Skip to content

G18 — Bazel build-evidence, validated end-to-end

Registry: UC-TC-bazel-build-evidence (modeled) Effort: M · Risk: medium (Bazel toolchain is heavy; jsonproto schema drift)

Problem

abicheck ships a Bazel L3 adapter (adapters/bazel.py) that turns bazel cquery/aquery jsonproto into BuildEvidence targets / compile+link units — but it has never been validated end-to-end on a real Bazel C++ project. The 2026-06 field evaluation hit this directly (P21): oneDAL (uxlfoundation/oneDAL) builds with Bazel + a legacy makefile, no CMake, so the L3/L4/L5 source path was unreachable without the project's full Intel toolchain (DPC++/oneMKL/TBB). The artifact (L0/L1) scan worked fine; the build/source tier could not be exercised. Status is therefore modeled: the parser exists, the coverage claim is not backed by a real run.

Pointers

  • abicheck/buildsource/adapters/bazel.py — the adapter (cquery/aquery jsonproto → compile/link units; live or pre-captured).
  • abicheck/buildsource/extractor.py — the action model; query_build_system is the opt-in tier a live aquery would fall under (ADR-032 D5).
  • abicheck/buildsource/inline.py _run_build_query — the --allow-build-query subprocess path (could drive bazel aquery).
  • Field-eval evidence: eval/FINDINGS.md P21; eval/FOLLOWUPS.md §E2.

Approach

  1. Capture a real bazel aquery --output=jsonproto 'mnemonic("CppCompile", //...)' (+ cquery) from a small, self-contained Bazel C++ project — pre-captured, so the test is non-executing and Bazel-free in CI (ADR-028 D6).
  2. Add a fixture + unit/integration test feeding that export through adapters/bazel.py and asserting non-empty compile/link units + a usable L3.
  3. Optionally wire a live --build-query 'bazel aquery …' recipe behind --allow-build-query for users who have Bazel.

Acceptance

  • A committed jsonproto fixture round-trips through the adapter to non-empty BuildEvidence (targets + compile units), asserted by a test.
  • dump --build-info <captured-pack> embeds the Bazel-derived L3.
  • Registry entry flips modeled → partial/complete with the test as evidence.

Status

modeled — adapter code exists; no real-project validation yet (this gap).