: application/vnd.github.raw+json' |
  rg -n -A 2 "'third_party/(highway/src|icu|dragonbox/src|fp16/src|fast_float/src|simdutf)'"
gh api "repos/v8/v8/contents/MODULE.bazel?ref=$v8_tag" \
  -H 'Accept: application/vnd.github.raw+json' | rg 'bazel_dep\(name = "abseil-cpp"'

• v8_highway, v8_icu, v8_dragonbox, v8_fp16: copy the respective DEPS commits. For Highway, use that commit in the archive URL and strip prefix.

• v8_fast_float: use the DEPS commit. If selecting a release-tag archive, verify the tag resolves to that commit, e.g. gh api repos/fastfloat/fast_float/git/ref/tags/v8.2.10 --jq '.object' (dereference annotated tags). Current v8.2.10 equals DEPS commit 34164f5.

• v8_simdutf: DEPS points at a Chromium vendoring repo, not the simdutf source repo. Read the metadata at that DEPS commit:

  simdutf_chromium=f7356eed293f8208c40b3c1b344a50bd70971983 # from new DEPS
  curl -fsSL "https://chromium.googlesource.com/chromium/src/third_party/simdutf/+/$simdutf_chromium/README.chromium?format=TEXT" |
    base64 -d
  

  Match its Version: to the upstream release tag and inspect Chromium's local changes before accepting the source form. Chromium vendors amalgamated files; we build upstream src/simdutf.cpp. At the current pin it reports 7.7.0, but its Revision: da645ece predates the v7.7.0 tag. This is a version match, not an assertion that the repositories' commits or bytes match.

• v8_abseil: follow abseil-cpp in V8's unpatched MODULE, not its Chromium DEPS vendor revision. Refresh abseil_repository.bzl release archive and its guarded inline namespace together; the V8 namespace must stay private.

For each http_archive/download_and_extract source, hash the exact new URL, not a different archive of the same commit:

source_url='https://github.com/google/highway/archive/2607d3b5b0113992fe84d3848859eae13b3b52c1.tar.gz'
curl -fsSL "$source_url" | shasum -a 256
# For an integrity= field (Abseil) instead:
curl -fsSL "$source_url" | openssl dgst -sha256 -binary | openssl base64 -A

Record the hex digest in sha256, or prefix the base64 digest with sha256- for integrity. Review each matching BUILD overlay against its stated source, then build the V8 generator and source/archive targets on supported platforms. The regular V8 canary/release checks validate the resulting engine; they do not prove upstream-version equivalence for a substituted source.