v0.1.1

Changed

• Resolve verifier credentials through OMP's model registry and inherit the calling session's candidate model and thinking level by default.
• Keep every catalog model and provider eligible. DeepSeek V4 Flash remains the recommended default because its native score-tag path is known-good and inexpensive.
• Correct the architecture and safety documentation: candidates are headless OMP subprocesses in detached worktrees, with repository isolation rather than host sandboxing or native Agent Hub integration.

Added

• Probe non-native verifier endpoints before candidate generation with the constrained score-token request required by llm-verifier==0.2.0.
• Add --no-verify and benchmark --generate-only modes for retaining candidate pools without tournament cost.

Fixed

• Reject endpoints that ignore or reject constrained prefill instead of paying for flat or arbitrary verifier scores.
• Stop generate-only benchmark runs from reporting the first candidate as a verifier selection.
• Make the official DeepSeek smoke test bypass stale caches and require a live verifier call.

v0.1.0

Initial release.

Run several Oh My Pi coding agents on one task in isolated git worktrees, rank their complete trajectories with LLM-as-a-Verifier, and apply only the winning patch.

Install:

omp plugin marketplace add wolfiesch/omp-best-of
omp plugin install best-of@omp-best-of

See CHANGELOG.md for the full list of what this release contains, and README.md for the safety model, artifact layout, and cost accounting.