PyPI Unreleased
unreleasedChanges on master since 2.0.0 that are not on PyPI yet (pyproject.toml already says 2.0.1). Entries are written from the merged pull requests; PR numbers are in parentheses.
Added
- Measured per-line coverage is the ground truth for gaps (#165); gaps are ranked and suppressed by actionability (#167),
quell findranks gaps and offers measured coverage (#168), and process-control exits are suppressed (#170). - Synthesis mines real construction sites for custom types (#162), builds guard-aware mocks for parameters you do not supply (#163), reuses the project's own
conftest.pyfixtures (#157) and generates native async tests instead ofasyncio.run(#158). - The verifier classifies gate failures and retries instead of discarding the test (#171).
- MCP server for the Claude connector with OAuth PKCE and 8 tools (#123), and cloud sync with the
quell synccommand and--syncflag (#122). - Benchmarks: an ablation harness for the synthesis ladder (#172) and a recorded measurement of quelltest on itself (#169).
Changed
- The Production Readiness Score scores the codebase, not quelltest's own output (#156).
- The rule engine types every previously silent skip (#164).
quell findreports verified tests as WRITTEN, uses the real PRS and no longer crashes on Windows (#161).- Licence: the project is relicensed from MIT to Apache-2.0 (#178; applies from the next release; releases up to 2.0.0 stay MIT).
Removed
- Breaking:
quell scanandquell checkare removed and now fail with an error; usequell find(#124).
Fixed
- The verifier runs candidate tests inside the project's test tree (#160).
- Boundary injection no longer corrupts mined constructions (#173); synthesis rung 3 is now actually called and composed with rung 2 (#174).
- Six defects found by adversarial review (#176); CI: the fake self-check is fixed, lint is wider, the synthesis ladder is gated and the CLI contract is pinned (#175); a mis-attributed benchmark claim for rung 1 is retracted and marginal contribution is measured instead (#177).