I build small, inspectable developer tools for regression testing and CI.

These projects include runnable synthetic examples, explicit limits, and MIT-licensed source.

Find duplicate JUnit test identities across shards and retries, including conflicting outcomes that a report consumer could hide. Python 3.12+ with Expat 2.6.0+.

Try the broken/control demo · Input/output contract · v0.1.0 preview

A clean result means no identity collisions under your chosen scope; keep the test runner's pass/fail gate.

Check saved AI SDK approval histories and catch approval metadata lost during a declared lossless save/reload. TypeScript library and CLI; Node.js 22/24; adapter pinned to ai@7.0.127 JSON ModelMessage.

Save/reload regression demo · Checkpoints and limits · v0.1.0 archive

It checks a supported approval projection; authorization, provider acceptance, and arbitrary conversation formats are outside its scope.

Test whether explicitly grouped tool or subagent calls actually overlap. Set concurrency, overlap, launch-skew, and duration limits around same-clock recorded intervals. Python 3.10+.

Run the blocking/cooperative demo · Trace format · v0.1.0 wheel

Measured overlap does not establish correct outputs, safe independence, or an end-to-end speedup.

All three tools run locally without API keys or a hosted service. Start from the linked source or release instructions; these versions are distributed on GitHub rather than npm/PyPI.

For a bug report, open an issue in the relevant repository with the version, expected result, and a minimal synthetic or sanitized example. Review logs and report output for private data before sharing.