v0.1.1

Added

• KernelBench GPU optimization scaffolding for all 270 upstream tasks, including
  a CUDA evaluation harness and container configuration.
• A five-iteration KernelBench aggregate: 3.12× median best-so-far speedup with
  Claude Opus 4.8 on an H200.
• The per-task aggregate and plot source behind the 154-task AlgoTune result.
• Advanced loop examples for research, implementation task queues, AI workflow
  optimization, and nested AutoHelix runs.
• CI coverage for Python 3.11 and 3.12, plus tests and wheel verification in the
  gated PyPI publishing workflow.

Fixed

• Preserve and scope-check changes when an agent creates commits inside its
  iteration worktree.
• Fail closed when an agent reports an infrastructure error, without merging
  changes or consuming the iteration number.
• Handle Codex turn.failed events and valid non-object JSON output.
• Reject baselines that omit configured metrics.
• Kill constraint and metric subprocess groups when a run is interrupted.

Changed

• Removed the older bin-packing and budget-demo examples in favor of examples
  organized around reusable loop patterns.

v0.1.0

First public release of AutoHelix — put an AI agent in a verified improvement loop.

See the README for what it does, how it works, and examples.