v0.22.1

Added

• Two-tier retrieval quality gates: Added a daily focused Ollama smoke evaluation and a weekly full-repository representative evaluation spanning TypeScript, Rust, Swift, and PHP, with manual tier selection, separate artifacts, and dataset-specific absolute budgets.

Changed

• Faster real-provider evaluation: Scoped the four-query scheduled Ollama smoke gate to the relevant indexer, OpenCode adapter, and configuration source areas, retaining real semantic competition and all expected files while avoiding full-repository embedding on CPU-only GitHub runners.

Fixed

• Evaluation quality gate: Replaced the retired GitHub Models fallback with a pinned local Ollama provider, refreshed stale smoke-dataset paths after the adapter and ranking refactors, explicitly scoped scheduled runs to the smoke dataset, and fail fast when CI reindexing produces no searchable vectors instead of reporting a misleading zero-quality score.
• Stabilized the linked-worktree config creation test under full-suite CPU contention without changing runtime behavior.

Validation

• Full build, typecheck, lint, and 1,328 tests passed.
• Packed clean-install smoke tests passed for open-codebase-index and opencode-codebase-index, including both MCP binary names.
• GitHub-hosted full-repository eval passed at Hit@5 84.62%, MRR@10 0.5987, and p95 105.778 ms.