A Rust port of Bazel 9.3.2 for Linux x86-64. Starlark runs on a vendored, patched starlark-rust; everything else ports Bazel's Java. The real Bazel 9.3.2 binary is the referee: Blaze's outputs are compared with it byte for byte or item by item.
Two real projects are the targets: brotli (simple C,
build //...) and re2 (C++ with abseil, a pybind11 Python
extension via rules_python, googletest; test //:all //python:all).
Score at commit ef66f55, from a binary built on a clean checkout, with Blaze's own
repository fetching and no oracle shortcuts (label blaze-ef66f55e904e; full table in
corpus/scores/blaze-ef66f55e904e.md):
| Surface | brotli | re2 |
|---|---|---|
| exit codes | 10/10 | 10/10 |
mod graph, MODULE.bazel.lock |
27/27, 193/193 | 29/29, 191/191 |
query (label_kind, streamed_jsonproto, incl. deps()) |
162/162, 162/162 | 1091/1091, 1091/1091 |
cquery (label_kind, jsonproto), including output order |
177/177, 18/18 | 3296/3296, 39/39 |
aquery actions, including order and ids |
66/66 (+2 allowed) | 1374/1374 (+54 allowed) |
| build stderr summary lines | 5/5 | 7/7, plus 1 extra line |
| execution log | 44/44 | 552/552 |
| output files (sha256) | 96/96 | 5043/5043 |
| test results | none | 21/21 |
| build event stream | 37/41 | 109/113 |
23 of 26 surfaces match; 13,850 of 13,858 compared items. "Allowed" items are runfiles-manifest
action keys and volatile build-metric counters, which also differ between two runs of real Bazel.
A third project, workerd, builds end to end with Blaze
on its untouched checkout: 9,713 actions, Bazel's exact summary line, and an output binary with
the same sha256 as Bazel's; query and cquery match in full. That was measured with a binary
built from an uncommitted tree, so workerd still counts as missing in the score file; see
docs/workerd.md.
Updated 2026-11-19. Three columns, three different claims:
- Ported: the code exists and runs.
- Compared with Bazel: outputs were compared with real Bazel 9.3.2 by a script; the
comparison is named. "Referee" is the audited
scripts/oracleon brotli and re2; the other tools were written by the implementers of each feature; an audit of their hardening returned STOP (docs/reviews/2026-11-19-audit-4.md), so those rows say exactly what was compared. - Reviewed and accepted: an independent reviewer compared the Rust with Bazel's Java and
returned PASS (
docs/reviews/), and the feature is in the accepted baseline (ef66f55). Nothing ported after that commit is accepted yet.
| Feature | Ported | Compared with Bazel | Reviewed and accepted |
|---|---|---|---|
Module resolution, MODULE.bazel.lock, mod graph |
yes | referee: identical | yes |
| Repository fetching, downloader, repository rules | yes | referee: identical outputs | yes |
Loading, BUILD and .bzl evaluation (starlark-rust, patched) |
yes | referee: identical | yes |
query (label_kind, streamed_jsonproto) |
yes | referee: identical, with order | yes |
| Analysis, configuration, toolchain resolution | yes | referee: identical | yes |
cquery (label_kind, jsonproto), aquery (jsonproto) |
yes | referee: identical, with order; runfiles action keys allow-listed | yes |
| Actions, action keys, C++ feature engine | yes | referee: identical | yes |
build, test: local and sandboxed execution, action cache |
yes | referee: execution log, output hashes and test results identical | yes |
| Evaluator (blocking Skyframe port) | yes | through every surface above | yes |
Build event stream, binary and JSON files, for build and test |
yes | referee: 37 of 41 and 109 of 113 items | yes, with those items open |
info (42 of 57 compared cases), clean, version |
yes | scripts/w1/compare-info |
yes |
| Feature | Ported | Compared with Bazel | Reviewed and accepted |
|---|---|---|---|
| Starlark aspects; full workerd build | yes | output binary sha256 equal to Bazel's; query 46,371 and cquery 38,892 lines identical | no |
| Bazel's byte-string semantics for non-ASCII text | yes | scripts/bytes: 114 identical, 1 different of 115; neither recording names its binary, so not a pass |
review PASS; not in the baseline |
| Disk cache, gRPC remote cache, remote execution, build without the bytes | yes | lead-measured against Bazel's own remote worker on brotli and re2 | review STOP in round 3; the blocking fix is in, re-review pending |
run |
yes | 20 cases: exit code and program output equal in 20; a filtered set of summary lines equal in 15 (7 of them declared failures); the rest of stderr is not compared and differs | no |
help topics, fetch, canonicalize-flags, license |
yes | 137 invocations, equal after normalisation: 120 identical (13 by declared exit code), 16 fail identically on both sides, 1 different; binary sha256 315c2147… |
no |
mod dump_repo_mapping, show_repo, tidy |
yes | 133 cases: exit code and stdout equal in all (44 of them equal failures); stderr differs in 96 | no |
Text-format output (--build_event_text_file, cquery --output=textproto) |
yes | on Bazel's own outputs: build events identical per record, cquery byte-identical, aquery by reprint only; Blaze's own aquery text output crashes | no |
| Invocation policy | yes | 63 invocations: 55 identical (14 by declared exit code), 8 fail identically on both sides | no |
| Symbolic macros, lazy macro expansion | yes | 45 of 52 cases by the tool before hardening (22 where Bazel succeeds, 23 equal failures); referee unchanged | no |
unused_inputs_list, C++ compile action templates |
yes | 11 of 12 probe steps equal; the twelfth differs in two runfiles action keys (the tool has no allow list); referee unchanged | no |
| Java rules (library, binary, test, plugins, deploy jars, toolchains) | yes | 18 of 18 steps equal in termination, stdout, actions, output hashes and a filtered set of stderr lines; other stderr differs; under --spawn_strategy=linux-sandbox only |
no |
Platform mappings, platform flags, --toolchain_resolution_debug, --target_environment, project flag scoping |
yes | 26 of 31 cases by the tool before hardening (15 where Bazel succeeds, 11 equal failures); referee unchanged | no |
Build events for run, fetch, query; target summaries, aspect events |
yes | scenario matrix; score rows unchanged | no |
| Starlark core error texts | mostly | implementer's probe (starlark_conformance probe-diff, cases recorded from Bazel's interpreter; dirty tree, no label): 456 of 533 equal |
no |
| Bazel-style crash handling (exit 37) | yes | one crash case compared | no |
| Feature | State |
|---|---|
Persistent server (client daemon, gRPC server, command lock, shutdown) |
being built; today every command is one process |
| Incremental builds from a kept graph | evaluator changes being built; behind a switch until proven |
Remaining query/cquery/aquery output formats, --query_file |
being ported |
| Console output: progress bar, critical path, summary lines | being ported; 130 of 777 error-scenario steps match Bazel in full |
| Persistent workers | being ported |
Exclusive tests, --test_output=streamed, test size/timeout/language filters |
being ported |
Build Event Service upload, events for cquery/aquery |
being ported |
Flag sets (--scl_config), remaining Starlark transition cases |
being ported |
| Reporting every syntax error of a file, as Bazel does | being ported |
| Remote: repo contents cache, HTTP cache, credential helpers, chunking | queued behind the remote review |
coverage; dump, config, print_action, mobile-install; native aspects; genquery,
extra actions; include scanning; dynamic execution; Docker sandbox; action rewinding; SkyQuery;
--profile / command.profile.gz; vendor mode; Apple and Android rules. The full list is
docs/missing-features.md; sizes per area are in docs/coverage/README.md (about 139k ± 15k
of 406k lines of needed Java were unported this morning).
Windows and macOS; JVM-only features (heap info keys, thread dumps); remote analysis caching.
- Bazel's own integration suite (
docs/upstream-tests.md): the first total for v0.1.0 was withdrawn after an audit found the run contaminated; a clean run is being recorded. - Error and UI transcripts (
scripts/w4/probe): 130 of 777 steps identical to Bazel in full. - Rust tests: 6,615 passed, 0 failed, 2,021 ignored at the last full run (remote lead,
dirty tree at
cc7373c); the accepted baseline is 5,487 passed.
Known limits: Linux x86-64 only, and scores are only comparable on a 16-CPU machine (Bazel's
cquery order depends on the CPU count); linux-sandbox, process-wrapper and @bazel_tools
come from Bazel 9.3.2's install base.
Needs Rust 1.97 (pinned in rust-toolchain.toml) and, at build time, Bazel 9.3.2's install base
to embed (see crates/blaze/build.rs; set BLAZE_ORACLE_INSTALL_BASE or run Bazel 9.3.2 once so
it exists under ~/.cache/blaze/oracle-out/ur/install/).
scripts/cargo-capped build -p blaze # -> target/debug/blazeBefore running the analysis tests: scripts/test-deps/prepare (materialises the rules_cc,
rules_java, rules_python, bazel_skylib, platforms and protobuf versions of the pinned Bazel into
~/.cache/blaze/test-deps, with a sha256 manifest; the tests fail without it).
Note: commits 0388d76 through 078b27a do not build (a partial commit). 0e68a79 and later
do; it was verified with cargo check --workspace --all-targets from a clean checkout.
target/debug/blaze version
cd /path/to/brotli && /path/to/blaze/target/debug/blaze info release output_base
# Run a command the way the referee does (workspace copy, clean env), on a corpus project:
scripts/w4a/run brotli query //... --output=label_kind
scripts/w4a/run brotli cquery //... --output=label_kind
scripts/w4a/run re2 aquery '//:re2' --output=jsonproto
# Score Blaze against Bazel on every surface:
scripts/oracle/record --binary target/debug/blaze --label demo --project brotli
scripts/oracle/compare oracle-9.3.2 democrates/: one crate per cluster of Bazel Java packages (docs/crates.md).third_party/starlark-rust: vendored starlark-rust with a numbered patch queue (PATCHES.md).scripts/oracle: the referee (record, compare, scoreboard).scripts/upstream-tests: runs Bazel's own shell/Python integration tests against any binary.AGENTS.md,PORTING.md: rules the porting agents follow.docs/research/: background research.docs/decisions.md: rulings.docs/STATUS.md: where things stand.