Repository navigation
TestDelta v0.1.0a3 — Existing PRs and isolated GitHub Action
Pre-releaseExisting test-only pull requests can now be compared with an external manifest, without rewriting their commits. The native GitHub Action runs TestDelta in its own environment and preserves the caller’s project Python and dependencies.
- Exact external-config bytes/hash are recorded and checked by replay; the original config file can be removed after comparison.
- Evidence includes bounded changed-test and actual mutant diffs, full-diff hashes, per-run UTC timestamps, and a conservative suite-call estimate.
testdelta exportand automatic job summaries omit raw logs, source/config text, diffs, and absolute runtime paths. Full local reports remain available.- A selected-interpreter regression was fixed: the collector is now loaded alone rather than exposing the supervisor’s site-packages to project tests.
Validation: 194 source tests and lint passed locally. The frozen 30-development/20-validation corpus matched all 50 declared outcomes: 17 loss,17 no-loss,16 invalid-baseline inconclusive. These hand-authored controls test classification behavior separately from automatic mutation quality. GitHub CI verifies the four supported Python/OS combinations, full corpus, Ubuntu/macOS native Action, and installed-wheel comparison/replay.
A fixed local pricing benchmark completed 20/20 repetitions with 10 Cosmic Ray mutants each, no exclusions/errors/retries, p95 27.697 seconds and maximum 27.963 seconds. Setup exclusions, per-run evidence, uncontrolled host load and exact scope are recorded in docs/performance.md. This is not an arbitrary-project or cross-platform performance bound.
This remains an early alpha, distributed here rather than assumed on PyPI. Trusted pytest code retains host filesystem/network permissions; no network-blocked execution profile or full dependency/OS replay image is provided. Mixed source/test changes, Windows and plugin/environment-dependent suites remain outside the supported initial scope. The two recorded Boltons compatibility checks remain inconclusive. No general accuracy, algorithmic novelty, adoption or star-growth claim is made. See docs/quality-status.md for deferred original-plan targets.
Assets contain the wheel, source distribution, and SHA256SUMS. Raw local evidence and virtual environments are excluded.