Skip to content

fix: install relay CLI from /v2 module path, pin to v2.0.4 - #50

Merged
SoundMatt merged 1 commit into
mainfrom
fix/relay-v2-install-path
Jul 30, 2026
Merged

fix: install relay CLI from /v2 module path, pin to v2.0.4#50
SoundMatt merged 1 commit into
mainfrom
fix/relay-v2-install-path

Conversation

@SoundMatt

Copy link
Copy Markdown
Owner

What

CI's relay-conform job installed the reference RELAY CLI with:

go install github.com/SoundMatt/RELAY/cmd/relay@latest

That bare (pre-v2) module path has been silently resolving to the last
v1.x tag (v1.14.0) ever since RELAY's v2.0 MAJOR bump — Go's semantic
import versioning rules require a /v2 path suffix for a module's v2.x+
tags to resolve at all (SoundMatt/RELAY#70). This repo's relay conform/
relay interop CI gates have therefore been running against a reference
tool three spec releases behind (v1.12, v1.13, v1.14, v2.0 all shipped
since) — and would have kept resolving to v1.14.0 forever, since v2.x
tags now live only under the new module path.

Fixed to:

go install github.com/SoundMatt/RELAY/v2/cmd/relay@v2.0.4

An explicit version pin rather than @latest, matching this repo's
existing convention of pinning external tool checkouts to a released ref
(e.g. cpp-FuSa is pinned to v0.15.0 in the same workflow, #16).

Conformance verification (done for real, not skipped)

Read RELAY's spec/CHANGELOG.md in full for every version between what
CI was actually running (v1.14.0) and the new pin (v2.0.4): v1.12, v1.13,
v1.14, v2.0, plus the v2.0.1–v2.0.4 patch releases.

  • v1.12 — adds "c" as a valid CLI language value. Doesn't affect
    cpp-LIN ("cpp" was already valid).
  • v1.13 — deep-audit fix pass. Most fixes are internal to the relay
    tool itself (crossbar routing, router mutex, SOME/IP, stdin bound,
    Ctrl+C handling) or affect other protocols. Two things worth checking
    for LIN specifically:
    • A go:embed glob bug had made spec/vectors/errors/*.json
      unreachable through RELAY's Vector/VectorNames API — meaning
      relay interop's reject-path comparison never actually ran for
      any implementation before this fix, LIN included. This is now fixed,
      so this PR is the first time cpp-LIN's error-handling has been
      checked against relay interop for real. Verified below — it passes.
    • §18.2 gained C++ HealthProvider/MetricsProvider/Drainer binding
      shapes. These are documented as optional interfaces ("applicable
      to all protocols; a C++ implementation exposing them MUST use these
      shapes") — not a new mandatory requirement, so no code change needed
      (cpp-LIN already implements them in this shape regardless).
  • v1.14 — §13.7.2 module-name registry expansion for RCP/DDS internal
    naming. Not applicable to LIN.
  • v2.0 — BREAKING: replaces RCP's canonical types entirely for TC18
    conformance. RCP-specific, as expected — no LIN impact.
  • v2.0.1–v2.0.4 — bug fixes/doc corrections, none touching LIN or C++
    bindings (v2.0.4 itself is the go.mod module-path fix this PR is
    reacting to).

Built RELAY v2.0.4 locally the same way CI now will (go install github.com/SoundMatt/RELAY/v2/cmd/relay@v2.0.4) and ran the real gates
against a fresh Release build of cpp-lin-cli:

$ relay conform --strict ./build/cli/cpp-lin-cli
PASS  §12.1  conforms to RELAY version document schema
PASS  §12.2  conforms to RELAY capabilities document schema
PASS  §12.3  conforms to RELAY status document schema
RESULT: PASS

$ relay interop --strict --protocol LIN ./build/cli/cpp-lin-cli
lin-diagnostic-wrong-checksum (LIN)  — cpp-lin-cli EQUIVALENT (correctly rejected)
lin-frame (LIN)                     — cpp-lin-cli EQUIVALENT
lin-id-overflow (LIN)               — cpp-lin-cli EQUIVALENT (correctly rejected)
RESULT: PASS

Both gates pass clean, including the two error vectors that were never
actually exercised by relay interop before v1.13's embed fix. No new
conformance gap — lin::validate_frame() already rejects both cases
(ID > 0x3F, and a diagnostic frame with a non-Classic checksum) with the
correct ErrInvalidFrame sentinel.

Version bump

None. No library code, headers, or test vectors changed — this is a pure
CI-pin fix with no behavior change, same category as #16 (which also
didn't bump the version).

Full local suite (re-run before pushing)

  • ctest (Release): 171/171 passed
  • ASan+UBSan (Debug, -fsanitize=address,undefined): 171/171 passed
  • clang-tidy / gcc-12: not available in this sandbox (macOS, no
    Ubuntu-pinned toolchain); CI runs these natively on ubuntu-22.04
    with the versions the workflow already pins, unaffected by this change.

Test plan

  • go install github.com/SoundMatt/RELAY/v2/cmd/relay@v2.0.4 resolves
    and installs relay 2.0.4 (spec 2.0) — verified locally.
  • relay conform --strict against a freshly built cpp-lin-cli — PASS.
  • relay interop --strict --protocol LIN against a freshly built
    cpp-lin-cli — PASS, including both LIN error vectors.
  • Full ctest suite — 171/171 green.
  • ASan+UBSan ctest suite — 171/171 green.
  • CI (this PR) — pending.

CI's `go install github.com/SoundMatt/RELAY/cmd/relay@latest` used the
bare (pre-v2) module path. Since RELAY v2.0.0, Go's semantic import
versioning rules require the /v2 suffix for that path to resolve at
all — the bare path has been silently resolving to the last v1.x tag
(v1.14.0) this whole time, three spec releases behind (v1.12, v1.13,
v1.14, v2.0 all shipped since), and would keep resolving there forever
now that v2.x tags live under the new path (RELAY#70).

Changed to `go install github.com/SoundMatt/RELAY/v2/cmd/relay@v2.0.4`,
an explicit pin rather than @latest, matching this repo's existing
"pin, don't float" convention for external tool checkouts (e.g.
cpp-FuSa is pinned to v0.15.0 in the same workflow, #16).

Verified conformance against the real RELAY v2.0.4 CLI (built the same
way CI now will) before pushing:
- `relay conform --strict ./build/cli/cpp-lin-cli` — PASS
- `relay interop --strict --protocol LIN ./build/cli/cpp-lin-cli` — PASS,
  including both LIN error vectors (lin-id-overflow,
  lin-diagnostic-wrong-checksum). Those reject-path vectors were never
  actually exercised by any implementation's CI before now — RELAY
  v1.13 fixed a `go:embed` glob bug that had made spec/vectors/errors/*
  unreachable via the Vector API, so `relay interop`'s error-path
  comparison silently only ever ran the happy path. cpp-LIN's
  validate_frame() already rejects both cases correctly, so no gap.

Read RELAY spec/CHANGELOG.md v1.12/v1.13/v1.14/v2.0 in full: v1.12 (add
"c" as a valid CLI language) and v1.14 (RCP/DDS module-name registry)
don't touch LIN or C++. v1.13's §18.2 HealthProvider/MetricsProvider/
Drainer C++ binding shapes are optional-interface documentation, not a
new mandatory requirement — no code change needed. v2.0's RCP
canonical-type replacement is RCP-specific, as expected, and doesn't
touch LIN.

No library code or test vectors changed, and no new conformance gap
was found, so no version bump — consistent with this repo's precedent
for pure CI-pin fixes (#16, which also didn't bump the version).

Full local suite also re-run and green: 171/171 ctest, ASan+UBSan
(171/171). clang-tidy/gcc-12 aren't available on this machine and
weren't run locally; CI runs them natively on Ubuntu with the pinned
versions this workflow already specifies.

Signed-off-by: Matt <47545907+SoundMatt@users.noreply.github.com>
@SoundMatt
SoundMatt merged commit 9b17111 into main Jul 30, 2026
15 checks passed
@SoundMatt
SoundMatt deleted the fix/relay-v2-install-path branch July 30, 2026 22:37
SoundMatt added a commit that referenced this pull request Jul 30, 2026
Both the fusa-asil-b and sarif jobs in .github/workflows/ci.yml were
pinned to cpp-FuSa v0.15.0 (#16), three releases behind the current
v0.18.0. Bump both checkouts.

Built cpp-FuSa v0.18.0 locally the same way CI does and ran the full
fusa-asil-b gating sequence (init, check, lint, trace, cyber, qualify,
hara init, iso26262, iec61508, boundary, tara, fmea, safety-case, sas,
sci, badge, vuln, metrics, report) plus the sarif job's check --format
sarif against cpp-LIN's own tree, for real, before pushing. No new
ERROR-level findings: check/lint/cyber/qualify all pass clean (0
errors), identical to v0.15.0's clean-pass status; qualify's 8/8 case
suite still passes.

cyber's scanned-file count dropped from 28 to 16 (findings 43 -> 33)
between the two binaries — traced to upstream commit 8564bf0 (part of
the v0.16-v0.18 range), which fixed source-walking commands to honour
.fusa.json's sourceDirs (["src", "include"]) per x-FuSa spec S1.2.1.
v0.15.0 was over-scanning cli/ and tests/, outside this project's
declared source scope; v0.18.0 correctly excludes them. This is a
upstream bug fix, not a suppression of real findings.

iso26262/iec61508 gap-report commands still exit 1 (gaps > 0, expected
for a project without full formal ISO 26262 documentation) but both
calls in ci.yml are already wrapped in `|| true`, unchanged from
v0.15.0's behavior.

Full local suite re-run clean: cmake/ninja Release build (171/171
ctest), ASan+UBSan Debug build (171/171 ctest). ThreadSanitizer could
not be exercised locally (this sandbox's arm64 toolchain lacks a
linkable libtsan runtime); CI runs it natively on ubuntu-22.04/gcc-12
where TSan is supported, unaffected by this change.

No version bump: pure CI-tool-pin fix with no behavior change and no
new findings, same category as #16 and #50 (neither bumped the
version or touched CHANGELOG.md).

Signed-off-by: Matt <47545907+SoundMatt@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant