**Anthropic SDK instrumentation** in Python and TypeScript: wrap any existing Anthropic client with `instrument_client` / `instrumentClient` to capture streaming and non-streaming production traces.
**TypeScript `autoctx solve` CLI** brings one-command scenario generation and execution to full parity with Python.
**`autoctx build-dataset` filters** (`--provider`, `--app`, `--env`, `--outcome`) turn captured production traces into scoped training datasets.
**CUDA training backend** alongside MLX, so distillation is no longer Apple Silicon only.
**Semantic prompt compaction** with tail-preserving reducers for longer sessions.
**Hierarchical investigation evidence** with artifact drill-down for richer diagnosis traces.
