JacobLinCool's picture
Build-and-Find v1.1.0 (camera-ready release)
3fc502b verified
|
Raw History Blame Contribute Delete
3.96 kB

Runtime tests of the released build artifacts

artifact_tests.csv records one run of each released build artifact's own test suite. scripts/run_artifact_tests.py ran cargo build --release and then cargo test in each of the 48 released build artifacts under runs_release/artifacts/ (24 per task) on 2026-07-24. The release build runs first because several integration tests locate the binary under target/release/.

The test suites are builder-written: the builder prompt asks for tests that pass with cargo test (src/bnf/prompts/builder.py), and each builder wrote its own. They indicate self-consistency, not specification conformance.

Results

quantity value
released build artifacts 48
compiled (cargo build --release succeeded) 41
compile-pass builds whose full suite passed 40 of 41
tests passed 2,060 of 2,061

The seven builds that did not compile have no root Cargo.toml; they are the seven builds that the release-mode compile probe excludes (compile_pass is false in runs_release/tables/build_runs.csv). The one failing test is cors_sets_origin in build-scratch_nanoweb-codex_gpt5_5_low-t1-s20260502-ce7e6480 (25 of 26 tests pass). report.md gives the per-builder and per-task counts.

Columns of artifact_tests.csv

  • build_run_id, task_id: the build run as recorded in runs_release/tables/build_runs.csv.
  • cargo_present: whether the artifact has a root Cargo.toml. Artifacts without one are not built; their build and test columns are empty or zero.
  • build_rc, build_ok, build_seconds: exit code, success, and wall-clock seconds of cargo build --release.
  • test_rc, test_ok, test_seconds: the same for cargo test. test_ok is true when every test binary passed.
  • test_binaries: the number of test result: lines in the cargo test output (unit-test, integration-test, and doc-test binaries).
  • tests_passed, tests_failed, tests_ignored: sums over those lines.
  • note: why an artifact was not built, or the failing tests (up to five) named in the cargo test output.

Booleans are written as True and False.

Run conditions

Each command had a 1200-second timeout. Builds of the same task shared one CARGO_TARGET_DIR so that dependencies compiled once per task. Every compile-pass artifact ships a Cargo.lock that pins its external dependencies, all of them crates.io packages. Eight compile-pass artifacts use Rust edition 2024, which needs Rust 1.85 or later. The CSV does not record the host or the Rust toolchain version.

Rerunning the tests

uv run python scripts/run_artifact_tests.py

The script needs cargo and network access to crates.io (or a populated cargo registry cache). It writes this directory's artifact_tests.csv by default (--out changes the path) and rewrites it after every artifact. Other options: --artifacts-root (default runs_release/artifacts), --target-root (shared target directory, default /tmp/bnf-artifact-tests-target, with one subdirectory per task), and --only <substring> (only artifacts whose build run ID contains the substring). The timing columns differ from run to run. The script now passes --locked to both cargo commands, so a rerun fails instead of rewriting a released Cargo.lock; the recorded run used the commands without the flag and left every released lock file unchanged.

Regenerating the summary

BNF_RUNS_ROOT=runs_release uv run python scripts/summarize_runtime_tests.py

make paper-analysis runs the same command. The script checks artifact_tests.csv against runs_release/tables/build_runs.csv (the same 48 build runs and tasks, and build_ok equal to compile_pass) and writes:

  • summary.json: inputs, input hashes, build, suite, and test counts, the failing suites with their failing tests, and per-builder and per-task counts.
  • report.md: the same values as tables.