Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 68 additions & 0 deletions docs/adr-004-pin-only-mutation-testing-contract.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
# Architectural decision record (ADR) 004: pin-only mutation-testing contract

## Status

Accepted (2026-07-22): Keep Whitaker's mutation-testing adoption pin-only and
contract-test its declared configuration rather than claiming parity with the
continuous integration (CI) test baseline.

## Date

2026-07-22.

## Context and problem statement

Whitaker delegates mutation testing to the shared
`leynos/shared-actions/.github/workflows/mutation-cargo.yml` workflow. The
caller supplies workspace paths, exclusion globs, and `--all-features`, but the
shared workflow cannot reproduce the per-invocation `RUSTFLAGS` used by the
repository's `test` and `typecheck` Makefile targets.

Those flags are required by the `whitaker`, `function_attrs_follow_docs`,
`module_max_lines`, and `no_expect_outside_tests` crates when the
`dylint-driver` feature enables `feature(rustc_private)`. Applying the flags
workspace-wide is not viable because, as documented in `.cargo/config.toml`,
they break `cargo install` for `whitaker-installer`.

The decision is whether the mutation workflow should claim a workspace baseline
equivalent to `make test`, or expose the closest safe mutation scope and test
only that declared configuration.

## Decision drivers

- Keep the mutation workflow usable without breaking installer builds.
- Describe the mutation scope honestly rather than imply CI parity.
- Detect accidental caller drift while permitting automated commit-pin updates.
- Keep mutation testing informational and independent of pull-request gates.

## Options considered

| Option | Shared-workflow feasibility | Mutation coverage | Operational risk | CI-baseline claim / verification | Rationale |
| ------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------- |
| A: require full CI-baseline parity | Unsafe and unimplementable: the current shared workflow cannot inject the required flags per crate. | Aims for full workspace coverage equivalent to `make test`. | Applying the flags workspace-wide breaks `cargo install` for `whitaker-installer`. | Would claim parity, but cannot verify it through the shared workflow. | Missing per-crate flags make this option unsafe and unimplementable. |
| B: exclude every crate that needs dynamic-linking flags | Runs reliably with the shared workflow by removing the affected crates. | Omits important production code, making the reported scope less useful. | Lower run risk, but accepts the risk of untested affected crates. | Makes no full-baseline claim; verification covers only the reduced scope. | Reliable execution deliberately omits important affected crates. |
| C: adopt a pin-only declared-configuration contract | Closest safe approximation supported by the shared workflow. | Retains useful informational coverage without claiming full workspace coverage. | Keeps mutation testing informational and independent of pull-request gates. | Verifies the caller's declared security and configuration shape, not a full workspace assertion. | A declared-shape contract is the closest safe approximation, not full workspace parity. |

_Table 1: Comparison of mutation-testing options._

## Decision outcome / proposed direction

Adopt Option C.

In the context of shared mutation testing, facing incompatible per-crate
compiler-flag requirements, Whitaker chooses a pin-only declared-configuration
contract over either claimed CI parity or excluding every affected crate, to
retain useful informational mutation coverage while accepting that the run
cannot reproduce `make test` crate-for-crate.

The caller remains pinned to a full commit SHA. Its contract test validates the
pin shape, permissions, concurrency, triggers, and declared inputs. It does not
hard-code the pin value or assert successful full-workspace mutation testing.

## Known risks and limitations

- Mutation results do not certify the same crate scope as `make test`.
- Crates needing the dynamic-linking flags may fail or provide incomplete
mutation results under the shared workflow.
- The decision must be revisited if the shared workflow gains a safe per-crate
mechanism for supplying the required flags.
4 changes: 4 additions & 0 deletions docs/contents.md
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,10 @@
pipeline](adr-003-formal-proof-strategy-for-clone-detector-pipeline.md)
records the formal verification direction for the clone detector pipeline and
its proof boundaries.
- [Architectural decision record (ADR) 004: pin-only mutation-testing
contract](adr-004-pin-only-mutation-testing-contract.md) records why the
shared mutation workflow tests its declared configuration without claiming
parity with the continuous integration test baseline.

## Requests for comments

Expand Down
85 changes: 85 additions & 0 deletions docs/developers-guide.md
Original file line number Diff line number Diff line change
Expand Up @@ -291,6 +291,91 @@ make check-fmt # Verify formatting
make fmt # Apply formatting
```

## Mutation-testing workflow contract tests

Whitaker runs scheduled, informational mutation testing through a thin caller
workflow, [`.github/workflows/mutation-testing.yml`](../.github/workflows/mutation-testing.yml),
which delegates to the shared reusable workflow
`leynos/shared-actions/.github/workflows/mutation-cargo.yml`. The heavy
lifting — running `cargo-mutants`, sharding, and summarizing survivors — lives
in `shared-actions`; this repository carries only declarative configuration.
The run is **informational only**: it never gates a pull request. Survivors
are reported through the job summary and downloadable artefacts so they can be
Comment thread
leynos marked this conversation as resolved.
triaged into tests, not enforced as a blocking check.

The workflow runs in two modes. A **daily schedule** (04:50 UTC) fires a
change-scoped run that mutates only the source files touched within the
detection window, so quiet days are cheap no-ops. A **manual dispatch** (the
Actions "Run workflow" control) mutates the whole workspace; select a branch
in that control to exercise a feature branch.

The caller passes a small set of configuration inputs, each carrying intent:

- `paths` — `src/,common/,crates/,installer/,suite/`, the workspace member
prefixes rooted at the repository root; there are no top-level `examples/`
or `benches/` directories to add.
- `exclude-globs` — scaffolding whose surviving mutants would be noise rather
than genuine test gaps: the `rustc_*` proxy crates and the `clippy_utils`
stub (both re-export compiler internals), each lint's `ui/` and `examples/`
dylint fixtures, and the shared test infrastructure in `src/testing` and
`common/src/test_support`.
- `extra-args` — `--all-features`, matching the feature baseline the
Makefile's `CARGO_FLAGS` uses for `make test`, so feature-gated code is not
reported as untested.

Unlike the excludes applied by `make test`'s `TEST_EXCLUDES`, the mutation
caller does **not** exclude the `whitaker` root crate or the
`function_attrs_follow_docs`, `module_max_lines`, and `no_expect_outside_tests`
lint crates from mutation scope (only their `ui/`/`examples/` fixtures are
excluded). Those crates enable `feature(rustc_private)` under the
`dylint-driver` feature and need the dynamic-linking `RUSTFLAGS`
(`-C prefer-dynamic -Z force-unstable-if-unmarked`) that the `test` and
`typecheck` Makefile targets inject per invocation — see the note in
[`.cargo/config.toml`](../.cargo/config.toml), which deliberately keeps that
flag out of workspace-wide configuration because it would break `cargo
install` for `whitaker-installer`. The shared mutation workflow has no
equivalent per-crate RUSTFLAGS step, so it cannot reproduce
`TEST_CARGO_FLAGS` faithfully across the whole workspace. Consequently this
adoption is **pin-only**: the caller declares the best approximation of the CI
scope it safely can, rather than a `--test-workspace` run that mirrors
`make test` crate-for-crate, and the contract test asserts only that this
declared configuration holds — not that a full workspace mutation baseline
passes. [ADR 004](adr-004-pin-only-mutation-testing-contract.md) records this
decision, its alternatives, and the accepted limitations.

The `uses:` reference pins the shared workflow to a full 40-character commit
SHA rather than a branch or tag, so a force-push upstream cannot silently
change what runs here. The contract test asserts only that the pin is a full
commit SHA, not a particular value, so Dependabot bumps it automatically
without any accompanying test edit.

### Workflow contract tests

Because the caller is configuration rather than code, a contract test suite,
[`tests/workflow_contracts/mutation_testing_test.py`](../tests/workflow_contracts/mutation_testing_test.py),
pins the shape it must uphold, failing the pull request when the caller
drifts — repointing the pin at a branch, widening the token scope, or
dropping a configuration input — rather than letting the breakage surface
only in a scheduled run. Run it locally with:

```sh
make test-workflow-contracts
```

which wraps `uv run --with 'pytest>=8' --with 'pyyaml>=6' pytest
tests/workflow_contracts -q`. The suite validates:

- the `uses:` reference targets `mutation-cargo.yml` pinned to a full commit
SHA;
- the `with:` block carries exactly the expected `paths`, `exclude-globs`,
and `extra-args` configuration described above;
- job permissions are least-privilege (`contents: read`, `id-token: write`)
and the workflow-level default token scope is empty;
- `concurrency` serializes runs per ref without cancelling one in progress;
and
- the triggers keep the daily schedule and a plain `workflow_dispatch` with no
legacy branch input.

## Proof workflows

Whitaker now ships repository-managed proof tooling for the formal verification
Expand Down
Loading